@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 +1,155 @@
1
+ {
2
+ "$schema": "https://json-schema.org/draft/2020-12/schema",
3
+ "$id": "https://aiwg.io/schemas/decision/DecisionFeatureSet.v1.schema.json",
4
+ "title": "AIWG Decision Feature Set v1",
5
+ "description": "Versioned probabilistic decision feature-set contract with source pins, lifecycle policy binding, deterministic columns, missingness, provenance and train/serve compatibility.",
6
+ "type": "object",
7
+ "required": ["schemaVersion", "kind", "metadata", "spec"],
8
+ "additionalProperties": false,
9
+ "properties": {
10
+ "schemaVersion": { "const": "decision-feature-export/v1" },
11
+ "kind": { "const": "DecisionFeatureSet" },
12
+ "metadata": {
13
+ "type": "object",
14
+ "required": ["id", "version", "description", "digest"],
15
+ "additionalProperties": false,
16
+ "properties": {
17
+ "id": { "type": "string", "minLength": 1 },
18
+ "version": { "type": "string", "minLength": 1 },
19
+ "description": { "type": "string" },
20
+ "digest": { "$ref": "#/$defs/sha256" }
21
+ }
22
+ },
23
+ "spec": {
24
+ "type": "object",
25
+ "required": ["pins", "compatibleServedModels", "missingValuePolicy", "ordering", "privacy", "columns"],
26
+ "additionalProperties": false,
27
+ "properties": {
28
+ "pins": {
29
+ "type": "object",
30
+ "required": ["ruleset", "binding", "definitions", "adapters", "calibration"],
31
+ "additionalProperties": false,
32
+ "properties": {
33
+ "ruleset": { "$ref": "#/$defs/pin" },
34
+ "binding": { "$ref": "#/$defs/pin" },
35
+ "definitions": {
36
+ "type": "object",
37
+ "additionalProperties": { "$ref": "#/$defs/pin" }
38
+ },
39
+ "adapters": {
40
+ "type": "object",
41
+ "additionalProperties": {
42
+ "type": "object",
43
+ "required": ["adapter", "adapterVersion", "requestedModel", "servedModel"],
44
+ "additionalProperties": false,
45
+ "properties": {
46
+ "adapter": { "type": "string", "minLength": 1 },
47
+ "adapterVersion": { "type": "string", "minLength": 1 },
48
+ "requestedModel": { "type": "string", "minLength": 1 },
49
+ "servedModel": { "type": ["string", "null"] }
50
+ }
51
+ }
52
+ },
53
+ "calibration": {
54
+ "type": "object",
55
+ "additionalProperties": {
56
+ "type": "object",
57
+ "required": ["calibrationRef", "maxAgeMs", "recordedAtEpochMs"],
58
+ "additionalProperties": false,
59
+ "properties": {
60
+ "calibrationRef": { "type": "string", "minLength": 1 },
61
+ "maxAgeMs": { "type": ["integer", "null"], "minimum": 1 },
62
+ "recordedAtEpochMs": { "type": ["integer", "null"], "minimum": 0 }
63
+ }
64
+ }
65
+ }
66
+ }
67
+ },
68
+ "compatibleServedModels": {
69
+ "type": "object",
70
+ "additionalProperties": {
71
+ "type": "array",
72
+ "items": { "type": "string", "minLength": 1 }
73
+ }
74
+ },
75
+ "missingValuePolicy": {
76
+ "type": "object",
77
+ "required": ["numeric", "categorical", "missingStateColumn", "missingNeverZero"],
78
+ "additionalProperties": false,
79
+ "properties": {
80
+ "numeric": { "const": "null" },
81
+ "categorical": { "const": "null" },
82
+ "missingStateColumn": { "const": true },
83
+ "missingNeverZero": { "const": true }
84
+ }
85
+ },
86
