@openpond/evals 0.11.16 → 0.11.18

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 (158) hide show
  1. package/dist/connected-evidence/contracts.js +72 -0
  2. package/dist/connected-evidence/imports.js +71 -0
  3. package/dist/connected-evidence/index.js +5 -0
  4. package/dist/connected-evidence/normalize.js +97 -0
  5. package/dist/connected-evidence/python-correctness.js +73 -0
  6. package/dist/connected-evidence/sources/claude-code.js +71 -0
  7. package/dist/connected-evidence/sources/codex.js +52 -0
  8. package/dist/connected-evidence/sources/history.js +86 -0
  9. package/dist/connected-evidence/sources/usage.js +15 -0
  10. package/dist/connected-evidence/usage-summary.js +52 -0
  11. package/dist/experiments.js +16 -4
  12. package/dist/graders.js +9 -6
  13. package/dist/human-review/comparison.js +161 -0
  14. package/dist/human-review/contracts.js +110 -0
  15. package/dist/human-review/execution.js +61 -0
  16. package/dist/human-review/handoff.js +18 -0
  17. package/dist/human-review/index.js +9 -0
  18. package/dist/human-review/local-publication.js +18 -0
  19. package/dist/human-review/repository.js +1 -0
  20. package/dist/human-review/results.js +18 -0
  21. package/dist/human-review/service.js +268 -0
  22. package/dist/human-review/task-proposals.js +39 -0
  23. package/dist/index.js +1 -0
  24. package/dist/javascript-verifier-contract.js +2 -2
  25. package/dist/learning/authoring.js +2 -1
  26. package/dist/learning/contracts.js +353 -81
  27. package/dist/learning/index.js +1 -0
  28. package/dist/learning/reward-authoring.js +5 -2
  29. package/dist/learning/training-handoff.js +32 -0
  30. package/dist/metrics-node.js +2 -0
  31. package/dist/profile-evaluation-target.js +7 -0
  32. package/dist/profile-evaluations.js +7 -10
  33. package/dist/profile-external-evaluation.js +35 -0
  34. package/dist/profile-source-candidate.js +5 -0
  35. package/dist/rewards.js +5 -3
  36. package/dist/taskset-run-contract.js +10 -0
  37. package/dist/tasksets.js +2 -0
  38. package/dist/types/builtin-benchmarks/harness-refiner.d.ts +75 -1
  39. package/dist/types/builtin-benchmarks/harness-refiner.d.ts.map +1 -1
  40. package/dist/types/conformance.d.ts +150 -2
  41. package/dist/types/conformance.d.ts.map +1 -1
  42. package/dist/types/connected-evidence/contracts.d.ts +317 -0
  43. package/dist/types/connected-evidence/contracts.d.ts.map +1 -0
  44. package/dist/types/connected-evidence/imports.d.ts +19 -0
  45. package/dist/types/connected-evidence/imports.d.ts.map +1 -0
  46. package/dist/types/connected-evidence/index.d.ts +6 -0
  47. package/dist/types/connected-evidence/index.d.ts.map +1 -0
  48. package/dist/types/connected-evidence/normalize.d.ts +25 -0
  49. package/dist/types/connected-evidence/normalize.d.ts.map +1 -0
  50. package/dist/types/connected-evidence/python-correctness.d.ts +196 -0
  51. package/dist/types/connected-evidence/python-correctness.d.ts.map +1 -0
  52. package/dist/types/connected-evidence/sources/claude-code.d.ts +74 -0
  53. package/dist/types/connected-evidence/sources/claude-code.d.ts.map +1 -0
  54. package/dist/types/connected-evidence/sources/codex.d.ts +74 -0
  55. package/dist/types/connected-evidence/sources/codex.d.ts.map +1 -0
  56. package/dist/types/connected-evidence/sources/history.d.ts +146 -0
  57. package/dist/types/connected-evidence/sources/history.d.ts.map +1 -0
  58. package/dist/types/connected-evidence/sources/usage.d.ts +3 -0
  59. package/dist/types/connected-evidence/sources/usage.d.ts.map +1 -0
  60. package/dist/types/connected-evidence/usage-summary.d.ts +162 -0
  61. package/dist/types/connected-evidence/usage-summary.d.ts.map +1 -0
  62. package/dist/types/evidence/conformance.d.ts +28 -28
  63. package/dist/types/evidence/contracts.d.ts +27 -27
  64. package/dist/types/execution-contracts.d.ts +159 -5
  65. package/dist/types/execution-contracts.d.ts.map +1 -1
  66. package/dist/types/experiments.d.ts +476 -7
  67. package/dist/types/experiments.d.ts.map +1 -1
  68. package/dist/types/graders.d.ts +21 -1
  69. package/dist/types/graders.d.ts.map +1 -1
  70. package/dist/types/human-review/comparison.d.ts +1959 -0
  71. package/dist/types/human-review/comparison.d.ts.map +1 -0
  72. package/dist/types/human-review/contracts.d.ts +2719 -0
  73. package/dist/types/human-review/contracts.d.ts.map +1 -0
  74. package/dist/types/human-review/execution.d.ts +298 -0
  75. package/dist/types/human-review/execution.d.ts.map +1 -0
  76. package/dist/types/human-review/handoff.d.ts +124 -0
  77. package/dist/types/human-review/handoff.d.ts.map +1 -0
  78. package/dist/types/human-review/index.d.ts +10 -0
  79. package/dist/types/human-review/index.d.ts.map +1 -0
  80. package/dist/types/human-review/local-publication.d.ts +165 -0
  81. package/dist/types/human-review/local-publication.d.ts.map +1 -0
  82. package/dist/types/human-review/repository.d.ts +43 -0
  83. package/dist/types/human-review/repository.d.ts.map +1 -0
  84. package/dist/types/human-review/results.d.ts +118 -0
  85. package/dist/types/human-review/results.d.ts.map +1 -0
  86. package/dist/types/human-review/service.d.ts +878 -0
  87. package/dist/types/human-review/service.d.ts.map +1 -0
  88. package/dist/types/human-review/task-proposals.d.ts +70 -0
  89. package/dist/types/human-review/task-proposals.d.ts.map +1 -0
  90. package/dist/types/index.d.ts +1 -0
  91. package/dist/types/index.d.ts.map +1 -1
  92. package/dist/types/javascript-environment-contract.d.ts +1 -1
  93. package/dist/types/javascript-verifier-contract.d.ts +1 -1
  94. package/dist/types/javascript-verifier-contract.d.ts.map +1 -1
  95. package/dist/types/learned-preference.d.ts +2 -2
  96. package/dist/types/learning/acceptance-group.d.ts +13 -13
  97. package/dist/types/learning/acceptance-plan.d.ts +1 -1
  98. package/dist/types/learning/admission.d.ts +78 -1
  99. package/dist/types/learning/admission.d.ts.map +1 -1
  100. package/dist/types/learning/authoring-service.d.ts +2 -0
  101. package/dist/types/learning/authoring-service.d.ts.map +1 -1
  102. package/dist/types/learning/authoring.d.ts +20 -0
  103. package/dist/types/learning/authoring.d.ts.map +1 -1
  104. package/dist/types/learning/batch-service.d.ts +76 -2
  105. package/dist/types/learning/batch-service.d.ts.map +1 -1
  106. package/dist/types/learning/contracts.d.ts +759 -24
  107. package/dist/types/learning/contracts.d.ts.map +1 -1
  108. package/dist/types/learning/index.d.ts +1 -0
  109. package/dist/types/learning/index.d.ts.map +1 -1
  110. package/dist/types/learning/iteration-candidate-decision.d.ts +2 -2
  111. package/dist/types/learning/iteration-dispatch-contracts.d.ts +2 -2
  112. package/dist/types/learning/operations.d.ts +1118 -8
  113. package/dist/types/learning/operations.d.ts.map +1 -1
  114. package/dist/types/learning/repository.d.ts +402 -15
  115. package/dist/types/learning/repository.d.ts.map +1 -1
  116. package/dist/types/learning/reward-authoring.d.ts.map +1 -1
  117. package/dist/types/learning/reward-calibration.d.ts +74 -0
  118. package/dist/types/learning/reward-calibration.d.ts.map +1 -1
  119. package/dist/types/learning/reward-checks.d.ts +14 -2
  120. package/dist/types/learning/reward-checks.d.ts.map +1 -1
  121. package/dist/types/learning/training-handoff.d.ts +201 -0
  122. package/dist/types/learning/training-handoff.d.ts.map +1 -0
  123. package/dist/types/learning/transport.d.ts +963 -21
  124. package/dist/types/learning/transport.d.ts.map +1 -1
  125. package/dist/types/metrics-node.d.ts +3 -1
  126. package/dist/types/metrics-node.d.ts.map +1 -1
  127. package/dist/types/model-improvement-qualification.d.ts +3 -3
  128. package/dist/types/native-harness-experiment-attempt.d.ts +2 -2
  129. package/dist/types/profile-evaluation-comparisons.d.ts +230 -0
  130. package/dist/types/profile-evaluation-comparisons.d.ts.map +1 -1
  131. package/dist/types/profile-evaluation-target.d.ts +14 -0
  132. package/dist/types/profile-evaluation-target.d.ts.map +1 -0
  133. package/dist/types/profile-evaluations.d.ts +117 -14
  134. package/dist/types/profile-evaluations.d.ts.map +1 -1
  135. package/dist/types/profile-external-evaluation.d.ts +404 -0
