@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 +1,758 @@
1
+ {
2
+ "$schema": "https://json-schema.org/draft/2020-12/schema",
3
+ "$id": "https://aiwg.io/schemas/decision/ContextPruningEvaluationReport.v1.schema.json",
4
+ "title": "Context Pruning Evaluation Report",
5
+ "type": "object",
6
+ "additionalProperties": false,
7
+ "required": [
8
+ "schemaVersion",
9
+ "preregistration",
10
+ "holdoutAccessedAt",
11
+ "integrity",
12
+ "metrics",
13
+ "derived",
14
+ "missingInputs",
15
+ "findings",
16
+ "upstreamDecision",
17
+ "decision",
18
+ "advisory",
19
+ "digest"
20
+ ],
21
+ "properties": {
22
+ "schemaVersion": {
23
+ "const": "decision-context-pruning-evaluation-report/v1"
24
+ },
25
+ "preregistration": {
26
+ "$ref": "#/$defs/preregistration"
27
+ },
28
+ "holdoutAccessedAt": {
29
+ "type": [
30
+ "string",
31
+ "null"
32
+ ],
33
+ "format": "date-time"
34
+ },
35
+ "integrity": {
36
+ "$ref": "#/$defs/integrity"
37
+ },
38
+ "metrics": {
39
+ "type": "object",
40
+ "additionalProperties": false,
41
+ "required": [
42
+ "pairs",
43
+ "quality",
44
+ "protectedRetentionBps",
45
+ "providerUsage",
46
+ "estimatorUsage",
47
+ "totalCalls",
48
+ "latencyMs",
49
+ "promptCache",
50
+ "sharedStateAccounting"
51
+ ],
52
+ "properties": {
53
+ "pairs": {
54
+ "type": "array",
55
+ "items": {
56
+ "type": "object",
57
+ "additionalProperties": false,
58
+ "required": [
59
+ "pairId",
60
+ "slice"
61
+ ],
62
+ "properties": {
63
+ "pairId": {
64
+ "type": "string",
65
+ "minLength": 1
66
+ },
67
+ "slice": {
68
+ "type": "string",
69
+ "minLength": 1
70
+ }
71
+ }
72
+ }
73
+ },
74
+ "quality": {
75
+ "type": "array",
76
+ "items": {
77
+ "type": "object",
78
+ "additionalProperties": false,
79
+ "required": [
80
+ "metric",
81
+ "pairs"
82
+ ],
83
+ "properties": {
84
+ "metric": {
85
+ "enum": [
86
+ "downstream-task-success",
87
+ "requirement-coverage",
88
+ "factual-coverage",
89
+ "citation-accuracy",
90
+ "human-preference"
91
+ ]
92
+ },
93
+ "pairs": {
94
+ "type": "array",
95
+ "items": {
96
+ "type": "object",
97
+ "additionalProperties": false,
98
+ "required": [
99
+ "pairId",
100
+ "baseline",
101
+ "candidate"
102
+ ],
103
+ "properties": {
104
+ "pairId": {
105
+ "type": "string",
106
+ "minLength": 1
107
+ },
108
+ "baseline": {
109
+ "type": "number",
110
+ "minimum": 0,
111
+ "maximum": 1
112
+ },
113
+ "candidate": {
114
+ "type": "number",
115
+ "minimum": 0,
116
+ "maximum": 1
117
+ }
118
+ }
119
+ }
120
+ }
121
+ }
122
+ }
123
+ },
124
+ "protectedRetentionBps": {
125
+ "type": "integer",
126
+ "minimum": 0,
127
+ "maximum": 10000
128
+ },
129
+ "providerUsage": {
130
+ "$ref": "#/$defs/usageAccounting"
131
+ },
132
+ "estimatorUsage": {
133
+ "$ref": "#/$defs/usageAccounting"
134
+ },
135
+ "totalCalls": {
136
+ "type": "integer",
137
+ "minimum": 0
138
+ },
139
+ "latencyMs": {
140
+ "type": "object",
141
+ "additionalProperties": false,
142
+ "required": [
