@aiwg/cli 2026.9.21 → 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.
- package/agentic/code/providers/capability-matrix.yaml +2 -1
- package/dist/src/artifacts/cli.js +23 -18
- package/dist/src/artifacts/discovery-shadow-evaluator.js +177 -0
- package/dist/src/artifacts/discovery-shadow.js +195 -0
- package/dist/src/artifacts/fortemi-core-query-adapter.js +6 -2
- package/dist/src/artifacts/index-builder.js +47 -9
- package/dist/src/artifacts/query-engine.js +108 -5
- package/dist/src/artifacts/types.js +1 -0
- package/dist/src/cli/handlers/decision.js +88 -0
- package/dist/src/cli/handlers/gates.js +154 -0
- package/dist/src/cli/handlers/index.js +5 -1
- package/dist/src/cli/handlers/sessions.js +10 -7
- package/dist/src/cli/handlers/utilities.js +20 -0
- package/dist/src/config/aiwg-config.js +15 -2
- package/dist/src/decision/acceptance.js +25 -0
- package/dist/src/decision/adapters/jev.js +24 -2
- package/dist/src/decision/calibration/registry.js +9 -0
- package/dist/src/decision/compose.js +74 -15
- package/dist/src/decision/context-live-promotion.js +344 -0
- package/dist/src/decision/context-live-qualification.js +591 -0
- package/dist/src/decision/context-pruning.js +847 -0
- package/dist/src/decision/driver.js +430 -0
- package/dist/src/decision/egress-live-qualification.js +874 -0
- package/dist/src/decision/ensemble/contract.js +64 -1
- package/dist/src/decision/ensemble/index.js +1 -0
- package/dist/src/decision/ensemble/report.js +8 -42
- package/dist/src/decision/ensemble/runtime.js +571 -0
- package/dist/src/decision/ensemble-study/artifacts.js +85 -0
- package/dist/src/decision/ensemble-study/calibration.js +443 -0
- package/dist/src/decision/ensemble-study/corpus.js +229 -0
- package/dist/src/decision/ensemble-study/multifact-artifacts.js +23 -0
- package/dist/src/decision/ensemble-study/multifact-study.js +596 -0
- package/dist/src/decision/ensemble-study/multifact.js +241 -0
- package/dist/src/decision/ensemble-study/protocol.js +77 -0
- package/dist/src/decision/ensemble-study/score.js +462 -0
- package/dist/src/decision/ensemble-study/staged.js +87 -0
- package/dist/src/decision/ensemble-study/statistics.js +84 -0
- package/dist/src/decision/evaluate.js +227 -14
- package/dist/src/decision/export.js +4 -0
- package/dist/src/decision/feature-export.js +1074 -0
- package/dist/src/decision/graph-index.js +16 -0
- package/dist/src/decision/graph-live-qualification.js +1041 -0
- package/dist/src/decision/graph-live-workload.js +245 -0
- package/dist/src/decision/graph-skill-bridge.js +21 -3
- package/dist/src/decision/heldout/calibration.js +140 -0
- package/dist/src/decision/heldout/collector.js +430 -0
- package/dist/src/decision/heldout/contract.js +309 -0
- package/dist/src/decision/heldout/d29-generator-ids.js +12 -0
- package/dist/src/decision/heldout/d29-passage-baseline-v2.js +93 -0
- package/dist/src/decision/heldout/d29-passage-baseline-v3.js +100 -0
- package/dist/src/decision/heldout/d29-passage-baseline.js +66 -0
- package/dist/src/decision/heldout/d29-pools-v8.js +575 -0
- package/dist/src/decision/heldout/d29-pools.js +288 -0
- package/dist/src/decision/heldout/d29-v6.js +169 -0
- package/dist/src/decision/heldout/d29-v7.js +202 -0
- package/dist/src/decision/heldout/d29-v8.js +586 -0
- package/dist/src/decision/heldout/generator-registry.js +67 -0
- package/dist/src/decision/heldout/generators.js +696 -0
- package/dist/src/decision/heldout/index.js +8 -0