  136. package/dist/types/profile-external-evaluation.d.ts.map +1 -0
  137. package/dist/types/profile-source-candidate.d.ts +15 -0
  138. package/dist/types/profile-source-candidate.d.ts.map +1 -0
  139. package/dist/types/review-conformance.d.ts +7 -7
  140. package/dist/types/rewards.d.ts +253 -0
  141. package/dist/types/rewards.d.ts.map +1 -1
  142. package/dist/types/rollouts.d.ts +3 -3
  143. package/dist/types/taskset-run-contract.d.ts +230 -0
  144. package/dist/types/taskset-run-contract.d.ts.map +1 -1
  145. package/dist/types/tasksets.d.ts +311 -3
  146. package/dist/types/tasksets.d.ts.map +1 -1
  147. package/dist/types/telemetry-analysis.d.ts +2 -2
  148. package/dist/types/telemetry-bundle.d.ts +6 -6
  149. package/dist/types/telemetry.d.ts +3 -3
  150. package/package.json +13 -1
  151. package/schemas/learning/v1/command-request.schema.json +1745 -409
  152. package/schemas/learning/v1/decision.schema.json +11 -0
  153. package/schemas/learning/v1/draft.schema.json +11 -0
  154. package/schemas/learning/v1/grade.schema.json +11 -0
  155. package/schemas/learning/v1/package.schema.json +532 -12
  156. package/schemas/learning/v1/policy.schema.json +836 -28
  157. package/schemas/learning/v1/reward.schema.json +520 -0
  158. package/schemas/learning/v1/reward_check.schema.json +9 -0
@@ -0,0 +1,1959 @@
1
+ import { z } from 'zod';
2
+ import { type ExperimentCaseIdentity } from '../experiments.js';
3
+ import { ReviewReleaseSchema, type HumanReview, type HumanAttemptRefSchema } from './contracts.js';
4
+ export declare const HumanComparisonRuleSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
5
+ kind: z.ZodLiteral<"normalized_score">;
6
+ minimum: z.ZodNumber;
7
+ maximum: z.ZodNumber;
8
+ passThreshold: z.ZodOptional<z.ZodNumber>;
9
+ feedbackKey: z.ZodString;
10
+ criterionId: z.ZodString;
11
+ }, z.core.$strict>, z.ZodObject<{
12
+ kind: z.ZodLiteral<"boolean">;
13
+ feedbackKey: z.ZodString;
14
+ criterionId: z.ZodString;
15
+ }, z.core.$strict>, z.ZodObject<{
16
+ kind: z.ZodLiteral<"category">;
17
+ feedbackKey: z.ZodString;
18
+ criterionId: z.ZodString;
19
+ }, z.core.$strict>, z.ZodObject<{
20
+ kind: z.ZodLiteral<"preference_score">;
21
+ tieScore: z.ZodNullable<z.ZodNumber>;
22
+ feedbackKey: z.ZodString;
23
+ criterionId: z.ZodString;
24
+ }, z.core.$strict>], "kind">;
25
+ export declare const HumanComparisonSelectionsSchema: z.ZodObject<{
26
+ rules: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
27
+ kind: z.ZodLiteral<"normalized_score">;
28
+ minimum: z.ZodNumber;
29
+ maximum: z.ZodNumber;
30
+ passThreshold: z.ZodOptional<z.ZodNumber>;
31
+ feedbackKey: z.ZodString;
32
+ criterionId: z.ZodString;
33
+ }, z.core.$strict>, z.ZodObject<{
34
+ kind: z.ZodLiteral<"boolean">;
35
+ feedbackKey: z.ZodString;
36
+ criterionId: z.ZodString;
37
+ }, z.core.$strict>, z.ZodObject<{
38
+ kind: z.ZodLiteral<"category">;
39
+ feedbackKey: z.ZodString;
40
+ criterionId: z.ZodString;
41
+ }, z.core.$strict>, z.ZodObject<{
42
+ kind: z.ZodLiteral<"preference_score">;
43
+ tieScore: z.ZodNullable<z.ZodNumber>;
44
+ feedbackKey: z.ZodString;
45
+ criterionId: z.ZodString;
46
+ }, z.core.$strict>], "kind">>;
47
+ baseline: z.ZodObject<{
48
+ execution: z.ZodObject<{
49
+ id: z.ZodString;
50
+ contentHash: z.ZodString;
51
+ }, z.core.$strict>;
52
+ resultHash: z.ZodString;
53
+ reviews: z.ZodArray<z.ZodObject<{
54
+ review: z.ZodObject<{
55
+ id: z.ZodString;
56
+ revision: z.ZodNumber;
57
+ contentHash: z.ZodString;
58
+ }, z.core.$strict>;
59
+ decision: z.ZodObject<{
60
+ id: z.ZodString;
61
+ revision: z.ZodNumber;
62
+ contentHash: z.ZodString;
63
+ }, z.core.$strict>;
64
+ }, z.core.$strict>>;
65
+ }, z.core.$strict>;
66
+ candidate: z.ZodObject<{
67
+ execution: z.ZodObject<{
68
+ id: z.ZodString;
69
+ contentHash: z.ZodString;
70
+ }, z.core.$strict>;
71
+ resultHash: z.ZodString;
72
+ reviews: z.ZodArray<z.ZodObject<{
73
+ review: z.ZodObject<{
74
+ id: z.ZodString;
75
+ revision: z.ZodNumber;
76
+ contentHash: z.ZodString;
77
+ }, z.core.$strict>;
78
+ decision: z.ZodObject<{
79
+ id: z.ZodString;
80
+ revision: z.ZodNumber;
81
+ contentHash: z.ZodString;
82
+ }, z.core.$strict>;
83
+ }, z.core.$strict>>;
84
+ }, z.core.$strict>;
85
+ }, z.core.$strict>;
86
+ export type HumanComparisonSelections = z.infer<typeof HumanComparisonSelectionsSchema>;
87
+ declare const Evidence: z.ZodObject<{
88
+ manifest: z.ZodObject<{
89
+ schemaVersion: z.ZodLiteral<"openpond.experimentManifest.v1">;
90
+ id: z.ZodString;
91
+ name: z.ZodString;
92
+ teamId: z.ZodString;
93
+ operationId: z.ZodString;
94
+ maximumCostUsd: z.ZodNullable<z.ZodNumber>;
95
+ dataset: z.ZodObject<{
96
+ id: z.ZodString;
97
+ contentHash: z.ZodString;
98
+ revision: z.ZodNumber;
99
+ }, z.core.$strict>;
100
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
101
+ kind: z.ZodLiteral<"model">;
102
+ modelId: z.ZodString;
103
+ configurationHash: z.ZodString;
104
+ }, z.core.$strict>, z.ZodObject<{
105
+ kind: z.ZodLiteral<"harness">;
106
+ source: z.ZodObject<{
107
+ profileId: z.ZodString;
108
+ sourceRevision: z.ZodString;
109
+ harnessRelease: z.ZodObject<{
110
+ id: z.ZodString;
111
+ contentHash: z.ZodString;
112
+ }, z.core.$strict>;
113
+ catalogHash: z.ZodString;
114
+ definitionId: z.ZodString;
115
+ definitionHash: z.ZodString;
116
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
117
+ kind: z.ZodLiteral<"profile">;
118
+ }, z.core.$strict>, z.ZodObject<{
119
+ kind: z.ZodLiteral<"workflow">;
120
+ workflowId: z.ZodString;
121
+ }, z.core.$strict>, z.ZodObject<{
122
+ kind: z.ZodLiteral<"skill">;
123
+ skillPath: z.ZodString;
124
+ }, z.core.$strict>, z.ZodObject<{
125
+ kind: z.ZodLiteral<"agent_action">;
126
+ actionId: z.ZodString;
127
+ }, z.core.$strict>], "kind">;
128
+ environmentHash: z.ZodString;
129
+ externalDatasetBinding: z.ZodOptional<z.ZodObject<{
130
+ schemaVersion: z.ZodLiteral<"openpond.profileExternalDatasetBinding.v1">;
131
+ dataset: z.ZodObject<{
132
+ id: z.ZodString;
133
+ contentHash: z.ZodString;
134
+ revision: z.ZodNumber;
135
+ }, z.core.$strict>;
136
+ packageHash: z.ZodString;
137
+ profileId: z.ZodString;
138
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
139
+ kind: z.ZodLiteral<"profile">;
140
+ }, z.core.$strict>, z.ZodObject<{
141
+ kind: z.ZodLiteral<"workflow">;
142
+ workflowId: z.ZodString;
143
+ }, z.core.$strict>, z.ZodObject<{
144
+ kind: z.ZodLiteral<"skill">;
145
+ skillPath: z.ZodString;
146
+ }, z.core.$strict>, z.ZodObject<{
147
+ kind: z.ZodLiteral<"agent_action">;
148
+ actionId: z.ZodString;
149
+ }, z.core.$strict>], "kind">;
150
+ declaredProfileCatalogHash: z.ZodNullable<z.ZodString>;
151
+ protectedProfileClosureHash: z.ZodString;
152
+ recordedOrigin: z.ZodOptional<z.ZodObject<{
153
+ execution: z.ZodObject<{
154
+ id: z.ZodString;
155
+ contentHash: z.ZodString;
156
+ }, z.core.$strict>;
157
+ originalEvidence: z.ZodObject<{
158
+ id: z.ZodString;
159
+ contentHash: z.ZodString;
160
+ }, z.core.$strict>;
161
+ members: z.ZodArray<z.ZodObject<{
162
+ caseId: z.ZodString;
163
+ seed: z.ZodString;
164
+ fixtureId: z.ZodString;
165
+ sourceId: z.ZodString;
166
+ snapshotHash: z.ZodString;
167
+ boundaryRevisionHash: z.ZodString;
168
+ }, z.core.$strict>>;
169
+ }, z.core.$strict>>;
170
+ gradingSource: z.ZodOptional<z.ZodObject<{
171
+ pass: z.ZodObject<{
172
+ id: z.ZodString;
173
+ contentHash: z.ZodString;
174
+ }, z.core.$strict>;
175
+ evidence: z.ZodObject<{
176
+ id: z.ZodString;
177
+ contentHash: z.ZodString;
178
+ }, z.core.$strict>;
179
+ }, z.core.$strict>>;
180
+ split: z.ZodEnum<{
181
+ train: "train";
182
+ validation: "validation";
183
+ test: "test";
184
+ frozen_eval: "frozen_eval";
185
+ }>;
186
+ population: z.ZodArray<z.ZodObject<{
187
+ taskId: z.ZodString;
188
+ seed: z.ZodString;
189
+ fixtureId: z.ZodNullable<z.ZodString>;
190
+ }, z.core.$strict>>;
191
+ evaluators: z.ZodArray<z.ZodObject<{
192
+ release: z.ZodObject<{
193
+ id: z.ZodString;
194
+ contentHash: z.ZodString;
195
+ revision: z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>;
196
+ }, z.core.$strict>;
197
+ feedbackKey: z.ZodString;
198
+ configurationHash: z.ZodString;
199
+ output: z.ZodEnum<{
200
+ boolean: "boolean";
201
+ score: "score";
