@aiwg/cli 2026.9.24 → 2026.10.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (207) hide show
  1. package/agentic/code/providers/capability-matrix.yaml +2 -1
  2. package/dist/src/artifacts/discovery-shadow-evaluator.js +177 -0
  3. package/dist/src/artifacts/discovery-shadow.js +195 -0
  4. package/dist/src/artifacts/index-builder.js +47 -9
  5. package/dist/src/artifacts/query-engine.js +94 -1
  6. package/dist/src/artifacts/types.js +1 -0
  7. package/dist/src/cli/handlers/decision.js +88 -0
  8. package/dist/src/cli/handlers/gates.js +154 -0
  9. package/dist/src/cli/handlers/index.js +5 -1
  10. package/dist/src/cli/handlers/sessions.js +10 -7
  11. package/dist/src/cli/handlers/utilities.js +20 -0
  12. package/dist/src/config/aiwg-config.js +15 -2
  13. package/dist/src/decision/acceptance.js +25 -0
  14. package/dist/src/decision/adapters/jev.js +24 -2
  15. package/dist/src/decision/calibration/registry.js +9 -0
  16. package/dist/src/decision/compose.js +74 -15
  17. package/dist/src/decision/context-live-promotion.js +344 -0
  18. package/dist/src/decision/context-live-qualification.js +591 -0
  19. package/dist/src/decision/context-pruning.js +847 -0
  20. package/dist/src/decision/driver.js +430 -0
  21. package/dist/src/decision/egress-live-qualification.js +874 -0
  22. package/dist/src/decision/ensemble/contract.js +64 -1
  23. package/dist/src/decision/ensemble/index.js +1 -0
  24. package/dist/src/decision/ensemble/report.js +8 -42
  25. package/dist/src/decision/ensemble/runtime.js +571 -0
  26. package/dist/src/decision/ensemble-study/artifacts.js +85 -0
  27. package/dist/src/decision/ensemble-study/calibration.js +443 -0
  28. package/dist/src/decision/ensemble-study/corpus.js +229 -0
  29. package/dist/src/decision/ensemble-study/multifact-artifacts.js +23 -0
  30. package/dist/src/decision/ensemble-study/multifact-study.js +596 -0
  31. package/dist/src/decision/ensemble-study/multifact.js +241 -0
  32. package/dist/src/decision/ensemble-study/protocol.js +77 -0
  33. package/dist/src/decision/ensemble-study/score.js +462 -0
  34. package/dist/src/decision/ensemble-study/staged.js +87 -0
  35. package/dist/src/decision/ensemble-study/statistics.js +84 -0
  36. package/dist/src/decision/evaluate.js +227 -14
  37. package/dist/src/decision/export.js +4 -0
  38. package/dist/src/decision/feature-export.js +1074 -0
  39. package/dist/src/decision/graph-index.js +16 -0
  40. package/dist/src/decision/graph-live-qualification.js +1041 -0
  41. package/dist/src/decision/graph-live-workload.js +245 -0
  42. package/dist/src/decision/graph-skill-bridge.js +21 -3
  43. package/dist/src/decision/heldout/calibration.js +140 -0
  44. package/dist/src/decision/heldout/collector.js +430 -0
  45. package/dist/src/decision/heldout/contract.js +309 -0
  46. package/dist/src/decision/heldout/d29-generator-ids.js +12 -0
  47. package/dist/src/decision/heldout/d29-passage-baseline-v2.js +93 -0
  48. package/dist/src/decision/heldout/d29-passage-baseline-v3.js +100 -0
  49. package/dist/src/decision/heldout/d29-passage-baseline.js +66 -0
  50. package/dist/src/decision/heldout/d29-pools-v8.js +575 -0
  51. package/dist/src/decision/heldout/d29-pools.js +288 -0
  52. package/dist/src/decision/heldout/d29-v6.js +169 -0
  53. package/dist/src/decision/heldout/d29-v7.js +202 -0
  54. package/dist/src/decision/heldout/d29-v8.js +586 -0
  55. package/dist/src/decision/heldout/generator-registry.js +67 -0
  56. package/dist/src/decision/heldout/generators.js +696 -0
  57. package/dist/src/decision/heldout/index.js +8 -0
  58. package/dist/src/decision/heldout/journal.js +346 -0
  59. package/dist/src/decision/heldout/study-generators.js +64 -0
  60. package/dist/src/decision/heldout/types.js +2 -0
  61. package/dist/src/decision/host-policy.js +61 -0
  62. package/dist/src/decision/index.js +11 -0
  63. package/dist/src/decision/interop-profile.v1.json +112 -0
  64. package/dist/src/decision/interop.js +653 -0
  65. package/dist/src/decision/issue-triage/index.js +3 -0
  66. package/dist/src/decision/issue-triage/runtime.js +1069 -0
  67. package/dist/src/decision/issue-triage/types.js +2 -0
  68. package/dist/src/decision/patterns/catalog.js +28 -8
  69. package/dist/src/decision/patterns/playground.js +28 -11
  70. package/dist/src/decision/patterns/specs.js +19 -10
  71. package/dist/src/decision/preprocessed-evidence.js +724 -0
  72. package/dist/src/decision/qualification/capture.js +8 -2
  73. package/dist/src/decision/qualification/gate-evidence.js +19 -4
  74. package/dist/src/decision/qualification/quality.js +57 -49
  75. package/dist/src/decision/qualification/release.js +76 -3
  76. package/dist/src/decision/qualification/runner.js +5 -2
  77. package/dist/src/decision/routing/contract.js +175 -0
  78. package/dist/src/decision/routing/index.js +5 -0
  79. package/dist/src/decision/routing/report.js +216 -0
  80. package/dist/src/decision/routing/runtime.js +685 -0
  81. package/dist/src/decision/routing/types.js +2 -0
  82. package/dist/src/decision/sdlc-screening.js +1157 -0
  83. package/dist/src/decision/sensitivity/comparative.js +154 -0
  84. package/dist/src/decision/sensitivity/contract.js +218 -0
  85. package/dist/src/decision/sensitivity/index.js +5 -0
  86. package/dist/src/decision/sensitivity/runtime.js +912 -0
  87. package/dist/src/decision/sensitivity/types.js +2 -0
  88. package/dist/src/decision/validate.js +26 -0
  89. package/dist/src/extensions/commands/definitions.js +38 -0
  90. package/dist/src/extensions/manifest.js +44 -0
  91. package/dist/src/gates/discovery.js +315 -0
  92. package/dist/src/gates/driver.js +388 -0
  93. package/dist/src/gates/evaluate.js +719 -0
  94. package/dist/src/gates/floors.js +255 -0
  95. package/dist/src/gates/index.js +17 -0
  96. package/dist/src/gates/providers/evidence.js +38 -0
  97. package/dist/src/gates/providers/index.js +21 -0