- package/dist/src/decision/heldout/journal.js +346 -0
- package/dist/src/decision/heldout/study-generators.js +64 -0
- package/dist/src/decision/heldout/types.js +2 -0
- package/dist/src/decision/host-policy.js +61 -0
- package/dist/src/decision/index.js +11 -0
- package/dist/src/decision/interop-profile.v1.json +112 -0
- package/dist/src/decision/interop.js +653 -0
- package/dist/src/decision/issue-triage/index.js +3 -0
- package/dist/src/decision/issue-triage/runtime.js +1069 -0
- package/dist/src/decision/issue-triage/types.js +2 -0
- package/dist/src/decision/patterns/catalog.js +28 -8
- package/dist/src/decision/patterns/playground.js +28 -11
- package/dist/src/decision/patterns/specs.js +19 -10
- package/dist/src/decision/preprocessed-evidence.js +724 -0
- package/dist/src/decision/qualification/capture.js +8 -2
- package/dist/src/decision/qualification/gate-evidence.js +19 -4
- package/dist/src/decision/qualification/load.js +8 -1
- package/dist/src/decision/qualification/quality.js +57 -49
- package/dist/src/decision/qualification/release.js +76 -3
- package/dist/src/decision/qualification/runner.js +5 -2
- package/dist/src/decision/routing/contract.js +175 -0
- package/dist/src/decision/routing/index.js +5 -0
- package/dist/src/decision/routing/report.js +216 -0
- package/dist/src/decision/routing/runtime.js +685 -0
- package/dist/src/decision/routing/types.js +2 -0
- package/dist/src/decision/sdlc-screening.js +1157 -0
- package/dist/src/decision/sensitivity/comparative.js +154 -0
- package/dist/src/decision/sensitivity/contract.js +218 -0
- package/dist/src/decision/sensitivity/index.js +5 -0
- package/dist/src/decision/sensitivity/runtime.js +912 -0
- package/dist/src/decision/sensitivity/types.js +2 -0
- package/dist/src/decision/validate.js +26 -0
- package/dist/src/extensions/commands/definitions.js +38 -0
- package/dist/src/extensions/manifest.js +44 -0
- package/dist/src/gates/discovery.js +315 -0
- package/dist/src/gates/driver.js +388 -0
- package/dist/src/gates/evaluate.js +719 -0
- package/dist/src/gates/floors.js +255 -0
- package/dist/src/gates/index.js +17 -0
- package/dist/src/gates/providers/evidence.js +38 -0
- package/dist/src/gates/providers/index.js +21 -0
- package/dist/src/gates/providers/isolated-child-source.js +446 -0
- package/dist/src/gates/providers/loader.js +1069 -0
- package/dist/src/gates/providers/paired.js +52 -0
- package/dist/src/gates/providers/proportion.js +60 -0
- package/dist/src/gates/providers/registry.js +40 -0
- package/dist/src/gates/providers/scalar.js +41 -0
- package/dist/src/gates/providers/screening.js +53 -0
- package/dist/src/gates/providers/types.js +25 -0
- package/dist/src/gates/registry.js +1003 -0
- package/dist/src/gates/report.js +102 -0
- package/dist/src/gates/schema.js +56 -0
- package/dist/src/gates/stats/binomial.js +26 -0
- package/dist/src/gates/stats/bootstrap.js +62 -0
- package/dist/src/gates/stats/clopper-pearson.js +95 -0
- package/dist/src/gates/stats/digest.js +26 -0
- package/dist/src/gates/stats/error.js +5 -0
- package/dist/src/gates/stats/index.js +9 -0
- package/dist/src/gates/stats/normal.js +35 -0
- package/dist/src/gates/stats/paired.js +120 -0
- package/dist/src/gates/stats/splits.js +30 -0
- package/dist/src/gates/types.js +8 -0
- package/dist/src/governance/redaction.js +14 -4
- package/dist/src/mcp/muse-native-config.mjs +123 -0
- package/dist/src/mcp/tools/decision.mjs +188 -0
- package/dist/src/mcp/tools/subsystems.mjs +2 -0
