{
  "freeze_hash": "e2b27f9c409d6c93adb46443bed833ac351f1037aeba72cb530e43cb9d36c021",
  "frozen_gate": "FIVE_FACTOR_TASK_CAPABILITY_NOT_ESTABLISHED",
  "mechanism_assay_authorized": false,
  "onset_descriptive": {
    "LL": {
      "complete_ladder": true,
      "first_observed_below_criterion": 2,
      "limitation": "Observed task-specific screening result, not a universal complexity threshold.",
      "nonmonotonic": false
    },
    "LN": {
      "complete_ladder": true,
      "first_observed_below_criterion": 2,
      "limitation": "Observed task-specific screening result, not a universal complexity threshold.",
      "nonmonotonic": false
    },
    "NL": {
      "complete_ladder": true,
      "first_observed_below_criterion": 2,
      "limitation": "Observed task-specific screening result, not a universal complexity threshold.",
      "nonmonotonic": false
    },
    "NN": {
      "complete_ladder": true,
      "first_observed_below_criterion": 3,
      "limitation": "Observed task-specific screening result, not a universal complexity threshold.",
      "nonmonotonic": false
    }
  },
  "protocol": "res-complexity-semantics-v1",
  "stages": {
    "2": {
      "cells": {
        "LL": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.6875,
          "generated_accuracy": 0.6875,
          "generated_correct": 22,
          "generated_exact_95_ci": [
            0.49992241836783036,
            0.8388152759117995
          ],
          "generated_logit_choice_agreement": 0.78125,
          "logit_accuracy": 0.71875,
          "logit_correct": 23,
          "logit_exact_95_ci": [
            0.5325289003570655,
            0.8625430967044982
          ],
          "median_input_tokens": 164.0,
          "n": 32,
          "near_boundary_fraction": 0.21875,
          "parser_coverage": 1.0,
          "raw_zero_margin": 2,
          "strict_parser_coverage": 0.84375
        },
        "LN": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.71875,
          "generated_accuracy": 0.71875,
          "generated_correct": 23,
          "generated_exact_95_ci": [
            0.5325289003570655,
            0.8625430967044982
          ],
          "generated_logit_choice_agreement": 0.65625,
          "logit_accuracy": 0.8125,
          "logit_correct": 26,
          "logit_exact_95_ci": [
            0.63560769013591,
            0.9279238345415024
          ],
          "median_input_tokens": 164.0,
          "n": 32,
          "near_boundary_fraction": 0.375,
          "parser_coverage": 1.0,
          "raw_zero_margin": 1,
          "strict_parser_coverage": 0.875
        },
        "NL": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.59375,
          "generated_accuracy": 0.59375,
          "generated_correct": 19,
          "generated_exact_95_ci": [
            0.4064492465764764,
            0.763015899448257
          ],
          "generated_logit_choice_agreement": 0.625,
          "logit_accuracy": 0.90625,
          "logit_correct": 29,
          "logit_exact_95_ci": [
            0.7497730493162054,
            0.9802328198267152
          ],
          "median_input_tokens": 167.5,
          "n": 32,
          "near_boundary_fraction": 0.3125,
          "parser_coverage": 1.0,
          "raw_zero_margin": 3,
          "strict_parser_coverage": 0.8125
        },
        "NN": {
          "descriptive_criterion_met": true,
          "generated_B_fraction": 0.5,
          "generated_accuracy": 0.875,
          "generated_correct": 28,
          "generated_exact_95_ci": [
            0.7100515798092444,
            0.9648693468966887
          ],
          "generated_logit_choice_agreement": 0.84375,
          "logit_accuracy": 0.90625,
          "logit_correct": 29,
          "logit_exact_95_ci": [
            0.7497730493162054,
            0.9802328198267152
          ],
          "median_input_tokens": 167.5,
          "n": 32,
          "near_boundary_fraction": 0.3125,
          "parser_coverage": 1.0,
          "raw_zero_margin": 2,
          "strict_parser_coverage": 0.8125
        }
      },
      "exploratory_strata": {
        "allow_code": {
          "LL": {
            "A": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.5,
              "generated_accuracy": 0.75,
              "generated_correct": 12,
              "generated_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 164.5,
              "n": 16,
              "near_boundary_fraction": 0.0625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.6875
            }
          },
          "LN": {
            "A": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.5625,
              "generated_accuracy": 0.8125,
              "generated_correct": 13,
              "generated_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 164.5,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 1.0,
              "logit_correct": 16,
              "logit_exact_95_ci": [
                0.7940927857921773,
                1.0
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.75
            }
          },
          "NL": {
            "A": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.5,
              "generated_accuracy": 0.875,
              "generated_correct": 14,
              "generated_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "generated_logit_choice_agreement": 0.9375,
              "logit_accuracy": 0.9375,
              "logit_correct": 15,
              "logit_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.0,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.3125,
              "generated_correct": 5,
              "generated_exact_95_ci": [
                0.11016995471711746,
                0.5866206364513586
              ],
              "generated_logit_choice_agreement": 0.3125,
              "logit_accuracy": 0.875,
              "logit_correct": 14,
              "logit_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 3,
              "strict_parser_coverage": 0.625
            }
          },
          "NN": {
            "A": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.5,
              "generated_accuracy": 1.0,
              "generated_correct": 16,
              "generated_exact_95_ci": [
                0.7940927857921773,
                1.0
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.875,
              "logit_correct": 14,
              "logit_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.5,
              "generated_accuracy": 0.75,
              "generated_correct": 12,
              "generated_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.9375,
              "logit_correct": 15,
              "logit_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.625
            }
          }
        },
        "difficulty": {
          "LL": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.8125
            },
            "extreme": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.8125,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            }
          },
          "LN": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.8125,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            },
            "extreme": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.75,
              "generated_correct": 12,
              "generated_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.8125,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            }
          },
          "NL": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.5,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.875,
              "logit_correct": 14,
              "logit_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.75
            },
            "extreme": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.9375,
              "logit_correct": 15,
              "logit_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            }
          },
          "NN": {
            "boundary": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.625,
              "generated_accuracy": 0.875,
              "generated_correct": 14,
              "generated_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.875,
              "logit_correct": 14,
              "logit_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.75
            },
            "extreme": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.375,
              "generated_accuracy": 0.875,
              "generated_correct": 14,
              "generated_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.9375,
              "logit_correct": 15,
              "logit_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            }
          }
        },
        "family": {
          "LL": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.5625,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.8125
            },
            "1": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.8125,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            }
          },
          "LN": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            },
            "1": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.75,
              "generated_correct": 12,
              "generated_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.875,
              "logit_correct": 14,
              "logit_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            }
          },
          "NL": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.5,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.5625,
              "logit_accuracy": 0.8125,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.8125
            },
            "1": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 1.0,
              "logit_correct": 16,
              "logit_exact_95_ci": [
                0.7940927857921773,
                1.0
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.8125
            }
          },
          "NN": {
            "0": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.5,
              "generated_accuracy": 0.875,
              "generated_correct": 14,
              "generated_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.8125,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.8125
            },
            "1": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.5,
              "generated_accuracy": 0.875,
              "generated_correct": 14,
              "generated_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 1.0,
              "logit_correct": 16,
              "logit_exact_95_ci": [
                0.7940927857921773,
                1.0
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.8125
            }
          }
        },
        "ideal": {
          "LL": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.5,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.875,
              "generated_correct": 14,
              "generated_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.6875
            }
          },
          "LN": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.5,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.5625,
              "logit_accuracy": 0.9375,
              "logit_correct": 15,
              "logit_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.9375,
              "generated_accuracy": 0.9375,
              "generated_correct": 15,
              "generated_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.75
            }
          },