202
+ category: "category";
203
+ }>;
204
+ categories: z.ZodArray<z.ZodString>;
205
+ }, z.core.$strict>>;
206
+ fieldMappingsHash: z.ZodString;
207
+ fieldMappings: z.ZodArray<z.ZodObject<{
208
+ graderId: z.ZodString;
209
+ fields: z.ZodArray<z.ZodObject<{
210
+ destination: z.ZodEnum<{
211
+ output: "output";
212
+ input: "input";
213
+ expectedOutput: "expectedOutput";
214
+ evaluatorContext: "evaluatorContext";
215
+ }>;
216
+ path: z.ZodArray<z.ZodString>;
217
+ source: z.ZodEnum<{
218
+ output: "output";
219
+ input: "input";
220
+ expectedOutput: "expectedOutput";
221
+ evaluatorContext: "evaluatorContext";
222
+ }>;
223
+ sourcePath: z.ZodArray<z.ZodString>;
224
+ }, z.core.$strict>>;
225
+ }, z.core.$strict>>;
226
+ environmentHash: z.ZodString;
227
+ privateDatasetClosureHash: z.ZodString;
228
+ criterion: z.ZodObject<{
229
+ minimumPassRate: z.ZodNumber;
230
+ requireComplete: z.ZodBoolean;
231
+ }, z.core.$strict>;
232
+ contentHash: z.ZodString;
233
+ }, z.core.$strict>>;
234
+ sourceCandidate: z.ZodOptional<z.ZodObject<{
235
+ candidateId: z.ZodString;
236
+ frozenRevision: z.ZodNumber;
237
+ frozenHash: z.ZodString;
238
+ harnessRelease: z.ZodObject<{
239
+ id: z.ZodString;
240
+ contentHash: z.ZodString;
241
+ }, z.core.$strict>;
242
+ protectedClosureHash: z.ZodString;
243
+ }, z.core.$strict>>;
244
+ }, z.core.$strict>;
245
+ model: z.ZodObject<{
246
+ modelId: z.ZodString;
247
+ configurationHash: z.ZodString;
248
+ }, z.core.$strict>;
249
+ }, z.core.$strict>, z.ZodObject<{
250
+ kind: z.ZodLiteral<"agent">;
251
+ source: z.ZodObject<{
252
+ harnessRelease: z.ZodObject<{
253
+ id: z.ZodString;
254
+ contentHash: z.ZodString;
255
+ }, z.core.$strict>;
256
+ agentSnapshot: z.ZodObject<{
257
+ id: z.ZodString;
258
+ contentHash: z.ZodString;
259
+ }, z.core.$strict>;
260
+ sourcePackageHash: z.ZodString;
261
+ }, z.core.$strict>;
262
+ model: z.ZodObject<{
263
+ modelId: z.ZodString;
264
+ configurationHash: z.ZodString;
265
+ }, z.core.$strict>;
266
+ }, z.core.$strict>, z.ZodObject<{
267
+ kind: z.ZodLiteral<"fixture">;
268
+ configurationHash: z.ZodString;
269
+ }, z.core.$strict>], "kind">;
270
+ execution: z.ZodOptional<z.ZodObject<{
271
+ packageHash: z.ZodString;
272
+ runtimeTargetHash: z.ZodString;
273
+ metricPolicyHash: z.ZodString;
274
+ compatibility: z.ZodOptional<z.ZodObject<{
275
+ protocol: z.ZodLiteral<"openpond.evaluation-execution.v1">;
276
+ targetKind: z.ZodEnum<{
277
+ agent: "agent";
278
+ model: "model";
279
+ fixture: "fixture";
280
+ harness: "harness";
281
+ }>;
282
+ }, z.core.$strict>>;
283
+ }, z.core.$strict>>;
284
+ lineage: z.ZodOptional<z.ZodObject<{
285
+ definition: z.ZodNullable<z.ZodObject<{
286
+ id: z.ZodString;
287
+ contentHash: z.ZodString;
288
+ revision: z.ZodNumber;
289
+ }, z.core.$strict>>;
290
+ execution: z.ZodObject<{
291
+ id: z.ZodString;
292
+ contentHash: z.ZodString;
293
+ }, z.core.$strict>;
294
+ scoringPassId: z.ZodNullable<z.ZodString>;
295
+ humanProjection: z.ZodOptional<z.ZodObject<{
296
+ originalManifest: z.ZodObject<{
297
+ id: z.ZodString;
298
+ contentHash: z.ZodString;
299
+ }, z.core.$strict>;
300
+ selectionHash: z.ZodString;
301
+ rulesHash: z.ZodString;
302
+ }, z.core.$strict>>;
303
+ }, z.core.$strict>>;
304
+ evaluators: z.ZodArray<z.ZodObject<{
305
+ release: z.ZodObject<{
306
+ id: z.ZodString;
307
+ contentHash: z.ZodString;
308
+ revision: z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>;
309
+ }, z.core.$strict>;
310
+ feedbackKey: z.ZodString;
311
+ configurationHash: z.ZodOptional<z.ZodString>;
312
+ output: z.ZodEnum<{
313
+ boolean: "boolean";
314
+ score: "score";
315
+ category: "category";
316
+ }>;
317
+ categories: z.ZodArray<z.ZodString>;
318
+ }, z.core.$strict>>;
319
+ population: z.ZodArray<z.ZodObject<{
320
+ caseId: z.ZodString;
321
+ seed: z.ZodString;
322
+ fixtureId: z.ZodNullable<z.ZodString>;
323
+ }, z.core.$strict>>;
324
+ createdAt: z.ZodString;
325
+ contentHash: z.ZodString;
326
+ }, z.core.$strict>;
327
+ result: z.ZodObject<{
328
+ schemaVersion: z.ZodLiteral<"openpond.experimentResult.v1">;
329
+ manifest: z.ZodObject<{
330
+ id: z.ZodString;
331
+ contentHash: z.ZodString;
332
+ }, z.core.$strict>;
333
+ status: z.ZodEnum<{
334
+ cancelled: "cancelled";
335
+ failed: "failed";
336
+ completed: "completed";
337
+ }>;
338
+ cases: z.ZodArray<z.ZodObject<{
339
+ identity: z.ZodObject<{
340
+ caseId: z.ZodString;
341
+ seed: z.ZodString;
342
+ fixtureId: z.ZodNullable<z.ZodString>;
343
+ }, z.core.$strict>;
344
+ status: z.ZodEnum<{
345
+ cancelled: "cancelled";
346
+ failed: "failed";
347
+ completed: "completed";
348
+ unavailable: "unavailable";
349
+ }>;
350
+ output: z.ZodNullable<z.ZodUnknown>;
351
+ error: z.ZodNullable<z.ZodObject<{
352
+ code: z.ZodString;
353
+ message: z.ZodString;
354
+ }, z.core.$strict>>;
355
+ feedback: z.ZodArray<z.ZodObject<{
356
+ feedbackKey: z.ZodString;
357
+ evaluator: z.ZodObject<{
358
+ id: z.ZodString;
359
+ contentHash: z.ZodString;
360
+ revision: z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>;
361
+ }, z.core.$strict>;
362
+ status: z.ZodEnum<{
363
+ pending: "pending";
364
+ failed: "failed";
365
+ scored: "scored";
366
+ unavailable: "unavailable";
367
+ }>;
368
+ value: z.ZodNullable<z.ZodUnion<readonly [z.ZodBoolean, z.ZodNumber, z.ZodString]>>;
369
+ passed: z.ZodOptional<z.ZodNullable<z.ZodBoolean>>;
370
+ reasoning: z.ZodNullable<z.ZodString>;
371
+ evidenceRefs: z.ZodArray<z.ZodObject<{
372
+ id: z.ZodString;
373
+ contentHash: z.ZodString;
374
+ mediaType: z.ZodDefault<z.ZodNullable<z.ZodString>>;
375
+ sizeBytes: z.ZodDefault<z.ZodNullable<z.ZodNumber>>;
376
+ }, z.core.$strict>>;
377
+ }, z.core.$strict>>;
378
+ usage: z.ZodObject<{
379
+ inputTokens: z.ZodNullable<z.ZodNumber>;
380
+ outputTokens: z.ZodNullable<z.ZodNumber>;
381
+ totalTokens: z.ZodNullable<z.ZodNumber>;
382
+ costUsd: z.ZodNullable<z.ZodNumber>;
383
+ latencyMs: z.ZodNullable<z.ZodNumber>;
384
+ }, z.core.$strict>;
385
+ traceRef: z.ZodNullable<z.ZodObject<{
386
+ id: z.ZodString;
387
+ contentHash: z.ZodString;
388
+ mediaType: z.ZodDefault<z.ZodNullable<z.ZodString>>;
389
+ sizeBytes: z.ZodDefault<z.ZodNullable<z.ZodNumber>>;
390
+ }, z.core.$strict>>;
391
+ startedAt: z.ZodNullable<z.ZodString>;
392
+ completedAt: z.ZodNullable<z.ZodString>;
393
+ }, z.core.$strict>>;
394
+ completedAt: z.ZodString;
395
+ contentHash: z.ZodString;
396
+ }, z.core.$strict>;
397
+ }, z.core.$strict>;
398
+ export declare const HumanComparisonProjectionSchema: z.ZodObject<{
399
+ schemaVersion: z.ZodLiteral<"openpond.humanComparisonProjection.v1">;
400
+ teamId: z.ZodString;
401
+ projectId: z.ZodString;
402
+ selections: z.ZodObject<{
403
+ rules: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
404
+ kind: z.ZodLiteral<"normalized_score">;
405
+ minimum: z.ZodNumber;
406
+ maximum: z.ZodNumber;
407
+ passThreshold: z.ZodOptional<z.ZodNumber>;
408
+ feedbackKey: z.ZodString;
409
+ criterionId: z.ZodString;
410
+ }, z.core.$strict>, z.ZodObject<{
411
+ kind: z.ZodLiteral<"boolean">;
412
+ feedbackKey: z.ZodString;
413
+ criterionId: z.ZodString;
414
+ }, z.core.$strict>, z.ZodObject<{
415
+ kind: z.ZodLiteral<"category">;
416
+ feedbackKey: z.ZodString;
417
+ criterionId: z.ZodString;
418
+ }, z.core.$strict>, z.ZodObject<{
419
+ kind: z.ZodLiteral<"preference_score">;
420
+ tieScore: z.ZodNullable<z.ZodNumber>;
421
+ feedbackKey: z.ZodString;
422
+ criterionId: z.ZodString;
423
+ }, z.core.$strict>], "kind">>;
424
+ baseline: z.ZodObject<{
425
+ execution: z.ZodObject<{
426
+ id: z.ZodString;
427
+ contentHash: z.ZodString;
428
+ }, z.core.$strict>;
429
+ resultHash: z.ZodString;
430
+ reviews: z.ZodArray<z.ZodObject<{
431
+ review: z.ZodObject<{
432
+ id: z.ZodString;
433
+ revision: z.ZodNumber;
434
+ contentHash: z.ZodString;
435
+ }, z.core.$strict>;
436
+ decision: z.ZodObject<{
437
+ id: z.ZodString;
438
+ revision: z.ZodNumber;
439
+ contentHash: z.ZodString;
440
+ }, z.core.$strict>;
441
+ }, z.core.$strict>>;
442
+ }, z.core.$strict>;
443
+ candidate: z.ZodObject<{
444
+ execution: z.ZodObject<{
445
+ id: z.ZodString;
446
+ contentHash: z.ZodString;
447
+ }, z.core.$strict>;
448
+ resultHash: z.ZodString;
449