- package/dist/src/providers/capability-matrix.yaml +2 -1
- package/dist/src/providers/provider-definitions.js +2 -1
- package/dist/src/sessions/adapters/muse.js +336 -39
- package/dist/src/sessions/batch-import.js +7 -0
- package/dist/src/sessions/workspace-discovery.js +13 -2
- package/package.json +1 -1
- package/schemas/decision/ConformalFrozenOpenData.v2.schema.json +371 -0
- package/schemas/decision/ConformalOpenDataReport.v2.schema.json +421 -0
- package/schemas/decision/ConformalPreregistration.v2.schema.json +529 -0
- package/schemas/decision/ContextPruningCandidate.v1.schema.json +47 -0
- package/schemas/decision/ContextPruningEvaluationReport.v1.schema.json +758 -0
- package/schemas/decision/ContextPruningPreregistration.v1.schema.json +175 -0
- package/schemas/decision/ContextPruningReceipt.v1.schema.json +72 -0
- package/schemas/decision/D17Multifact.v1.schema.json +1851 -0
- package/schemas/decision/D17StudyAnalysis.v1.schema.json +42 -0
- package/schemas/decision/D17StudyAnalysis.v2.schema.json +51 -0
- package/schemas/decision/D17StudyCalibration.v1.schema.json +819 -0
- package/schemas/decision/D17StudyDryRun.v1.schema.json +169 -0
- package/schemas/decision/D17StudyDryRun.v2.schema.json +239 -0
- package/schemas/decision/D17StudyGold.v1.schema.json +108 -0
- package/schemas/decision/D17StudyGuide.v1.schema.json +63 -0
- package/schemas/decision/D17StudyNativeTemplates.v1.schema.json +619 -0
- package/schemas/decision/D17StudyProtocol.v1.schema.json +241 -0
- package/schemas/decision/D17StudyReport.v1.schema.json +902 -0
- package/schemas/decision/D17StudyReport.v2.schema.json +1319 -0
- package/schemas/decision/D17StudyReview.v1.schema.json +175 -0
- package/schemas/decision/D17StudySplits.v1.schema.json +217 -0
- package/schemas/decision/D29Study.v1.schema.json +1903 -0
- package/schemas/decision/D29Study.v2.schema.json +1344 -0
- package/schemas/decision/D29Study.v3.schema.json +609 -0
- package/schemas/decision/D29Study.v4.schema.json +1133 -0
- package/schemas/decision/D29Study.v6.schema.json +1205 -0
- package/schemas/decision/D29Study.v7.schema.json +1471 -0
- package/schemas/decision/D29Study.v8.schema.json +671 -0
- package/schemas/decision/D29Study.v9.schema.json +778 -0
- package/schemas/decision/DecisionComparativeReplayAuditSample.v1.schema.json +137 -0
- package/schemas/decision/DecisionComparativeReplayCorpus.v1.schema.json +258 -0
- package/schemas/decision/DecisionComparativeReplayFreeze.v1.schema.json +74 -0
- package/schemas/decision/DecisionComparativeReplayGold.v1.schema.json +50 -0
- package/schemas/decision/DecisionComparativeReplayGoldAsset.v1.schema.json +37 -0
- package/schemas/decision/DecisionComparativeReplayIntegrity.v1.schema.json +81 -0
- package/schemas/decision/DecisionComparativeReplayMembers.v1.schema.json +52 -0
- package/schemas/decision/DecisionComparativeReplayPreregistration.v1.schema.json +113 -0
- package/schemas/decision/DecisionComparativeReplayReport.v1.schema.json +407 -0
- package/schemas/decision/DecisionComparativeReplayReviewPackets.v1.schema.json +260 -0
- package/schemas/decision/DecisionComparativeReplaySplit.v1.schema.json +84 -0
- package/schemas/decision/DecisionFeatureObservation.v1.schema.json +564 -0
- package/schemas/decision/DecisionFeatureSet.v1.schema.json +155 -0
- package/schemas/decision/DecisionInteropMapping.v1.schema.json +85 -0
- package/schemas/decision/DecisionIssueTriageEvaluationManifest.v1.schema.json +174 -0