          "NL": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.5,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.5625,
              "logit_accuracy": 0.9375,
              "logit_correct": 15,
              "logit_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.875,
              "logit_correct": 14,
              "logit_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.625
            }
          },
          "NN": {
            "A": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.125,
              "generated_accuracy": 0.875,
              "generated_correct": 14,
              "generated_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "generated_logit_choice_agreement": 0.9375,
              "logit_accuracy": 0.9375,
              "logit_correct": 15,
              "logit_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.875,
              "generated_correct": 14,
              "generated_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.875,
              "logit_correct": 14,
              "logit_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.75
            }
          }
        },
        "peer_congruent": {
          "LL": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.9375
            },
            "True": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.8125,
              "generated_correct": 13,
              "generated_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.75
            }
          },
          "LN": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 0.5625,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.5,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 1.0
            },
            "True": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.8125,
              "generated_correct": 13,
              "generated_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.9375,
              "logit_correct": 15,
              "logit_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "median_input_tokens": 164.0,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.75
            }
          },
          "NL": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.5625,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.9375,
              "logit_correct": 15,
              "logit_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.75
            },
            "True": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.625,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.875,
              "logit_correct": 14,
              "logit_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 3,
              "strict_parser_coverage": 0.875
            }
          },
          "NN": {
            "False": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.4375,
              "generated_accuracy": 0.8125,
              "generated_correct": 13,
              "generated_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.875,
              "logit_correct": 14,
              "logit_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.75
            },
            "True": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.5625,
              "generated_accuracy": 0.9375,
              "generated_correct": 15,
              "generated_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.9375,
              "logit_correct": 15,
              "logit_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "median_input_tokens": 167.5,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            }
          }
        },
        "persona": {
          "LL": {
            "EMBER": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.75,
              "generated_correct": 12,
              "generated_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 163.0,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.8125
            },
            "ORCHID": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.625,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 165.0,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            }
          },
          "LN": {
            "EMBER": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.5,
              "logit_accuracy": 0.8125,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "median_input_tokens": 163.0,
              "n": 16,
              "near_boundary_fraction": 0.5,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            },
            "ORCHID": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.625,
              "generated_accuracy": 0.75,
              "generated_correct": 12,
              "generated_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.8125,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "median_input_tokens": 165.0,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            }
          },
          "NL": {
            "EMBER": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.5,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.875,
              "logit_correct": 14,
              "logit_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "median_input_tokens": 167.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.8125
            },
            "ORCHID": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.5,
              "logit_accuracy": 0.9375,
              "logit_correct": 15,
              "logit_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "median_input_tokens": 168.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.8125
            }
          },
          "NN": {
            "EMBER": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.5625,
              "generated_accuracy": 0.8125,
              "generated_correct": 13,
              "generated_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.875,
              "logit_correct": 14,
              "logit_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "median_input_tokens": 167.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.8125
            },
            "ORCHID": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.4375,
              "generated_accuracy": 0.9375,
              "generated_correct": 15,
              "generated_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.9375,
              "logit_correct": 15,
              "logit_exact_95_ci": [
                0.6976792615654681,
                0.998418888277234
              ],
              "median_input_tokens": 168.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.8125
            }
          }
        }
      },
      "factor_count": 2,
      "generated_pairs_descriptive": {
        "LL_to_LN": {
          "accuracy_difference_right_minus_left": 0.03125,
          "both_correct": 21,
          "both_wrong": 8,
          "corrections": 2,
          "regressions": 1
        },
        "LL_to_NL": {
          "accuracy_difference_right_minus_left": -0.09375,
          "both_correct": 16,
          "both_wrong": 7,
          "corrections": 3,
          "regressions": 6
        },
        "LL_to_NN": {
          "accuracy_difference_right_minus_left": 0.1875,
          "both_correct": 20,
          "both_wrong": 2,
          "corrections": 8,
          "regressions": 2
        },
        "NL_to_LN": {
          "accuracy_difference_right_minus_left": 0.125,
          "both_correct": 16,
          "both_wrong": 6,
          "corrections": 7,
          "regressions": 3
        },
        "NN_to_LN": {
          "accuracy_difference_right_minus_left": -0.15625,
          "both_correct": 21,
          "both_wrong": 2,
          "corrections": 2,
          "regressions": 7
        },
        "NN_to_NL": {
          "accuracy_difference_right_minus_left": -0.28125,
          "both_correct": 18,
          "both_wrong": 3,
          "corrections": 1,
          "regressions": 10
        }
      },
      "logit_pair_descriptive": {
        "accuracy_difference_right_minus_left": 0.1875,
        "both_correct": 22,
        "both_wrong": 2,
        "corrections": 7,
        "regressions": 1
      },
      "near_boundary": {
        "fraction_of_regressions_near_boundary": 1.0,
        "interpretation": "Preregistered output-margin diagnostic, not evidence of an internal mechanism.",
        "median_LL_abs_margin": {
          "concordant": 2.25,
          "discordant": 1.0,
          "regressions": 0.0
        },
        "regressions": 1,
        "regressions_with_LL_abs_margin_at_most_threshold": 1,
        "threshold_abs_raw_LL_margin": 0.75
      },
      "primary": {
        "status": "NOT_CONFIRMATORY_AT_THIS_LEVEL"
      },
      "secondary_model": {
        "status": "NOT_FIT"
      },
      "status": "COMPLETE_DIAGNOSTIC"
    },
    "3": {
      "cells": {
        "LL": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.84375,
          "generated_accuracy": 0.53125,
          "generated_correct": 17,
          "generated_exact_95_ci": [
            0.34743680928639287,
            0.7090601770939848
          ],
          "generated_logit_choice_agreement": 0.78125,
          "logit_accuracy": 0.625,
          "logit_correct": 20,
          "logit_exact_95_ci": [
            0.4369224550969266,
            0.7889996835760071
          ],
          "median_input_tokens": 177.0,
          "n": 32,
          "near_boundary_fraction": 0.21875,
          "parser_coverage": 1.0,
          "raw_zero_margin": 1,
          "strict_parser_coverage": 0.84375
        },
        "LN": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.78125,
          "generated_accuracy": 0.46875,
          "generated_correct": 15,
          "generated_exact_95_ci": [
            0.2909398229060153,
            0.6525631907136071
          ],
          "generated_logit_choice_agreement": 0.78125,
          "logit_accuracy": 0.625,
          "logit_correct": 20,
          "logit_exact_95_ci": [
            0.4369224550969266,
            0.7889996835760071
          ],
          "median_input_tokens": 177.0,
          "n": 32,
          "near_boundary_fraction": 0.0625,
          "parser_coverage": 1.0,
          "raw_zero_margin": 0,
          "strict_parser_coverage": 0.90625
        },
        "NL": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.78125,
          "generated_accuracy": 0.53125,
          "generated_correct": 17,
          "generated_exact_95_ci": [
            0.34743680928639287,
            0.7090601770939848
          ],
          "generated_logit_choice_agreement": 0.65625,
          "logit_accuracy": 0.625,
          "logit_correct": 20,
          "logit_exact_95_ci": [
            0.4369224550969266,
            0.7889996835760071
          ],
          "median_input_tokens": 181.5,
          "n": 32,
          "near_boundary_fraction": 0.40625,
          "parser_coverage": 1.0,
          "raw_zero_margin": 1,
          "strict_parser_coverage": 0.84375
        },
        "NN": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.75,
          "generated_accuracy": 0.625,
          "generated_correct": 20,
          "generated_exact_95_ci": [
            0.4369224550969266,
            0.7889996835760071
          ],
          "generated_logit_choice_agreement": 0.65625,
          "logit_accuracy": 0.65625,
          "logit_correct": 21,
          "logit_exact_95_ci": [
            0.4680689637197313,
            0.8142809082834339
          ],
          "median_input_tokens": 181.5,
          "n": 32,
          "near_boundary_fraction": 0.25,
          "parser_coverage": 1.0,
          "raw_zero_margin": 1,
          "strict_parser_coverage": 0.875