+ "ordering": { "const": "definition-pins-declared-order" },
87
+ "privacy": {
88
+ "type": "object",
89
+ "required": ["classification", "retentionMs", "accessScope", "exportPolicyId", "deletion", "lifecycleSurface"],
90
+ "additionalProperties": false,
91
+ "properties": {
92
+ "classification": { "enum": ["internal", "confidential", "restricted"] },
93
+ "retentionMs": { "type": "integer", "minimum": 1 },
94
+ "accessScope": { "type": "string", "minLength": 1 },
95
+ "exportPolicyId": { "type": "string", "minLength": 1 },
96
+ "deletion": { "enum": ["erase", "tombstone"] },
97
+ "lifecycleSurface": { "const": "export" }
98
+ }
99
+ },
100
+ "columns": {
101
+ "type": "array",
102
+ "minItems": 1,
103
+ "items": {
104
+ "type": "object",
105
+ "required": ["name", "alias", "primitive", "valueType", "nullable", "order", "source", "uncertaintyProfile", "calibrationRef", "derivation", "description"],
106
+ "additionalProperties": false,
107
+ "properties": {
108
+ "name": { "type": "string", "minLength": 1 },
109
+ "alias": { "type": "string", "minLength": 1 },
110
+ "primitive": { "enum": ["choice", "ordinal-score", "truth-probability", "ruleset"] },
111
+ "valueType": { "enum": ["number", "string", "boolean"] },
112
+ "nullable": { "type": "boolean" },
113
+ "order": { "type": "integer", "minimum": 0 },
114
+ "domain": { "type": "array", "items": { "type": "string" } },
115
+ "source": { "$ref": "#/$defs/provenance" },
116
+ "uncertaintyProfile": { "type": ["string", "null"] },
117
+ "calibrationRef": { "type": ["string", "null"] },
118
+ "derivation": {
119
+ "anyOf": [
120
+ { "type": "null" },
121
+ {
122
+ "type": "object",
123
+ "required": ["algorithm", "version"],
124
+ "additionalProperties": false,
125
+ "properties": {
126
+ "algorithm": { "type": "string", "minLength": 1 },
127
+ "version": { "type": "string", "minLength": 1 }
128
+ }
129
+ }
130
+ ]
131
+ },
132
+ "description": { "type": "string" }
133
+ }
134
+ }
135
+ }
136
+ }
137
+ }
138
+ },
139
+ "$defs": {
140
+ "sha256": { "type": "string", "pattern": "^sha256:[a-f0-9]{64}$" },
141
+ "pin": {
142
+ "type": "object",
143
+ "required": ["id", "version", "digest"],
144
+ "additionalProperties": false,
145
+ "properties": {
146
+ "id": { "type": "string", "minLength": 1 },
147
+ "version": { "type": "string", "minLength": 1 },
148
+ "digest": { "$ref": "#/$defs/sha256" }
149
+ }
150
+ },
151
+ "provenance": {
152
+ "enum": ["provider-native", "model-self-report", "adapter-one-hot", "derived", "calibrated", "missing"]
153
+ }
154
+ }
155
+ }
@@ -0,0 +1,85 @@
1
+ {
2
+ "$schema": "https://json-schema.org/draft/2020-12/schema",
3
+ "$id": "https://aiwg.io/schemas/decision/DecisionInteropMapping.v1.schema.json",
4
+ "type": "object",
5
+ "properties": {
6
+ "apiVersion": { "const": "decision-interop.aiwg.io/v1" },
7
+ "kind": { "const": "DecisionInteropMapping" },
8
+ "metadata": {
9
+ "type": "object",
10
+ "properties": {
11
+ "id": { "type": "string", "minLength": 1 },
12
+ "version": { "type": "string", "pattern": "^\\d+\\.\\d+\\.\\d+$" },
13
+ "description": { "type": "string", "minLength": 1 }
14
+ },
15
+ "required": ["id", "version", "description"],
16
+ "additionalProperties": false
