@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 +1,163 @@
1
+ {
2
+ "$schema": "https://json-schema.org/draft/2020-12/schema",
3
+ "$id": "https://aiwg.io/schemas/decision/DecisionSensitivityReport.v1.schema.json",
4
+ "title": "Decision Sensitivity Report",
5
+ "description": "Immutable D23 counterfactual sensitivity report. It is associative reviewer evidence, not causal explanation, correctness proof, or action authority.",
6
+ "type": "object",
7
+ "additionalProperties": false,
8
+ "required": [
9
+ "schemaVersion", "id", "plan", "tenantId", "workspaceId", "projectId", "sourceSubject",
10
+ "analysisKind", "semantics", "actionAuthorization", "status", "generatedAt", "privacy",
11
+ "retention", "budget", "baseline", "rows", "summary", "warnings", "digest"
12
+ ],
13
+ "properties": {
14
+ "schemaVersion": { "const": "decision-sensitivity-report/v1" },
15
+ "id": { "$ref": "#/$defs/name" },
16
+ "plan": { "$ref": "#/$defs/pin" },
17
+ "tenantId": { "$ref": "#/$defs/name" },
18
+ "workspaceId": { "$ref": "#/$defs/name" },
19
+ "projectId": { "$ref": "#/$defs/name" },
20
+ "sourceSubject": {
21
+ "type": "object",
22
+ "additionalProperties": false,
23
+ "required": ["tenantId", "workspaceId", "projectId", "subjectRef"],
24
+ "properties": {
25
+ "tenantId": { "$ref": "#/$defs/name" },
26
+ "workspaceId": { "$ref": "#/$defs/name" },
27
+ "projectId": { "$ref": "#/$defs/name" },
28
+ "subjectRef": { "$ref": "#/$defs/name" }
29
+ }
30
+ },
31
+ "analysisKind": { "enum": ["policy-replay", "input-reevaluation"] },
32
+ "semantics": { "const": "associative-sensitivity-not-causal" },
33
+ "actionAuthorization": { "const": "not-authorized" },
34
+ "status": { "enum": ["completed", "partial", "budget-exhausted", "rejected", "failed"] },
35
+ "generatedAt": { "$ref": "#/$defs/dateTime" },
36
+ "privacy": {
37
+ "type": "object",
38
+ "additionalProperties": false,
39
+ "required": ["redaction", "summaryPrecisionBps"],
40
+ "properties": {
41
+ "redaction": { "enum": ["hash-values", "opaque-ids"] },
42
+ "summaryPrecisionBps": { "type": "integer", "minimum": 1, "maximum": 10000 }
43
+ }
44
+ },
45
+ "retention": {
46
+ "type": "object",
47
+ "additionalProperties": false,
48
+ "required": ["class", "deleteAfterEpochMs", "legalHold"],
49
+ "properties": {
50
+ "class": { "enum": ["ephemeral", "review-evidence"] },
51
+ "deleteAfterEpochMs": { "type": "integer", "minimum": 0 },
52
+ "legalHold": { "type": "boolean" }
53
+ }
54
+ },
55
+ "budget": {
56
+ "type": "object",
57
+ "additionalProperties": false,
58
+ "required": ["maxVariants", "processedVariants", "backendCalls", "tokens", "costMicros", "exhausted"],
59
+ "properties": {
60
+ "maxVariants": { "type": "integer", "minimum": 1 },
61
+ "processedVariants": { "type": "integer", "minimum": 0 },
62
+ "backendCalls": { "type": "integer", "minimum": 0 },
63
+ "tokens": { "type": "integer", "minimum": 0 },
64
+ "costMicros": { "type": "integer", "minimum": 0 },
65
+ "exhausted": { "type": "boolean" }
66
+ }
67
+ },
68
+ "baseline": {
69
+ "type": "object",
70
+ "additionalProperties": false,
71
+ "required": ["resultRef", "outcomeDigest", "status", "reason"],
72
+ "properties": {
73
+ "resultRef": { "$ref": "#/$defs/pin" },
74
+ "outcomeDigest": { "$ref": "#/$defs/digest" },
75
+ "status": { "type": "string", "minLength": 1, "maxLength": 80 },