        }
      },
      "exploratory_strata": {
        "allow_code": {
          "LL": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 1.0,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.0,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.4375,
              "generated_correct": 7,
              "generated_exact_95_ci": [
                0.19753414053266782,
                0.7012231009168226
              ],
              "generated_logit_choice_agreement": 0.5625,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.75
            }
          },
          "LN": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 1.0,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.0,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.3125,
              "generated_correct": 5,
              "generated_exact_95_ci": [
                0.11016995471711746,
                0.5866206364513586
              ],
              "generated_logit_choice_agreement": 0.5625,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.8125
            }
          },
          "NL": {
            "A": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.75,
              "generated_correct": 12,
              "generated_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "generated_logit_choice_agreement": 0.9375,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.3125,
              "generated_correct": 5,
              "generated_exact_95_ci": [
                0.11016995471711746,
                0.5866206364513586
              ],
              "generated_logit_choice_agreement": 0.375,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.6875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.6875
            }
          },
          "NN": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.9375,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.0625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.375,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.75
            }
          }
        },
        "difficulty": {
          "LL": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            },
            "extreme": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.8125
            }
          },
          "LN": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.375,
              "generated_correct": 6,
              "generated_exact_95_ci": [
                0.15198367508112198,
                0.6456539056979218
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.5,
              "logit_correct": 8,
              "logit_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.0625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            },
            "extreme": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.0625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            }
          },
          "NL": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.9375,
              "generated_accuracy": 0.4375,
              "generated_correct": 7,
              "generated_exact_95_ci": [
                0.19753414053266782,
                0.7012231009168226
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.4375,
              "logit_correct": 7,
              "logit_exact_95_ci": [
                0.19753414053266782,
                0.7012231009168226
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.5,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            },
            "extreme": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.625,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 0.5625,
              "logit_accuracy": 0.8125,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.75
            }
          },
          "NN": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 1.0,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.5,
              "logit_accuracy": 0.5,
              "logit_correct": 8,
              "logit_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "extreme": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.5,
              "generated_accuracy": 0.75,
              "generated_correct": 12,
              "generated_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.8125,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.75
            }
          }
        },
        "family": {
          "LL": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.8125
            },
            "1": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.9375,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            }
          },
          "LN": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.375,
              "generated_correct": 6,
              "generated_exact_95_ci": [
                0.15198367508112198,
                0.6456539056979218
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.0,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            },
            "1": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            }
          },
          "NL": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.4375,
              "generated_correct": 7,
              "generated_exact_95_ci": [
                0.19753414053266782,
                0.7012231009168226
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.8125
            },
            "1": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            }
          },
          "NN": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            },
            "1": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            }
          }
        },
        "ideal": {
          "LL": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.1875,
              "generated_correct": 3,
              "generated_exact_95_ci": [
                0.04047373390594587,
                0.4564565462316118
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.4375,
              "logit_correct": 7,
              "logit_exact_95_ci": [
                0.19753414053266782,
                0.7012231009168226
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.875,
              "generated_correct": 14,
              "generated_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.8125,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.8125
            }
          },
          "LN": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.1875,
              "generated_correct": 3,
              "generated_exact_95_ci": [
                0.04047373390594587,
                0.4564565462316118
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.0,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.75,
              "generated_correct": 12,
              "generated_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "generated_logit_choice_agreement": 0.9375,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.8125
            }
          },
          "NL": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.25,
              "generated_correct": 4,
              "generated_exact_95_ci": [
                0.0726620382528822,
                0.5237708198961277
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.8125,
              "generated_correct": 13,
              "generated_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.6875
            }
          },
          "NN": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.625,
              "generated_accuracy": 0.375,
              "generated_correct": 6,
              "generated_exact_95_ci": [
                0.15198367508112198,
                0.6456539056979218
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.875,
              "generated_correct": 14,
              "generated_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.75
            }
          }
        },
        "peer_congruent": {
          "LL": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.9375,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.9375
            },
            "True": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.75
            }
          },
          "LN": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.0625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "True": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.4375,
              "generated_correct": 7,
              "generated_exact_95_ci": [
                0.19753414053266782,
                0.7012231009168226
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.0625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.8125
            }
          },
          "NL": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            },
            "True": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.8125
            }
          },
          "NN": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            },
            "True": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.5625,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 181.5,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            }
          }
        },
        "persona": {
          "LL": {
            "EMBER": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            },
            "ORCHID": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.4375,
              "generated_correct": 7,
              "generated_exact_95_ci": [
                0.19753414053266782,
                0.7012231009168226
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 178.0,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.8125
            }
          },
          "LN": {
            "EMBER": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.4375,
              "generated_correct": 7,
              "generated_exact_95_ci": [
                0.19753414053266782,
                0.7012231009168226
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 177.0,
              "n": 16,
              "near_boundary_fraction": 0.0625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            },
            "ORCHID": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 178.0,
              "n": 16,
              "near_boundary_fraction": 0.0625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            }
          },
          "NL": {
            "EMBER": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.5,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 181.0,
              "n": 16,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            },
            "ORCHID": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 182.0,
              "n": 16,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.8125
            }
          },
          "NN": {
            "EMBER": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 181.0,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            },
            "ORCHID": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 182.0,
              "n": 16,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            }
          }
        }
      },
      "factor_count": 3,
      "generated_pairs_descriptive": {
        "LL_to_LN": {
          "accuracy_difference_right_minus_left": -0.0625,
          "both_correct": 14,
          "both_wrong": 14,
          "corrections": 1,
          "regressions": 3
        },
        "LL_to_NL": {
          "accuracy_difference_right_minus_left": 0.0,
          "both_correct": 15,
          "both_wrong": 13,
          "corrections": 2,
          "regressions": 2
        },
        "LL_to_NN": {