143
+ "p50",
144
+ "p95",
145
+ "p99"
146
+ ],
147
+ "properties": {
148
+ "p50": {
149
+ "type": [
150
+ "number",
151
+ "null"
152
+ ]
153
+ },
154
+ "p95": {
155
+ "type": [
156
+ "number",
157
+ "null"
158
+ ]
159
+ },
160
+ "p99": {
161
+ "type": [
162
+ "number",
163
+ "null"
164
+ ]
165
+ }
166
+ }
167
+ },
168
+ "promptCache": {
169
+ "type": "object",
170
+ "additionalProperties": false,
171
+ "required": [
172
+ "baseline",
173
+ "prunedDownstream"
174
+ ],
175
+ "properties": {
176
+ "baseline": {
177
+ "$ref": "#/$defs/promptCacheArm"
178
+ },
179
+ "prunedDownstream": {
180
+ "$ref": "#/$defs/promptCacheArm"
181
+ }
182
+ }
183
+ },
184
+ "sharedStateAccounting": {
185
+ "const": "not-applicable"
186
+ }
187
+ }
188
+ },
189
+ "derived": {
190
+ "$ref": "#/$defs/derived"
191
+ },
192
+ "missingInputs": {
193
+ "type": "array",
194
+ "uniqueItems": true,
195
+ "items": {
196
+ "type": "string",
197
+ "minLength": 1
198
+ }
199
+ },
200
+ "findings": {
201
+ "type": "array",
202
+ "uniqueItems": true,
203
+ "items": {
204
+ "type": "string",
205
+ "minLength": 1
206
+ }
207
+ },
208
+ "upstreamDecision": {
209
+ "$ref": "#/$defs/gateDecision"
210
+ },
211
+ "decision": {
212
+ "$ref": "#/$defs/gateDecision"
213
+ },
214
+ "advisory": {
215
+ "enum": [
216
+ "INSUFFICIENT EVIDENCE",
217
+ null
218
+ ]
219
+ },
220
+ "digest": {
221
+ "$ref": "#/$defs/digest"
222
+ }
223
+ },
224
+ "$defs": {
225
+ "digest": {
226
+ "type": "string",
227
+ "pattern": "^sha256:[0-9a-f]{64}$"
228
+ },
229
+ "gateDecision": {
230
+ "enum": [
231
+ "PROMOTE",
232
+ "HOLD",
233
+ "ROLLBACK"
234
+ ]
235
+ },
236
+ "integrity": {
237
+ "type": "object",
238
+ "additionalProperties": false,
239
+ "required": [
240
+ "sample_n",
241
+ "uncertainty",
242
+ "paired_baseline",
243
+ "integrity_mode",
244
+ "fresh_workspace_required",
245
+ "fresh_workspace_verified",
246
+ "integrity_state",
247
+ "trusted_score_source",
248
+ "compromise_labels",
249
+ "weak_signal_reason",
250
+ "release_gate"
251
+ ],
252
+ "properties": {
253
+ "sample_n": {
254
+ "type": "integer",
255
+ "minimum": 0
256
+ },
257
+ "uncertainty": {},
258
+ "paired_baseline": {},
259
+ "integrity_mode": {
260
+ "type": "string",
261
+ "minLength": 1
262
+ },
263
+ "fresh_workspace_required": {
264
+ "type": "boolean"
265
+ },
266
+ "fresh_workspace_verified": {
267
+ "type": "boolean"
268
+ },
269
+ "integrity_state": {
270
+ "type": "string",
271
+ "minLength": 1
272
+ },
273
+ "trusted_score_source": {
274
+ "type": "string",
275
+ "minLength": 1
276
+ },
277
+ "compromise_labels": {
278
+ "type": "array",
279
+ "items": {
280
+ "type": "string"
281
+ }
282
+ },
283
+ "weak_signal_reason": {
284
+ "type": [
285
+ "string",
286
+ "null"
287
+ ]
288
+ },
289
+ "release_gate": {
290
+ "type": "object",
291
+ "additionalProperties": false,
292
+ "required": [
293
+ "decision",
294
+ "reasons"
295
+ ],
296
+ "properties": {
297
+ "decision": {
