{
  "schema_version": 1,
  "simulation_id": "heybop-creative-plan-simulation-v1",
  "simulation_sha256": "764315bb44e429427fba966cca0906cb374d5e152142e4c3ef181ce73ba2d0c3",
  "source_corpus_sha256": "8bf1887ca237008c3976500225f3829c3fe7a2c73f26a434324bc011dbd10d42",
  "auditor_sha256": "2af90edb0cd73588876e642a9a11438c85c31e271592e717c461b357872a6a95",
  "audit": {
    "valid": true,
    "scope": "Provenance disclosures, source references, full position coverage and arithmetic of authored simulation labels only. No planner performance or causal effect is measured."
  },
  "simulation_counts": {
    "distinct_source_briefs": 24,
    "source_families": 6,
    "scenario_versions": 2,
    "scenario_display_names": {
      "simulated_baseline": "Example A",
      "simulated_revision": "Example B"
    },
    "attempts_per_brief_per_version": 3,
    "simulated_outcome_records": 144,
    "paired_attempt_positions": 72,
    "simulated_baseline": {
      "records": 72,
      "pass": 41,
      "fail": 31,
      "pass_percent": 56.9,
      "fail_percent": 43.1
    },
    "simulated_revision": {
      "records": 72,
      "pass": 57,
      "fail": 15,
      "pass_percent": 79.2,
      "fail_percent": 20.8
    },
    "paired_transitions": {
      "both_pass": 38,
      "fail_to_pass": 19,
      "pass_to_fail": 3,
      "both_fail": 12
    },
    "simulated_pass_count_difference": 16,
    "simulated_pass_percentage_point_difference": 22.2,
    "by_family": {
      "ordinary": {
        "simulated_baseline": {
          "records": 12,
          "pass": 9,
          "fail": 3,
          "pass_percent": 75,
          "fail_percent": 25
        },
        "simulated_revision": {
          "records": 12,
          "pass": 10,
          "fail": 2,
          "pass_percent": 83.3,
          "fail_percent": 16.7
        },
        "paired_transitions": {
          "both_pass": 8,
          "fail_to_pass": 2,
          "pass_to_fail": 1,
          "both_fail": 1
        },
        "simulated_pass_count_difference": 1,
        "simulated_pass_percentage_point_difference": 8.3
      },
      "sparse_unknowns": {
        "simulated_baseline": {
          "records": 12,
          "pass": 5,
          "fail": 7,
          "pass_percent": 41.7,
          "fail_percent": 58.3
        },
        "simulated_revision": {
          "records": 12,
          "pass": 9,
          "fail": 3,
          "pass_percent": 75,
          "fail_percent": 25
        },
        "paired_transitions": {
          "both_pass": 5,
          "fail_to_pass": 4,
          "pass_to_fail": 0,
          "both_fail": 3
        },
        "simulated_pass_count_difference": 4,
        "simulated_pass_percentage_point_difference": 33.3
      },
      "selected_passage": {
        "simulated_baseline": {
          "records": 12,
          "pass": 6,
          "fail": 6,
          "pass_percent": 50,
          "fail_percent": 50
        },
        "simulated_revision": {
          "records": 12,
          "pass": 9,
          "fail": 3,
          "pass_percent": 75,
          "fail_percent": 25
        },
        "paired_transitions": {
          "both_pass": 5,
          "fail_to_pass": 4,
          "pass_to_fail": 1,
          "both_fail": 2
        },
        "simulated_pass_count_difference": 3,
        "simulated_pass_percentage_point_difference": 25
      },
      "continuity": {
        "simulated_baseline": {
          "records": 12,
          "pass": 7,
          "fail": 5,
          "pass_percent": 58.3,
          "fail_percent": 41.7
        },
        "simulated_revision": {
          "records": 12,
          "pass": 9,
          "fail": 3,
          "pass_percent": 75,
          "fail_percent": 25
        },
        "paired_transitions": {
          "both_pass": 6,
          "fail_to_pass": 3,
          "pass_to_fail": 1,
          "both_fail": 2
        },