          "accuracy_difference_right_minus_left": 0.09375,
          "both_correct": 15,
          "both_wrong": 10,
          "corrections": 5,
          "regressions": 2
        },
        "NL_to_LN": {
          "accuracy_difference_right_minus_left": -0.0625,
          "both_correct": 13,
          "both_wrong": 13,
          "corrections": 2,
          "regressions": 4
        },
        "NN_to_LN": {
          "accuracy_difference_right_minus_left": -0.15625,
          "both_correct": 14,
          "both_wrong": 11,
          "corrections": 1,
          "regressions": 6
        },
        "NN_to_NL": {
          "accuracy_difference_right_minus_left": -0.09375,
          "both_correct": 16,
          "both_wrong": 11,
          "corrections": 1,
          "regressions": 4
        }
      },
      "logit_pair_descriptive": {
        "accuracy_difference_right_minus_left": 0.03125,
        "both_correct": 16,
        "both_wrong": 7,
        "corrections": 5,
        "regressions": 4
      },
      "near_boundary": {
        "fraction_of_regressions_near_boundary": 0.5,
        "interpretation": "Preregistered output-margin diagnostic, not evidence of an internal mechanism.",
        "median_LL_abs_margin": {
          "concordant": 3.25,
          "discordant": 2.75,
          "regressions": 3.5
        },
        "regressions": 4,
        "regressions_with_LL_abs_margin_at_most_threshold": 2,
        "threshold_abs_raw_LL_margin": 0.75
      },
      "primary": {
        "status": "NOT_CONFIRMATORY_AT_THIS_LEVEL"
      },
      "secondary_model": {
        "status": "NOT_FIT"
      },
      "status": "COMPLETE_DIAGNOSTIC"
    },
    "4": {
      "cells": {
        "LL": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.84375,
          "generated_accuracy": 0.53125,
          "generated_correct": 17,
          "generated_exact_95_ci": [
            0.34743680928639287,
            0.7090601770939848
          ],
          "generated_logit_choice_agreement": 0.875,
          "logit_accuracy": 0.59375,
          "logit_correct": 19,
          "logit_exact_95_ci": [
            0.4064492465764764,
            0.763015899448257
          ],
          "median_input_tokens": 189.0,
          "n": 32,
          "near_boundary_fraction": 0.15625,
          "parser_coverage": 1.0,
          "raw_zero_margin": 0,
          "strict_parser_coverage": 0.875
        },
        "LN": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.9375,
          "generated_accuracy": 0.5625,
          "generated_correct": 18,
          "generated_exact_95_ci": [
            0.3766257315900938,
            0.7363618755890153
          ],
          "generated_logit_choice_agreement": 0.65625,
          "logit_accuracy": 0.65625,
          "logit_correct": 21,
          "logit_exact_95_ci": [
            0.4680689637197313,
            0.8142809082834339
          ],
          "median_input_tokens": 189.0,
          "n": 32,
          "near_boundary_fraction": 0.28125,
          "parser_coverage": 1.0,
          "raw_zero_margin": 1,
          "strict_parser_coverage": 0.875
        },
        "NL": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.75,
          "generated_accuracy": 0.625,
          "generated_correct": 20,
          "generated_exact_95_ci": [
            0.4369224550969266,
            0.7889996835760071
          ],
          "generated_logit_choice_agreement": 0.65625,
          "logit_accuracy": 0.65625,
          "logit_correct": 21,
          "logit_exact_95_ci": [
            0.4680689637197313,
            0.8142809082834339
          ],
          "median_input_tokens": 195.5,
          "n": 32,
          "near_boundary_fraction": 0.21875,
          "parser_coverage": 1.0,
          "raw_zero_margin": 1,
          "strict_parser_coverage": 0.96875
        },
        "NN": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.6875,
          "generated_accuracy": 0.625,
          "generated_correct": 20,
          "generated_exact_95_ci": [
            0.4369224550969266,
            0.7889996835760071
          ],
          "generated_logit_choice_agreement": 0.71875,
          "logit_accuracy": 0.65625,
          "logit_correct": 21,
          "logit_exact_95_ci": [
            0.4680689637197313,
            0.8142809082834339
          ],
          "median_input_tokens": 195.5,
          "n": 32,
          "near_boundary_fraction": 0.1875,
          "parser_coverage": 1.0,
          "raw_zero_margin": 1,
          "strict_parser_coverage": 0.96875
        }
      },
      "exploratory_strata": {
        "allow_code": {
          "LL": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 1.0,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.0,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.4375,
              "generated_correct": 7,
              "generated_exact_95_ci": [
                0.19753414053266782,
                0.7012231009168226
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 189.5,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.75
            }
          },
          "LN": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 1.0,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.0,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 1.0,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.3125,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 189.5,
              "n": 16,
              "near_boundary_fraction": 0.5625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.75
            }
          },
          "NL": {
            "A": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.5625,
              "generated_accuracy": 0.8125,
              "generated_correct": 13,
              "generated_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "generated_logit_choice_agreement": 0.9375,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.9375,
              "generated_accuracy": 0.4375,
              "generated_correct": 7,
              "generated_exact_95_ci": [
                0.19753414053266782,
                0.7012231009168226
              ],
              "generated_logit_choice_agreement": 0.375,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.9375
            }
          },
          "NN": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.9375,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.5,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.9375
            }
          }
        },
        "difficulty": {
          "LL": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.5,
              "logit_correct": 8,
              "logit_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.0625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            },
            "extreme": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            }
          },
          "LN": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 1.0,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.5,
              "logit_accuracy": 0.5,
              "logit_correct": 8,
              "logit_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            },
            "extreme": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.8125,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            }
          },
          "NL": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.5625,
              "logit_accuracy": 0.5,
              "logit_correct": 8,
              "logit_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 1.0
            },
            "extreme": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.8125,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            }
          },
          "NN": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.375,
              "generated_correct": 6,
              "generated_exact_95_ci": [
                0.15198367508112198,
                0.6456539056979218
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.4375,
              "logit_correct": 7,
              "logit_exact_95_ci": [
                0.19753414053266782,
                0.7012231009168226
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 1.0
            },
            "extreme": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.625,
              "generated_accuracy": 0.875,
              "generated_correct": 14,
              "generated_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.875,
              "logit_correct": 14,
              "logit_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            }
          }
        },
        "family": {
          "LL": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.8125
            },
            "1": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.9375,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 1.0,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            }
          },
          "LN": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.9375,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.8125
            },
            "1": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.9375,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.9375
            }
          },
          "NL": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.5,
              "logit_correct": 8,
              "logit_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.9375
            },
            "1": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.75,
              "generated_correct": 12,
              "generated_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.8125,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.0625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            }
          },
          "NN": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            },
            "1": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 1.0
            }
          }
        },
        "ideal": {
          "LL": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.1875,
              "generated_correct": 3,
              "generated_exact_95_ci": [
                0.04047373390594587,
                0.4564565462316118
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.3125,
              "logit_correct": 5,
              "logit_exact_95_ci": [
                0.11016995471711746,
                0.5866206364513586
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.875,
              "generated_correct": 14,
              "generated_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.875,
              "logit_correct": 14,
              "logit_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.0625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.8125
            }
          },
          "LN": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.125,
              "generated_correct": 2,
              "generated_exact_95_ci": [
                0.015513603815413893,
                0.3834762368492636
              ],
              "generated_logit_choice_agreement": 0.5625,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.9375
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 1.0,
              "generated_accuracy": 1.0,
              "generated_correct": 16,
              "generated_exact_95_ci": [
                0.7940927857921773,
                1.0
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.8125
            }
          },
          "NL": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.625,
              "generated_accuracy": 0.375,