  98. package/dist/src/gates/providers/isolated-child-source.js +446 -0
  99. package/dist/src/gates/providers/loader.js +1069 -0
  100. package/dist/src/gates/providers/paired.js +52 -0
  101. package/dist/src/gates/providers/proportion.js +60 -0
  102. package/dist/src/gates/providers/registry.js +40 -0
  103. package/dist/src/gates/providers/scalar.js +41 -0
  104. package/dist/src/gates/providers/screening.js +53 -0
  105. package/dist/src/gates/providers/types.js +25 -0
  106. package/dist/src/gates/registry.js +1003 -0
  107. package/dist/src/gates/report.js +102 -0
  108. package/dist/src/gates/schema.js +56 -0
  109. package/dist/src/gates/stats/binomial.js +26 -0
  110. package/dist/src/gates/stats/bootstrap.js +62 -0
  111. package/dist/src/gates/stats/clopper-pearson.js +95 -0
  112. package/dist/src/gates/stats/digest.js +26 -0
  113. package/dist/src/gates/stats/error.js +5 -0
  114. package/dist/src/gates/stats/index.js +9 -0
  115. package/dist/src/gates/stats/normal.js +35 -0
  116. package/dist/src/gates/stats/paired.js +120 -0
  117. package/dist/src/gates/stats/splits.js +30 -0
  118. package/dist/src/gates/types.js +8 -0
  119. package/dist/src/governance/redaction.js +14 -4
  120. package/dist/src/mcp/muse-native-config.mjs +123 -0
  121. package/dist/src/mcp/tools/decision.mjs +188 -0
  122. package/dist/src/mcp/tools/subsystems.mjs +2 -0
  123. package/dist/src/providers/capability-matrix.yaml +2 -1
  124. package/dist/src/providers/provider-definitions.js +2 -1
  125. package/dist/src/sessions/adapters/muse.js +336 -39
  126. package/dist/src/sessions/batch-import.js +7 -0
  127. package/dist/src/sessions/workspace-discovery.js +13 -2
  128. package/package.json +1 -1
  129. package/schemas/decision/ConformalFrozenOpenData.v2.schema.json +371 -0
  130. package/schemas/decision/ConformalOpenDataReport.v2.schema.json +421 -0
  131. package/schemas/decision/ConformalPreregistration.v2.schema.json +529 -0
  132. package/schemas/decision/ContextPruningCandidate.v1.schema.json +47 -0
  133. package/schemas/decision/ContextPruningEvaluationReport.v1.schema.json +758 -0
  134. package/schemas/decision/ContextPruningPreregistration.v1.schema.json +175 -0
  135. package/schemas/decision/ContextPruningReceipt.v1.schema.json +72 -0
  136. package/schemas/decision/D17Multifact.v1.schema.json +1851 -0
  137. package/schemas/decision/D17StudyAnalysis.v1.schema.json +42 -0
  138. package/schemas/decision/D17StudyAnalysis.v2.schema.json +51 -0
  139. package/schemas/decision/D17StudyCalibration.v1.schema.json +819 -0
  140. package/schemas/decision/D17StudyDryRun.v1.schema.json +169 -0
  141. package/schemas/decision/D17StudyDryRun.v2.schema.json +239 -0
  142. package/schemas/decision/D17StudyGold.v1.schema.json +108 -0
  143. package/schemas/decision/D17StudyGuide.v1.schema.json +63 -0
  144. package/schemas/decision/D17StudyNativeTemplates.v1.schema.json +619 -0
  145. package/schemas/decision/D17StudyProtocol.v1.schema.json +241 -0
  146. package/schemas/decision/D17StudyReport.v1.schema.json +902 -0
  147. package/schemas/decision/D17StudyReport.v2.schema.json +1319 -0
  148. package/schemas/decision/D17StudyReview.v1.schema.json +175 -0
  149. package/schemas/decision/D17StudySplits.v1.schema.json +217 -0
  150. package/schemas/decision/D29Study.v1.schema.json +1903 -0
  151. package/schemas/decision/D29Study.v2.schema.json +1344 -0
  152. package/schemas/decision/D29Study.v3.schema.json +609 -0
  153. package/schemas/decision/D29Study.v4.schema.json +1133 -0
  154. package/schemas/decision/D29Study.v6.schema.json +1205 -0
  155. package/schemas/decision/D29Study.v7.schema.json +1471 -0
  156. package/schemas/decision/D29Study.v8.schema.json +671 -0
  157. package/schemas/decision/D29Study.v9.schema.json +778 -0
  158. package/schemas/decision/DecisionComparativeReplayAuditSample.v1.schema.json +137 -0
  159. package/schemas/decision/DecisionComparativeReplayCorpus.v1.schema.json +258 -0
  160. package/schemas/decision/DecisionComparativeReplayFreeze.v1.schema.json +74 -0
  161. package/schemas/decision/DecisionComparativeReplayGold.v1.schema.json +50 -0
  162. package/schemas/decision/DecisionComparativeReplayGoldAsset.v1.schema.json +37 -0
  163. package/schemas/decision/DecisionComparativeReplayIntegrity.v1.schema.json +81 -0
  164. package/schemas/decision/DecisionComparativeReplayMembers.v1.schema.json +52 -0
  165. package/schemas/decision/DecisionComparativeReplayPreregistration.v1.schema.json +113 -0
  166. package/schemas/decision/DecisionComparativeReplayReport.v1.schema.json +407 -0
  167. package/schemas/decision/DecisionComparativeReplayReviewPackets.v1.schema.json +260 -0
  168. package/schemas/decision/DecisionComparativeReplaySplit.v1.schema.json +84 -0
  169. package/schemas/decision/DecisionFeatureObservation.v1.schema.json +564 -0
  170. package/schemas/decision/DecisionFeatureSet.v1.schema.json +155 -0
  171. package/schemas/decision/DecisionInteropMapping.v1.schema.json +85 -0
  172. package/schemas/decision/DecisionIssueTriageEvaluationManifest.v1.schema.json +174 -0
  173. package/schemas/decision/DecisionIssueTriageEvaluationReport.v1.schema.json +280 -0
  174. package/schemas/decision/DecisionIssueTriagePilotPack.v1.schema.json +114 -0
  175. package/schemas/decision/DecisionRouteReceipt.v1.schema.json +155 -0
  176. package/schemas/decision/DecisionRoutingPolicy.v1.schema.json +132 -0
  177. package/schemas/decision/DecisionRoutingShadowReport.v1.schema.json +170 -0
  178. package/schemas/decision/DecisionRoutingTask.v1.schema.json +49 -0
  179. package/schemas/decision/DecisionRuleset.schema.json +8 -2
  180. package/schemas/decision/DecisionRuleset.v1alpha2.schema.json +8 -2
  181. package/schemas/decision/DecisionSensitivityPlan.v1.schema.json +191 -0
  182. package/schemas/decision/DecisionSensitivityReport.v1.schema.json +163 -0