17
+ },
18
+ "spec": {
19
+ "type": "object",
20
+ "properties": {
21
+ "source": { "$ref": "#/$defs/source" },
22
+ "dryRun": { "const": true },
23
+ "activation": { "const": "requires-review-publish" },
24
+ "profile": { "$ref": "#/$defs/dmnProfile" },
25
+ "normalizedRuleset": { "$ref": "https://aiwg.io/schemas/decision/DecisionRuleset.v1alpha2.schema.json" },
26
+ "normalizedDigest": { "type": "string", "pattern": "^sha256:[0-9a-f]{64}$" },
27
+ "warnings": { "type": "array", "items": { "type": "string" } },
28
+ "errors": { "type": "array", "items": { "type": "string" } },
29
+ "roundTripStatus": { "enum": ["not-run", "equivalent", "not-equivalent"] }
30
+ },
31
+ "required": ["source", "dryRun", "activation", "profile", "normalizedRuleset", "normalizedDigest", "warnings", "errors", "roundTripStatus"],
32
+ "additionalProperties": false
33
+ }
34
+ },
35
+ "required": ["apiVersion", "kind", "metadata", "spec"],
36
+ "additionalProperties": false,
37
+ "$defs": {
38
+ "source": {
39
+ "type": "object",
40
+ "properties": {
41
+ "format": { "enum": ["dmn-xml", "opa-profile", "aiwg-profile"] },
42
+ "version": { "type": "string", "minLength": 1 },
43
+ "origin": { "type": "string", "minLength": 1 },
44
+ "digest": { "type": "string", "pattern": "^sha256:[0-9a-f]{64}$" },
45
+ "signature": {
46
+ "type": "object",
47
+ "properties": {
48
+ "state": { "enum": ["not-provided", "verified", "untrusted"] },
49
+ "keyId": { "type": "string", "minLength": 1 }
50
+ },
51
+ "required": ["state"],
52
+ "additionalProperties": false
53
+ },
54
+ "parser": {
55
+ "type": "object",
56
+ "properties": {
57
+ "id": { "const": "aiwg-decision-interop" },
58
+ "version": { "const": "decision-interop.aiwg.io/v1" }
59
+ },
60
+ "required": ["id", "version"],
61
+ "additionalProperties": false
62
+ },
63
+ "bounds": { "type": "object" }
64
+ },
65
+ "required": ["format", "version", "origin", "digest", "signature", "parser", "bounds"],
66
+ "additionalProperties": false
67
+ },
68
+ "dmnProfile": {
69
+ "type": "object",
70
+ "properties": {
71
+ "profileVersion": { "const": "decision-interop.aiwg.io/v1" },
72
+ "dmnVersion": { "const": "1.6" },
73
+ "namespace": { "const": "https://www.omg.org/spec/DMN/20240513/MODEL/" },
74
+ "hitPolicy": { "enum": ["UNIQUE", "ANY", "FIRST", "RULE ORDER", "COLLECT"] },
75
+ "aggregation": { "enum": ["LIST", "SUM", "MIN", "MAX", "COUNT", null] },
76
+ "ruleOrder": { "type": "array", "items": { "type": "string", "minLength": 1 }, "minItems": 1 },
77
+ "supportedUnaryTests": { "type": "array", "items": { "enum": ["dash", "literal", "eq", "ne", "lt", "lte", "gt", "gte"] }, "minItems": 1 },
78
+ "unsupportedConstructs": { "type": "array", "items": { "type": "string" } },
79
+ "extensions": { "type": "array", "items": { "enum": ["aiwg:order", "aiwg:defer-on-conflict"] } }
80
+ },
81
+ "required": ["profileVersion", "dmnVersion", "namespace", "hitPolicy", "aggregation", "ruleOrder", "supportedUnaryTests", "unsupportedConstructs", "extensions"],
82
+ "additionalProperties": false
83
+ }
84
+ }
85
+ }
@@ -0,0 +1,174 @@
1
+ {
2
+ "$schema": "https://json-schema.org/draft/2020-12/schema",
3