              "generated_correct": 6,
              "generated_exact_95_ci": [
                0.15198367508112198,
                0.6456539056979218
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.875,
              "generated_correct": 14,
              "generated_exact_95_ci": [
                0.6165237631507365,
                0.984486396184586
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            }
          },
          "NN": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.5625,
              "generated_accuracy": 0.4375,
              "generated_correct": 7,
              "generated_exact_95_ci": [
                0.19753414053266782,
                0.7012231009168226
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 1.0
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.8125,
              "generated_correct": 13,
              "generated_exact_95_ci": [
                0.5435434537683883,
                0.9595262660940542
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            }
          }
        },
        "peer_congruent": {
          "LL": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.9375,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            },
            "True": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.8125
            }
          },
          "LN": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 1.0,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            },
            "True": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.8125
            }
          },
          "NL": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "True": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.9375
            }
          },
          "NN": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 1.0
            },
            "True": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.625,
              "generated_accuracy": 0.75,
              "generated_correct": 12,
              "generated_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 195.5,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            }
          }
        },
        "persona": {
          "LL": {
            "EMBER": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.625,
              "generated_correct": 10,
              "generated_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            },
            "ORCHID": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.4375,
              "generated_correct": 7,
              "generated_exact_95_ci": [
                0.19753414053266782,
                0.7012231009168226
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 190.0,
              "n": 16,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            }
          },
          "LN": {
            "EMBER": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.9375,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.625,
              "logit_correct": 10,
              "logit_exact_95_ci": [
                0.3543460943020781,
                0.8480163249188779
              ],
              "median_input_tokens": 189.0,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            },
            "ORCHID": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.9375,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.6875,
              "logit_correct": 11,
              "logit_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "median_input_tokens": 190.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            }
          },
          "NL": {
            "EMBER": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.6875,
              "generated_correct": 11,
              "generated_exact_95_ci": [
                0.41337936354864124,
                0.8898300452828825
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 195.0,
              "n": 16,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "ORCHID": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.5625,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 196.0,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.9375
            }
          },
          "NN": {
            "EMBER": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.75,
              "generated_correct": 12,
              "generated_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.75,
              "logit_correct": 12,
              "logit_exact_95_ci": [
                0.4762291801038723,
                0.9273379617471178
              ],
              "median_input_tokens": 195.0,
              "n": 16,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 1.0
            },
            "ORCHID": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.625,
              "generated_accuracy": 0.5,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.24651011149057533,
                0.7534898885094248
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.5625,
              "logit_correct": 9,
              "logit_exact_95_ci": [
                0.2987768990831774,
                0.8024658594673322
              ],
              "median_input_tokens": 196.0,
              "n": 16,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.9375
            }
          }
        }
      },
      "factor_count": 4,
      "generated_pairs_descriptive": {
        "LL_to_LN": {
          "accuracy_difference_right_minus_left": 0.03125,
          "both_correct": 16,
          "both_wrong": 13,
          "corrections": 2,
          "regressions": 1
        },
        "LL_to_NL": {
          "accuracy_difference_right_minus_left": 0.09375,
          "both_correct": 15,
          "both_wrong": 10,
          "corrections": 5,
          "regressions": 2
        },
        "LL_to_NN": {
          "accuracy_difference_right_minus_left": 0.09375,
          "both_correct": 14,
          "both_wrong": 9,
          "corrections": 6,
          "regressions": 3
        },
        "NL_to_LN": {
          "accuracy_difference_right_minus_left": -0.0625,
          "both_correct": 16,
          "both_wrong": 10,
          "corrections": 2,
          "regressions": 4
        },
        "NN_to_LN": {
          "accuracy_difference_right_minus_left": -0.0625,
          "both_correct": 15,
          "both_wrong": 9,
          "corrections": 3,
          "regressions": 5
        },
        "NN_to_NL": {
          "accuracy_difference_right_minus_left": 0.0,
          "both_correct": 16,
          "both_wrong": 8,
          "corrections": 4,
          "regressions": 4
        }
      },
      "logit_pair_descriptive": {
        "accuracy_difference_right_minus_left": 0.0625,
        "both_correct": 14,
        "both_wrong": 6,
        "corrections": 7,
        "regressions": 5
      },
      "near_boundary": {
        "fraction_of_regressions_near_boundary": 0.2,
        "interpretation": "Preregistered output-margin diagnostic, not evidence of an internal mechanism.",
        "median_LL_abs_margin": {
          "concordant": 4.375,
          "discordant": 1.0,
          "regressions": 1.0
        },
        "regressions": 5,
        "regressions_with_LL_abs_margin_at_most_threshold": 1,
        "threshold_abs_raw_LL_margin": 0.75
      },
      "primary": {
        "status": "NOT_CONFIRMATORY_AT_THIS_LEVEL"
      },
      "secondary_model": {
        "status": "NOT_FIT"
      },
      "status": "COMPLETE_DIAGNOSTIC"
    },
    "5": {
      "cells": {
        "LL": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.828125,
          "generated_accuracy": 0.546875,
          "generated_correct": 35,
          "generated_exact_95_ci": [
            0.41747631531604745,
            0.671792885496282
          ],
          "generated_logit_choice_agreement": 0.8125,
          "logit_accuracy": 0.640625,
          "logit_correct": 41,
          "logit_exact_95_ci": [
            0.5109625981612298,
            0.7568097089687873
          ],
          "median_input_tokens": 201.5,
          "n": 64,
          "near_boundary_fraction": 0.25,
          "parser_coverage": 1.0,
          "raw_zero_margin": 5,
          "strict_parser_coverage": 0.828125
        },
        "LN": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.859375,
          "generated_accuracy": 0.515625,
          "generated_correct": 33,
          "generated_exact_95_ci": [
            0.3872588082204236,
            0.6424980032802374
          ],
          "generated_logit_choice_agreement": 0.640625,
          "logit_accuracy": 0.6875,
          "logit_correct": 44,
          "logit_exact_95_ci": [
            0.5594056331825381,
            0.7975785274198701
          ],
          "median_input_tokens": 201.5,
          "n": 64,
          "near_boundary_fraction": 0.3125,
          "parser_coverage": 1.0,
          "raw_zero_margin": 4,
          "strict_parser_coverage": 0.796875
        },
        "NL": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.84375,
          "generated_accuracy": 0.640625,
          "generated_correct": 41,
          "generated_exact_95_ci": [
            0.5109625981612298,
            0.7568097089687873
          ],
          "generated_logit_choice_agreement": 0.6875,
          "logit_accuracy": 0.640625,
          "logit_correct": 41,
          "logit_exact_95_ci": [
            0.5109625981612298,
            0.7568097089687873
          ],
          "median_input_tokens": 209.5,
          "n": 64,
          "near_boundary_fraction": 0.390625,
          "parser_coverage": 0.984375,
          "raw_zero_margin": 4,
          "strict_parser_coverage": 0.828125
        },
        "NN": {
          "descriptive_criterion_met": false,
          "generated_B_fraction": 0.84375,
          "generated_accuracy": 0.59375,
          "generated_correct": 38,
          "generated_exact_95_ci": [
            0.4636768777077156,
            0.7148529871891394
          ],
          "generated_logit_choice_agreement": 0.671875,
          "logit_accuracy": 0.640625,
          "logit_correct": 41,
          "logit_exact_95_ci": [
            0.5109625981612298,
            0.7568097089687873
          ],
          "median_input_tokens": 209.5,
          "n": 64,
          "near_boundary_fraction": 0.359375,
          "parser_coverage": 1.0,
          "raw_zero_margin": 3,
          "strict_parser_coverage": 0.703125
        }
      },
      "exploratory_strata": {
        "allow_code": {
          "LL": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.78125,
              "generated_accuracy": 0.71875,
              "generated_correct": 23,
              "generated_exact_95_ci": [
                0.5325289003570655,
                0.8625430967044982
              ],
              "generated_logit_choice_agreement": 1.0,
              "logit_accuracy": 0.71875,
              "logit_correct": 23,
              "logit_exact_95_ci": [
                0.5325289003570655,
                0.8625430967044982
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.0,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.96875
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.375,
              "generated_correct": 12,
              "generated_exact_95_ci": [
                0.2110003164239928,
                0.5630775449030734
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.5625,
              "logit_correct": 18,
              "logit_exact_95_ci": [
                0.3766257315900938,
                0.7363618755890153
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.5,
              "parser_coverage": 1.0,