298
+ "$ref": "#/$defs/gateDecision"
299
+ },
300
+ "reasons": {
301
+ "type": "array",
302
+ "items": {
303
+ "type": "string"
304
+ }
305
+ }
306
+ }
307
+ }
308
+ }
309
+ },
310
+ "usage": {
311
+ "type": "object",
312
+ "additionalProperties": false,
313
+ "required": [
314
+ "inputTokens",
315
+ "cachedInputTokens",
316
+ "outputTokens",
317
+ "costUsd"
318
+ ],
319
+ "properties": {
320
+ "inputTokens": {
321
+ "type": "integer",
322
+ "minimum": 0
323
+ },
324
+ "cachedInputTokens": {
325
+ "type": [
326
+ "integer",
327
+ "null"
328
+ ],
329
+ "minimum": 0
330
+ },
331
+ "outputTokens": {
332
+ "type": "integer",
333
+ "minimum": 0
334
+ },
335
+ "costUsd": {
336
+ "type": [
337
+ "number",
338
+ "null"
339
+ ],
340
+ "minimum": 0
341
+ }
342
+ }
343
+ },
344
+ "usageAccounting": {
345
+ "type": "object",
346
+ "additionalProperties": false,
347
+ "required": [
348
+ "baseline",
349
+ "prunedDownstream",
350
+ "decisionCalls",
351
+ "fallbackCalls",
352
+ "transformationCalls"
353
+ ],
354
+ "properties": {
355
+ "baseline": {
356
+ "$ref": "#/$defs/usage"
357
+ },
358
+ "prunedDownstream": {
359
+ "$ref": "#/$defs/usage"
360
+ },
361
+ "decisionCalls": {
362
+ "$ref": "#/$defs/usage"
363
+ },
364
+ "fallbackCalls": {
365
+ "$ref": "#/$defs/usage"
366
+ },
367
+ "transformationCalls": {
368
+ "$ref": "#/$defs/usage"
369
+ }
370
+ }
371
+ },
372
+ "preregistration": {
373
+ "type": "object",
374
+ "additionalProperties": false,
375
+ "required": [
376
+ "schemaVersion",
377
+ "id",
378
+ "registeredAt",
379
+ "pairSetDigest",
380
+ "thresholds",
381
+ "digest"
382
+ ],
383
+ "properties": {
384
+ "schemaVersion": {
385
+ "const": "decision-context-pruning-preregistration/v1"
386
+ },
387
+ "id": {
388
+ "type": "string",
389
+ "minLength": 1
390
+ },
391
+ "registeredAt": {
392
+ "type": "string",
393
+ "format": "date-time"
394
+ },
395
+ "pairSetDigest": {
396
+ "$ref": "#/$defs/digest"
397
+ },
398
+ "thresholds": {
399
+ "type": "object",
400
+ "additionalProperties": false,
401
+ "required": [
402
+ "protectedRetentionBps",
403
+ "confidenceInterval",
404
+ "qualityMetrics",
405
+ "slices",
406
+ "minimumOverallN",
407
+ "minimumSliceN",
408
+ "minimumZeroVarianceN",
409
+ "minimumProtectedN",
410
+ "powerRule",
411
+ "qualityNonInferiorityMarginBps",
412
+ "positiveTotalTokenTarget",
413
+ "positiveTotalCostTargetUsd"
414
+ ],
415
+ "properties": {
416
+ "protectedRetentionBps": {
417
+ "const": 10000
418
+ },
419
+ "confidenceInterval": {
420
+ "type": "object",
421
+ "additionalProperties": false,
422
+ "required": [
423
+ "levelBps",
424
+ "binaryMethod",
425
+ "boundedMethod",
426
+ "bootstrapSeed",
427
+ "bootstrapResamples"
428
+ ],
429
+ "properties": {
430
+ "levelBps": {
431
+ "type": "integer",
432
+ "exclusiveMinimum": 5000,
433
+ "exclusiveMaximum": 9999
434
+ },
435
+ "binaryMethod": {
436
+ "enum": [
437
+ "newcombe-10",
438
+ "tango"
439
+ ]
440
+ },
441