+ "$id": "https://aiwg.io/schemas/decision/DecisionIssueTriageEvaluationManifest.v1.schema.json",
4
+ "title": "DecisionIssueTriageEvaluationManifest.v1",
5
+ "description": "Preregistered offline/holdout manifest for the issue triage shadow pilot.",
6
+ "type": "object",
7
+ "additionalProperties": false,
8
+ "required": [
9
+ "schemaVersion",
10
+ "id",
11
+ "pilotPack",
12
+ "calibration",
13
+ "dataset",
14
+ "slices",
15
+ "thresholds",
16
+ "holdout",
17
+ "integrity"
18
+ ],
19
+ "properties": {
20
+ "schemaVersion": { "const": "decision-issue-triage-evaluation-manifest/v1" },
21
+ "id": { "type": "string", "minLength": 1 },
22
+ "pilotPack": { "$ref": "#/$defs/pin" },
23
+ "calibration": {
24
+ "description": "Calibration registry, policy and per-sample request rule frozen before holdout access; all null pins no registry.",
25
+ "type": "object",
26
+ "additionalProperties": false,
27
+ "required": ["registryDigest", "policyDigest", "requestRule"],
28
+ "properties": {
29
+ "registryDigest": { "type": ["string", "null"], "pattern": "^sha256:[a-f0-9]{64}$" },
30
+ "policyDigest": { "type": ["string", "null"], "pattern": "^sha256:[a-f0-9]{64}$" },
31
+ "requestRule": {
32
+ "type": ["object", "null"],
33
+ "additionalProperties": false,
34
+ "required": ["requestedAlias", "runIdPrefix", "calibrationArtifactId", "notBefore"],
35
+ "properties": {
36
+ "requestedAlias": { "type": "string", "minLength": 1 },
37
+ "runIdPrefix": { "type": "string", "minLength": 1 },
38
+ "calibrationArtifactId": { "type": ["string", "null"], "minLength": 1 },
39
+ "notBefore": { "type": "string", "format": "date-time" }
40
+ }
41
+ }
42
+ }
43
+ },
44
+ "dataset": {
45
+ "type": "object",
46
+ "additionalProperties": false,
47
+ "required": ["id", "digest", "splitMethod", "minimumTotal", "minimumPerSupportedSlice", "splitDigests"],
48
+ "properties": {
49
+ "id": { "type": "string", "minLength": 1 },
50
+ "digest": { "$ref": "#/$defs/digest" },
51
+ "splitMethod": { "enum": ["time", "source"] },
52
+ "minimumTotal": { "type": "integer", "minimum": 1 },
53
+ "minimumPerSupportedSlice": { "type": "integer", "minimum": 1 },
54
+ "splitDigests": {
55
+ "description": "Digests of the frozen tuning, calibration and test memberships, pinned before holdout access.",
56
+ "type": "object",
57
+ "additionalProperties": false,
58
+ "required": ["tuning", "calibration", "test"],
59
+ "properties": {
60
+ "tuning": { "$ref": "#/$defs/digest" },
61
+ "calibration": { "$ref": "#/$defs/digest" },
62
+ "test": { "$ref": "#/$defs/digest" }
63
+ }
64
+ }
65
+ }
66
+ },
67
+ "slices": {
68
+ "type": "array",
69
+ "minItems": 5,
70
+ "items": {
71
+ "type": "object",
72
+ "additionalProperties": false,
73
+ "required": ["id", "dimension", "minimumSupport", "whenInsufficient", "aggregation"],
74
+ "properties": {
75
+ "id": { "type": "string", "minLength": 1 },
76
+ "dimension": {
77
+ "enum": ["issue-type", "area", "state-completeness", "adversarial-authority", "new-or-unseen-category"]
78
+ },
79
+ "minimumSupport": { "type": "integer", "minimum": 1 },
80
+ "whenInsufficient": { "const": "label-insufficient" },
81
+ "aggregation": { "enum": ["report-supported", "suppress-small-n"] }