  183. package/schemas/decision/HeldoutApproval.v1.schema.json +312 -0
  184. package/schemas/decision/HeldoutAttempt.v1.schema.json +211 -0
  185. package/schemas/decision/HeldoutBaseline.v1.schema.json +90 -0
  186. package/schemas/decision/HeldoutCalibrationPhase.v1.schema.json +87 -0
  187. package/schemas/decision/HeldoutCorpus.v1.schema.json +167 -0
  188. package/schemas/decision/HeldoutEvent.v1.schema.json +41 -0
  189. package/schemas/decision/HeldoutFrozen.v1.schema.json +142 -0
  190. package/schemas/decision/HeldoutPreregistration.v1.schema.json +222 -0
  191. package/schemas/decision/HeldoutSpendEvent.v1.schema.json +122 -0
  192. package/schemas/decision/HeldoutSpendHead.v1.schema.json +25 -0
  193. package/schemas/decision/HeldoutSpendHead.v2.schema.json +46 -0
  194. package/schemas/decision/HeldoutSummary.v1.schema.json +101 -0
  195. package/schemas/decision/OpaDecisionLogExport.v1.schema.json +65 -0
  196. package/schemas/decision/PreprocessedEvidence.v1.schema.json +306 -0
  197. package/schemas/decision/RulesetResult.v1alpha2.schema.json +175 -0
  198. package/schemas/decision/SdlcEvidenceScreening.v1.schema.json +499 -0
  199. package/schemas/decision/SdlcScreeningPreregistration.v1.schema.json +132 -0
  200. package/schemas/decision/SdlcScreeningPreregistration.v2.schema.json +138 -0
  201. package/schemas/decision/SdlcScreeningRelease.v1.schema.json +654 -0
  202. package/schemas/decision/SdlcScreeningRelease.v2.schema.json +654 -0
  203. package/tools/agents/deploy-agents.mjs +1 -1
  204. package/tools/agents/providers/base.mjs +26 -7
  205. package/tools/agents/providers/factory.mjs +10 -2
  206. package/tools/agents/providers/muse-hooks.mjs +1 -29
  207. package/tools/agents/providers/muse.mjs +6 -3
@@ -0,0 +1,1851 @@
1
+ {
2
+ "$schema": "https://json-schema.org/draft/2020-12/schema",
3
+ "$id": "https://aiwg.io/schemas/decision/D17Multifact.v1.schema.json",
4
+ "title": "D17Multifact",
5
+ "description": "D17-MF calibrated multi-fact probe (#2850): analysis and protocol, development review (template and completed), shortcut audit and report.",
6
+ "$defs": {
7
+ "analysis": {
8
+ "type": "object",
9
+ "additionalProperties": false,
10
+ "required": [
11
+ "schemaVersion",
12
+ "protocol",
13
+ "sourceDigests",
14
+ "auditDigest",
15
+ "dispatchOrderDigest",
16
+ "cells"
17
+ ],
18
+ "properties": {
19
+ "schemaVersion": {
20
+ "const": "decision-d17mf-analysis/v2"
21
+ },
22
+ "protocol": {
23
+ "type": "object",
24
+ "additionalProperties": false,
25
+ "required": [
26
+ "schemaVersion",
27
+ "issue",
28
+ "study",
29
+ "generator",
30
+ "design",
31
+ "metrics",
32
+ "calibration",
33
+ "interval",
34
+ "multiplicity",
35
+ "coverage",
36
+ "scoringClock",
37
+ "hypotheses",
38
+ "decisionRules",
39
+ "shortcutAudit",
40
+ "budget",
41
+ "promotion"
42
+ ],
43
+ "properties": {
44
+ "schemaVersion": {
45
+ "const": "decision-d17mf-protocol/v2"
46
+ },
47
+ "issue": {
48
+ "const": "roctinam/aiwg#2850"
49
+ },
50
+ "study": {
51
+ "const": "D17"
52
+ },
53
+ "generator": {
54
+ "const": "d17-multifact/v1"
55
+ },
56
+ "design": {
57
+ "type": "object"
58
+ },
59
+ "metrics": {
60
+ "type": "object"
61
+ },
62
+ "calibration": {
63
+ "type": "object"
64
+ },
65
+ "interval": {
66
+ "type": "object"
67
+ },
68
+ "multiplicity": {
69
+ "type": "object"
70
+ },
71
+ "coverage": {
72
+ "type": "object",
73
+ "additionalProperties": false,
74
+ "required": [
75
+ "minimumObservedFraction",
76
+ "rule"
77
+ ],
78
+ "properties": {
79
+ "minimumObservedFraction": {
80
+ "const": 0.95
81
+ },
82
+ "rule": {
83
+ "type": "string"
84
+ }
85
+ }
86
+ },
87
+ "scoringClock": {
88
+ "type": "string"
89
+ },
90
+ "hypotheses": {
91
+ "type": "array",
92
+ "minItems": 5,
93
+ "maxItems": 5
94
+ },
95
+ "decisionRules": {
96
+ "type": "object"
97
+ },
98
+ "shortcutAudit": {
99
+ "type": "object"
100
+ },
101
+ "budget": {
102
+ "type": "string"
103
+ },
104
+ "promotion": {
105
+ "const": "none: a diagnostic probe of the provider; the study decision is always HOLD"
106
+ }
107
+ }
108
+ },
109
+ "sourceDigests": {
110
+ "type": "object",
111
+ "minProperties": 1,
112
+ "additionalProperties": {
113
+ "type": "string",
114
+ "pattern": "^sha256:[a-f0-9]{64}$"
115
+ }
116
+ },
117
+ "auditDigest": {
118
+ "type": "string",
119
+ "pattern": "^sha256:[a-f0-9]{64}$"
120
+ },
121
+ "dispatchOrderDigest": {
122
+ "type": "string",
123
+ "pattern": "^sha256:[a-f0-9]{64}$"
124
+ },
125
+ "cells": {
126
+ "type": "array",
127
+ "minItems": 78,
128
+ "maxItems": 78,
129
+ "items": {
130
+ "type": "object",
131
+ "additionalProperties": false,
132
+ "required": [
133
+ "split",
134
+ "key",
135
+ "rule",
136
+ "count"
137
+ ],
138
+ "properties": {
139
+ "split": {
140
+ "enum": [
141
+ "tuning",
142
+ "test"
143
+ ]
144
+ },
145
+ "key": {
146
+ "type": "string"
147
+ },
148
+ "rule": {
149
+ "enum": [
150
+ "exhaustive",
151
+ "d17"
152
+ ]
153
+ },
154
+ "count": {
155
+ "type": "integer",
156
+ "minimum": 1
157
+ }
158
+ }
159
+ }
160
+ }
161
+ }
162
+ },
163
+ "reviewTemplate": {
164
+ "type": "object",
165
+ "additionalProperties": false,
166
+ "required": [
167
+ "schemaVersion",
168
+ "reviewer",
169
+ "preregistrationReview",
170
+ "assessments",
171
+ "instructions"
172
+ ],
173
+ "properties": {
174
+ "schemaVersion": {
175
+ "const": "decision-d17mf-review/v1"
176
+ },
177
+ "reviewer": {
178
+ "const": null
179
+ },
180
+ "preregistrationReview": {
181
+ "const": null
182
+ },
183
+ "assessments": {
184
+ "type": "array",
185
+ "minItems": 40,
186
+ "maxItems": 40,
187
+ "items": {
188
+ "type": "object",
189
+ "additionalProperties": false,
190
+ "required": [
191
+ "assessmentId",
192
+ "rowId",
193
+ "cell",
194
+ "inputDigest",
195
+ "payload",
196
+ "reviewedAt",
197
+ "goldAuditLabel",
198
+ "goldAmbiguousOrIncorrect",
199
+ "rationale"
200
+ ],
201