- package/schemas/decision/DecisionIssueTriageEvaluationReport.v1.schema.json +280 -0
- package/schemas/decision/DecisionIssueTriagePilotPack.v1.schema.json +114 -0
- package/schemas/decision/DecisionRouteReceipt.v1.schema.json +155 -0
- package/schemas/decision/DecisionRoutingPolicy.v1.schema.json +132 -0
- package/schemas/decision/DecisionRoutingShadowReport.v1.schema.json +170 -0
- package/schemas/decision/DecisionRoutingTask.v1.schema.json +49 -0
- package/schemas/decision/DecisionRuleset.schema.json +8 -2
- package/schemas/decision/DecisionRuleset.v1alpha2.schema.json +8 -2
- package/schemas/decision/DecisionSensitivityPlan.v1.schema.json +191 -0
- package/schemas/decision/DecisionSensitivityReport.v1.schema.json +163 -0
- package/schemas/decision/HeldoutApproval.v1.schema.json +312 -0
- package/schemas/decision/HeldoutAttempt.v1.schema.json +211 -0
- package/schemas/decision/HeldoutBaseline.v1.schema.json +90 -0
- package/schemas/decision/HeldoutCalibrationPhase.v1.schema.json +87 -0
- package/schemas/decision/HeldoutCorpus.v1.schema.json +167 -0
- package/schemas/decision/HeldoutEvent.v1.schema.json +41 -0
- package/schemas/decision/HeldoutFrozen.v1.schema.json +142 -0
- package/schemas/decision/HeldoutPreregistration.v1.schema.json +222 -0
- package/schemas/decision/HeldoutSpendEvent.v1.schema.json +122 -0
- package/schemas/decision/HeldoutSpendHead.v1.schema.json +25 -0
- package/schemas/decision/HeldoutSpendHead.v2.schema.json +46 -0
- package/schemas/decision/HeldoutSummary.v1.schema.json +101 -0
- package/schemas/decision/OpaDecisionLogExport.v1.schema.json +65 -0
- package/schemas/decision/PreprocessedEvidence.v1.schema.json +306 -0
- package/schemas/decision/RulesetResult.v1alpha2.schema.json +175 -0
- package/schemas/decision/SdlcEvidenceScreening.v1.schema.json +499 -0
- package/schemas/decision/SdlcScreeningPreregistration.v1.schema.json +132 -0
- package/schemas/decision/SdlcScreeningPreregistration.v2.schema.json +138 -0
- package/schemas/decision/SdlcScreeningRelease.v1.schema.json +654 -0
- package/schemas/decision/SdlcScreeningRelease.v2.schema.json +654 -0
- package/tools/agents/deploy-agents.mjs +1 -1
- package/tools/agents/providers/base.mjs +26 -7
- package/tools/agents/providers/factory.mjs +10 -2
- package/tools/agents/providers/muse-hooks.mjs +1 -29
- package/tools/agents/providers/muse.mjs +6 -3
|
@@ -0,0 +1,169 @@
|
|
|
1
|
+
{
|
|
2
|
+
"$schema": "https://json-schema.org/draft/2020-12/schema",
|
|
3
|
+
"$id": "https://aiwg.io/schemas/decision/D17StudyDryRun.v1.schema.json",
|
|
4
|
+
"title": "D17StudyDryRun",
|
|
5
|
+
"type": "object",
|
|
6
|
+
"additionalProperties": false,
|
|
7
|
+
"required": [
|
|
8
|
+
"schemaVersion",
|
|
9
|
+
"providerCalls",
|
|
10
|
+
"syntheticOnly",
|
|
11
|
+
"subjects",
|
|
12
|
+
"splits",
|
|
13
|
+
"expected",
|
|
14
|
+
"worstCase",
|
|
15
|
+
"approvalCeilings",
|
|
16
|
+
"portfolioUsd",
|
|
17
|
+
"stopFraction",
|
|
18
|
+
"fitsStudyCapBeforeStop",
|
|
19
|
+
"sessionLimitMs",
|
|
20
|
+
"requiresResumableSessions",
|
|
21
|
+
"reviewAssessments",
|
|
22
|
+
"missingInputs"
|
|
23
|
+
],
|
|
24
|
+
"properties": {
|
|
25
|
+
"schemaVersion": {
|
|
26
|
+
"const": "decision-d17-dry-run/v1"
|
|
27
|
+
},
|
|
28
|
+
"providerCalls": {
|
|
29
|
+
"const": 0
|
|
30
|
+
},
|
|
31
|
+
"syntheticOnly": {
|
|
32
|
+
"const": true
|
|
33
|
+
},
|
|
34
|
+
"subjects": {
|
|
35
|
+
"const": 1800
|
|
36
|
+
},
|
|
37
|
+
"splits": {