76
+ "reason": { "type": "string", "minLength": 1, "maxLength": 80 }
77
+ }
78
+ },
79
+ "rows": { "type": "array", "items": { "$ref": "#/$defs/row" } },
80
+ "summary": {
81
+ "type": "object",
82
+ "additionalProperties": false,
83
+ "required": ["totalRows", "changedOutcomes", "changedAcceptance", "changedRules", "minimumThresholdChangeBps"],
84
+ "properties": {
85
+ "totalRows": { "type": "integer", "minimum": 0 },
86
+ "changedOutcomes": { "type": "integer", "minimum": 0 },
87
+ "changedAcceptance": { "type": "integer", "minimum": 0 },
88
+ "changedRules": { "type": "integer", "minimum": 0 },
89
+ "minimumThresholdChangeBps": { "type": ["integer", "null"], "minimum": 0, "maximum": 10000 }
90
+ }
91
+ },
92
+ "warnings": { "type": "array", "uniqueItems": true, "items": { "type": "string", "minLength": 1, "maxLength": 160 } },
93
+ "digest": { "$ref": "#/$defs/digest" }
94
+ },
95
+ "$defs": {
96
+ "name": { "type": "string", "minLength": 1, "maxLength": 160, "pattern": "^[A-Za-z0-9][A-Za-z0-9._:/-]*$" },
97
+ "digest": { "type": "string", "pattern": "^sha256:[0-9a-f]{64}$" },
98
+ "dateTime": { "type": "string", "format": "date-time" },
99
+ "pin": {
100
+ "type": "object",
101
+ "additionalProperties": false,
102
+ "required": ["id", "version", "digest"],
103
+ "properties": {
104
+ "id": { "$ref": "#/$defs/name" },
105
+ "version": { "type": "string", "minLength": 1, "maxLength": 80 },
106
+ "digest": { "$ref": "#/$defs/digest" }
107
+ }
108
+ },
109
+ "redactedChange": {
110
+ "type": "object",
111
+ "additionalProperties": false,
112
+ "required": ["path", "valueDigest", "sensitivity"],
113
+ "properties": {
114
+ "path": { "type": "string", "minLength": 0, "maxLength": 512 },
115
+ "valueDigest": { "$ref": "#/$defs/digest" },
116
+ "sensitivity": { "enum": ["public", "internal", "confidential", "restricted"] }
117
+ }
118
+ },
119
+ "row": {
120
+ "type": "object",
121
+ "additionalProperties": false,
122
+ "required": [
123
+ "variantId", "kind", "changes", "inference", "sourceReceipt", "counterfactualReceipt",
124
+ "resultDigest", "freshInvocationId", "deltas", "resourceUse", "warnings"
125
+ ],
126
+ "properties": {
127
+ "variantId": { "$ref": "#/$defs/name" },
128
+ "kind": { "enum": ["variant", "unchanged-control", "baseline-stability"] },
129
+ "changes": { "type": "array", "items": { "$ref": "#/$defs/redactedChange" } },
130
+ "inference": { "enum": ["reused-stored-evidence", "new-invocation", "deduplicated-control", "unreplayable"] },
131
+ "sourceReceipt": { "$ref": "#/$defs/pin" },
132
+ "counterfactualReceipt": { "anyOf": [{ "type": "null" }, { "$ref": "#/$defs/pin" }] },
133
+ "resultDigest": { "$ref": "#/$defs/digest" },
134
+ "freshInvocationId": { "type": ["string", "null"], "maxLength": 160 },
135
+ "deltas": {
136
+ "type": "object",
137
+ "additionalProperties": false,
138
+ "required": ["outcomeChanged", "acceptanceChanged", "ruleChanged", "firstChangedRule", "distributionDistanceBps", "status", "reason"],
139
+ "properties": {
140
+ "outcomeChanged": { "type": "boolean" },
141
+ "acceptanceChanged": { "type": "boolean" },
142
+ "ruleChanged": { "type": "boolean" },
143
+ "firstChangedRule": { "type": ["string", "null"], "maxLength": 160 },
144
+ "distributionDistanceBps": { "type": ["integer", "null"], "minimum": 0, "maximum": 10000 },