+ "properties": {
202
+ "assessmentId": {
203
+ "type": "string",
204
+ "pattern": "^mf-assessment-[0-9]{2}$"
205
+ },
206
+ "rowId": {
207
+ "type": "string",
208
+ "pattern": "^[0-9a-f]{64}$"
209
+ },
210
+ "cell": {
211
+ "type": "string",
212
+ "pattern": "^[A-Za-z0-9][A-Za-z0-9_-]{0,63}$"
213
+ },
214
+ "inputDigest": {
215
+ "type": "string",
216
+ "pattern": "^sha256:[a-f0-9]{64}$"
217
+ },
218
+ "payload": {
219
+ "type": "string",
220
+ "minLength": 1,
221
+ "maxLength": 8192
222
+ },
223
+ "reviewedAt": {
224
+ "anyOf": [
225
+ {
226
+ "type": "null"
227
+ },
228
+ {
229
+ "type": "string",
230
+ "pattern": "^\\d{4}-\\d{2}-\\d{2}T\\d{2}:\\d{2}:\\d{2}(\\.\\d{3})?Z$"
231
+ }
232
+ ]
233
+ },
234
+ "goldAuditLabel": {
235
+ "anyOf": [
236
+ {
237
+ "type": "null"
238
+ },
239
+ {
240
+ "enum": [
241
+ "yes",
242
+ "no"
243
+ ]
244
+ }
245
+ ]
246
+ },
247
+ "goldAmbiguousOrIncorrect": {
248
+ "anyOf": [
249
+ {
250
+ "type": "null"
251
+ },
252
+ {
253
+ "type": "boolean"
254
+ }
255
+ ]
256
+ },
257
+ "rationale": {
258
+ "anyOf": [
259
+ {
260
+ "type": "null"
261
+ },
262
+ {
263
+ "type": "string",
264
+ "minLength": 1,
265
+ "maxLength": 8192
266
+ }
267
+ ]
268
+ }
269
+ }
270
+ }
271
+ },
272
+ "instructions": {
273
+ "type": "string",
274
+ "minLength": 1
275
+ }
276
+ }
277
+ },
278
+ "review": {
279
+ "type": "object",
280
+ "additionalProperties": false,
281
+ "required": [
282
+ "schemaVersion",
283
+ "reviewer",
284
+ "preregistrationReview",
285
+ "assessments",
286
+ "instructions"
287
+ ],
288
+ "properties": {
289
+ "schemaVersion": {
290
+ "const": "decision-d17mf-review/v1"
291
+ },
292
+ "reviewer": {
293
+ "type": "string",
294
+ "minLength": 1,
295
+ "maxLength": 512
296
+ },
297
+ "preregistrationReview": {
298
+ "type": "string",
299
+ "minLength": 1,
300
+ "maxLength": 8192
301
+ },
302
+ "assessments": {
303
+ "type": "array",
304
+ "minItems": 40,
305
+ "maxItems": 40,
306
+ "items": {
307
+ "type": "object",
308
+ "additionalProperties": false,
309
+ "required": [
310
+ "assessmentId",
311
+ "rowId",
312
+ "cell",
313
+ "inputDigest",
314
+ "payload",
315
+ "reviewedAt",
316
+ "goldAuditLabel",
317
+ "goldAmbiguousOrIncorrect",
318
+ "rationale"
319
+ ],
320
+ "properties": {
321
+ "assessmentId": {
322
+ "type": "string",
323
+ "pattern": "^mf-assessment-[0-9]{2}$"
324
+ },
325
+ "rowId": {
326
+ "type": "string",
327
+ "pattern": "^[0-9a-f]{64}$"
328
+ },
329
+ "cell": {
330
+ "type": "string",
331
+ "pattern": "^[A-Za-z0-9][A-Za-z0-9_-]{0,63}$"
332
+ },
333
+ "inputDigest": {
334
+ "type": "string",
335
+ "pattern": "^sha256:[a-f0-9]{64}$"
336
+ },
337
+ "payload": {
338
+ "type": "string",
339
+ "minLength": 1,
340
+ "maxLength": 8192
341
+ },
342
+ "reviewedAt": {
343
+ "anyOf": [
344
+ {
345
+ "type": "null"
346
+ },
347
+ {
348
+ "type": "string",
349
+ "pattern": "^\\d{4}-\\d{2}-\\d{2}T\\d{2}:\\d{2}:\\d{2}(\\.\\d{3})?Z$"
350
+ }
351
+ ]
352
+ },
353
+ "goldAuditLabel": {
354
+ "anyOf": [
355
+ {
356
+ "type": "null"
357
+ },
358
+ {
359
+ "enum": [
360
+ "yes",
361
+ "no"
362
+ ]
363
+ }
364
+ ]
365
+ },
366
+ "goldAmbiguousOrIncorrect": {
367
+ "anyOf": [
368
+ {
369
+ "type": "null"
370
+ },
371
+ {
372
+ "type": "boolean"
373
+ }
374
+ ]
375
+ },
376
+ "rationale": {
377
+ "anyOf": [
378
+ {
379
+ "type": "null"
380
+ },
381
+ {
382
+ "type": "string",
383
+ "minLength": 1,
384
+ "maxLength": 8192
385
+ }
386
+ ]
387
+ }
388
+ }
389
+ }
390
+ },
391
+ "instructions": {
392
+ "type": "string",
393
+ "minLength": 1
394
+ }
395
+ }
396
+ },
397
+ "report": {
398
+ "type": "object",
399
+ "additionalProperties": false,
400
+ "required": [
401
+ "schemaVersion",
402
+ "issue",
403
+ "scope",
404
+ "calibrated",
405
+ "primaryMetric",
406
+ "calibration",
407
+ "corpusDigest",
408
+ "preregistrationDigest",
409
+ "analysisDigest",
410
+ "integrityDigest",
411
+ "auditDigest",
412
+ "coverage",
413
+ "overall",
414
+ "cells",
415
+ "byFailurePosition",
416
+ "auroc",
417
+ "hypotheses",
418
+ "decision",
419
+ "promotion"
420
+ ],
421
+ "properties": {
422
+ "schemaVersion": {
423
+ "const": "decision-d17mf-report/v2"
424
+ },
425
+ "issue": {
426
+ "const": "roctinam/aiwg#2850"
427
+ },
428
+ "scope": {
429
+ "const": "synthetic-calibrated-probe"
430
+ },
431
+ "calibrated": {
432
+ "const": true
433
+ },
434
+ "primaryMetric": {
435
+ "type": "string"
436
+ },
437
+ "calibration": {
438
+ "type": "object",
439
+ "additionalProperties": false,
440
+ "required": [
441
+ "mode",
442
+ "calibrationSetDigest",
443
+ "memberArtifactDigest",
444
+ "mappingDigest",
445
+ "compatibility",
446
+ "qualifiedAt",
447
+ "collectionEndedAt",
448
+ "expiresAt"
449
+ ],
450
+ "properties": {
451
+ "mode": {
452
+ "const": "artifact"
453
+ },
454
+ "calibrationSetDigest": {
455
+ "type": "string",
456
+ "pattern": "^sha256:[a-f0-9]{64}$"
457
+ },
458
+ "memberArtifactDigest": {
459
+ "type": "string",
460
+ "pattern": "^sha256:[a-f0-9]{64}$"
461
+ },
462
+ "mappingDigest": {
463
+ "type": "string",
464
+ "pattern": "^sha256:[a-f0-9]{64}$"
465
+ },
466
+ "compatibility": {
467
+ "type": "object",
468
+ "additionalProperties": false,
469
+ "required": [
470
+ "action",
471
+ "state"
472
+ ],
473
+ "properties": {
474
+ "action": {
475
+ "const": "allow"
476
+ },
477
+ "state": {
478
+ "enum": [
479
+ "exact",
480
+ "approved-compatible"
481
+ ]
482
+ }
483
+ }
484
+ },
485
+ "qualifiedAt": {
486
+ "type": "string",
487
+ "pattern": "^\\d{4}-\\d{2}-\\d{2}T\\d{2}:\\d{2}:\\d{2}(\\.\\d{3})?Z$"
488
+ },
489
+ "collectionEndedAt": {
490
+ "type": "string",
491
+ "pattern": "^\\d{4}-\\d{2}-\\d{2}T\\d{2}:\\d{2}:\\d{2}(\\.\\d{3})?Z$"
492
+ },
493
+ "expiresAt": {
494
+ "type": "string",
495
+ "pattern": "^\\d{4}-\\d{2}-\\d{2}T\\d{2}:\\d{2}:\\d{2}(\\.\\d{3})?Z$"
496
+ }
497
+ }
498
+ },
499
+ "corpusDigest": {
500
+ "type": "string",
501
+ "pattern": "^sha256:[a-f0-9]{64}$"
502