        "simulated_pass_count_difference": 2,
        "simulated_pass_percentage_point_difference": 16.7
      },
      "local_edit": {
        "simulated_baseline": {
          "records": 12,
          "pass": 7,
          "fail": 5,
          "pass_percent": 58.3,
          "fail_percent": 41.7
        },
        "simulated_revision": {
          "records": 12,
          "pass": 10,
          "fail": 2,
          "pass_percent": 83.3,
          "fail_percent": 16.7
        },
        "paired_transitions": {
          "both_pass": 7,
          "fail_to_pass": 3,
          "pass_to_fail": 0,
          "both_fail": 2
        },
        "simulated_pass_count_difference": 3,
        "simulated_pass_percentage_point_difference": 25
      },
      "untrusted_reference": {
        "simulated_baseline": {
          "records": 12,
          "pass": 7,
          "fail": 5,
          "pass_percent": 58.3,
          "fail_percent": 41.7
        },
        "simulated_revision": {
          "records": 12,
          "pass": 10,
          "fail": 2,
          "pass_percent": 83.3,
          "fail_percent": 16.7
        },
        "paired_transitions": {
          "both_pass": 7,
          "fail_to_pass": 3,
          "pass_to_fail": 0,
          "both_fail": 2
        },
        "simulated_pass_count_difference": 3,
        "simulated_pass_percentage_point_difference": 25
      }
    },
    "by_source_split": {
      "development": {
        "simulated_baseline": {
          "records": 36,
          "pass": 20,
          "fail": 16,
          "pass_percent": 55.6,
          "fail_percent": 44.4
        },
        "simulated_revision": {
          "records": 36,
          "pass": 28,
          "fail": 8,
          "pass_percent": 77.8,
          "fail_percent": 22.2
        },
        "paired_transitions": {
          "both_pass": 18,
          "fail_to_pass": 10,
          "pass_to_fail": 2,
          "both_fail": 6
        },
        "simulated_pass_count_difference": 8,
        "simulated_pass_percentage_point_difference": 22.2
      },
      "reserved": {
        "simulated_baseline": {
          "records": 36,
          "pass": 21,
          "fail": 15,
          "pass_percent": 58.3,
          "fail_percent": 41.7
        },
        "simulated_revision": {
          "records": 36,
          "pass": 29,
          "fail": 7,
          "pass_percent": 80.6,
          "fail_percent": 19.4
        },
        "paired_transitions": {
          "both_pass": 20,
          "fail_to_pass": 9,
          "pass_to_fail": 1,
          "both_fail": 6
        },
        "simulated_pass_count_difference": 8,
        "simulated_pass_percentage_point_difference": 22.2
      }
    },
    "authored_failed_checks_by_criterion": {
      "simulated_baseline": {
        "causality": 2,
        "cast": 2,
        "direction": 6,
        "continuity": 8,
        "unknowns": 3,
        "frame": 1,
        "scope": 5,
        "passage": 1,
        "local_edit": 2,
        "format": 1
      },
      "simulated_revision": {
        "continuity": 4,
        "causality": 2,
        "direction": 3,
        "frame": 1,
        "scope": 1,
        "passage": 1,
        "props": 1,
        "untrusted": 1,
        "format": 1
      }
    }
  },
  "observed_evidence": {
    "generated_planner_responses": 0,
    "hamming_runs": 0,
    "independent_human_ratings": 0,
    "actual_prompt_versions_tested": 0,
    "model_performance_measured": false,
    "hamming_caused_improvement_established": false,
    "reserved_cases_are_unseen_holdout": false
  },
  "interpretation": "All rates and before/after differences summarize deliberately authored scenario labels. They are not product performance, generated-response scores, human preferences or evidence of an effect caused by Hamming. There are 24 distinct briefs; repeated positions are not independent briefs."
}