145
+ "status": { "type": "string", "minLength": 1, "maxLength": 80 },
146
+ "reason": { "type": "string", "minLength": 1, "maxLength": 80 }
147
+ }
148
+ },
149
+ "resourceUse": {
150
+ "type": "object",
151
+ "additionalProperties": false,
152
+ "required": ["backendCalls", "tokens", "costMicros"],
153
+ "properties": {
154
+ "backendCalls": { "type": "integer", "minimum": 0 },
155
+ "tokens": { "type": "integer", "minimum": 0 },
156
+ "costMicros": { "type": "integer", "minimum": 0 }
157
+ }
158
+ },
159
+ "warnings": { "type": "array", "uniqueItems": true, "items": { "type": "string", "minLength": 1, "maxLength": 160 } }
160
+ }
161
+ }
162
+ }
163
+ }
@@ -0,0 +1,312 @@
1
+ {
2
+ "$schema": "https://json-schema.org/draft/2020-12/schema",
3
+ "$id": "https://aiwg.io/schemas/decision/HeldoutApproval.v1.schema.json",
4
+ "title": "HeldoutApproval",
5
+ "type": "object",
6
+ "additionalProperties": false,
7
+ "required": [
8
+ "schemaVersion",
9
+ "approved",
10
+ "study",
11
+ "runId",
12
+ "reviewer",
13
+ "approvalReference",
14
+ "sourceCommit",
15
+ "exactHeadCi",
16
+ "stagingHost",
17
+ "stagingWorkspace",
18
+ "model",
19
+ "servedModel",
20
+ "region",
21
+ "credentialRef",
22
+ "credentialResolverDigest",
23
+ "corpusDigest",
24
+ "preregistrationDigest",
25
+ "executionDigest",
26
+ "calibration",
27
+ "providerTermsReference",
28
+ "priceBound",
29
+ "budget",
30
+ "priorStudySpendUsd",
31
+ "priorPortfolioSpendUsd"
32
+ ],
33
+ "properties": {
34
+ "schemaVersion": {
35
+ "const": "decision-heldout-approval/v1"
36
+ },
37
+ "approved": {
38
+ "const": true
39
+ },
40
+ "study": {
41
+ "enum": [
42
+ "D17",
43
+ "D29"
44
+ ]
45
+ },
46
+ "runId": {
47
+ "type": "string",
48
+ "pattern": "^[A-Za-z0-9][A-Za-z0-9_-]{0,63}$"
49
+ },
50
+ "reviewer": {
51
+ "type": "string",
52
+ "minLength": 1,
53
+ "maxLength": 512
54
+ },
55
+ "approvalReference": {
56
+ "type": "string",
57
+ "minLength": 1,
58
+ "maxLength": 512
59
+ },
60
+ "sourceCommit": {
61
+ "type": "string",
62
+ "pattern": "^[a-f0-9]{40}$"
63
+ },
64
+ "exactHeadCi": {
65
+ "type": "string",
66
+ "minLength": 1,
67
+ "maxLength": 512
68
+ },
69
+ "stagingHost": {
70
+ "const": "titan"
71
+ },
72
+ "stagingWorkspace": {
73
+ "type": "string",
74
+ "minLength": 1,
75
+ "maxLength": 512
76
+ },
77
+ "model": {
78
+ "const": "jev-1.13.0"
79
+ },
80
+ "servedModel": {
81
+ "const": "jev-1.13.0"
82
+ },
83
+ "region": {
84
+ "type": "string",
85
+ "minLength": 1,
86
+ "maxLength": 512
87
+ },
88
+ "credentialRef": {
89
+ "type": "string",
90
+ "pattern": "^openbao-approle\\.[a-z0-9][a-z0-9-]{0,62}\\.typesafe-jev$"
91
+ },
92
+ "credentialResolverDigest": {
93
+ "type": "string",
94
+ "pattern": "^sha256:[a-f0-9]{64}$"
95
+ },
96
+ "corpusDigest": {
97
+ "type": "string",
98
+ "pattern": "^sha256:[a-f0-9]{64}$"
99
+ },
100
+ "preregistrationDigest": {
101
+ "type": "string",
102
+ "pattern": "^sha256:[a-f0-9]{64}$"
103
+ },
104
+ "executionDigest": {
105
+ "type": "string",
106
+ "pattern": "^sha256:[a-f0-9]{64}$"
107
+ },
108
+ "gateBindingDigest": {
109
+ "type": "string",
110
+ "pattern": "^sha256:[a-f0-9]{64}$",
111
+ "description": "Explicit digest of the preregistered GateBinding under evaluation (D29 v8 gate adoption; transitive via preregistration studyAnalysisDigest)."