+ },
503
+ "preregistrationDigest": {
504
+ "type": "string",
505
+ "pattern": "^sha256:[a-f0-9]{64}$"
506
+ },
507
+ "analysisDigest": {
508
+ "type": "string",
509
+ "pattern": "^sha256:[a-f0-9]{64}$"
510
+ },
511
+ "integrityDigest": {
512
+ "type": "string",
513
+ "pattern": "^sha256:[a-f0-9]{64}$"
514
+ },
515
+ "auditDigest": {
516
+ "type": "string",
517
+ "pattern": "^sha256:[a-f0-9]{64}$"
518
+ },
519
+ "coverage": {
520
+ "type": "object",
521
+ "additionalProperties": false,
522
+ "required": [
523
+ "testRows",
524
+ "observed",
525
+ "developmentRows",
526
+ "minimumObservedFraction",
527
+ "complete"
528
+ ],
529
+ "properties": {
530
+ "testRows": {
531
+ "type": "integer"
532
+ },
533
+ "observed": {
534
+ "type": "integer"
535
+ },
536
+ "developmentRows": {
537
+ "type": "integer"
538
+ },
539
+ "minimumObservedFraction": {
540
+ "type": "number"
541
+ },
542
+ "complete": {
543
+ "type": "boolean"
544
+ }
545
+ }
546
+ },
547
+ "overall": {
548
+ "type": "object",
549
+ "additionalProperties": false,
550
+ "required": [
551
+ "n",
552
+ "observed",
553
+ "coverage",
554
+ "raw",
555
+ "calibrated"
556
+ ],
557
+ "properties": {
558
+ "n": {
559
+ "type": "integer",
560
+ "minimum": 0
561
+ },
562
+ "observed": {
563
+ "type": "integer",
564
+ "minimum": 0
565
+ },
566
+ "coverage": {
567
+ "anyOf": [
568
+ {
569
+ "type": "null"
570
+ },
571
+ {
572
+ "type": "number"
573
+ }
574
+ ]
575
+ },
576
+ "raw": {
577
+ "type": "object",
578
+ "additionalProperties": false,
579
+ "required": [
580
+ "correct",
581
+ "accuracy",
582
+ "wilson95",
583
+ "saidYesRate"
584
+ ],
585
+ "properties": {
586
+ "correct": {
587
+ "type": "integer",
588
+ "minimum": 0
589
+ },
590
+ "accuracy": {
591
+ "anyOf": [
592
+ {
593
+ "type": "null"
594
+ },
595
+ {
596
+ "type": "number"
597
+ }
598
+ ]
599
+ },
600
+ "wilson95": {
601
+ "anyOf": [
602
+ {
603
+ "type": "null"
604
+ },
605
+ {
606
+ "type": "array",
607
+ "minItems": 2,
608
+ "maxItems": 2,
609
+ "items": {
610
+ "type": "number"
611
+ }
612
+ }
613
+ ]
614
+ },
615
+ "saidYesRate": {
616
+ "anyOf": [
617
+ {
618
+ "type": "null"
619
+ },
620
+ {
621
+ "type": "number"
622
+ }
623
+ ]
624
+ }
625
+ }
626
+ },
627
+ "calibrated": {
628
+ "type": "object",
629
+ "additionalProperties": false,
630
+ "required": [
631
+ "correct",
632
+ "accuracy",
633
+ "wilson95",
634
+ "meanConfidence",
635
+ "confidenceMinusAccuracy",
636
+ "ece",
637
+ "calibrationUnreliable"
638
+ ],
639
+ "properties": {
640
+ "correct": {
641
+ "type": "integer",
642
+ "minimum": 0
643
+ },
644
+ "accuracy": {
645
+ "anyOf": [
646
+ {
647
+ "type": "null"
648
+ },
649
+ {
650
+ "type": "number"
651
+ }
652
+ ]
653
+ },
654
+ "wilson95": {
655
+ "anyOf": [
656
+ {
657
+ "type": "null"
658
+ },
659
+ {
660
+ "type": "array",
661
+ "minItems": 2,
662
+ "maxItems": 2,
663
+ "items": {
664
+ "type": "number"
665
+ }
666
+ }
667
+ ]
668
+ },
669
+ "meanConfidence": {
670
+ "anyOf": [
671
+ {
672
+ "type": "null"
673
+ },
674
+ {
675
+ "type": "number"
676
+ }
677
+ ]
678
+ },
679
+ "confidenceMinusAccuracy": {
680
+ "anyOf": [
681
+ {
682
+ "type": "null"
683
+ },
684
+ {
685
+ "type": "number"
686
+ }
687
+ ]
688
+ },
689
+ "ece": {
690
+ "anyOf": [
691
+ {
692
+ "type": "null"
693
+ },
694
+ {
695
+ "type": "number"
696
+ }
697
+ ]
698
+ },
699
+ "calibrationUnreliable": {
700
+ "anyOf": [
701
+ {
702
+ "type": "null"
703
+ },
704
+ {
705
+ "type": "boolean"
706
+ }
707
+ ]
708
+ }
709
+ }
710
+ }
711
+ }
712
+ },
713
+ "cells": {
714
+ "type": "object",
715
+ "minProperties": 39,
716
+ "maxProperties": 39,
717
+ "additionalProperties": {
718
+ "type": "object",
719
+ "additionalProperties": false,
720
+ "required": [
721
+ "n",
722
+ "observed",
723
+ "coverage",
724
+ "raw",
725
+ "calibrated"
726
+ ],
727
+ "properties": {
728
+ "n": {
729
+ "type": "integer",
730
+ "minimum": 0
731
+ },
732
+ "observed": {
733
+ "type": "integer",
734
+ "minimum": 0
735
+ },
736
+ "coverage": {
737
+ "anyOf": [
738
+ {
739
+ "type": "null"
740
+ },
741
+ {
742
+ "type": "number"
743
+ }
744
+ ]
745
+ },
746
+ "raw": {
747
+ "type": "object",
748
+ "additionalProperties": false,
749
+ "required": [
750
+ "correct",
751
+ "accuracy",
752
+ "wilson95",
753
+ "saidYesRate"
754
+ ],
755
+ "properties": {
756
+ "correct": {
757
+ "type": "integer",
758
+ "minimum": 0
759
+ },
760
+ "accuracy": {
761
+ "anyOf": [
762
+ {
763
+ "type": "null"
764
+ },
765
+ {
766
+ "type": "number"
767
+ }
768
+ ]
769
+ },
770
+ "wilson95": {
771
+ "anyOf": [
772
+ {
773
+ "type": "null"
774
+ },
775
+ {
776
+ "type": "array",
777
+ "minItems": 2,
778
+ "maxItems": 2,
779
+ "items": {
780
+ "type": "number"
781
+ }
782
+ }
783
+ ]
784
+ },
785
+ "saidYesRate": {
786
+ "anyOf": [
787
+ {
788
+ "type": "null"
789
+ },
790
+ {
791
+ "type": "number"
792
+ }
793
+ ]
794
+ }
795
+ }
796
+ },
797
+ "calibrated": {
798
+ "type": "object",
799
+ "additionalProperties": false,
800
+ "required": [
801
+ "correct",
802
+ "accuracy",
803
+ "wilson95",
804
+ "meanConfidence",
805
+ "confidenceMinusAccuracy",
806
+ "ece",
807
+ "calibrationUnreliable"
808
+ ],
809
+ "properties": {
810
+ "correct": {
811
+ "type": "integer",
812
+ "minimum": 0
813
+ },
814
+ "accuracy": {
815
+ "anyOf": [
816
+ {
817
+ "type": "null"
818
+ },
819
+ {
820
+ "type": "number"
821
+ }
822
+ ]
823
+ },
824
+ "wilson95": {
825
+ "anyOf": [
826
+ {
827
+ "type": "null"
828
+ },
829
+ {
830
+ "type": "array",
831
+ "minItems": 2,
832
+ "maxItems": 2,
833
+ "items": {
834
+ "type": "number"
835
+ }
836
+ }
837
+ ]
838
+ },
839
+ "meanConfidence": {
840
+ "anyOf": [
841
+ {
842
+ "type": "null"
843
+ },
844
+ {
845
+ "type": "number"
846
+ }
847
+ ]
848
+ },
849
+ "confidenceMinusAccuracy": {
850
+ "anyOf": [
851
+ {
852
+ "type": "null"
853
+ },
854
+ {
855
+ "type": "number"
856
+ }
857
+ ]
858
+ },
859
+ "ece": {
860
+ "anyOf": [