+ "boundedMethod": {
442
+ "const": "percentile-bootstrap"
443
+ },
444
+ "bootstrapSeed": {
445
+ "type": "integer",
446
+ "minimum": 0,
447
+ "maximum": 4294967295
448
+ },
449
+ "bootstrapResamples": {
450
+ "type": "integer",
451
+ "minimum": 1,
452
+ "maximum": 1000000
453
+ }
454
+ }
455
+ },
456
+ "qualityMetrics": {
457
+ "type": "array",
458
+ "minItems": 5,
459
+ "maxItems": 5,
460
+ "uniqueItems": true,
461
+ "items": {
462
+ "type": "object",
463
+ "additionalProperties": false,
464
+ "required": [
465
+ "metric",
466
+ "scale"
467
+ ],
468
+ "properties": {
469
+ "metric": {
470
+ "enum": [
471
+ "downstream-task-success",
472
+ "requirement-coverage",
473
+ "factual-coverage",
474
+ "citation-accuracy",
475
+ "human-preference"
476
+ ]
477
+ },
478
+ "scale": {
479
+ "enum": [
480
+ "binary",
481
+ "bounded"
482
+ ]
483
+ }
484
+ }
485
+ }
486
+ },
487
+ "slices": {
488
+ "type": "array",
489
+ "minItems": 1,
490
+ "uniqueItems": true,
491
+ "items": {
492
+ "type": "string",
493
+ "minLength": 1
494
+ }
495
+ },
496
+ "minimumOverallN": {
497
+ "type": "integer",
498
+ "minimum": 2
499
+ },
500
+ "minimumSliceN": {
501
+ "type": "integer",
502
+ "minimum": 1
503
+ },
504
+ "minimumZeroVarianceN": {
505
+ "type": "integer",
506
+ "minimum": 2
507
+ },
508
+ "minimumProtectedN": {
509
+ "type": "integer",
510
+ "minimum": 2
511
+ },
512
+ "powerRule": {
513
+ "type": [
514
+ "string",
515
+ "null"
516
+ ]
517
+ },
518
+ "qualityNonInferiorityMarginBps": {
519
+ "type": "integer",
520
+ "minimum": -10000,
521
+ "maximum": 0
522
+ },
523
+ "positiveTotalTokenTarget": {
524
+ "type": "integer",
525
+ "minimum": 1
526
+ },
527
+ "positiveTotalCostTargetUsd": {
528
+ "type": "number",
529
+ "exclusiveMinimum": 0
530
+ }
531
+ }
532
+ },
533
+ "digest": {
534
+ "$ref": "#/$defs/digest"
535
+ }
536
+ }
537
+ },
538
+ "promptCacheArm": {
539
+ "type": "object",
540
+ "additionalProperties": false,
541
+ "required": [
542
+ "hits",
543
+ "misses",
544
+ "cachedInputTokens"
545
+ ],
546
+ "properties": {
547
+ "hits": {
548
+ "type": "integer",
549
+ "minimum": 0
550
+ },
551
+ "misses": {
552
+ "type": "integer",
553
+ "minimum": 0
554
+ },
555
+ "cachedInputTokens": {
556
+ "type": "integer",
557
+ "minimum": 0
558
+ }
559
+ }
560
+ },
561
+ "derived": {
562
+ "type": "object",
563
+ "additionalProperties": false,
564
+ "required": [
565
+ "sampleN",
566
+ "sliceSupport",
567
+ "quality",
568
+ "providerSavings",
569
+ "estimatorSavings",
570
+ "protectedRetention"
571
+ ],
572
+ "properties": {
573
+ "sampleN": {
574
+ "type": "integer",
575
+ "minimum": 0
576
+ },
577
+ "sliceSupport": {
578
+ "type": "object",
579
+ "additionalProperties": {
580
+ "type": "integer",
581
+ "minimum": 0
582
+ }
583
+ },
584
+ "quality": {
585
+ "type": "array",
586
+ "items": {
587
+ "type": "object",
588
+ "additionalProperties": false,
589
+ "required": [
590
+ "metric",
591
+ "scale",
592
+ "n",
593