+ reviews: z.ZodArray<z.ZodObject<{
450
+ review: z.ZodObject<{
451
+ id: z.ZodString;
452
+ revision: z.ZodNumber;
453
+ contentHash: z.ZodString;
454
+ }, z.core.$strict>;
455
+ decision: z.ZodObject<{
456
+ id: z.ZodString;
457
+ revision: z.ZodNumber;
458
+ contentHash: z.ZodString;
459
+ }, z.core.$strict>;
460
+ }, z.core.$strict>>;
461
+ }, z.core.$strict>;
462
+ }, z.core.$strict>;
463
+ original: z.ZodObject<{
464
+ baseline: z.ZodObject<{
465
+ id: z.ZodString;
466
+ manifestHash: z.ZodString;
467
+ resultHash: z.ZodString;
468
+ }, z.core.$strict>;
469
+ candidate: z.ZodObject<{
470
+ id: z.ZodString;
471
+ manifestHash: z.ZodString;
472
+ resultHash: z.ZodString;
473
+ }, z.core.$strict>;
474
+ }, z.core.$strict>;
475
+ humanResults: z.ZodObject<{
476
+ baseline: z.ZodObject<{
477
+ schemaVersion: z.ZodLiteral<"openpond.humanResultView.v1">;
478
+ executionId: z.ZodString;
479
+ executionHash: z.ZodString;
480
+ selection: z.ZodArray<z.ZodObject<{
481
+ id: z.ZodString;
482
+ revision: z.ZodNumber;
483
+ contentHash: z.ZodString;
484
+ }, z.core.$strict>>;
485
+ coverage: z.ZodObject<{
486
+ pending: z.ZodNumber;
487
+ submitted: z.ZodNumber;
488
+ accepted: z.ZodNumber;
489
+ rejected: z.ZodNumber;
490
+ cancelled: z.ZodNumber;
491
+ }, z.core.$strict>;
492
+ results: z.ZodArray<z.ZodObject<{
493
+ review: z.ZodObject<{
494
+ id: z.ZodString;
495
+ revision: z.ZodNumber;
496
+ contentHash: z.ZodString;
497
+ }, z.core.$strict>;
498
+ decision: z.ZodNullable<z.ZodObject<{
499
+ id: z.ZodString;
500
+ revision: z.ZodNumber;
501
+ contentHash: z.ZodString;
502
+ }, z.core.$strict>>;
503
+ attemptIds: z.ZodArray<z.ZodString>;
504
+ mode: z.ZodEnum<{
505
+ individual: "individual";
506
+ pairwise: "pairwise";
507
+ }>;
508
+ origin: z.ZodEnum<{
509
+ owner_attested_local: "owner_attested_local";
510
+ native_local: "native_local";
511
+ recorded_evidence: "recorded_evidence";
512
+ hosted_retained: "hosted_retained";
513
+ }>;
514
+ status: z.ZodEnum<{
515
+ submitted: "submitted";
516
+ accepted: "accepted";
517
+ rejected: "rejected";
518
+ cancelled: "cancelled";
519
+ pending: "pending";
520
+ }>;
521
+ grader: z.ZodObject<{
522
+ id: z.ZodString;
523
+ revision: z.ZodNumber;
524
+ contentHash: z.ZodString;
525
+ }, z.core.$strict>;
526
+ answers: z.ZodArray<z.ZodObject<{
527
+ criterionId: z.ZodString;
528
+ value: z.ZodUnion<readonly [z.ZodNumber, z.ZodBoolean, z.ZodString, z.ZodNull]>;
529
+ abstain: z.ZodBoolean;
530
+ note: z.ZodString;
531
+ strength: z.ZodOptional<z.ZodEnum<{
532
+ slight: "slight";
533
+ moderate: "moderate";
534
+ strong: "strong";
535
+ }>>;
536
+ }, z.core.$strict>>;
537
+ }, z.core.$strict>>;
538
+ contentHash: z.ZodString;
539
+ }, z.core.$strict>;
540
+ candidate: z.ZodObject<{
541
+ schemaVersion: z.ZodLiteral<"openpond.humanResultView.v1">;
542
+ executionId: z.ZodString;
543
+ executionHash: z.ZodString;
544
+ selection: z.ZodArray<z.ZodObject<{
545
+ id: z.ZodString;
546
+ revision: z.ZodNumber;
547
+ contentHash: z.ZodString;
548
+ }, z.core.$strict>>;
549
+ coverage: z.ZodObject<{
550
+ pending: z.ZodNumber;
551
+ submitted: z.ZodNumber;
552
+ accepted: z.ZodNumber;
553
+ rejected: z.ZodNumber;
554
+ cancelled: z.ZodNumber;
555
+ }, z.core.$strict>;
556
+ results: z.ZodArray<z.ZodObject<{
557
+ review: z.ZodObject<{
558
+ id: z.ZodString;
559
+ revision: z.ZodNumber;
560
+ contentHash: z.ZodString;
561
+ }, z.core.$strict>;
562
+ decision: z.ZodNullable<z.ZodObject<{
563
+ id: z.ZodString;
564
+ revision: z.ZodNumber;
565
+ contentHash: z.ZodString;
566
+ }, z.core.$strict>>;
567
+ attemptIds: z.ZodArray<z.ZodString>;
568
+ mode: z.ZodEnum<{
569
+ individual: "individual";
570
+ pairwise: "pairwise";
571
+ }>;
572
+ origin: z.ZodEnum<{
573
+ owner_attested_local: "owner_attested_local";
574
+ native_local: "native_local";
575
+ recorded_evidence: "recorded_evidence";
576
+ hosted_retained: "hosted_retained";
577
+ }>;
578
+ status: z.ZodEnum<{
579
+ submitted: "submitted";
580
+ accepted: "accepted";
581
+ rejected: "rejected";
582
+ cancelled: "cancelled";
583
+ pending: "pending";
584
+ }>;
585
+ grader: z.ZodObject<{
586
+ id: z.ZodString;
587
+ revision: z.ZodNumber;
588
+ contentHash: z.ZodString;
589
+ }, z.core.$strict>;
590
+ answers: z.ZodArray<z.ZodObject<{
591
+ criterionId: z.ZodString;
592
+ value: z.ZodUnion<readonly [z.ZodNumber, z.ZodBoolean, z.ZodString, z.ZodNull]>;
593
+ abstain: z.ZodBoolean;
594
+ note: z.ZodString;
595
+ strength: z.ZodOptional<z.ZodEnum<{
596
+ slight: "slight";
597
+ moderate: "moderate";
598
+ strong: "strong";
599
+ }>>;
600
+ }, z.core.$strict>>;
601
+ }, z.core.$strict>>;
602
+ contentHash: z.ZodString;
603
+ }, z.core.$strict>;
604
+ }, z.core.$strict>;
605
+ baseline: z.ZodObject<{
606
+ manifest: z.ZodObject<{
607
+ schemaVersion: z.ZodLiteral<"openpond.experimentManifest.v1">;
608
+ id: z.ZodString;
609
+ name: z.ZodString;
610
+ teamId: z.ZodString;
611
+ operationId: z.ZodString;
612
+ maximumCostUsd: z.ZodNullable<z.ZodNumber>;
613
+ dataset: z.ZodObject<{
614
+ id: z.ZodString;
615
+ contentHash: z.ZodString;
616
+ revision: z.ZodNumber;
617
+ }, z.core.$strict>;
618
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
619
+ kind: z.ZodLiteral<"model">;
620
+ modelId: z.ZodString;
621
+ configurationHash: z.ZodString;
622
+ }, z.core.$strict>, z.ZodObject<{
623
+ kind: z.ZodLiteral<"harness">;
624
+ source: z.ZodObject<{
625
+ profileId: z.ZodString;
626
+ sourceRevision: z.ZodString;
627
+ harnessRelease: z.ZodObject<{
628
+ id: z.ZodString;
629
+ contentHash: z.ZodString;
630
+ }, z.core.$strict>;
631
+ catalogHash: z.ZodString;
632
+ definitionId: z.ZodString;
633
+ definitionHash: z.ZodString;
634
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
635
+ kind: z.ZodLiteral<"profile">;
636
+ }, z.core.$strict>, z.ZodObject<{
637
+ kind: z.ZodLiteral<"workflow">;
638
+ workflowId: z.ZodString;
639
+ }, z.core.$strict>, z.ZodObject<{
640
+ kind: z.ZodLiteral<"skill">;
641
+ skillPath: z.ZodString;
642
+ }, z.core.$strict>, z.ZodObject<{
643
+ kind: z.ZodLiteral<"agent_action">;
644
+ actionId: z.ZodString;
645
+ }, z.core.$strict>], "kind">;
646
+ environmentHash: z.ZodString;
647
+ externalDatasetBinding: z.ZodOptional<z.ZodObject<{
648
+ schemaVersion: z.ZodLiteral<"openpond.profileExternalDatasetBinding.v1">;
649
+ dataset: z.ZodObject<{
650
+ id: z.ZodString;
651
+ contentHash: z.ZodString;
652
+ revision: z.ZodNumber;
653
+ }, z.core.$strict>;
654
+ packageHash: z.ZodString;
655
+ profileId: z.ZodString;
656
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
657
+ kind: z.ZodLiteral<"profile">;
658
+ }, z.core.$strict>, z.ZodObject<{
659
+ kind: z.ZodLiteral<"workflow">;
660
+ workflowId: z.ZodString;
661
+ }, z.core.$strict>, z.ZodObject<{
662
+ kind: z.ZodLiteral<"skill">;
663
+ skillPath: z.ZodString;
664
+ }, z.core.$strict>, z.ZodObject<{
665
+ kind: z.ZodLiteral<"agent_action">;
666
+ actionId: z.ZodString;
667
+ }, z.core.$strict>], "kind">;
668
+ declaredProfileCatalogHash: z.ZodNullable<z.ZodString>;
669
+ protectedProfileClosureHash: z.ZodString;
670
+ recordedOrigin: z.ZodOptional<z.ZodObject<{
671
+ execution: z.ZodObject<{
672
+ id: z.ZodString;
673
+ contentHash: z.ZodString;
674
+ }, z.core.$strict>;
675
+ originalEvidence: z.ZodObject<{
676
+ id: z.ZodString;
677
+ contentHash: z.ZodString;
678
+ }, z.core.$strict>;
679
+ members: z.ZodArray<z.ZodObject<{
680
+ caseId: z.ZodString;
681
+ seed: z.ZodString;
682
+ fixtureId: z.ZodString;
683
+ sourceId: z.ZodString;
684
+ snapshotHash: z.ZodString;
685
+ boundaryRevisionHash: z.ZodString;
686
+ }, z.core.$strict>>;
687
+ }, z.core.$strict>>;
688
+ gradingSource: z.ZodOptional<z.ZodObject<{
689
+ pass: z.ZodObject<{
690
+ id: z.ZodString;
691
+ contentHash: z.ZodString;
692
+ }, z.core.$strict>;
693
+ evidence: z.ZodObject<{
694
+ id: z.ZodString;
695
+ contentHash: z.ZodString;
696
+ }, z.core.$strict>;
697
+ }, z.core.$strict>>;
698
+ split: z.ZodEnum<{
699
+ train: "train";
700
+ validation: "validation";
701
+ test: "test";
702
+ frozen_eval: "frozen_eval";
703
+ }>;
704
+ population: z.ZodArray<z.ZodObject<{
705
+ taskId: z.ZodString;
706
+ seed: z.ZodString;
707
+ fixtureId: z.ZodNullable<z.ZodString>;
708