861
+ {
862
+ "type": "null"
863
+ },
864
+ {
865
+ "type": "number"
866
+ }
867
+ ]
868
+ },
869
+ "calibrationUnreliable": {
870
+ "anyOf": [
871
+ {
872
+ "type": "null"
873
+ },
874
+ {
875
+ "type": "boolean"
876
+ }
877
+ ]
878
+ }
879
+ }
880
+ }
881
+ }
882
+ }
883
+ },
884
+ "byFailurePosition": {
885
+ "type": "object",
886
+ "additionalProperties": {
887
+ "type": "object",
888
+ "additionalProperties": false,
889
+ "required": [
890
+ "n",
891
+ "observed",
892
+ "coverage",
893
+ "raw",
894
+ "calibrated"
895
+ ],
896
+ "properties": {
897
+ "n": {
898
+ "type": "integer",
899
+ "minimum": 0
900
+ },
901
+ "observed": {
902
+ "type": "integer",
903
+ "minimum": 0
904
+ },
905
+ "coverage": {
906
+ "anyOf": [
907
+ {
908
+ "type": "null"
909
+ },
910
+ {
911
+ "type": "number"
912
+ }
913
+ ]
914
+ },
915
+ "raw": {
916
+ "type": "object",
917
+ "additionalProperties": false,
918
+ "required": [
919
+ "correct",
920
+ "accuracy",
921
+ "wilson95",
922
+ "saidYesRate"
923
+ ],
924
+ "properties": {
925
+ "correct": {
926
+ "type": "integer",
927
+ "minimum": 0
928
+ },
929
+ "accuracy": {
930
+ "anyOf": [
931
+ {
932
+ "type": "null"
933
+ },
934
+ {
935
+ "type": "number"
936
+ }
937
+ ]
938
+ },
939
+ "wilson95": {
940
+ "anyOf": [
941
+ {
942
+ "type": "null"
943
+ },
944
+ {
945
+ "type": "array",
946
+ "minItems": 2,
947
+ "maxItems": 2,
948
+ "items": {
949
+ "type": "number"
950
+ }
951
+ }
952
+ ]
953
+ },
954
+ "saidYesRate": {
955
+ "anyOf": [
956
+ {
957
+ "type": "null"
958
+ },
959
+ {
960
+ "type": "number"
961
+ }
962
+ ]
963
+ }
964
+ }
965
+ },
966
+ "calibrated": {
967
+ "type": "object",
968
+ "additionalProperties": false,
969
+ "required": [
970
+ "correct",
971
+ "accuracy",
972
+ "wilson95",
973
+ "meanConfidence",
974
+ "confidenceMinusAccuracy",
975
+ "ece",
976
+ "calibrationUnreliable"
977
+ ],
978
+ "properties": {
979
+ "correct": {
980
+ "type": "integer",
981
+ "minimum": 0
982
+ },
983
+ "accuracy": {
984
+ "anyOf": [
985
+ {
986
+ "type": "null"
987
+ },
988
+ {
989
+ "type": "number"
990
+ }
991
+ ]
992
+ },
993
+ "wilson95": {
994
+ "anyOf": [
995
+ {
996
+ "type": "null"
997
+ },
998
+ {
999
+ "type": "array",
1000
+ "minItems": 2,
1001
+ "maxItems": 2,
1002
+ "items": {
1003
+ "type": "number"
1004
+ }
1005
+ }
1006
+ ]
1007
+ },
1008
+ "meanConfidence": {
1009
+ "anyOf": [
1010
+ {
1011
+ "type": "null"
1012
+ },
1013
+ {
1014
+ "type": "number"
1015
+ }
1016
+ ]
1017
+ },
1018
+ "confidenceMinusAccuracy": {
1019
+ "anyOf": [
1020
+ {
1021
+ "type": "null"
1022
+ },
1023
+ {
1024
+ "type": "number"
1025
+ }
1026
+ ]
1027
+ },
1028
+ "ece": {
1029
+ "anyOf": [
1030
+ {
1031
+ "type": "null"
1032
+ },
1033
+ {
1034
+ "type": "number"
1035
+ }
1036
+ ]
1037
+ },
1038
+ "calibrationUnreliable": {
1039
+ "anyOf": [
1040
+ {
1041
+ "type": "null"
1042
+ },
1043
+ {
1044
+ "type": "boolean"
1045
+ }
1046
+ ]
1047
+ }
1048
+ }
1049
+ }
1050
+ }
1051
+ }
1052
+ },
1053
+ "auroc": {
1054
+ "type": "array"
1055
+ },
1056
+ "hypotheses": {
1057
+ "type": "object",
1058
+ "additionalProperties": false,
1059
+ "required": [
1060
+ "H1",
1061
+ "H2",
1062
+ "H3",
1063
+ "H3a",
1064
+ "H4",
1065
+ "replication",
1066
+ "calibrationUnderShift"
1067
+ ],
1068
+ "properties": {
1069
+ "H1": {
1070
+ "type": "object",
1071
+ "additionalProperties": false,
1072
+ "required": [
1073
+ "verdict",
1074
+ "byEncoding",
1075
+ "tests"
1076
+ ],
1077
+ "properties": {
1078
+ "verdict": {
1079
+ "anyOf": [
1080
+ {
1081
+ "type": "boolean"
1082
+ },
1083
+ {
1084
+ "const": "insufficient"
1085
+ }
1086
+ ]
1087
+ },
1088
+ "byEncoding": {
1089
+ "type": "array"
1090
+ },
1091
+ "tests": {
1092
+ "type": "array"
1093
+ }
1094
+ }
1095
+ },
1096
+ "H2": {
1097
+ "type": "object",
1098
+ "additionalProperties": false,
1099
+ "required": [
1100
+ "verdict",
1101
+ "byOutcome"
1102
+ ],
1103
+ "properties": {
1104
+ "verdict": {
1105
+ "anyOf": [
1106
+ {
1107
+ "type": "boolean"
1108
+ },
1109
+ {
1110
+ "const": "insufficient"
1111
+ }
1112
+ ]
1113
+ },
1114
+ "byOutcome": {
1115
+ "type": "array"
1116
+ }
1117
+ }
1118
+ },
1119
+ "H3": {
1120
+ "type": "object",
1121
+ "additionalProperties": false,
1122
+ "required": [
1123
+ "verdict",
1124
+ "byPool"
1125
+ ],
1126
+ "properties": {
1127
+ "verdict": {
1128
+ "anyOf": [
1129
+ {
1130
+ "type": "boolean"
1131
+ },
1132
+ {
1133
+ "const": "insufficient"
1134
+ }
1135
+ ]
1136
+ },
1137
+ "byPool": {
1138
+ "type": "array"
1139
+ }
1140
+ }
1141
+ },
1142
+ "H3a": {
1143
+ "type": "object",
1144
+ "additionalProperties": false,
1145
+ "required": [
1146
+ "verdict",
1147
+ "byPool"
1148
+ ],
1149
+ "properties": {
1150
+ "verdict": {
1151
+ "enum": [
1152
+ "contrastive-specific",
1153
+ "general",
1154
+ "mixed",
1155
+ "insufficient"
1156
+ ]
1157
+ },
1158
+ "byPool": {
1159
+ "type": "array"
1160
+ }
1161
+ }
1162
+ },
1163
+ "H4": {
1164
+ "type": "object",
1165
+ "additionalProperties": false,
1166
+ "required": [
1167
+ "verdict",
1168
+ "byPool",
1169
+ "differences"
1170
+ ],
1171
+ "properties": {
1172
+ "verdict": {
1173
+ "anyOf": [
1174
+ {
1175
+ "type": "boolean"
1176
+ },
1177
+ {
1178
+ "const": "insufficient"
1179
+ }
1180
+ ]
1181
+ },
1182
+ "byPool": {
1183
+ "type": "array"
1184
+ },
1185
+ "differences": {
1186
+ "type": "array",
1187
+ "minItems": 4,
1188
+ "items": {
1189
+ "type": "string"
1190
+ }
1191
+ }
1192
+ }
1193
+ },
1194
+ "replication": {
1195
+ "type": "object",
1196
+ "additionalProperties": false,
1197
+ "required": [
1198
+ "verdict",
1199
+ "cell"
1200
+ ],
1201
+ "properties": {
1202
+ "verdict": {
1203
+ "anyOf": [
1204
+ {
1205
+ "type": "boolean"
1206
+ },
1207
+ {
1208
+ "const": "insufficient"
1209
+ }
1210
+ ]
1211
+ },
1212
+ "cell": {
1213
+ "type": "object",
1214
+ "additionalProperties": false,
1215
+ "required": [
1216
+ "n",
1217
+ "observed",
1218
+ "coverage",