|
|
38
|
+
"type": "object",
|
|
39
|
+
"additionalProperties": false,
|
|
40
|
+
"required": [
|
|
41
|
+
"tuning",
|
|
42
|
+
"calibration",
|
|
43
|
+
"test"
|
|
44
|
+
],
|
|
45
|
+
"properties": {
|
|
46
|
+
"tuning": {
|
|
47
|
+
"const": 200
|
|
48
|
+
},
|
|
49
|
+
"calibration": {
|
|
50
|
+
"const": 400
|
|
51
|
+
},
|
|
52
|
+
"test": {
|
|
53
|
+
"const": 1200
|
|
54
|
+
}
|
|
55
|
+
}
|
|
56
|
+
},
|
|
57
|
+
"expected": {
|
|
58
|
+
"type": "object",
|
|
59
|
+
"additionalProperties": false,
|
|
60
|
+
"required": [
|
|
61
|
+
"firstAttempts",
|
|
62
|
+
"retryRateBps",
|
|
63
|
+
"attempts",
|
|
64
|
+
"inputTokens",
|
|
65
|
+
"outputTokens",
|
|
66
|
+
"usd",
|
|
67
|
+
"costSource"
|
|
68
|
+
],
|
|
69
|
+
"properties": {
|
|
70
|
+
"firstAttempts": {
|
|
71
|
+
"const": 7200
|
|
72
|
+
},
|
|
73
|
+
"retryRateBps": {
|
|
74
|
+
"const": 250
|
|
75
|
+
},
|
|
76
|
+
"attempts": {
|
|
77
|
+
"const": 7380
|
|
78
|
+
},
|
|
79
|
+
"inputTokens": {
|
|
80
|
+
"const": 27630720
|
|
81
|
+
},
|
|
82
|
+
"outputTokens": {
|
|
83
|
+
"type": "null"
|
|
84
|
+
},
|
|
85
|
+
"usd": {
|
|
86
|
+
"const": 2.7675
|
|
87
|
+
},
|
|
88
|
+
"costSource": {
|
|
89
|
+
"const": "planning-bound-derived-not-provider-reported"
|
|
90
|
+
}
|
|
91
|
+
}
|
|
92
|
+
},
|
|
93
|
+
"worstCase": {
|
|
94
|
+
"type": "object",
|
|
95
|
+
"additionalProperties": false,
|
|
96
|
+
"required": [
|
|
97
|
+
"attempts",
|
|
98
|
+
"totalTokens",
|
|
99
|
+
"reservedUsd",
|
|
100
|
+
"priceFloorUsdPerMTok"
|
|
101
|
+
],
|
|
102
|
+
"properties": {
|
|
103
|
+
"attempts": {
|
|
104
|
+
"const": 14400
|
|
105
|
+
},
|
|
106
|
+
"totalTokens": {
|
|
107
|
+
"const": 57600000
|
|
108
|
+
},
|
|
109
|
+
"reservedUsd": {
|
|
110
|
+
"const": 5.4
|
|
111
|
+
},
|
|
112
|
+
"priceFloorUsdPerMTok": {
|
|
113
|
+
"const": 0.1
|
|
114
|
+
}
|
|
115
|
+
}
|
|
116
|
+
},
|
|
117
|
+
"approvalCeilings": {
|
|
118
|
+
"type": "object",
|
|
119
|
+
"additionalProperties": false,
|
|
120
|
+
"required": [
|
|
121
|
+
"calls",
|
|
122
|
+
"tokens",
|
|
123
|
+
"usd"
|
|
124
|
+
],
|
|
125
|
+
"properties": {
|
|
126
|
+
"calls": {
|
|
127
|
+
"const": 18000
|
|
128
|
+
},
|
|
129
|
+
"tokens": {
|
|
130
|
+
"const": 72000000
|
|
131
|
+
},
|
|
132
|
+
"usd": {
|
|
133
|
+
"const": 8
|
|
134
|
+
}
|
|
135
|
+
}
|
|
136
|
+
},
|
|
137
|
+
"portfolioUsd": {
|
|
138
|
+
"const": 48
|
|
139
|
+
},
|
|
140
|
+
"stopFraction": {
|
|
141
|
+
"const": 0.8
|
|
142
|
+
},
|
|
143
|
+
"fitsStudyCapBeforeStop": {
|
|
144
|
+
"const": true
|
|
145
|
+
},
|
|
146
|
+
"sessionLimitMs": {
|
|
147
|
+
"const": 1800000
|
|
148
|
+
},
|
|
149
|
+
"requiresResumableSessions": {
|
|
150
|
+
"const": true
|
|
151
|
+
},
|
|
152
|
+
"reviewAssessments": {
|
|
153
|
+
"const": 88
|
|
154
|
+
},
|
|
155
|
+
"missingInputs": {
|
|
156
|
+
"type": "array",
|
|
157
|
+
"const": [
|
|
158
|
+
"operator approval",
|
|
159
|
+
"genuine D09 calibration for AC7/AC14 (outside diagnostic scope)",
|
|
160
|
+
"exact-source CI evidence",
|
|
161
|
+
"provider terms",
|
|
162
|
+
"prior study and portfolio spend",
|
|
163
|
+
"region and credential resolver pin",
|
|
164
|
+
"live observations",
|
|
165
|
+
"blind human review"
|
|
166
|
+
]
|
|
167
|
+
}
|
|
168
|
+
}
|
|
169
|
+
}
|
|
@@ -0,0 +1,239 @@
|
|
|
1
|
+
{
|
|
2
|
+
"$schema": "https://json-schema.org/draft/2020-12/schema",
|
|
3
|
+
"$id": "https://aiwg.io/schemas/decision/D17StudyDryRun.v2.schema.json",
|
|
4