              "raw_zero_margin": 5,
              "strict_parser_coverage": 0.6875
            }
          },
          "LN": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.84375,
              "generated_accuracy": 0.65625,
              "generated_correct": 21,
              "generated_exact_95_ci": [
                0.4680689637197313,
                0.8142809082834339
              ],
              "generated_logit_choice_agreement": 0.96875,
              "logit_accuracy": 0.6875,
              "logit_correct": 22,
              "logit_exact_95_ci": [
                0.49992241836783036,
                0.8388152759117995
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.0,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.9375
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.375,
              "generated_correct": 12,
              "generated_exact_95_ci": [
                0.2110003164239928,
                0.5630775449030734
              ],
              "generated_logit_choice_agreement": 0.3125,
              "logit_accuracy": 0.6875,
              "logit_correct": 22,
              "logit_exact_95_ci": [
                0.49992241836783036,
                0.8388152759117995
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 4,
              "strict_parser_coverage": 0.65625
            }
          },
          "NL": {
            "A": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.71875,
              "generated_accuracy": 0.75,
              "generated_correct": 24,
              "generated_exact_95_ci": [
                0.5659506250777573,
                0.8853839948974118
              ],
              "generated_logit_choice_agreement": 0.875,
              "logit_accuracy": 0.75,
              "logit_correct": 24,
              "logit_exact_95_ci": [
                0.5659506250777573,
                0.8853839948974118
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.15625,
              "parser_coverage": 0.96875,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.96875
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.96875,
              "generated_accuracy": 0.53125,
              "generated_correct": 17,
              "generated_exact_95_ci": [
                0.34743680928639287,
                0.7090601770939848
              ],
              "generated_logit_choice_agreement": 0.5,
              "logit_accuracy": 0.53125,
              "logit_correct": 17,
              "logit_exact_95_ci": [
                0.34743680928639287,
                0.7090601770939848
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 3,
              "strict_parser_coverage": 0.6875
            }
          },
          "NN": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.78125,
              "generated_accuracy": 0.65625,
              "generated_correct": 21,
              "generated_exact_95_ci": [
                0.4680689637197313,
                0.8142809082834339
              ],
              "generated_logit_choice_agreement": 0.90625,
              "logit_accuracy": 0.6875,
              "logit_correct": 22,
              "logit_exact_95_ci": [
                0.49992241836783036,
                0.8388152759117995
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.96875
            },
            "B": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.90625,
              "generated_accuracy": 0.53125,
              "generated_correct": 17,
              "generated_exact_95_ci": [
                0.34743680928639287,
                0.7090601770939848
              ],
              "generated_logit_choice_agreement": 0.4375,
              "logit_accuracy": 0.59375,
              "logit_correct": 19,
              "logit_exact_95_ci": [
                0.4064492465764764,
                0.763015899448257
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.59375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.4375
            }
          }
        },
        "difficulty": {
          "LL": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.90625,
              "generated_accuracy": 0.40625,
              "generated_correct": 13,
              "generated_exact_95_ci": [
                0.23698410055174313,
                0.5935507534235236
              ],
              "generated_logit_choice_agreement": 0.71875,
              "logit_accuracy": 0.5,
              "logit_correct": 16,
              "logit_exact_95_ci": [
                0.31887496423416395,
                0.681125035765836
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.28125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 4,
              "strict_parser_coverage": 0.875
            },
            "extreme": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.6875,
              "generated_correct": 22,
              "generated_exact_95_ci": [
                0.49992241836783036,
                0.8388152759117995
              ],
              "generated_logit_choice_agreement": 0.90625,
              "logit_accuracy": 0.78125,
              "logit_correct": 25,
              "logit_exact_95_ci": [
                0.6002717362878924,
                0.9072284677152807
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.21875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.78125
            }
          },
          "LN": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.9375,
              "generated_accuracy": 0.4375,
              "generated_correct": 14,
              "generated_exact_95_ci": [
                0.26363812441098466,
                0.6233742684099062
              ],
              "generated_logit_choice_agreement": 0.59375,
              "logit_accuracy": 0.53125,
              "logit_correct": 17,
              "logit_exact_95_ci": [
                0.34743680928639287,
                0.7090601770939848
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.875
            },
            "extreme": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.78125,
              "generated_accuracy": 0.59375,
              "generated_correct": 19,
              "generated_exact_95_ci": [
                0.4064492465764764,
                0.763015899448257
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.84375,
              "logit_correct": 27,
              "logit_exact_95_ci": [
                0.6721212343179581,
                0.947249435597866
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.71875
            }
          },
          "NL": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.96875,
              "generated_accuracy": 0.5,
              "generated_correct": 16,
              "generated_exact_95_ci": [
                0.31887496423416395,
                0.681125035765836
              ],
              "generated_logit_choice_agreement": 0.65625,
              "logit_accuracy": 0.53125,
              "logit_correct": 17,
              "logit_exact_95_ci": [
                0.34743680928639287,
                0.7090601770939848
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.40625,
              "parser_coverage": 0.96875,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.90625
            },
            "extreme": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.71875,
              "generated_accuracy": 0.78125,
              "generated_correct": 25,
              "generated_exact_95_ci": [
                0.6002717362878924,
                0.9072284677152807
              ],
              "generated_logit_choice_agreement": 0.71875,
              "logit_accuracy": 0.75,
              "logit_correct": 24,
              "logit_exact_95_ci": [
                0.5659506250777573,
                0.8853839948974118
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.75
            }
          },
          "NN": {
            "boundary": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.96875,
              "generated_accuracy": 0.46875,
              "generated_correct": 15,
              "generated_exact_95_ci": [
                0.2909398229060153,
                0.6525631907136071
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.46875,
              "logit_correct": 15,
              "logit_exact_95_ci": [
                0.2909398229060153,
                0.6525631907136071
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.875
            },
            "extreme": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.71875,
              "generated_accuracy": 0.71875,
              "generated_correct": 23,
              "generated_exact_95_ci": [
                0.5325289003570655,
                0.8625430967044982
              ],
              "generated_logit_choice_agreement": 0.65625,
              "logit_accuracy": 0.8125,
              "logit_correct": 26,
              "logit_exact_95_ci": [
                0.63560769013591,
                0.9279238345415024
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.34375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.53125
            }
          }
        },
        "family": {
          "LL": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.84375,
              "generated_accuracy": 0.53125,
              "generated_correct": 17,
              "generated_exact_95_ci": [
                0.34743680928639287,
                0.7090601770939848
              ],
              "generated_logit_choice_agreement": 0.65625,
              "logit_accuracy": 0.6875,
              "logit_correct": 22,
              "logit_exact_95_ci": [
                0.49992241836783036,
                0.8388152759117995
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.40625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 4,
              "strict_parser_coverage": 0.84375
            },
            "1": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.5625,
              "generated_correct": 18,
              "generated_exact_95_ci": [
                0.3766257315900938,
                0.7363618755890153
              ],
              "generated_logit_choice_agreement": 0.96875,
              "logit_accuracy": 0.59375,
              "logit_correct": 19,
              "logit_exact_95_ci": [
                0.4064492465764764,
                0.763015899448257
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.09375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.8125
            }
          },
          "LN": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.5,
              "generated_correct": 16,
              "generated_exact_95_ci": [
                0.31887496423416395,
                0.681125035765836
              ],
              "generated_logit_choice_agreement": 0.59375,
              "logit_accuracy": 0.65625,
              "logit_correct": 21,
              "logit_exact_95_ci": [
                0.4680689637197313,
                0.8142809082834339
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 3,
              "strict_parser_coverage": 0.8125
            },
            "1": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.84375,
              "generated_accuracy": 0.53125,
              "generated_correct": 17,
              "generated_exact_95_ci": [
                0.34743680928639287,
                0.7090601770939848
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.71875,
              "logit_correct": 23,
              "logit_exact_95_ci": [
                0.5325289003570655,
                0.8625430967044982