82
+ }
83
+ }
84
+ },
85
+ "thresholds": {
86
+ "type": "object",
87
+ "additionalProperties": false,
88
+ "required": [
89
+ "maximumFalseDuplicateRate",
90
+ "maximumFalseAutoRate",
91
+ "qualityNonInferiorityMargin",
92
+ "minimumAcceptedCoverage",
93
+ "confidenceInterval",
94
+ "minimumTotalSamples",
95
+ "minimumPerSliceSamples",
96
+ "minimumDuplicateSamples",
97
+ "minimumDuplicateRecall",
98
+ "benefit"
99
+ ],
100
+ "properties": {
101
+ "maximumFalseDuplicateRate": { "$ref": "#/$defs/unit" },
102
+ "maximumFalseAutoRate": { "$ref": "#/$defs/unit" },
103
+ "qualityNonInferiorityMargin": {
104
+ "description": "Cascade-minus-baseline accuracy margin; non-positive. -0.02 allows the cascade to be at most two points worse. Superiority (positive) margins are not supported.",
105
+ "type": "number",
106
+ "minimum": -1,
107
+ "maximum": 0
108
+ },
109
+ "minimumAcceptedCoverage": { "$ref": "#/$defs/unit" },
110
+ "confidenceInterval": {
111
+ "type": "object",
112
+ "additionalProperties": false,
113
+ "required": ["method", "level"],
114
+ "properties": {
115
+ "method": { "enum": ["wilson", "bootstrap", "exact"] },
116
+ "level": { "type": "number", "exclusiveMinimum": 0, "exclusiveMaximum": 1 }
117
+ }
118
+ },
119
+ "minimumTotalSamples": { "type": "integer", "minimum": 1 },
120
+ "minimumPerSliceSamples": { "type": "integer", "minimum": 1 },
121
+ "minimumDuplicateSamples": {
122
+ "description": "Minimum labeled duplicate cases before duplicate metrics can support PROMOTE.",
123
+ "type": "integer",
124
+ "minimum": 1
125
+ },
126
+ "minimumDuplicateRecall": {
127
+ "description": "Minimum Wilson lower bound of duplicate recall at the preregistered level.",
128
+ "$ref": "#/$defs/unit"
129
+ },
130
+ "benefit": {
131
+ "type": "object",
132
+ "additionalProperties": false,
133
+ "required": ["mustBePositive", "metric"],
134
+ "properties": {
135
+ "mustBePositive": { "const": true },
136
+ "metric": { "enum": ["total-task-token-cost", "reviewer-time"] }
137
+ }
138
+ }
139
+ }
140
+ },
141
+ "holdout": {
142
+ "type": "object",
143
+ "additionalProperties": false,
144
+ "required": ["thresholdsRegisteredAt", "holdoutAccessedAt"],
145
+ "properties": {
146
+ "thresholdsRegisteredAt": { "type": "string", "format": "date-time" },
147
+ "holdoutAccessedAt": { "type": ["string", "null"], "format": "date-time" }
148
+ }
149
+ },
150
+ "integrity": {
151
+ "type": "object",
152
+ "additionalProperties": false,
153
+ "required": ["extendsEvalIntegrity", "preserveGateDecision"],
154
+ "properties": {
155
+ "extendsEvalIntegrity": { "const": true },
156
+ "preserveGateDecision": { "const": true }
157
+ }
158
+ }
159
+ },
160
+ "$defs": {
161
+ "digest": { "type": "string", "pattern": "^sha256:[a-f0-9]{64}$" },
162
+ "pin": {
163
+ "type": "object",
164
+ "additionalProperties": false,
165
+ "required": ["id", "version", "digest"],
166
+ "properties": {
167
+ "id": { "type": "string", "minLength": 1 },
168
+ "version": { "type": "string", "minLength": 1 },
169
+ "digest": { "$ref": "#/$defs/digest" }
170
+ }
171
+ },
172
+ "unit": { "type": "number", "minimum": 0, "maximum": 1 }
173
+ }
174
+ }