|
+
"title": "D17StudyDryRun",
|
|
5
|
+
"type": "object",
|
|
6
|
+
"additionalProperties": false,
|
|
7
|
+
"required": [
|
|
8
|
+
"schemaVersion",
|
|
9
|
+
"providerCalls",
|
|
10
|
+
"syntheticOnly",
|
|
11
|
+
"subjects",
|
|
12
|
+
"splits",
|
|
13
|
+
"expected",
|
|
14
|
+
"worstCase",
|
|
15
|
+
"approvalCeilings",
|
|
16
|
+
"portfolioUsd",
|
|
17
|
+
"stopFraction",
|
|
18
|
+
"fitsStudyCapBeforeStop",
|
|
19
|
+
"sessionLimitMs",
|
|
20
|
+
"requiresResumableSessions",
|
|
21
|
+
"reviewAssessments",
|
|
22
|
+
"missingInputs",
|
|
23
|
+
"phases"
|
|
24
|
+
],
|
|
25
|
+
"properties": {
|
|
26
|
+
"schemaVersion": {
|
|
27
|
+
"const": "decision-d17-dry-run/v2"
|
|
28
|
+
},
|
|
29
|
+
"providerCalls": {
|
|
30
|
+
"const": 0
|
|
31
|
+
},
|
|
32
|
+
"syntheticOnly": {
|
|
33
|
+
"const": true
|
|
34
|
+
},
|
|
35
|
+
"subjects": {
|
|
36
|
+
"const": 1800
|
|
37
|
+
},
|
|
38
|
+
"splits": {
|
|
39
|
+
"type": "object",
|
|
40
|
+
"additionalProperties": false,
|
|
41
|
+
"required": [
|
|
42
|
+
"tuning",
|
|
43
|
+
"calibration",
|
|
44
|
+
"test"
|
|
45
|
+
],
|
|
46
|
+
"properties": {
|
|
47
|
+
"tuning": {
|
|
48
|
+
"const": 200
|
|
49
|
+
},
|
|
50
|
+
"calibration": {
|
|
51
|
+
"const": 400
|
|
52
|
+
},
|
|
53
|
+
"test": {
|
|
54
|
+
"const": 1200
|
|
55
|
+
}
|
|
56
|
+
}
|
|
57
|
+
},
|
|
58
|
+
"expected": {
|
|
59
|
+
"type": "object",
|
|
60
|
+
"additionalProperties": false,
|
|
61
|
+
"required": [
|
|
62
|
+
"firstAttempts",
|
|
63
|
+
"retryRateBps",
|
|
64
|
+
"attempts",
|
|
65
|
+
"inputTokens",
|
|
66
|
+
"outputTokens",
|
|
67
|
+
"usd",
|
|
68
|
+
"costSource"
|
|
69
|
+
],
|
|
70
|
+
"properties": {
|
|
71
|
+
"firstAttempts": {
|
|
72
|
+
"const": 7200
|
|
73
|
+
},
|
|
74
|
+
"retryRateBps": {
|
|
75
|
+
"const": 250
|
|
76
|
+
},
|
|
77
|
+
"attempts": {
|
|
78
|
+
"const": 7380
|
|
79
|
+
},
|
|
80
|
+
"inputTokens": {
|
|
81
|
+
"const": 27630720
|
|
82
|
+
},
|
|
83
|
+
"outputTokens": {
|
|
84
|
+
"type": "null"
|
|
85
|
+
},
|
|
86
|
+
"usd": {
|
|
87
|
+
"const": 2.7675
|
|
88
|
+
},
|
|
89
|
+
"costSource": {
|
|
90
|
+
"const": "planning-bound-derived-not-provider-reported"
|
|
91
|
+
}
|
|
92
|
+
}
|
|
93
|
+
},
|
|
94
|
+
"worstCase": {
|
|
95
|
+
"type": "object",
|
|
96
|
+
"additionalProperties": false,
|
|
97
|
+
"required": [
|
|
98
|
+
"attempts",
|
|
99
|
+
"totalTokens",
|
|
100
|
+
"reservedUsd",
|
|
101
|
+
"priceFloorUsdPerMTok"
|
|
102
|
+
],
|
|
103
|
+
"properties": {
|
|
104
|
+
"attempts": {
|
|
105
|
+
"const": 14400
|
|
106
|
+
},
|
|
107
|
+
"totalTokens": {
|
|
108
|
+
"const": 57600000
|
|
109
|
+
},
|
|
110
|
+
"reservedUsd": {
|
|
111
|
+
"const": 5.4
|
|
112
|
+
},
|
|
113
|
+
"priceFloorUsdPerMTok": {
|
|
114
|
+
"const": 0.1
|
|
115
|
+
}
|
|
116
|
+
}
|
|
117
|
+
},
|
|
118
|
+
"approvalCeilings": {
|
|
119
|
+
"type": "object",
|
|
120
|
+
"additionalProperties": false,
|
|
121
|
+
"required": [
|
|
122
|
+
"calls",
|
|
123
|
+
"tokens",
|
|
124
|
+
"usd"
|
|
125
|
+
],
|
|
126
|
+
"properties": {
|
|
127
|
+
"calls": {
|
|
128
|
+
"const": 24000
|
|
129
|
+
},
|
|
130
|
+
"tokens": {
|
|
131
|
+
"const": 96000000
|
|
132
|
+
},
|
|
133
|
+
"usd": {
|
|
134
|
+
"const": 8
|
|
135
|
+
}
|
|
136
|
+
}
|
|
137
|
+
},
|
|
138
|
+
"portfolioUsd": {
|
|
139
|
+
"const": 48
|
|
140
|
+
},
|
|
141
|
+
"stopFraction": {
|
|
142
|
+
"const": 0.8
|
|
143
|
+
},
|
|
144
|
+
"fitsStudyCapBeforeStop": {
|
|
145
|
+
"const": true
|
|