+ }, z.core.$strict>>;
709
+ evaluators: z.ZodArray<z.ZodObject<{
710
+ release: z.ZodObject<{
711
+ id: z.ZodString;
712
+ contentHash: z.ZodString;
713
+ revision: z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>;
714
+ }, z.core.$strict>;
715
+ feedbackKey: z.ZodString;
716
+ configurationHash: z.ZodString;
717
+ output: z.ZodEnum<{
718
+ boolean: "boolean";
719
+ score: "score";
720
+ category: "category";
721
+ }>;
722
+ categories: z.ZodArray<z.ZodString>;
723
+ }, z.core.$strict>>;
724
+ fieldMappingsHash: z.ZodString;
725
+ fieldMappings: z.ZodArray<z.ZodObject<{
726
+ graderId: z.ZodString;
727
+ fields: z.ZodArray<z.ZodObject<{
728
+ destination: z.ZodEnum<{
729
+ output: "output";
730
+ input: "input";
731
+ expectedOutput: "expectedOutput";
732
+ evaluatorContext: "evaluatorContext";
733
+ }>;
734
+ path: z.ZodArray<z.ZodString>;
735
+ source: z.ZodEnum<{
736
+ output: "output";
737
+ input: "input";
738
+ expectedOutput: "expectedOutput";
739
+ evaluatorContext: "evaluatorContext";
740
+ }>;
741
+ sourcePath: z.ZodArray<z.ZodString>;
742
+ }, z.core.$strict>>;
743
+ }, z.core.$strict>>;
744
+ environmentHash: z.ZodString;
745
+ privateDatasetClosureHash: z.ZodString;
746
+ criterion: z.ZodObject<{
747
+ minimumPassRate: z.ZodNumber;
748
+ requireComplete: z.ZodBoolean;
749
+ }, z.core.$strict>;
750
+ contentHash: z.ZodString;
751
+ }, z.core.$strict>>;
752
+ sourceCandidate: z.ZodOptional<z.ZodObject<{
753
+ candidateId: z.ZodString;
754
+ frozenRevision: z.ZodNumber;
755
+ frozenHash: z.ZodString;
756
+ harnessRelease: z.ZodObject<{
757
+ id: z.ZodString;
758
+ contentHash: z.ZodString;
759
+ }, z.core.$strict>;
760
+ protectedClosureHash: z.ZodString;
761
+ }, z.core.$strict>>;
762
+ }, z.core.$strict>;
763
+ model: z.ZodObject<{
764
+ modelId: z.ZodString;
765
+ configurationHash: z.ZodString;
766
+ }, z.core.$strict>;
767
+ }, z.core.$strict>, z.ZodObject<{
768
+ kind: z.ZodLiteral<"agent">;
769
+ source: z.ZodObject<{
770
+ harnessRelease: z.ZodObject<{
771
+ id: z.ZodString;
772
+ contentHash: z.ZodString;
773
+ }, z.core.$strict>;
774
+ agentSnapshot: z.ZodObject<{
775
+ id: z.ZodString;
776
+ contentHash: z.ZodString;
777
+ }, z.core.$strict>;
778
+ sourcePackageHash: z.ZodString;
779
+ }, z.core.$strict>;
780
+ model: z.ZodObject<{
781
+ modelId: z.ZodString;
782
+ configurationHash: z.ZodString;
783
+ }, z.core.$strict>;
784
+ }, z.core.$strict>, z.ZodObject<{
785
+ kind: z.ZodLiteral<"fixture">;
786
+ configurationHash: z.ZodString;
787
+ }, z.core.$strict>], "kind">;
788
+ execution: z.ZodOptional<z.ZodObject<{
789
+ packageHash: z.ZodString;
790
+ runtimeTargetHash: z.ZodString;
791
+ metricPolicyHash: z.ZodString;
792
+ compatibility: z.ZodOptional<z.ZodObject<{
793
+ protocol: z.ZodLiteral<"openpond.evaluation-execution.v1">;
794
+ targetKind: z.ZodEnum<{
795
+ agent: "agent";
796
+ model: "model";
797
+ fixture: "fixture";
798
+ harness: "harness";
799
+ }>;
800
+ }, z.core.$strict>>;
801
+ }, z.core.$strict>>;
802
+ lineage: z.ZodOptional<z.ZodObject<{
803
+ definition: z.ZodNullable<z.ZodObject<{
804
+ id: z.ZodString;
805
+ contentHash: z.ZodString;
806
+ revision: z.ZodNumber;
807
+ }, z.core.$strict>>;
808
+ execution: z.ZodObject<{
809
+ id: z.ZodString;
810
+ contentHash: z.ZodString;
811
+ }, z.core.$strict>;
812
+ scoringPassId: z.ZodNullable<z.ZodString>;
813
+ humanProjection: z.ZodOptional<z.ZodObject<{
814
+ originalManifest: z.ZodObject<{
815
+ id: z.ZodString;
816
+ contentHash: z.ZodString;
817
+ }, z.core.$strict>;
818
+ selectionHash: z.ZodString;
819
+ rulesHash: z.ZodString;
820
+ }, z.core.$strict>>;
821
+ }, z.core.$strict>>;
822
+ evaluators: z.ZodArray<z.ZodObject<{
823
+ release: z.ZodObject<{
824
+ id: z.ZodString;
825
+ contentHash: z.ZodString;
826
+ revision: z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>;
827
+ }, z.core.$strict>;
828
+ feedbackKey: z.ZodString;
829
+ configurationHash: z.ZodOptional<z.ZodString>;
830
+ output: z.ZodEnum<{
831
+ boolean: "boolean";
832
+ score: "score";
833
+ category: "category";
834
+ }>;
835
+ categories: z.ZodArray<z.ZodString>;
836
+ }, z.core.$strict>>;
837
+ population: z.ZodArray<z.ZodObject<{
838
+ caseId: z.ZodString;
839
+ seed: z.ZodString;
840
+ fixtureId: z.ZodNullable<z.ZodString>;
841
+ }, z.core.$strict>>;
842
+ createdAt: z.ZodString;
843
+ contentHash: z.ZodString;
844
+ }, z.core.$strict>;
845
+ result: z.ZodObject<{
846
+ schemaVersion: z.ZodLiteral<"openpond.experimentResult.v1">;
847
+ manifest: z.ZodObject<{
848
+ id: z.ZodString;
849
+ contentHash: z.ZodString;
850
+ }, z.core.$strict>;
851
+ status: z.ZodEnum<{
852
+ cancelled: "cancelled";
853
+ failed: "failed";
854
+ completed: "completed";
855
+ }>;
856
+ cases: z.ZodArray<z.ZodObject<{
857
+ identity: z.ZodObject<{
858
+ caseId: z.ZodString;
859
+ seed: z.ZodString;
860
+ fixtureId: z.ZodNullable<z.ZodString>;
861
+ }, z.core.$strict>;
862
+ status: z.ZodEnum<{
863
+ cancelled: "cancelled";
864
+ failed: "failed";
865
+ completed: "completed";
866
+ unavailable: "unavailable";
867
+ }>;
868
+ output: z.ZodNullable<z.ZodUnknown>;
869
+ error: z.ZodNullable<z.ZodObject<{
870
+ code: z.ZodString;
871
+ message: z.ZodString;
872
+ }, z.core.$strict>>;
873
+ feedback: z.ZodArray<z.ZodObject<{
874
+ feedbackKey: z.ZodString;
875
+ evaluator: z.ZodObject<{
876
+ id: z.ZodString;
877
+ contentHash: z.ZodString;
878
+ revision: z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>;
879
+ }, z.core.$strict>;
880
+ status: z.ZodEnum<{
881
+ pending: "pending";
882
+ failed: "failed";
883
+ scored: "scored";
884
+ unavailable: "unavailable";
885
+ }>;
886
+ value: z.ZodNullable<z.ZodUnion<readonly [z.ZodBoolean, z.ZodNumber, z.ZodString]>>;
887
+ passed: z.ZodOptional<z.ZodNullable<z.ZodBoolean>>;
888
+ reasoning: z.ZodNullable<z.ZodString>;
889
+ evidenceRefs: z.ZodArray<z.ZodObject<{
890
+ id: z.ZodString;
891
+ contentHash: z.ZodString;
892
+ mediaType: z.ZodDefault<z.ZodNullable<z.ZodString>>;
893
+ sizeBytes: z.ZodDefault<z.ZodNullable<z.ZodNumber>>;
894
+ }, z.core.$strict>>;
895
+ }, z.core.$strict>>;
896
+ usage: z.ZodObject<{
897
+ inputTokens: z.ZodNullable<z.ZodNumber>;
898
+ outputTokens: z.ZodNullable<z.ZodNumber>;
899
+ totalTokens: z.ZodNullable<z.ZodNumber>;
900
+ costUsd: z.ZodNullable<z.ZodNumber>;
901
+ latencyMs: z.ZodNullable<z.ZodNumber>;
902
+ }, z.core.$strict>;
903
+ traceRef: z.ZodNullable<z.ZodObject<{
904
+ id: z.ZodString;
905
+ contentHash: z.ZodString;
906
+ mediaType: z.ZodDefault<z.ZodNullable<z.ZodString>>;
907
+ sizeBytes: z.ZodDefault<z.ZodNullable<z.ZodNumber>>;
908
+ }, z.core.$strict>>;
909
+ startedAt: z.ZodNullable<z.ZodString>;
910
+ completedAt: z.ZodNullable<z.ZodString>;
911
+ }, z.core.$strict>>;
912
+ completedAt: z.ZodString;
913
+ contentHash: z.ZodString;
914
+ }, z.core.$strict>;
915
+ }, z.core.$strict>;
916
+ candidate: z.ZodObject<{
917
+ manifest: z.ZodObject<{
918
+ schemaVersion: z.ZodLiteral<"openpond.experimentManifest.v1">;
919
+ id: z.ZodString;
920
+ name: z.ZodString;
921
+ teamId: z.ZodString;
922
+ operationId: z.ZodString;
923
+ maximumCostUsd: z.ZodNullable<z.ZodNumber>;
924
+ dataset: z.ZodObject<{
925
+ id: z.ZodString;
926
+ contentHash: z.ZodString;
927
+ revision: z.ZodNumber;
928
+ }, z.core.$strict>;
929
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
930
+ kind: z.ZodLiteral<"model">;
931
+ modelId: z.ZodString;
932
+ configurationHash: z.ZodString;
933
+ }, z.core.$strict>, z.ZodObject<{
934
+ kind: z.ZodLiteral<"harness">;
935
+ source: z.ZodObject<{
936
+ profileId: z.ZodString;
937
+ sourceRevision: z.ZodString;
938
+ harnessRelease: z.ZodObject<{
939
+ id: z.ZodString;
940
+ contentHash: z.ZodString;
941
+ }, z.core.$strict>;
942
+ catalogHash: z.ZodString;
943
+ definitionId: z.ZodString;
944
+ definitionHash: z.ZodString;
945
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
946
+ kind: z.ZodLiteral<"profile">;
947
+ }, z.core.$strict>, z.ZodObject<{
948
+ kind: z.ZodLiteral<"workflow">;
949
+ workflowId: z.ZodString;
950
+ }, z.core.$strict>, z.ZodObject<{
951
+ kind: z.ZodLiteral<"skill">;
952
+ skillPath: z.ZodString;
953
+ }, z.core.$strict>, z.ZodObject<{
954
+ kind: z.ZodLiteral<"agent_action">;
955
+ actionId: z.ZodString;