1219
+ "raw",
1220
+ "calibrated"
1221
+ ],
1222
+ "properties": {
1223
+ "n": {
1224
+ "type": "integer",
1225
+ "minimum": 0
1226
+ },
1227
+ "observed": {
1228
+ "type": "integer",
1229
+ "minimum": 0
1230
+ },
1231
+ "coverage": {
1232
+ "anyOf": [
1233
+ {
1234
+ "type": "null"
1235
+ },
1236
+ {
1237
+ "type": "number"
1238
+ }
1239
+ ]
1240
+ },
1241
+ "raw": {
1242
+ "type": "object",
1243
+ "additionalProperties": false,
1244
+ "required": [
1245
+ "correct",
1246
+ "accuracy",
1247
+ "wilson95",
1248
+ "saidYesRate"
1249
+ ],
1250
+ "properties": {
1251
+ "correct": {
1252
+ "type": "integer",
1253
+ "minimum": 0
1254
+ },
1255
+ "accuracy": {
1256
+ "anyOf": [
1257
+ {
1258
+ "type": "null"
1259
+ },
1260
+ {
1261
+ "type": "number"
1262
+ }
1263
+ ]
1264
+ },
1265
+ "wilson95": {
1266
+ "anyOf": [
1267
+ {
1268
+ "type": "null"
1269
+ },
1270
+ {
1271
+ "type": "array",
1272
+ "minItems": 2,
1273
+ "maxItems": 2,
1274
+ "items": {
1275
+ "type": "number"
1276
+ }
1277
+ }
1278
+ ]
1279
+ },
1280
+ "saidYesRate": {
1281
+ "anyOf": [
1282
+ {
1283
+ "type": "null"
1284
+ },
1285
+ {
1286
+ "type": "number"
1287
+ }
1288
+ ]
1289
+ }
1290
+ }
1291
+ },
1292
+ "calibrated": {
1293
+ "type": "object",
1294
+ "additionalProperties": false,
1295
+ "required": [
1296
+ "correct",
1297
+ "accuracy",
1298
+ "wilson95",
1299
+ "meanConfidence",
1300
+ "confidenceMinusAccuracy",
1301
+ "ece",
1302
+ "calibrationUnreliable"
1303
+ ],
1304
+ "properties": {
1305
+ "correct": {
1306
+ "type": "integer",
1307
+ "minimum": 0
1308
+ },
1309
+ "accuracy": {
1310
+ "anyOf": [
1311
+ {
1312
+ "type": "null"
1313
+ },
1314
+ {
1315
+ "type": "number"
1316
+ }
1317
+ ]
1318
+ },
1319
+ "wilson95": {
1320
+ "anyOf": [
1321
+ {
1322
+ "type": "null"
1323
+ },
1324
+ {
1325
+ "type": "array",
1326
+ "minItems": 2,
1327
+ "maxItems": 2,
1328
+ "items": {
1329
+ "type": "number"
1330
+ }
1331
+ }
1332
+ ]
1333
+ },
1334
+ "meanConfidence": {
1335
+ "anyOf": [
1336
+ {
1337
+ "type": "null"
1338
+ },
1339
+ {
1340
+ "type": "number"
1341
+ }
1342
+ ]
1343
+ },
1344
+ "confidenceMinusAccuracy": {
1345
+ "anyOf": [
1346
+ {
1347
+ "type": "null"
1348
+ },
1349
+ {
1350
+ "type": "number"
1351
+ }
1352
+ ]
1353
+ },
1354
+ "ece": {
1355
+ "anyOf": [
1356
+ {
1357
+ "type": "null"
1358
+ },
1359
+ {
1360
+ "type": "number"
1361
+ }
1362
+ ]
1363
+ },
1364
+ "calibrationUnreliable": {
1365
+ "anyOf": [
1366
+ {
1367
+ "type": "null"
1368
+ },
1369
+ {
1370
+ "type": "boolean"
1371
+ }
1372
+ ]
1373
+ }
1374
+ }
1375
+ }
1376
+ }
1377
+ }
1378
+ }
1379
+ },
1380
+ "calibrationUnderShift": {
1381
+ "type": "object",
1382
+ "additionalProperties": false,
1383
+ "required": [
1384
+ "unreliableCells"
1385
+ ],
1386
+ "properties": {
1387
+ "unreliableCells": {
1388
+ "type": "array",
1389
+ "items": {
1390
+ "type": "string"
1391
+ }
1392
+ }
1393
+ }
1394
+ }
1395
+ }
1396
+ },
1397
+ "decision": {
1398
+ "const": "HOLD"
1399
+ },
1400
+ "promotion": {
1401
+ "const": "none: diagnostic probe"
1402
+ }
1403
+ }
1404
+ },
1405
+ "audit": {
1406
+ "type": "object",
1407
+ "additionalProperties": false,
1408
+ "required": [
1409
+ "schemaVersion",
1410
+ "limit",
1411
+ "folds",
1412
+ "groups",
1413
+ "equalizedPasses",
1414
+ "equalizedMax",
1415
+ "controlMax"
1416
+ ],
1417
+ "properties": {
1418
+ "schemaVersion": {
1419
+ "const": "decision-d17mf-shortcut-audit/v2"
1420
+ },
1421
+ "limit": {
1422
+ "const": 0.6
1423
+ },
1424
+ "folds": {
1425
+ "const": 5
1426
+ },
1427
+ "groups": {
1428
+ "type": "array",
1429
+ "minItems": 1,
1430
+ "items": {
1431
+ "type": "object",
1432
+ "additionalProperties": false,
1433
+ "required": [
1434
+ "group",
1435
+ "n",
1436
+ "yes",
1437
+ "scores",
1438
+ "max",
1439
+ "passes"
1440
+ ],
1441
+ "properties": {
1442
+ "group": {
1443
+ "type": "string"
1444
+ },
1445
+ "n": {
1446
+ "type": "integer"
1447
+ },
1448
+ "yes": {
1449
+ "type": "integer"
1450
+ },
1451
+ "scores": {
1452
+ "type": "object",
1453
+ "additionalProperties": false,
1454
+ "required": [
1455
+ "scalarThreshold",
1456
+ "tokenThreshold",
1457
+ "depth2Tree",
1458
+ "logistic"
1459
+ ],
1460
+ "properties": {
1461
+ "scalarThreshold": {
1462
+ "type": "number"
1463
+ },
1464
+ "tokenThreshold": {
1465
+ "type": "number"
1466
+ },
1467
+ "depth2Tree": {
1468
+ "type": "number"
1469
+ },
1470
+ "logistic": {
1471
+ "type": "number"
1472
+ }
1473
+ }
1474
+ },
1475
+ "max": {
1476
+ "type": "number"
1477
+ },
1478
+ "passes": {
1479
+ "anyOf": [
1480
+ {
1481
+ "type": "null"
1482
+ },
1483
+ {
1484
+ "type": "boolean"
1485
+ }
1486
+ ]
1487
+ }
1488
+ }
1489
+ }
1490
+ },
1491
+ "equalizedPasses": {
1492
+ "type": "boolean"
1493
+ },
1494
+ "equalizedMax": {
1495
+ "type": "number"
1496
+ },
1497
+ "controlMax": {
1498
+ "type": "number"
1499
+ }
1500
+ }
1501
+ },
1502
+ "reviewMaterial": {
1503
+ "type": "object",
1504
+ "additionalProperties": false,
1505
+ "required": [
1506
+ "schemaVersion",
1507
+ "study",
1508
+ "scope",
1509
+ "preparedDirectory",
1510
+ "seedDigest",
1511
+ "reviewTemplateDigest",
1512
+ "corpusDigest",
1513
+ "goldDigest",
1514
+ "preregistrationDigest",
1515
+ "splitOfItems",
1516
+ "itemCount",
1517
+ "protocol",
1518
+ "guide",
1519
+ "items"
1520
+ ],
1521
+ "properties": {
1522
+ "schemaVersion": {
1523
+ "const": "d17mf-development-review-material/v1"
1524
+ },
1525
+ "study": {
1526
+ "const": "D17-MF"
1527
+ },
1528
+ "scope": {
1529
+ "type": "string",
1530
+ "minLength": 1,
1531
+ "maxLength": 2000
1532
+ },
1533
+ "preparedDirectory": {
1534
+ "type": "string",
1535
+ "minLength": 1,
1536
+ "maxLength": 4096
1537
+ },
1538
+ "seedDigest": {
1539
+ "type": "string",
1540
+ "pattern": "^sha256:[a-f0-9]{64}$"
1541
+ },
1542
+ "reviewTemplateDigest": {
1543
+ "type": "string",
1544
+ "pattern": "^sha256:[a-f0-9]{64}$"
1545
+ },
1546
+ "corpusDigest": {
1547