146
|
+
},
|
|
147
|
+
"sessionLimitMs": {
|
|
148
|
+
"const": 1800000
|
|
149
|
+
},
|
|
150
|
+
"requiresResumableSessions": {
|
|
151
|
+
"const": true
|
|
152
|
+
},
|
|
153
|
+
"reviewAssessments": {
|
|
154
|
+
"const": 88
|
|
155
|
+
},
|
|
156
|
+
"missingInputs": {
|
|
157
|
+
"type": "array",
|
|
158
|
+
"const": [
|
|
159
|
+
"completed 40-item development review",
|
|
160
|
+
"calibration-phase operator approval",
|
|
161
|
+
"exact-source CI evidence",
|
|
162
|
+
"provider terms",
|
|
163
|
+
"prior study and portfolio spend",
|
|
164
|
+
"region and credential resolver pin",
|
|
165
|
+
"calibration-phase live observations",
|
|
166
|
+
"fitted and operator-reviewed D09 calibration set",
|
|
167
|
+
"test-phase operator approval",
|
|
168
|
+
"test-phase live observations",
|
|
169
|
+
"blind human review"
|
|
170
|
+
]
|
|
171
|
+
},
|
|
172
|
+
"phases": {
|
|
173
|
+
"type": "object",
|
|
174
|
+
"additionalProperties": false,
|
|
175
|
+
"required": [
|
|
176
|
+
"calibration",
|
|
177
|
+
"test"
|
|
178
|
+
],
|
|
179
|
+
"properties": {
|
|
180
|
+
"calibration": {
|
|
181
|
+
"type": "object",
|
|
182
|
+
"additionalProperties": false,
|
|
183
|
+
"required": [
|
|
184
|
+
"splits",
|
|
185
|
+
"rows",
|
|
186
|
+
"firstAttempts",
|
|
187
|
+
"worstCaseAttempts"
|
|
188
|
+
],
|
|
189
|
+
"properties": {
|
|
190
|
+
"splits": {
|
|
191
|
+
"type": "array",
|
|
192
|
+
"const": [
|
|
193
|
+
"tuning",
|
|
194
|
+
"calibration"
|
|
195
|
+
]
|
|
196
|
+
},
|
|
197
|
+
"rows": {
|
|
198
|
+
"const": 600
|
|
199
|
+
},
|
|
200
|
+
"firstAttempts": {
|
|
201
|
+
"const": 2400
|
|
202
|
+
},
|
|
203
|
+
"worstCaseAttempts": {
|
|
204
|
+
"const": 4800
|
|
205
|
+
}
|
|
206
|
+
}
|
|
207
|
+
},
|
|
208
|
+
"test": {
|
|
209
|
+
"type": "object",
|
|
210
|
+
"additionalProperties": false,
|
|
211
|
+
"required": [
|
|
212
|
+
"splits",
|
|
213
|
+
"rows",
|
|
214
|
+
"firstAttempts",
|
|
215
|
+
"worstCaseAttempts"
|
|
216
|
+
],
|
|
217
|
+
"properties": {
|
|
218
|
+
"splits": {
|
|
219
|
+
"type": "array",
|
|
220
|
+
"const": [
|
|
221
|
+
"test"
|
|
222
|
+
]
|
|
223
|
+
},
|
|
224
|
+
"rows": {
|
|
225
|
+
"const": 1200
|
|
226
|
+
},
|
|
227
|
+
"firstAttempts": {
|
|
228
|
+
"const": 4800
|
|
229
|
+
},
|
|
230
|
+
"worstCaseAttempts": {
|
|
231
|
+
"const": 9600
|
|
232
|
+
}
|
|
233
|
+
}
|
|
234
|
+
}
|
|
235
|
+
}
|
|
236
|
+
}
|
|
237
|
+
},
|
|
238
|
+
"description": "Staged-calibration D17 dry run with per-phase planning bounds. v1 remains the diagnostic dry run."
|
|
239
|
+
}
|
|
@@ -0,0 +1,108 @@
|
|
|
1
|
+
{
|
|
2
|
+
"$schema": "https://json-schema.org/draft/2020-12/schema",
|
|
3
|
+
"$id": "https://aiwg.io/schemas/decision/D17StudyGold.v1.schema.json",
|
|
4
|
+
"title": "D17StudyGold",
|
|
5
|
+
"type": "object",
|
|
6
|
+
"additionalProperties": false,
|
|
7
|
+
"required": [
|
|
8
|
+
"schemaVersion",
|
|
9
|
+
"labels",
|
|
10
|
+
"worlds"
|
|
11
|
+
],
|
|
12
|
+
"properties": {
|
|
13
|
+
"schemaVersion": {
|
|
14
|
+
"const": "decision-d17-gold/v1"
|
|
15
|
+
},
|
|
16
|
+
"labels": {
|
|
17
|
+
"type": "object",
|
|
18
|
+
"minProperties": 1800,
|
|
19
|
+
"maxProperties": 1800,
|
|
20
|
+
"additionalProperties": false,
|
|
21
|
+
"patternProperties": {
|
|
22
|
+
"^[0-9a-f]{64}$": {
|
|
23
|
+
"enum": [
|
|
24
|
+
"yes",
|
|
25
|
+
"no"
|
|
26
|
+
]
|
|
27
|
+
}
|
|
28
|
+
}
|
|
29
|
+
},
|
|
30
|
+
"worlds": {
|
|
31
|
+
"type": "object",
|
|
32
|
+