112
+ },
113
+ "gatePackDigests": {
114
+ "type": "array",
115
+ "description": "Authored pack pins resolved for the gateBindingDigest binding.",
116
+ "items": {
117
+ "type": "object",
118
+ "additionalProperties": false,
119
+ "required": [
120
+ "id",
121
+ "version",
122
+ "digest"
123
+ ],
124
+ "properties": {
125
+ "id": {
126
+ "type": "string",
127
+ "minLength": 1
128
+ },
129
+ "version": {
130
+ "type": "string",
131
+ "minLength": 1
132
+ },
133
+ "digest": {
134
+ "type": "string",
135
+ "pattern": "^sha256:[a-f0-9]{64}$"
136
+ },
137
+ "resolvedDigest": {
138
+ "type": "string",
139
+ "pattern": "^sha256:[a-f0-9]{64}$"
140
+ }
141
+ }
142
+ }
143
+ },
144
+ "providerTermsReference": {
145
+ "type": "string",
146
+ "minLength": 1,
147
+ "maxLength": 512
148
+ },
149
+ "priceBound": {
150
+ "type": "object",
151
+ "additionalProperties": false,
152
+ "required": [
153
+ "inputUsdPerMTok",
154
+ "outputUsdPerMTok",
155
+ "perRequestUsd",
156
+ "evidenceReferences",
157
+ "approvalReference"
158
+ ],
159
+ "properties": {
160
+ "inputUsdPerMTok": {
161
+ "type": "number",
162
+ "minimum": 0,
163
+ "maximum": 1000
164
+ },
165
+ "outputUsdPerMTok": {
166
+ "const": 0
167
+ },
168
+ "perRequestUsd": {
169
+ "type": "number",
170
+ "minimum": 0,
171
+ "maximum": 1000
172
+ },
173
+ "evidenceReferences": {
174
+ "type": "array",
175
+ "minItems": 1,
176
+ "maxItems": 8,
177
+ "items": {
178
+ "type": "string",
179
+ "minLength": 1,
180
+ "maxLength": 512
181
+ }
182
+ },
183
+ "approvalReference": {
184
+ "type": "string",
185
+ "minLength": 1,
186
+ "maxLength": 512
187
+ },
188
+ "outputTokenBound": {
189
+ "type": "integer",
190
+ "minimum": 0,
191
+ "maximum": 1000000
192
+ }
193
+ }
194
+ },
195
+ "budget": {
196
+ "type": "object",
197
+ "additionalProperties": false,
198
+ "required": [
199
+ "calls",
200
+ "tokens",
201
+ "usd"
202
+ ],
203
+ "properties": {
204
+ "calls": {
205
+ "type": "integer",
206
+ "minimum": 1,
207
+ "maximum": 9007199254740991
208
+ },
209
+ "tokens": {
210
+ "type": "integer",
211
+ "minimum": 1,
212
+ "maximum": 9007199254740991
213
+ },
214
+ "usd": {
215
+ "type": "number",
216
+ "exclusiveMinimum": 0,
217
+ "maximum": 8
218
+ }
219
+ }
220
+ },
221
+ "priorStudySpendUsd": {
222
+ "type": "number",
223
+ "minimum": 0,
224
+ "maximum": 1000
225
+ },
226
+ "priorPortfolioSpendUsd": {
227
+ "type": "number",
228
+ "minimum": 0,
229
+ "maximum": 1000
230
+ },
231
+ "calibration": {
232
+ "oneOf": [
233
+ {
234
+ "type": "object",
235
+ "additionalProperties": false,
236
+ "required": [
237
+ "mode"
238
+ ],
239
+ "properties": {
240
+ "mode": {
241