956
+ }, z.core.$strict>], "kind">;
957
+ environmentHash: z.ZodString;
958
+ externalDatasetBinding: z.ZodOptional<z.ZodObject<{
959
+ schemaVersion: z.ZodLiteral<"openpond.profileExternalDatasetBinding.v1">;
960
+ dataset: z.ZodObject<{
961
+ id: z.ZodString;
962
+ contentHash: z.ZodString;
963
+ revision: z.ZodNumber;
964
+ }, z.core.$strict>;
965
+ packageHash: z.ZodString;
966
+ profileId: z.ZodString;
967
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
968
+ kind: z.ZodLiteral<"profile">;
969
+ }, z.core.$strict>, z.ZodObject<{
970
+ kind: z.ZodLiteral<"workflow">;
971
+ workflowId: z.ZodString;
972
+ }, z.core.$strict>, z.ZodObject<{
973
+ kind: z.ZodLiteral<"skill">;
974
+ skillPath: z.ZodString;
975
+ }, z.core.$strict>, z.ZodObject<{
976
+ kind: z.ZodLiteral<"agent_action">;
977
+ actionId: z.ZodString;
978
+ }, z.core.$strict>], "kind">;
979
+ declaredProfileCatalogHash: z.ZodNullable<z.ZodString>;
980
+ protectedProfileClosureHash: z.ZodString;
981
+ recordedOrigin: z.ZodOptional<z.ZodObject<{
982
+ execution: z.ZodObject<{
983
+ id: z.ZodString;
984
+ contentHash: z.ZodString;
985
+ }, z.core.$strict>;
986
+ originalEvidence: z.ZodObject<{
987
+ id: z.ZodString;
988
+ contentHash: z.ZodString;
989
+ }, z.core.$strict>;
990
+ members: z.ZodArray<z.ZodObject<{
991
+ caseId: z.ZodString;
992
+ seed: z.ZodString;
993
+ fixtureId: z.ZodString;
994
+ sourceId: z.ZodString;
995
+ snapshotHash: z.ZodString;
996
+ boundaryRevisionHash: z.ZodString;
997
+ }, z.core.$strict>>;
998
+ }, z.core.$strict>>;
999
+ gradingSource: z.ZodOptional<z.ZodObject<{
1000
+ pass: z.ZodObject<{
1001
+ id: z.ZodString;
1002
+ contentHash: z.ZodString;
1003
+ }, z.core.$strict>;
1004
+ evidence: z.ZodObject<{
1005
+ id: z.ZodString;
1006
+ contentHash: z.ZodString;
1007
+ }, z.core.$strict>;
1008
+ }, z.core.$strict>>;
1009
+ split: z.ZodEnum<{
1010
+ train: "train";
1011
+ validation: "validation";
1012
+ test: "test";
1013
+ frozen_eval: "frozen_eval";
1014
+ }>;
1015
+ population: z.ZodArray<z.ZodObject<{
1016
+ taskId: z.ZodString;
1017
+ seed: z.ZodString;
1018
+ fixtureId: z.ZodNullable<z.ZodString>;
1019
+ }, z.core.$strict>>;
1020
+ evaluators: z.ZodArray<z.ZodObject<{
1021
+ release: z.ZodObject<{
1022
+ id: z.ZodString;
1023
+ contentHash: z.ZodString;
1024
+ revision: z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>;
1025
+ }, z.core.$strict>;
1026
+ feedbackKey: z.ZodString;
1027
+ configurationHash: z.ZodString;
1028
+ output: z.ZodEnum<{
1029
+ boolean: "boolean";
1030
+ score: "score";
1031
+ category: "category";
1032
+ }>;
1033
+ categories: z.ZodArray<z.ZodString>;
1034
+ }, z.core.$strict>>;
1035
+ fieldMappingsHash: z.ZodString;
1036
+ fieldMappings: z.ZodArray<z.ZodObject<{
1037
+ graderId: z.ZodString;
1038
+ fields: z.ZodArray<z.ZodObject<{
1039
+ destination: z.ZodEnum<{
1040
+ output: "output";
1041
+ input: "input";
1042
+ expectedOutput: "expectedOutput";
1043
+ evaluatorContext: "evaluatorContext";
1044
+ }>;
1045
+ path: z.ZodArray<z.ZodString>;
1046
+ source: z.ZodEnum<{
1047
+ output: "output";
1048
+ input: "input";
1049
+ expectedOutput: "expectedOutput";
1050
+ evaluatorContext: "evaluatorContext";
1051
+ }>;
1052
+ sourcePath: z.ZodArray<z.ZodString>;
1053
+ }, z.core.$strict>>;
1054
+ }, z.core.$strict>>;
1055
+ environmentHash: z.ZodString;
1056
+ privateDatasetClosureHash: z.ZodString;
1057
+ criterion: z.ZodObject<{
1058
+ minimumPassRate: z.ZodNumber;
1059
+ requireComplete: z.ZodBoolean;
1060
+ }, z.core.$strict>;
1061
+ contentHash: z.ZodString;
1062
+ }, z.core.$strict>>;
1063
+ sourceCandidate: z.ZodOptional<z.ZodObject<{
1064
+ candidateId: z.ZodString;
1065
+ frozenRevision: z.ZodNumber;
1066
+ frozenHash: z.ZodString;
1067
+ harnessRelease: z.ZodObject<{
1068
+ id: z.ZodString;
1069
+ contentHash: z.ZodString;
1070
+ }, z.core.$strict>;
1071
+ protectedClosureHash: z.ZodString;
1072
+ }, z.core.$strict>>;
1073
+ }, z.core.$strict>;
1074
+ model: z.ZodObject<{
1075
+ modelId: z.ZodString;
1076
+ configurationHash: z.ZodString;
1077
+ }, z.core.$strict>;
1078
+ }, z.core.$strict>, z.ZodObject<{
1079
+ kind: z.ZodLiteral<"agent">;
1080
+ source: z.ZodObject<{
1081
+ harnessRelease: z.ZodObject<{
1082
+ id: z.ZodString;
1083
+ contentHash: z.ZodString;
1084
+ }, z.core.$strict>;
1085
+ agentSnapshot: z.ZodObject<{
1086
+ id: z.ZodString;
1087
+ contentHash: z.ZodString;
1088
+ }, z.core.$strict>;
1089
+ sourcePackageHash: z.ZodString;
1090
+ }, z.core.$strict>;
1091
+ model: z.ZodObject<{
1092
+ modelId: z.ZodString;
1093
+ configurationHash: z.ZodString;
1094
+ }, z.core.$strict>;
1095
+ }, z.core.$strict>, z.ZodObject<{
1096
+ kind: z.ZodLiteral<"fixture">;
1097
+ configurationHash: z.ZodString;
1098
+ }, z.core.$strict>], "kind">;
1099
+ execution: z.ZodOptional<z.ZodObject<{
1100
+ packageHash: z.ZodString;
1101
+ runtimeTargetHash: z.ZodString;
1102
+ metricPolicyHash: z.ZodString;
1103
+ compatibility: z.ZodOptional<z.ZodObject<{
1104
+ protocol: z.ZodLiteral<"openpond.evaluation-execution.v1">;
1105
+ targetKind: z.ZodEnum<{
1106
+ agent: "agent";
1107
+ model: "model";
1108
+ fixture: "fixture";
1109
+ harness: "harness";
1110
+ }>;
1111
+ }, z.core.$strict>>;
1112
+ }, z.core.$strict>>;
1113
+ lineage: z.ZodOptional<z.ZodObject<{
1114
+ definition: z.ZodNullable<z.ZodObject<{
1115
+ id: z.ZodString;
1116
+ contentHash: z.ZodString;
1117
+ revision: z.ZodNumber;
1118
+ }, z.core.$strict>>;
1119
+ execution: z.ZodObject<{
1120
+ id: z.ZodString;
1121
+ contentHash: z.ZodString;
1122
+ }, z.core.$strict>;
1123
+ scoringPassId: z.ZodNullable<z.ZodString>;
1124
+ humanProjection: z.ZodOptional<z.ZodObject<{
1125
+ originalManifest: z.ZodObject<{
1126
+ id: z.ZodString;
1127
+ contentHash: z.ZodString;
1128
+ }, z.core.$strict>;
1129
+ selectionHash: z.ZodString;
1130
+ rulesHash: z.ZodString;
1131
+ }, z.core.$strict>>;
1132
+ }, z.core.$strict>>;
1133
+ evaluators: z.ZodArray<z.ZodObject<{
1134
+ release: z.ZodObject<{
1135
+ id: z.ZodString;
1136
+ contentHash: z.ZodString;
1137
+ revision: z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>;
1138
+ }, z.core.$strict>;
1139
+ feedbackKey: z.ZodString;
1140
+ configurationHash: z.ZodOptional<z.ZodString>;
1141
+ output: z.ZodEnum<{
1142
+ boolean: "boolean";
1143
+ score: "score";
1144
+ category: "category";
1145
+ }>;
1146
+ categories: z.ZodArray<z.ZodString>;
1147
+ }, z.core.$strict>>;
1148
+ population: z.ZodArray<z.ZodObject<{
1149
+ caseId: z.ZodString;
1150
+ seed: z.ZodString;
1151
+ fixtureId: z.ZodNullable<z.ZodString>;
1152
+ }, z.core.$strict>>;
1153
+ createdAt: z.ZodString;
1154
+ contentHash: z.ZodString;
1155
+ }, z.core.$strict>;
1156
+ result: z.ZodObject<{
1157
+ schemaVersion: z.ZodLiteral<"openpond.experimentResult.v1">;
1158
+ manifest: z.ZodObject<{
1159
+ id: z.ZodString;
1160
+ contentHash: z.ZodString;
1161
+ }, z.core.$strict>;
1162
+ status: z.ZodEnum<{
1163
+ cancelled: "cancelled";
1164
+ failed: "failed";
1165
+ completed: "completed";
1166
+ }>;
1167
+ cases: z.ZodArray<z.ZodObject<{
1168
+ identity: z.ZodObject<{
1169
+ caseId: z.ZodString;
1170
+ seed: z.ZodString;
1171
+ fixtureId: z.ZodNullable<z.ZodString>;
1172
+ }, z.core.$strict>;
1173
+ status: z.ZodEnum<{
1174
+ cancelled: "cancelled";
1175
+ failed: "failed";
1176
+ completed: "completed";
1177
+ unavailable: "unavailable";
1178
+ }>;
1179
+ output: z.ZodNullable<z.ZodUnknown>;
1180
+ error: z.ZodNullable<z.ZodObject<{
1181
+ code: z.ZodString;
1182
+ message: z.ZodString;
1183
+ }, z.core.$strict>>;
1184
+ feedback: z.ZodArray<z.ZodObject<{
1185
+ feedbackKey: z.ZodString;
1186
+ evaluator: z.ZodObject<{
1187
+ id: z.ZodString;
1188
+ contentHash: z.ZodString;
1189
+ revision: z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>;
1190
+ }, z.core.$strict>;
1191
+ status: z.ZodEnum<{
1192
+ pending: "pending";
1193
+ failed: "failed";
1194
+ scored: "scored";
1195
+ unavailable: "unavailable";
1196
+ }>;
1197
+ value: z.ZodNullable<z.ZodUnion<readonly [z.ZodBoolean, z.ZodNumber, z.ZodString]>>;
1198
+ passed: z.ZodOptional<z.ZodNullable<z.ZodBoolean>>;
1199
+ reasoning: z.ZodNullable<z.ZodString>;
1200
+ evidenceRefs: z.ZodArray<z.ZodObject<{
1201
+ id: z.ZodString;
1202
+ contentHash: z.ZodString;