"minProperties": 1800,
|
|
33
|
+
"maxProperties": 1800,
|
|
34
|
+
"additionalProperties": false,
|
|
35
|
+
"patternProperties": {
|
|
36
|
+
"^[0-9a-f]{64}$": {
|
|
37
|
+
"type": "object",
|
|
38
|
+
"additionalProperties": false,
|
|
39
|
+
"required": [
|
|
40
|
+
"slice",
|
|
41
|
+
"subject",
|
|
42
|
+
"object",
|
|
43
|
+
"destination",
|
|
44
|
+
"decoy",
|
|
45
|
+
"enabled",
|
|
46
|
+
"hasLink",
|
|
47
|
+
"explicitNegative",
|
|
48
|
+
"unknown",
|
|
49
|
+
"quantity",
|
|
50
|
+
"day"
|
|
51
|
+
],
|
|
52
|
+
"properties": {
|
|
53
|
+
"slice": {
|
|
54
|
+
"enum": [
|
|
55
|
+
"direct-facts",
|
|
56
|
+
"multi-fact",
|
|
57
|
+
"negation",
|
|
58
|
+
"authority"
|
|
59
|
+
]
|
|
60
|
+
},
|
|
61
|
+
"subject": {
|
|
62
|
+
"type": "string",
|
|
63
|
+
"minLength": 1,
|
|
64
|
+
"maxLength": 128
|
|
65
|
+
},
|
|
66
|
+
"object": {
|
|
67
|
+
"type": "string",
|
|
68
|
+
"minLength": 1,
|
|
69
|
+
"maxLength": 128
|
|
70
|
+
},
|
|
71
|
+
"destination": {
|
|
72
|
+
"type": "string",
|
|
73
|
+
"minLength": 1,
|
|
74
|
+
"maxLength": 128
|
|
75
|
+
},
|
|
76
|
+
"decoy": {
|
|
77
|
+
"type": "string",
|
|
78
|
+
"minLength": 1,
|
|
79
|
+
"maxLength": 128
|
|
80
|
+
},
|
|
81
|
+
"enabled": {
|
|
82
|
+
"type": "boolean"
|
|
83
|
+
},
|
|
84
|
+
"hasLink": {
|
|
85
|
+
"type": "boolean"
|
|
86
|
+
},
|
|
87
|
+
"explicitNegative": {
|
|
88
|
+
"type": "boolean"
|
|
89
|
+
},
|
|
90
|
+
"unknown": {
|
|
91
|
+
"type": "boolean"
|
|
92
|
+
},
|
|
93
|
+
"quantity": {
|
|
94
|
+
"type": "integer",
|
|
95
|
+
"minimum": 1,
|
|
96
|
+
"maximum": 99
|
|
97
|
+
},
|
|
98
|
+
"day": {
|
|
99
|
+
"type": "integer",
|
|
100
|
+
"minimum": 1,
|
|
101
|
+
"maximum": 365
|
|
102
|
+
}
|
|
103
|
+
}
|
|
104
|
+
}
|
|
105
|
+
}
|
|
106
|
+
}
|
|
107
|
+
}
|
|
108
|
+
}
|
|
@@ -0,0 +1,63 @@
|
|
|
1
|
+
{
|
|
2
|
+
"$schema": "https://json-schema.org/draft/2020-12/schema",
|
|
3
|
+
"$id": "https://aiwg.io/schemas/decision/D17StudyGuide.v1.schema.json",
|
|
4
|
+
"title": "D17StudyGuide",
|
|
5
|
+
"type": "object",
|
|
6
|
+
"additionalProperties": false,
|
|
7
|
+
"required": [
|
|
8
|
+
"schemaVersion",
|
|
9
|
+
"population",
|
|
10
|
+
"yes",
|
|
11
|
+
"no",
|
|
12
|
+
"authority",
|
|
13
|
+
"oracleAudit",
|
|
14
|
+
"invalidGold",
|
|
15
|
+
"timing",
|
|
16
|
+
"independence"
|
|
17
|
+
],
|
|
18
|
+
"properties": {
|
|
19
|
+
"schemaVersion": {
|
|
20
|
+
"const": "decision-d17-guide/v1"
|
|
21
|
+
},
|
|
22
|
+
"population": {
|
|
23
|
+
"type": "string",
|
|
24
|
+
"minLength": 1,
|
|
25
|
+
"maxLength": 8192
|
|
26
|
+
},
|
|
27
|
+
"yes": {
|
|
28
|
+
"type": "string",
|
|
29
|
+
"minLength": 1,
|
|
30
|
+
"maxLength": 8192
|
|
31
|
+
},
|
|
32
|
+
"no": {
|
|
33
|
+
"type": "string",
|
|
34
|
+
"minLength": 1,
|
|
35
|
+
"maxLength": 8192
|
|
36
|
+
},
|
|
37
|
+
"authority": {
|
|
38
|
+
"type": "string",
|
|
39
|
+
"minLength": 1,
|
|
40
|
+
"maxLength": 8192
|
|
41
|
+
},
|
|
42
|
+
"oracleAudit": {
|
|
43
|
+
"type": "string",
|
|
44
|
+
"minLength": 1,
|
|
45
|
+
"maxLength": 8192
|
|
46
|
+
},
|
|
47
|
+
"invalidGold": {
|
|
48
|
+
"type": "string",
|
|
49
|
+
"minLength": 1,
|
|
50
|
+
"maxLength": 8192
|
|
51
|
+
},
|
|
52
|
+
"timing": {
|
|
53
|
+
"type": "string",
|
|
54
|
+
"minLength": 1,
|
|
55
|
+
"maxLength": 8192
|
|
56
|
+
},
|
|
57
|
+
"independence": {
|
|
58
|
+
"type": "string",
|
|
59
|
+
"minLength": 1,
|
|
60
|
+
"maxLength": 8192
|
|
61
|
+
}
|
|
62
|
+
}
|
|
63
|
+
}
|