+ "const": "uncalibrated-diagnostic"
242
+ }
243
+ }
244
+ },
245
+ {
246
+ "type": "object",
247
+ "additionalProperties": false,
248
+ "required": [
249
+ "mode",
250
+ "phase"
251
+ ],
252
+ "properties": {
253
+ "mode": {
254
+ "const": "staged"
255
+ },
256
+ "phase": {
257
+ "const": "calibration"
258
+ }
259
+ }
260
+ },
261
+ {
262
+ "type": "object",
263
+ "additionalProperties": false,
264
+ "required": [
265
+ "mode",
266
+ "phase",
267
+ "calibrationArtifactDigest",
268
+ "calibrationPhaseRecordDigest",
269
+ "priorApprovalDigest"
270
+ ],
271
+ "properties": {
272
+ "mode": {
273
+ "const": "staged"
274
+ },
275
+ "phase": {
276
+ "const": "test"
277
+ },
278
+ "calibrationArtifactDigest": {
279
+ "type": "string",
280
+ "pattern": "^sha256:[a-f0-9]{64}$"
281
+ },
282
+ "calibrationPhaseRecordDigest": {
283
+ "type": "string",
284
+ "pattern": "^sha256:[a-f0-9]{64}$"
285
+ },
286
+ "priorApprovalDigest": {
287
+ "type": "string",
288
+ "pattern": "^sha256:[a-f0-9]{64}$"
289
+ }
290
+ }
291
+ },
292
+ {
293
+ "type": "object",
294
+ "additionalProperties": false,
295
+ "required": [
296
+ "mode",
297
+ "calibrationArtifactDigest"
298
+ ],
299
+ "properties": {
300
+ "mode": {
301
+ "const": "artifact"
302
+ },
303
+ "calibrationArtifactDigest": {
304
+ "type": "string",
305
+ "pattern": "^sha256:[a-f0-9]{64}$"
306
+ }
307
+ }
308
+ }
309
+ ]
310
+ }
311
+ }
312
+ }
@@ -0,0 +1,211 @@
1
+ {
2
+ "$schema": "https://json-schema.org/draft/2020-12/schema",
3
+ "$id": "https://aiwg.io/schemas/decision/HeldoutAttempt.v1.schema.json",
4
+ "title": "HeldoutAttempt",
5
+ "type": "object",
6
+ "additionalProperties": false,
7
+ "required": [
8
+ "schemaVersion",
9
+ "study",
10
+ "runId",
11
+ "corpusDigest",
12
+ "preregistrationDigest",
13
+ "approvalDigest",
14
+ "rowId",
15
+ "requestId",
16
+ "ordinal",
17
+ "reservedUsdMicros",
18
+ "reservedTokens",
19
+ "requestDigest",
20
+ "result"
21
+ ],
22
+ "properties": {
23
+ "schemaVersion": {
24
+ "const": "decision-heldout-attempt/v1"
25
+ },
26
+ "study": {
27
+ "enum": [
28
+ "D17",
29
+ "D29"
30
+ ]
31
+ },
32
+ "runId": {
33
+ "type": "string",
34
+ "pattern": "^[A-Za-z0-9][A-Za-z0-9_-]{0,63}$"
35
+ },
36
+ "corpusDigest": {
37
+ "type": "string",
38
+ "pattern": "^sha256:[a-f0-9]{64}$"
39
+ },
40
+ "preregistrationDigest": {
41
+ "type": "string",
42
+ "pattern": "^sha256:[a-f0-9]{64}$"
43
+ },
44
+ "approvalDigest": {
45
+ "type": "string",
46
+ "pattern": "^sha256:[a-f0-9]{64}$"
47
+ },
48
+ "rowId": {
49
+ "type": "string",
50
+ "pattern": "^[A-Za-z0-9][A-Za-z0-9_-]{0,63}$"
51
+ },
52
+ "requestId": {
53
+ "type": "string",
54
+ "pattern": "^[A-Za-z0-9][A-Za-z0-9_-]{0,63}$"