1203
+ mediaType: z.ZodDefault<z.ZodNullable<z.ZodString>>;
1204
+ sizeBytes: z.ZodDefault<z.ZodNullable<z.ZodNumber>>;
1205
+ }, z.core.$strict>>;
1206
+ }, z.core.$strict>>;
1207
+ usage: z.ZodObject<{
1208
+ inputTokens: z.ZodNullable<z.ZodNumber>;
1209
+ outputTokens: z.ZodNullable<z.ZodNumber>;
1210
+ totalTokens: z.ZodNullable<z.ZodNumber>;
1211
+ costUsd: z.ZodNullable<z.ZodNumber>;
1212
+ latencyMs: z.ZodNullable<z.ZodNumber>;
1213
+ }, z.core.$strict>;
1214
+ traceRef: z.ZodNullable<z.ZodObject<{
1215
+ id: z.ZodString;
1216
+ contentHash: z.ZodString;
1217
+ mediaType: z.ZodDefault<z.ZodNullable<z.ZodString>>;
1218
+ sizeBytes: z.ZodDefault<z.ZodNullable<z.ZodNumber>>;
1219
+ }, z.core.$strict>>;
1220
+ startedAt: z.ZodNullable<z.ZodString>;
1221
+ completedAt: z.ZodNullable<z.ZodString>;
1222
+ }, z.core.$strict>>;
1223
+ completedAt: z.ZodString;
1224
+ contentHash: z.ZodString;
1225
+ }, z.core.$strict>;
1226
+ }, z.core.$strict>;
1227
+ contentHash: z.ZodString;
1228
+ }, z.core.$strict>;
1229
+ export type HumanComparisonProjection = z.infer<typeof HumanComparisonProjectionSchema>;
1230
+ type Attempt = z.infer<typeof HumanAttemptRefSchema>;
1231
+ type Input = {
1232
+ teamId: string;
1233
+ projectId: string;
1234
+ baseline: z.infer<typeof Evidence>;
1235
+ candidate: z.infer<typeof Evidence>;
1236
+ selections: HumanComparisonSelections;
1237
+ authorize(): Promise<void>;
1238
+ readReview(ref: z.infer<typeof ReviewReleaseSchema>): Promise<HumanReview>;
1239
+ caseIdentity(attempt: Attempt): Promise<ExperimentCaseIdentity>;
1240
+ };
1241
+ /** The host supplies current resource/subject owners and real retained receipt
1242
+ * identities. This creates another evidence view; it never changes a pass. */
1243
+ export declare function projectAcceptedHumanComparison(input: Input): Promise<HumanComparisonProjection>;
1244
+ export declare function verifyHumanComparisonProjection(raw: unknown, originals?: {
1245
+ baseline: z.infer<typeof Evidence>;
1246
+ candidate: z.infer<typeof Evidence>;
1247
+ }): {
1248
+ schemaVersion: "openpond.humanComparisonProjection.v1";
1249
+ teamId: string;
1250
+ projectId: string;
1251
+ selections: {
1252
+ rules: ({
1253
+ kind: "normalized_score";
1254
+ minimum: number;
1255
+ maximum: number;
1256
+ feedbackKey: string;
1257
+ criterionId: string;
1258
+ passThreshold?: number | undefined;
1259
+ } | {
1260
+ kind: "boolean";
1261
+ feedbackKey: string;
1262
+ criterionId: string;
1263
+ } | {
1264
+ kind: "category";
1265
+ feedbackKey: string;
1266
+ criterionId: string;
1267
+ } | {
1268
+ kind: "preference_score";
1269
+ tieScore: number | null;
1270
+ feedbackKey: string;
1271
+ criterionId: string;
1272
+ })[];
1273
+ baseline: {
1274
+ execution: {
1275
+ id: string;
1276
+ contentHash: string;
1277
+ };
1278
+ resultHash: string;
1279
+ reviews: {
1280
+ review: {
1281
+ id: string;
1282
+ revision: number;
1283
+ contentHash: string;
1284
+ };
1285
+ decision: {
1286
+ id: string;
1287
+ revision: number;
1288
+ contentHash: string;
1289
+ };
1290
+ }[];
1291
+ };
1292
+ candidate: {
1293
+ execution: {
1294
+ id: string;
1295
+ contentHash: string;
1296
+ };
1297
+ resultHash: string;
1298
+ reviews: {
1299
+ review: {
1300
+ id: string;
1301
+ revision: number;
1302
+ contentHash: string;
1303
+ };
1304
+ decision: {
1305
+ id: string;
1306
+ revision: number;
1307
+ contentHash: string;
1308
+ };
1309
+ }[];
1310
+ };
1311
+ };
1312
+ original: {
1313
+ baseline: {
1314
+ id: string;
1315
+ manifestHash: string;
1316
+ resultHash: string;
1317
+ };
1318
+ candidate: {
1319
+ id: string;
1320
+ manifestHash: string;
1321
+ resultHash: string;
1322
+ };
1323
+ };
1324
+ humanResults: {
1325
+ baseline: {
1326
+ schemaVersion: "openpond.humanResultView.v1";
1327
+ executionId: string;
1328
+ executionHash: string;
1329
+ selection: {
1330
+ id: string;
1331
+ revision: number;
1332
+ contentHash: string;
1333
+ }[];
1334
+ coverage: {
1335
+ pending: number;
1336
+ submitted: number;
1337
+ accepted: number;
1338
+ rejected: number;
1339
+ cancelled: number;
1340
+ };
1341
+ results: {
1342
+ review: {
1343
+ id: string;
1344
+ revision: number;
1345
+ contentHash: string;
1346
+ };
1347
+ decision: {
1348
+ id: string;
1349
+ revision: number;
1350
+ contentHash: string;
1351
+ } | null;
1352
+ attemptIds: string[];
1353
+ mode: "individual" | "pairwise";
1354
+ origin: "owner_attested_local" | "native_local" | "recorded_evidence" | "hosted_retained";
1355
+ status: "submitted" | "accepted" | "rejected" | "cancelled" | "pending";
1356
+ grader: {
1357
+ id: string;
1358
+ revision: number;
1359
+ contentHash: string;
1360
+ };
1361
+ answers: {
1362
+ criterionId: string;
1363
+ value: string | number | boolean | null;
1364
+ abstain: boolean;
1365
+ note: string;
1366
+ strength?: "slight" | "moderate" | "strong" | undefined;
1367
+ }[];
1368
+ }[];
1369
+ contentHash: string;
1370
+ };
1371
+ candidate: {
1372
+ schemaVersion: "openpond.humanResultView.v1";
1373
+ executionId: string;
1374
+ executionHash: string;
1375
+ selection: {
1376
+ id: string;
1377
+ revision: number;
1378
+ contentHash: string;
1379
+ }[];
1380
+ coverage: {
1381
+ pending: number;
1382
+ submitted: number;
1383
+ accepted: number;
1384
+ rejected: number;
1385
+ cancelled: number;
1386
+ };
1387
+ results: {
1388
+ review: {
1389
+ id: string;
1390
+ revision: number;
1391
+ contentHash: string;
1392
+ };
1393
+ decision: {
1394
+ id: string;
1395
+ revision: number;
1396
+ contentHash: string;
1397
+ } | null;
1398
+ attemptIds: string[];
1399
+ mode: "individual" | "pairwise";
1400
+ origin: "owner_attested_local" | "native_local" | "recorded_evidence" | "hosted_retained";
1401
+ status: "submitted" | "accepted" | "rejected" | "cancelled" | "pending";
1402
+ grader: {
1403
+ id: string;
1404
+ revision: number;
1405
+ contentHash: string;
1406
+ };
1407
+ answers: {
1408
+ criterionId: string;
1409
+ value: string | number | boolean | null;
1410
+ abstain: boolean;
1411
+ note: string;
1412
+ strength?: "slight" | "moderate" | "strong" | undefined;
1413
+ }[];
1414
+ }[];
1415
+ contentHash: string;
1416
+ };
1417
+ };
1418
+ baseline: {
1419
+ manifest: {
1420
+ schemaVersion: "openpond.experimentManifest.v1";
1421
+ id: string;
1422
+ name: string;
1423
+ teamId: string;
1424
+ operationId: string;
1425
+ maximumCostUsd: number | null;
1426
+ dataset: {
1427
+ id: string;
1428
+ contentHash: string;
1429
+ revision: number;
1430
+ };
1431
+ target: {
1432
+ kind: "model";
1433
+ modelId: string;
1434
+ configurationHash: string;
1435
+ } | {
1436
+ kind: "harness";
1437
+ source: {
1438
+ profileId: string;
1439
+ sourceRevision: string;
1440
+ harnessRelease: {
1441
+ id: string;
1442
+ contentHash: string;
1443
+ };
1444
+ catalogHash: string;
1445
+ definitionId: string;
1446
+ definitionHash: string;
1447
+ target: {
1448
+ kind: "profile";
1449
+ } | {
1450
+ kind: "workflow";
1451
+ workflowId: string;
1452
+ } | {
1453
+ kind: "skill";
1454
+ skillPath: string;
1455
+ } | {
1456
+ kind: "agent_action";
1457
+ actionId: string;
1458
+ };
1459
+ environmentHash: string;
1460
+ externalDatasetBinding?: {
1461
+ schemaVersion: "openpond.profileExternalDatasetBinding.v1";
1462
+ dataset: {
1463
+ id: string;
1464
+ contentHash: string;
1465
+ revision: number;
1466
+ };
1467
+ packageHash: string;
1468
+ profileId: string;
1469
+ target: {
1470
+ kind: "profile";
1471
+ } | {
1472
+ kind: "workflow";
1473
+ workflowId: string;
1474
+ } | {
1475
+ kind: "skill";
1476
+ skillPath: string;
1477
+ } | {
1478
+ kind: "agent_action";
1479
+ actionId: string;
1480
+ };
1481
+ declaredProfileCatalogHash: string | null;
1482
+ protectedProfileClosureHash: string;
1483
+ split: "train" | "validation" | "test" | "frozen_eval";
1484
+ population: {
1485
+ taskId: string;
1486
+ seed: string;
1487
+ fixtureId: string | null;
1488
+ }[];
1489
+ evaluators: {
1490
+ release: {
1491
+ id: string;
1492
+ contentHash: string;
1493
+ revision: string | number;
1494
+ };
1495
+ feedbackKey: string;
1496
+ configurationHash: string;
1497
+ output: "boolean" | "score" | "category";
1498
+ categories: string[];
1499
+ }[];
1500
+ fieldMappingsHash: string;