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.78125
            }
          },
          "NL": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.84375,
              "generated_accuracy": 0.65625,
              "generated_correct": 21,
              "generated_exact_95_ci": [
                0.4680689637197313,
                0.8142809082834339
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.59375,
              "logit_correct": 19,
              "logit_exact_95_ci": [
                0.4064492465764764,
                0.763015899448257
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.875
            },
            "1": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.84375,
              "generated_accuracy": 0.625,
              "generated_correct": 20,
              "generated_exact_95_ci": [
                0.4369224550969266,
                0.7889996835760071
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.6875,
              "logit_correct": 22,
              "logit_exact_95_ci": [
                0.49992241836783036,
                0.8388152759117995
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.40625,
              "parser_coverage": 0.96875,
              "raw_zero_margin": 4,
              "strict_parser_coverage": 0.78125
            }
          },
          "NN": {
            "0": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.84375,
              "generated_accuracy": 0.65625,
              "generated_correct": 21,
              "generated_exact_95_ci": [
                0.4680689637197313,
                0.8142809082834339
              ],
              "generated_logit_choice_agreement": 0.65625,
              "logit_accuracy": 0.625,
              "logit_correct": 20,
              "logit_exact_95_ci": [
                0.4369224550969266,
                0.7889996835760071
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 0,
              "strict_parser_coverage": 0.71875
            },
            "1": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.84375,
              "generated_accuracy": 0.53125,
              "generated_correct": 17,
              "generated_exact_95_ci": [
                0.34743680928639287,
                0.7090601770939848
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.65625,
              "logit_correct": 21,
              "logit_exact_95_ci": [
                0.4680689637197313,
                0.8142809082834339
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.28125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 3,
              "strict_parser_coverage": 0.6875
            }
          }
        },
        "ideal": {
          "LL": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.78125,
              "generated_accuracy": 0.21875,
              "generated_correct": 7,
              "generated_exact_95_ci": [
                0.09277153228471932,
                0.3997282637121077
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.40625,
              "logit_correct": 13,
              "logit_exact_95_ci": [
                0.23698410055174313,
                0.5935507534235236
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.21875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 3,
              "strict_parser_coverage": 0.9375
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.875,
              "generated_correct": 28,
              "generated_exact_95_ci": [
                0.7100515798092444,
                0.9648693468966887
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.875,
              "logit_correct": 28,
              "logit_exact_95_ci": [
                0.7100515798092444,
                0.9648693468966887
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.28125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.71875
            }
          },
          "LN": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.84375,
              "generated_accuracy": 0.15625,
              "generated_correct": 5,
              "generated_exact_95_ci": [
                0.05275056440213409,
                0.3278787656820418
              ],
              "generated_logit_choice_agreement": 0.59375,
              "logit_accuracy": 0.5625,
              "logit_correct": 18,
              "logit_exact_95_ci": [
                0.3766257315900938,
                0.7363618755890153
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.40625,
              "parser_coverage": 1.0,
              "raw_zero_margin": 3,
              "strict_parser_coverage": 0.90625
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.875,
              "generated_correct": 28,
              "generated_exact_95_ci": [
                0.7100515798092444,
                0.9648693468966887
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.8125,
              "logit_correct": 26,
              "logit_exact_95_ci": [
                0.63560769013591,
                0.9279238345415024
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.21875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.6875
            }
          },
          "NL": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.6875,
              "generated_accuracy": 0.28125,
              "generated_correct": 9,
              "generated_exact_95_ci": [
                0.1374569032955017,
                0.46747109964293465
              ],
              "generated_logit_choice_agreement": 0.65625,
              "logit_accuracy": 0.5625,
              "logit_correct": 18,
              "logit_exact_95_ci": [
                0.3766257315900938,
                0.7363618755890153
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.46875,
              "parser_coverage": 0.96875,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.9375
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 1.0,
              "generated_accuracy": 1.0,
              "generated_correct": 32,
              "generated_exact_95_ci": [
                0.8911188393206471,
                1.0
              ],
              "generated_logit_choice_agreement": 0.71875,
              "logit_accuracy": 0.71875,
              "logit_correct": 23,
              "logit_exact_95_ci": [
                0.5325289003570655,
                0.8625430967044982
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.71875
            }
          },
          "NN": {
            "A": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.75,
              "generated_accuracy": 0.25,
              "generated_correct": 8,
              "generated_exact_95_ci": [
                0.11461600510258818,
                0.4340493749222428
              ],
              "generated_logit_choice_agreement": 0.5625,
              "logit_accuracy": 0.5625,
              "logit_correct": 18,
              "logit_exact_95_ci": [
                0.3766257315900938,
                0.7363618755890153
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.4375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.8125
            },
            "B": {
              "descriptive_criterion_met": true,
              "generated_B_fraction": 0.9375,
              "generated_accuracy": 0.9375,
              "generated_correct": 30,
              "generated_exact_95_ci": [
                0.7919305701049926,
                0.9923392636535435
              ],
              "generated_logit_choice_agreement": 0.78125,
              "logit_accuracy": 0.71875,
              "logit_correct": 23,
              "logit_exact_95_ci": [
                0.5325289003570655,
                0.8625430967044982
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.28125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.59375
            }
          }
        },
        "peer_congruent": {
          "LL": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.84375,
              "generated_accuracy": 0.53125,
              "generated_correct": 17,
              "generated_exact_95_ci": [
                0.34743680928639287,
                0.7090601770939848
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.65625,
              "logit_correct": 21,
              "logit_exact_95_ci": [
                0.4680689637197313,
                0.8142809082834339
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.1875,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.8125
            },
            "True": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.5625,
              "generated_correct": 18,
              "generated_exact_95_ci": [
                0.3766257315900938,
                0.7363618755890153
              ],
              "generated_logit_choice_agreement": 0.8125,
              "logit_accuracy": 0.625,
              "logit_correct": 20,
              "logit_exact_95_ci": [
                0.4369224550969266,
                0.7889996835760071
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 3,
              "strict_parser_coverage": 0.84375
            }
          },
          "LN": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.84375,
              "generated_accuracy": 0.40625,
              "generated_correct": 13,
              "generated_exact_95_ci": [
                0.23698410055174313,
                0.5935507534235236
              ],
              "generated_logit_choice_agreement": 0.53125,
              "logit_accuracy": 0.75,
              "logit_correct": 24,
              "logit_exact_95_ci": [
                0.5659506250777573,
                0.8853839948974118
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.78125
            },
            "True": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.625,
              "generated_correct": 20,
              "generated_exact_95_ci": [
                0.4369224550969266,
                0.7889996835760071
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.625,
              "logit_correct": 20,
              "logit_exact_95_ci": [
                0.4369224550969266,
                0.7889996835760071
              ],
              "median_input_tokens": 201.5,
              "n": 32,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.8125
            }
          },
          "NL": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.84375,
              "generated_accuracy": 0.625,
              "generated_correct": 20,
              "generated_exact_95_ci": [
                0.4369224550969266,
                0.7889996835760071
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.6875,
              "logit_correct": 22,
              "logit_exact_95_ci": [
                0.49992241836783036,
                0.8388152759117995
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.46875,
              "parser_coverage": 0.96875,
              "raw_zero_margin": 3,
              "strict_parser_coverage": 0.875
            },
            "True": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.84375,
              "generated_accuracy": 0.65625,
              "generated_correct": 21,
              "generated_exact_95_ci": [
                0.4680689637197313,
                0.8142809082834339
              ],
              "generated_logit_choice_agreement": 0.75,
              "logit_accuracy": 0.59375,
              "logit_correct": 19,
              "logit_exact_95_ci": [
                0.4064492465764764,
                0.763015899448257
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.3125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.78125
            }