+ "type": "string",
1548
+ "pattern": "^sha256:[a-f0-9]{64}$"
1549
+ },
1550
+ "goldDigest": {
1551
+ "type": "string",
1552
+ "pattern": "^sha256:[a-f0-9]{64}$"
1553
+ },
1554
+ "preregistrationDigest": {
1555
+ "type": "string",
1556
+ "pattern": "^sha256:[a-f0-9]{64}$"
1557
+ },
1558
+ "splitOfItems": {
1559
+ "type": "string",
1560
+ "minLength": 1,
1561
+ "maxLength": 2000
1562
+ },
1563
+ "itemCount": {
1564
+ "const": 40
1565
+ },
1566
+ "protocol": {
1567
+ "type": "array",
1568
+ "minItems": 1,
1569
+ "maxItems": 8,
1570
+ "items": {
1571
+ "type": "string",
1572
+ "minLength": 1,
1573
+ "maxLength": 2000
1574
+ }
1575
+ },
1576
+ "guide": {
1577
+ "type": "object",
1578
+ "additionalProperties": false,
1579
+ "required": [
1580
+ "population",
1581
+ "yes",
1582
+ "no",
1583
+ "authority",
1584
+ "invalidGold",
1585
+ "timing",
1586
+ "independence"
1587
+ ],
1588
+ "properties": {
1589
+ "population": {
1590
+ "type": "string",
1591
+ "minLength": 1,
1592
+ "maxLength": 2000
1593
+ },
1594
+ "yes": {
1595
+ "type": "string",
1596
+ "minLength": 1,
1597
+ "maxLength": 2000
1598
+ },
1599
+ "no": {
1600
+ "type": "string",
1601
+ "minLength": 1,
1602
+ "maxLength": 2000
1603
+ },
1604
+ "authority": {
1605
+ "type": "string",
1606
+ "minLength": 1,
1607
+ "maxLength": 2000
1608
+ },
1609
+ "invalidGold": {
1610
+ "type": "string",
1611
+ "minLength": 1,
1612
+ "maxLength": 2000
1613
+ },
1614
+ "timing": {
1615
+ "type": "string",
1616
+ "minLength": 1,
1617
+ "maxLength": 2000
1618
+ },
1619
+ "independence": {
1620
+ "type": "string",
1621
+ "minLength": 1,
1622
+ "maxLength": 2000
1623
+ }
1624
+ }
1625
+ },
1626
+ "items": {
1627
+ "type": "array",
1628
+ "minItems": 40,
1629
+ "maxItems": 40,
1630
+ "items": {
1631
+ "type": "object",
1632
+ "additionalProperties": false,
1633
+ "required": [
1634
+ "assessmentId",
1635
+ "rowId",
1636
+ "cell",
1637
+ "inputDigest",
1638
+ "blind",
1639
+ "unblind"
1640
+ ],
1641
+ "properties": {
1642
+ "assessmentId": {
1643
+ "type": "string",
1644
+ "pattern": "^mf-assessment-[0-9]{2}$"
1645
+ },
1646
+ "rowId": {
1647
+ "type": "string",
1648
+ "pattern": "^[0-9a-f]{64}$"
1649
+ },
1650
+ "cell": {
1651
+ "type": "string",
1652
+ "pattern": "^[A-Za-z0-9][A-Za-z0-9_-]{0,63}$"
1653
+ },
1654
+ "inputDigest": {
1655
+ "type": "string",
1656
+ "pattern": "^sha256:[a-f0-9]{64}$"
1657
+ },
1658
+ "blind": {
1659
+ "type": "object",
1660
+ "additionalProperties": false,
1661
+ "required": [
1662
+ "payload"
1663
+ ],
1664
+ "properties": {
1665
+ "payload": {
1666
+ "type": "string",
1667
+ "minLength": 1,
1668
+ "maxLength": 8192
1669
+ }
1670
+ }
1671
+ },
1672
+ "unblind": {
1673
+ "type": "object",
1674
+ "additionalProperties": false,
1675
+ "required": [
1676
+ "goldLabel",
1677
+ "textOracleLabel",
1678
+ "variant",
1679
+ "cell",
1680
+ "factors",
1681
+ "latentWorld"
1682
+ ],
1683
+ "properties": {
1684
+ "goldLabel": {
1685
+ "enum": [
1686
+ "yes",
1687
+ "no"
1688
+ ]
1689
+ },
1690
+ "textOracleLabel": {
1691
+ "enum": [
1692
+ "yes",
1693
+ "no"
1694
+ ]
1695
+ },
1696
+ "variant": {
1697
+ "type": "string",
1698
+ "pattern": "^[A-Za-z0-9][A-Za-z0-9_-]{0,63}$"
1699
+ },
1700
+ "cell": {
1701
+ "type": "string",
1702
+ "pattern": "^[A-Za-z0-9][A-Za-z0-9_-]{0,63}$"
1703
+ },
1704
+ "factors": {
1705
+ "type": "object",
1706
+ "additionalProperties": false,
1707
+ "required": [
1708
+ "pool",
1709
+ "hops",
1710
+ "outcome",
1711
+ "layout",
1712
+ "rule"
1713
+ ],
1714
+ "properties": {
1715
+ "pool": {
1716
+ "enum": [
1717
+ "pool-a",
1718
+ "pool-b",
1719
+ "pool-c"
1720
+ ]
1721
+ },
1722
+ "hops": {
1723
+ "enum": [
1724
+ 1,
1725
+ 3
1726
+ ]
1727
+ },
1728
+ "outcome": {
1729
+ "enum": [
1730
+ "yes",
1731
+ "relay-off",
1732
+ "missing-contrastive",
1733
+ "missing-negation",
1734
+ "missing-unrelated",
1735
+ "missing-absence"
1736
+ ]
1737
+ },
1738
+ "layout": {
1739
+ "enum": [
1740
+ "equalized",
1741
+ "d17-faithful"
1742
+ ]
1743
+ },
1744
+ "rule": {
1745
+ "enum": [
1746
+ "exhaustive",
1747
+ "d17"
1748
+ ]
1749
+ }
1750
+ }
1751
+ },
1752
+ "latentWorld": {
1753
+ "type": "object",
1754
+ "additionalProperties": false,
1755
+ "required": [
1756
+ "chain",
1757
+ "decoys",
1758
+ "failureKind",
1759
+ "failurePosition",
1760
+ "targets",
1761
+ "day",
1762
+ "quantity"
1763
+ ],
1764
+ "properties": {
1765
+ "chain": {
1766
+ "type": "array",
1767
+ "minItems": 1,
1768
+ "maxItems": 16,
1769
+ "items": {
1770
+ "type": "string",
1771
+ "pattern": "^[A-Za-z0-9][A-Za-z0-9_-]{0,31}$"
1772
+ }
1773
+ },
1774
+ "decoys": {
1775
+ "type": "array",
1776
+ "minItems": 0,
1777
+ "maxItems": 16,
1778
+ "items": {
1779
+ "type": "string",
1780
+ "pattern": "^[A-Za-z0-9][A-Za-z0-9_-]{0,31}$"
1781
+ }
1782
+ },
1783
+ "failureKind": {
1784
+ "enum": [
1785
+ "link",
1786
+ "relay",
1787
+ null
1788
+ ]
1789
+ },
1790
+ "failurePosition": {
1791
+ "anyOf": [
1792
+ {
1793
+ "type": "null"
1794
+ },
1795
+ {
1796
+ "type": "integer",
1797
+ "minimum": 0,
1798
+ "maximum": 8
1799
+ }
1800
+ ]
1801
+ },
1802
+ "targets": {
1803
+ "type": "object",
1804
+ "maxProperties": 16,
1805
+ "additionalProperties": {
1806
+ "type": "integer",
1807
+ "minimum": 0,
1808
+ "maximum": 64
1809
+ }
1810
+ },
1811
+ "day": {
1812
+ "type": "integer",
1813
+ "minimum": 0,
1814
+ "maximum": 100000
1815
+ },
1816
+ "quantity": {
1817
+ "type": "integer",
1818
+ "minimum": 0,
1819
+ "maximum": 100000
1820
+ }
1821
+ }
1822
+ }
1823
+ }
1824
+ }
1825
+ }
1826
+ }
1827
+ }
1828
+ }
1829
+ }
1830
+ },
1831
+ "oneOf": [
1832
+ {
1833
+ "$ref": "#/$defs/analysis"
1834
+ },
1835
+ {
1836
+ "$ref": "#/$defs/reviewTemplate"
1837
+ },
1838
+ {
1839
+ "$ref": "#/$defs/review"
1840
+ },
1841
+ {
1842
+ "$ref": "#/$defs/report"
1843
+ },
1844
+ {
1845
+ "$ref": "#/$defs/audit"
1846
+ },
1847
+ {
1848
+ "$ref": "#/$defs/reviewMaterial"
1849
+ }
1850
+ ]
1851
+ }