+ "missingPairs",
594
+ "harm",
595
+ "sliceSupport",
596
+ "interval",
597
+ "decision"
598
+ ],
599
+ "properties": {
600
+ "metric": {
601
+ "enum": [
602
+ "downstream-task-success",
603
+ "requirement-coverage",
604
+ "factual-coverage",
605
+ "citation-accuracy",
606
+ "human-preference"
607
+ ]
608
+ },
609
+ "scale": {
610
+ "enum": [
611
+ "binary",
612
+ "bounded"
613
+ ]
614
+ },
615
+ "n": {
616
+ "type": "integer",
617
+ "minimum": 0
618
+ },
619
+ "missingPairs": {
620
+ "type": "integer",
621
+ "minimum": 0
622
+ },
623
+ "harm": {
624
+ "type": "boolean"
625
+ },
626
+ "sliceSupport": {
627
+ "type": "object",
628
+ "additionalProperties": {
629
+ "type": "integer",
630
+ "minimum": 0
631
+ }
632
+ },
633
+ "interval": {
634
+ "oneOf": [
635
+ {
636
+ "type": "null"
637
+ },
638
+ {
639
+ "type": "object",
640
+ "additionalProperties": false,
641
+ "required": [
642
+ "lowerBps",
643
+ "upperBps",
644
+ "estimateBps",
645
+ "n",
646
+ "method"
647
+ ],
648
+ "properties": {
649
+ "lowerBps": {
650
+ "type": "integer",
651
+ "minimum": -10000,
652
+ "maximum": 10000
653
+ },
654
+ "upperBps": {
655
+ "type": "integer",
656
+ "minimum": -10000,
657
+ "maximum": 10000
658
+ },
659
+ "estimateBps": {
660
+ "type": "integer",
661
+ "minimum": -10000,
662
+ "maximum": 10000
663
+ },
664
+ "n": {
665
+ "type": "integer",
666
+ "minimum": 1
667
+ },
668
+ "method": {
669
+ "enum": [
670
+ "newcombe-hybrid-score",
671
+ "tango-score",
672
+ "percentile-bootstrap"
673
+ ]
674
+ }
675
+ }
676
+ }
677
+ ]
678
+ },
679
+ "decision": {
680
+ "enum": [
681
+ "non-inferior",
682
+ "not-non-inferior",
683
+ "insufficient"
684
+ ]
685
+ }
686
+ }
687
+ }
688
+ },
689
+ "providerSavings": {
690
+ "type": "object",
691
+ "additionalProperties": false,
692
+ "required": [
693
+ "totalTokens",
694
+ "uncachedTokens",
695
+ "costUsd",
696
+ "cachedInputTokensDelta"
697
+ ],
698
+ "properties": {
699
+ "totalTokens": {
700
+ "type": "integer"
701
+ },
702
+ "uncachedTokens": {
703
+ "type": "integer"
704
+ },
705
+ "costUsd": {
706
+ "type": [
707
+ "number",
708
+ "null"
709
+ ]
710
+ },
711
+ "cachedInputTokensDelta": {
712
+ "type": "integer"
713
+ }
714
+ }
715
+ },
716
+ "estimatorSavings": {
717
+ "type": "object",
718
+ "additionalProperties": false,
719
+ "required": [
720
+ "totalTokens"
721
+ ],
722
+ "properties": {
723
+ "totalTokens": {
724
+ "type": "integer"
725
+ }
726
+ }
727
+ },
728
+ "protectedRetention": {
729
+ "type": "object",
730
+ "additionalProperties": false,
731
+ "required": [
732
+ "protectedItems",
733
+ "retained",
734
+ "bps"
735
+ ],
736
+ "properties": {
737
+ "protectedItems": {
738
+ "type": "integer",
739
+ "minimum": 0
740
+ },
741
+ "retained": {
742
+ "type": "integer",
743
+ "minimum": 0
744
+ },
745
+ "bps": {
746
+ "type": [
747
+ "integer",
748
+ "null"
749
+ ],
750
+ "minimum": 0,
751
+ "maximum": 10000
752
+ }
753
+ }
754
+ }
755
+ }
756
+ }
757
+ }
758
+ }