          },
          "NN": {
            "False": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.5,
              "generated_correct": 16,
              "generated_exact_95_ci": [
                0.31887496423416395,
                0.681125035765836
              ],
              "generated_logit_choice_agreement": 0.71875,
              "logit_accuracy": 0.65625,
              "logit_correct": 21,
              "logit_exact_95_ci": [
                0.4680689637197313,
                0.8142809082834339
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.6875
            },
            "True": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.6875,
              "generated_correct": 22,
              "generated_exact_95_ci": [
                0.49992241836783036,
                0.8388152759117995
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.625,
              "logit_correct": 20,
              "logit_exact_95_ci": [
                0.4369224550969266,
                0.7889996835760071
              ],
              "median_input_tokens": 209.5,
              "n": 32,
              "near_boundary_fraction": 0.34375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.71875
            }
          }
        },
        "persona": {
          "LL": {
            "EMBER": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.625,
              "generated_correct": 20,
              "generated_exact_95_ci": [
                0.4369224550969266,
                0.7889996835760071
              ],
              "generated_logit_choice_agreement": 0.84375,
              "logit_accuracy": 0.65625,
              "logit_correct": 21,
              "logit_exact_95_ci": [
                0.4680689637197313,
                0.8142809082834339
              ],
              "median_input_tokens": 201.0,
              "n": 32,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.875
            },
            "ORCHID": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.78125,
              "generated_accuracy": 0.46875,
              "generated_correct": 15,
              "generated_exact_95_ci": [
                0.2909398229060153,
                0.6525631907136071
              ],
              "generated_logit_choice_agreement": 0.78125,
              "logit_accuracy": 0.625,
              "logit_correct": 20,
              "logit_exact_95_ci": [
                0.4369224550969266,
                0.7889996835760071
              ],
              "median_input_tokens": 202.0,
              "n": 32,
              "near_boundary_fraction": 0.25,
              "parser_coverage": 1.0,
              "raw_zero_margin": 4,
              "strict_parser_coverage": 0.78125
            }
          },
          "LN": {
            "EMBER": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.5,
              "generated_correct": 16,
              "generated_exact_95_ci": [
                0.31887496423416395,
                0.681125035765836
              ],
              "generated_logit_choice_agreement": 0.59375,
              "logit_accuracy": 0.71875,
              "logit_correct": 23,
              "logit_exact_95_ci": [
                0.5325289003570655,
                0.8625430967044982
              ],
              "median_input_tokens": 201.0,
              "n": 32,
              "near_boundary_fraction": 0.28125,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.8125
            },
            "ORCHID": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.84375,
              "generated_accuracy": 0.53125,
              "generated_correct": 17,
              "generated_exact_95_ci": [
                0.34743680928639287,
                0.7090601770939848
              ],
              "generated_logit_choice_agreement": 0.6875,
              "logit_accuracy": 0.65625,
              "logit_correct": 21,
              "logit_exact_95_ci": [
                0.4680689637197313,
                0.8142809082834339
              ],
              "median_input_tokens": 202.0,
              "n": 32,
              "near_boundary_fraction": 0.34375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 3,
              "strict_parser_coverage": 0.78125
            }
          },
          "NL": {
            "EMBER": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.625,
              "generated_correct": 20,
              "generated_exact_95_ci": [
                0.4369224550969266,
                0.7889996835760071
              ],
              "generated_logit_choice_agreement": 0.71875,
              "logit_accuracy": 0.65625,
              "logit_correct": 21,
              "logit_exact_95_ci": [
                0.4680689637197313,
                0.8142809082834339
              ],
              "median_input_tokens": 209.0,
              "n": 32,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 3,
              "strict_parser_coverage": 0.875
            },
            "ORCHID": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.65625,
              "generated_correct": 21,
              "generated_exact_95_ci": [
                0.4680689637197313,
                0.8142809082834339
              ],
              "generated_logit_choice_agreement": 0.65625,
              "logit_accuracy": 0.625,
              "logit_correct": 20,
              "logit_exact_95_ci": [
                0.4369224550969266,
                0.7889996835760071
              ],
              "median_input_tokens": 210.0,
              "n": 32,
              "near_boundary_fraction": 0.40625,
              "parser_coverage": 0.96875,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.78125
            }
          },
          "NN": {
            "EMBER": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.8125,
              "generated_accuracy": 0.625,
              "generated_correct": 20,
              "generated_exact_95_ci": [
                0.4369224550969266,
                0.7889996835760071
              ],
              "generated_logit_choice_agreement": 0.625,
              "logit_accuracy": 0.625,
              "logit_correct": 20,
              "logit_exact_95_ci": [
                0.4369224550969266,
                0.7889996835760071
              ],
              "median_input_tokens": 209.0,
              "n": 32,
              "near_boundary_fraction": 0.375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 2,
              "strict_parser_coverage": 0.6875
            },
            "ORCHID": {
              "descriptive_criterion_met": false,
              "generated_B_fraction": 0.875,
              "generated_accuracy": 0.5625,
              "generated_correct": 18,
              "generated_exact_95_ci": [
                0.3766257315900938,
                0.7363618755890153
              ],
              "generated_logit_choice_agreement": 0.71875,
              "logit_accuracy": 0.65625,
              "logit_correct": 21,
              "logit_exact_95_ci": [
                0.4680689637197313,
                0.8142809082834339
              ],
              "median_input_tokens": 210.0,
              "n": 32,
              "near_boundary_fraction": 0.34375,
              "parser_coverage": 1.0,
              "raw_zero_margin": 1,
              "strict_parser_coverage": 0.71875
            }
          }
        }
      },
      "factor_count": 5,
      "generated_pairs_descriptive": {
        "LL_to_LN": {
          "accuracy_difference_right_minus_left": -0.03125,
          "both_correct": 31,
          "both_wrong": 27,
          "corrections": 2,
          "regressions": 4
        },
        "LL_to_NL": {
          "accuracy_difference_right_minus_left": 0.09375,
          "both_correct": 35,
          "both_wrong": 23,
          "corrections": 6,
          "regressions": 0
        },
        "LL_to_NN": {
          "accuracy_difference_right_minus_left": 0.046875,
          "both_correct": 32,
          "both_wrong": 23,
          "corrections": 6,
          "regressions": 3
        },
        "NL_to_LN": {
          "accuracy_difference_right_minus_left": -0.125,
          "both_correct": 33,
          "both_wrong": 23,
          "corrections": 0,
          "regressions": 8
        },
        "NN_to_LN": {
          "accuracy_difference_right_minus_left": -0.078125,
          "both_correct": 30,
          "both_wrong": 23,
          "corrections": 3,
          "regressions": 8
        },
        "NN_to_NL": {
          "accuracy_difference_right_minus_left": 0.046875,
          "both_correct": 37,
          "both_wrong": 22,
          "corrections": 4,
          "regressions": 1
        }
      },
      "logit_pair_descriptive": {
        "accuracy_difference_right_minus_left": 0.0,
        "both_correct": 33,
        "both_wrong": 15,
        "corrections": 8,
        "regressions": 8
      },
      "near_boundary": {
        "fraction_of_regressions_near_boundary": 0.625,
        "interpretation": "Preregistered output-margin diagnostic, not evidence of an internal mechanism.",
        "median_LL_abs_margin": {
          "concordant": 6.0,
          "discordant": 1.25,
          "regressions": 0.625
        },
        "regressions": 8,
        "regressions_with_LL_abs_margin_at_most_threshold": 5,
        "threshold_abs_raw_LL_margin": 0.75
      },
      "primary": {
        "alpha": 0.05,
        "contrast": "LL vs NN",
        "endpoint": "generated accuracy",
        "neutral_benefit_replicated": false,
        "p_value": 0.5078125,
        "reject_equal_accuracy": false,
        "status": "COMPLETE",
        "test": "exact two-sided McNemar"
      },
      "secondary_model": {
        "effects": {
          "Intercept": {
            "approximate_95_interval": [
              0.4906867646637331,
              1.4514310861933153
            ],
            "log_odds_mean": 0.9710589254285242,
            "posterior_sd": 0.24508783712489343
          },
          "label_loaded": {
            "approximate_95_interval": [
              -1.761534396266891,
              -0.44948228931477263
            ],
            "log_odds_mean": -1.1055083427908319,
            "posterior_sd": 0.33470717014084655
          },
          "label_loaded:task_loaded": {
            "approximate_95_interval": [
              -1.2421077703819687,
              0.6038975650863437
            ],
            "log_odds_mean": -0.31910510264781256,
            "posterior_sd": 0.47091972843579405
          },
          "task_loaded": {
            "approximate_95_interval": [
              0.09664395422504446,
              1.4632032880470875
            ],
            "log_odds_mean": 0.7799236211360661,
            "posterior_sd": 0.3486120749546029
          }
        },
        "fixed_effect_prior_sd": 2.0,
        "fixture_log_sd_mean": 1.4598761193336864,
        "formula": "correct ~ label_loaded * task_loaded + (1 | fixture_id)",
        "method": "BinomialBayesMixedGLM.fit_vb (mean-field variational Bayes)",
        "note": "Secondary localization; approximate marginal posterior intervals, not confirmatory tests.",
        "random_effect_log_sd_prior_sd": 0.5,
        "status": "FIT_OK",
        "warnings": []
      },
      "status": "COMPLETE_DIAGNOSTIC"
    }
  },
  "status": "COMPLETE_DIAGNOSTIC",
  "token_ledger": {
    "completed_prompt_attempts": 640,
    "excludes": "Interrupted uncheckpointed batches; generated slots may include padding after EOS.",
    "generated_token_slots": 1648,
    "padded_input_token_slots": 123328,
    "unpadded_input_tokens": 121388
  }
}