55
+ },
56
+ "ordinal": {
57
+ "enum": [
58
+ 1,
59
+ 2
60
+ ]
61
+ },
62
+ "reservedUsdMicros": {
63
+ "type": "integer",
64
+ "minimum": 1,
65
+ "maximum": 9007199254740991
66
+ },
67
+ "reservedTokens": {
68
+ "type": "integer",
69
+ "minimum": 1,
70
+ "maximum": 9007199254740991
71
+ },
72
+ "requestDigest": {
73
+ "type": "string",
74
+ "pattern": "^sha256:[a-f0-9]{64}$"
75
+ },
76
+ "result": {
77
+ "anyOf": [
78
+ {
79
+ "type": "object",
80
+ "additionalProperties": false,
81
+ "required": [
82
+ "disposition",
83
+ "reason",
84
+ "servedModel",
85
+ "inputTokens",
86
+ "outputTokens",
87
+ "providerCostUsd",
88
+ "responseDigest",
89
+ "receipt",
90
+ "receiptDigest",
91
+ "traceDigest",
92
+ "latencyMs",
93
+ "accountedUsdMicros"
94
+ ],
95
+ "properties": {
96
+ "disposition": {
97
+ "enum": [
98
+ "success",
99
+ "retryable",
100
+ "measurement-failure",
101
+ "stop"
102
+ ]
103
+ },
104
+ "reason": {
105
+ "type": "string",
106
+ "minLength": 1,
107
+ "maxLength": 512
108
+ },
109
+ "servedModel": {
110
+ "anyOf": [
111
+ {
112
+ "type": "string",
113
+ "minLength": 1,
114
+ "maxLength": 512
115
+ },
116
+ {
117
+ "type": "null"
118
+ }
119
+ ]
120
+ },
121
+ "inputTokens": {
122
+ "anyOf": [
123
+ {
124
+ "type": "integer",
125
+ "minimum": 0,
126
+ "maximum": 9007199254740991
127
+ },
128
+ {
129
+ "type": "null"
130
+ }
131
+ ]
132
+ },
133
+ "outputTokens": {
134
+ "anyOf": [
135
+ {
136
+ "type": "integer",
137
+ "minimum": 0,
138
+ "maximum": 9007199254740991
139
+ },
140
+ {
141
+ "type": "null"
142
+ }
143
+ ]
144
+ },
145
+ "providerCostUsd": {
146
+ "anyOf": [
147
+ {
148
+ "type": "number",
149
+ "minimum": 0,
150
+ "maximum": 1000
151
+ },
152
+ {
153
+ "type": "null"
154
+ }
155
+ ]
156
+ },
157
+ "responseDigest": {
158
+ "anyOf": [
159
+ {
160
+ "type": "string",
161
+ "pattern": "^sha256:[a-f0-9]{64}$"
162
+ },
163
+ {
164
+ "type": "null"
165
+ }
166
+ ]
167
+ },
168
+ "receipt": {
169
+ "anyOf": [
170
+ {
171
+ "type": "object"
172
+ },
173
+ {
174
+ "type": "null"
175
+ }
176
+ ]
177
+ },
178
+ "receiptDigest": {
179
+ "anyOf": [
180
+ {
181
+ "type": "string",
182
+ "pattern": "^sha256:[a-f0-9]{64}$"
183
+ },
184
+ {
185
+ "type": "null"
186
+ }
187
+ ]
188
+ },
189
+ "traceDigest": {
190
+ "type": "string",
191
+ "pattern": "^sha256:[a-f0-9]{64}$"
192
+ },
193
+ "latencyMs": {
194
+ "type": "integer",
195
+ "minimum": 0,
196
+ "maximum": 9007199254740991
197
+ },
198
+ "accountedUsdMicros": {
199
+ "type": "integer",
200
+ "minimum": 1,
201
+ "maximum": 9007199254740991
202
+ }
203
+ }
204
+ },
205
+ {
206
+ "type": "null"
207
+ }
208
+ ]
209
+ }
210
+ }
211
+ }