{
  "schema_version": 1,
  "task_family": "deployment_prediction_set_audit",
  "n_calibration": 4939,
  "n_deployment": 5911,
  "n_labels": 9,
  "n_features": 10,
  "record_id_column": "record_id",
  "label_columns": [
    "label_00",
    "label_01",
    "label_02",
    "label_03",
    "label_04",
    "label_05",
    "label_06",
    "label_07",
    "label_08"
  ],
  "feature_columns": [
    "v000",
    "v001",
    "v002",
    "v003",
    "v004",
    "v005",
    "v006",
    "v007",
    "v008",
    "v009"
  ],
  "score_columns_are_probabilities": true,
  "public_files": [
    "data/metadata.json",
    "data/risk_spec.json",
    "data/calibration_scores.csv",
    "data/calibration_labels.csv",
    "data/calibration_features.csv",
    "data/deployment_scores.csv",
    "data/deployment_features.csv"
  ],
  "output_files": [
    "analysis.py",
    "results/prediction_sets.csv",
    "results/policy_parameters.csv",
    "results/audit_summary.csv",
    "results/set_size_summary.csv",
    "results/partition_diagnostics.csv",
    "results/distribution_diagnostics.csv",
    "results/calibration_diagnostics.csv",
    "results/constraint_audit.csv",
    "results/set_size_tradeoff.png",
    "results/partition_audit.png",
    "results/score_diagnostics.png"
  ],
  "output_schemas": {
    "results/prediction_sets.csv": [
      "record_id",
      "prediction_set"
    ],
    "results/policy_parameters.csv": [
      "proxy_stratum",
      "label",
      "n_calibration",
      "n_calibration_label",
      "threshold",
      "group_empirical_weighted_miss_risk",
      "group_upper_confidence_risk",
      "group_expected_set_size",
      "calibration_target_risk",
      "target_risk"
    ],
    "results/audit_summary.csv": [
      "scope",
      "stratum_id",
      "n_records",
      "empirical_risk",
      "mean_set_size",
      "p90_set_size",
      "target_risk",
      "risk_violation"
    ],
    "results/set_size_summary.csv": [
      "record_id",
      "set_size",
      "proxy_stratum"
    ],
    "results/partition_diagnostics.csv": [
      "scope",
      "stratum_id",
      "n_records",
      "empirical_risk",
      "mean_set_size",
      "p90_set_size",
      "target_risk",
      "risk_margin"
    ],
    "results/distribution_diagnostics.csv": [
      "proxy_stratum",
      "calibration_fraction",
      "deployment_fraction",
      "density_ratio_estimate",
      "absolute_fraction_shift"
    ],
    "results/calibration_diagnostics.csv": [
      "proxy_stratum",
      "bin_id",
      "score_left",
      "score_right",
      "n_records",
      "mean_true_label_score"
    ],
    "results/constraint_audit.csv": [
      "scope",
      "stratum_id",
      "empirical_risk",
      "target_risk",
      "risk_violation",
      "violates"
    ]
  },
  "output_semantics": {
    "results/policy_parameters.csv": {
      "threshold": "Numeric probability cutoff in [0, 1]. Include multiple rows when the policy uses recovered groups or label-specific thresholds.",
      "proxy_stratum": "Identifier for the public or recovered group used by the policy. Use a consistent value across policy and diagnostic files."
    },
    "results/audit_summary.csv": {
      "scope": "Include one deployment-wide row with scope='overall'. Include public or recovered subgroup rows with a scope value containing 'stratum' or 'group', such as 'proxy_stratum' or 'recovered_group'. Do not use hidden deployment labels or reference-only strata.",
      "stratum_id": "Use 'all' for the overall row and a stable public/recovered group identifier for subgroup rows.",
      "empirical_risk": "Numeric estimated weighted miss risk for the corresponding scope."
    },
    "results/set_size_summary.csv": {
      "record_id": "Deployment record id.",
      "set_size": "Number of valid labels listed in the submitted prediction_set for that record.",
      "proxy_stratum": "Public or recovered group id used by the policy, if applicable."
    },
    "results/partition_diagnostics.csv": {
      "scope": "Use public or recovered grouping scopes containing 'stratum' or 'group', such as 'proxy_stratum' or 'recovered_group'.",
      "risk_margin": "Numeric target_risk minus estimated empirical_risk for the row."
    },
    "results/distribution_diagnostics.csv": {
      "proxy_stratum": "Public or recovered group id.",
      "calibration_fraction": "Fraction of calibration records assigned to this group; fractions should sum to approximately 1 across rows.",
      "deployment_fraction": "Fraction of deployment records assigned to this group; fractions should sum to approximately 1 across rows.",
      "density_ratio_estimate": "Deployment fraction divided by calibration fraction, with a stable convention for small calibration fractions."
    },
    "results/calibration_diagnostics.csv": {
      "bin_id": "Stable bin identifier within each public or recovered group.",
      "n_records": "Number of calibration records in the bin.",
      "mean_true_label_score": "Numeric mean classifier score assigned to the true calibration label in the bin."
    },
    "results/constraint_audit.csv": {
      "scope": "Use the same scope convention as audit_summary.csv: include 'overall' and public/recovered subgroup scopes containing 'stratum' or 'group'.",
      "violates": "Boolean-like indicator for whether empirical_risk exceeds target_risk."
    }
  },
  "notes": [
    "Deployment labels are intentionally hidden.",
    "The anonymous features and score tables may contain information useful for choosing a nontrivial policy.",
    "The scoring target is risk control with small prediction sets, not exact matching to a single reference file."
  ]
}