File size: 5,664 Bytes
1834e53
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
{
  "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."
  ]
}