1501
+ fieldMappings: {
1502
+ graderId: string;
1503
+ fields: {
1504
+ destination: "output" | "input" | "expectedOutput" | "evaluatorContext";
1505
+ path: string[];
1506
+ source: "output" | "input" | "expectedOutput" | "evaluatorContext";
1507
+ sourcePath: string[];
1508
+ }[];
1509
+ }[];
1510
+ environmentHash: string;
1511
+ privateDatasetClosureHash: string;
1512
+ criterion: {
1513
+ minimumPassRate: number;
1514
+ requireComplete: boolean;
1515
+ };
1516
+ contentHash: string;
1517
+ recordedOrigin?: {
1518
+ execution: {
1519
+ id: string;
1520
+ contentHash: string;
1521
+ };
1522
+ originalEvidence: {
1523
+ id: string;
1524
+ contentHash: string;
1525
+ };
1526
+ members: {
1527
+ caseId: string;
1528
+ seed: string;
1529
+ fixtureId: string;
1530
+ sourceId: string;
1531
+ snapshotHash: string;
1532
+ boundaryRevisionHash: string;
1533
+ }[];
1534
+ } | undefined;
1535
+ gradingSource?: {
1536
+ pass: {
1537
+ id: string;
1538
+ contentHash: string;
1539
+ };
1540
+ evidence: {
1541
+ id: string;
1542
+ contentHash: string;
1543
+ };
1544
+ } | undefined;
1545
+ } | undefined;
1546
+ sourceCandidate?: {
1547
+ candidateId: string;
1548
+ frozenRevision: number;
1549
+ frozenHash: string;
1550
+ harnessRelease: {
1551
+ id: string;
1552
+ contentHash: string;
1553
+ };
1554
+ protectedClosureHash: string;
1555
+ } | undefined;
1556
+ };
1557
+ model: {
1558
+ modelId: string;
1559
+ configurationHash: string;
1560
+ };
1561
+ } | {
1562
+ kind: "agent";
1563
+ source: {
1564
+ harnessRelease: {
1565
+ id: string;
1566
+ contentHash: string;
1567
+ };
1568
+ agentSnapshot: {
1569
+ id: string;
1570
+ contentHash: string;
1571
+ };
1572
+ sourcePackageHash: string;
1573
+ };
1574
+ model: {
1575
+ modelId: string;
1576
+ configurationHash: string;
1577
+ };
1578
+ } | {
1579
+ kind: "fixture";
1580
+ configurationHash: string;
1581
+ };
1582
+ evaluators: {
1583
+ release: {
1584
+ id: string;
1585
+ contentHash: string;
1586
+ revision: string | number;
1587
+ };
1588
+ feedbackKey: string;
1589
+ output: "boolean" | "score" | "category";
1590
+ categories: string[];
1591
+ configurationHash?: string | undefined;
1592
+ }[];
1593
+ population: {
1594
+ caseId: string;
1595
+ seed: string;
1596
+ fixtureId: string | null;
1597
+ }[];
1598
+ createdAt: string;
1599
+ contentHash: string;
1600
+ execution?: {
1601
+ packageHash: string;
1602
+ runtimeTargetHash: string;
1603
+ metricPolicyHash: string;
1604
+ compatibility?: {
1605
+ protocol: "openpond.evaluation-execution.v1";
1606
+ targetKind: "agent" | "model" | "fixture" | "harness";
1607
+ } | undefined;
1608
+ } | undefined;
1609
+ lineage?: {
1610
+ definition: {
1611
+ id: string;
1612
+ contentHash: string;
1613
+ revision: number;
1614
+ } | null;
1615
+ execution: {
1616
+ id: string;
1617
+ contentHash: string;
1618
+ };
1619
+ scoringPassId: string | null;
1620
+ humanProjection?: {
1621
+ originalManifest: {
1622
+ id: string;
1623
+ contentHash: string;
1624
+ };
1625
+ selectionHash: string;
1626
+ rulesHash: string;
1627
+ } | undefined;
1628
+ } | undefined;
1629
+ };
1630
+ result: {
1631
+ schemaVersion: "openpond.experimentResult.v1";
1632
+ manifest: {
1633
+ id: string;
1634
+ contentHash: string;
1635
+ };
1636
+ status: "cancelled" | "failed" | "completed";
1637
+ cases: {
1638
+ identity: {
1639
+ caseId: string;
1640
+ seed: string;
1641
+ fixtureId: string | null;
1642
+ };
1643
+ status: "cancelled" | "failed" | "completed" | "unavailable";
1644
+ output: unknown;
1645
+ error: {
1646
+ code: string;
1647
+ message: string;
1648
+ } | null;
1649
+ feedback: {
1650
+ feedbackKey: string;
1651
+ evaluator: {
1652
+ id: string;
1653
+ contentHash: string;
1654
+ revision: string | number;
1655
+ };
1656
+ status: "pending" | "failed" | "scored" | "unavailable";
1657
+ value: string | number | boolean | null;
1658
+ reasoning: string | null;
1659
+ evidenceRefs: {
1660
+ id: string;
1661
+ contentHash: string;
1662
+ mediaType: string | null;
1663
+ sizeBytes: number | null;
1664
+ }[];
1665
+ passed?: boolean | null | undefined;
1666
+ }[];
1667
+ usage: {
1668
+ inputTokens: number | null;
1669
+ outputTokens: number | null;
1670
+ totalTokens: number | null;
1671
+ costUsd: number | null;
1672
+ latencyMs: number | null;
1673
+ };
1674
+ traceRef: {
1675
+ id: string;
1676
+ contentHash: string;
1677
+ mediaType: string | null;
1678
+ sizeBytes: number | null;
1679
+ } | null;
1680
+ startedAt: string | null;
1681
+ completedAt: string | null;
1682
+ }[];
1683
+ completedAt: string;
1684
+ contentHash: string;
1685
+ };
1686
+ };
1687
+ candidate: {
1688
+ manifest: {
1689
+ schemaVersion: "openpond.experimentManifest.v1";
1690
+ id: string;
1691
+ name: string;
1692
+ teamId: string;
1693
+ operationId: string;
1694
+ maximumCostUsd: number | null;
1695
+ dataset: {
1696
+ id: string;
1697
+ contentHash: string;
1698
+ revision: number;
1699
+ };
1700
+ target: {
1701
+ kind: "model";
1702
+ modelId: string;
1703
+ configurationHash: string;
1704
+ } | {
1705
+ kind: "harness";
1706
+ source: {
1707
+ profileId: string;
1708
+ sourceRevision: string;
1709
+ harnessRelease: {
1710
+ id: string;
1711
+ contentHash: string;
1712
+ };
1713
+ catalogHash: string;
1714
+ definitionId: string;
1715
+ definitionHash: string;
1716
+ target: {
1717
+ kind: "profile";
1718
+ } | {
1719
+ kind: "workflow";
1720
+ workflowId: string;
1721
+ } | {
1722
+ kind: "skill";
1723
+ skillPath: string;
1724
+ } | {
1725
+ kind: "agent_action";
1726
+ actionId: string;
1727
+ };
1728
+ environmentHash: string;
1729
+ externalDatasetBinding?: {
1730
+ schemaVersion: "openpond.profileExternalDatasetBinding.v1";
1731
+ dataset: {
1732
+ id: string;
1733
+ contentHash: string;
1734
+ revision: number;
1735
+ };
1736
+ packageHash: string;
1737
+ profileId: string;
1738
+ target: {
1739
+ kind: "profile";
1740
+ } | {
1741
+ kind: "workflow";
1742
+ workflowId: string;
1743
+ } | {
1744
+ kind: "skill";
1745
+ skillPath: string;
1746
+ } | {
1747
+ kind: "agent_action";
1748
+ actionId: string;
1749
+ };
1750
+ declaredProfileCatalogHash: string | null;
1751
+ protectedProfileClosureHash: string;
1752
+ split: "train" | "validation" | "test" | "frozen_eval";
1753
+ population: {
1754
+ taskId: string;
1755
+ seed: string;
1756
+ fixtureId: string | null;
1757
+ }[];
1758
+ evaluators: {
1759
+ release: {
1760
+ id: string;
1761
+ contentHash: string;
1762
+ revision: string | number;
1763
+ };
1764
+ feedbackKey: string;
1765
+ configurationHash: string;
1766
+ output: "boolean" | "score" | "category";
1767
+ categories: string[];
1768
+ }[];
1769
+ fieldMappingsHash: string;
1770
+ fieldMappings: {
1771
+ graderId: string;
1772
+ fields: {
1773
+ destination: "output" | "input" | "expectedOutput" | "evaluatorContext";
1774
+ path: string[];
1775
+ source: "output" | "input" | "expectedOutput" | "evaluatorContext";
1776
+ sourcePath: string[];
1777
+ }[];
1778
+ }[];
1779
+ environmentHash: string;
1780
+ privateDatasetClosureHash: string;
1781
+ criterion: {
1782
+ minimumPassRate: number;
1783
+ requireComplete: boolean;
1784
+ };
1785
+ contentHash: string;
1786
+ recordedOrigin?: {
1787
+ execution: {
1788
+ id: string;
1789
+ contentHash: string;
1790
+ };
1791
+ originalEvidence: {
1792
+ id: string;
1793
+ contentHash: string;
1794
+ };
1795
+ members: {
1796
+ caseId: string;
1797
+ seed: string;
1798
+ fixtureId: string;
1799
+ sourceId: string;
1800
+ snapshotHash: string;
1801
+ boundaryRevisionHash: string;
1802
+ }[];
1803
+ } | undefined;
1804
+ gradingSource?: {
1805
+ pass: {
1806
+ id: string;
1807
+ contentHash: string;
1808
+ };
1809
+ evidence: {
1810
+ id: string;
1811
+ contentHash: string;
1812
+ };
1813
+ } | undefined;
1814
+ } | undefined;
1815
+ sourceCandidate?: {
1816
+ candidateId: string;
1817
+ frozenRevision: number;
1818
+ frozenHash: string;
1819
+ harnessRelease: {
1820
+ id: string;
1821
+ contentHash: string;
1822
+ };
1823
+ protectedClosureHash: string;
1824
+ } | undefined;
1825
+ };
1826
+ model: {
1827
+ modelId: string;
1828
+ configurationHash: string;
1829
+ };
1830
+ } | {
1831
+ kind: "agent";
1832
+ source: {
1833
+ harnessRelease: {
1834
+ id: string;
1835
+ contentHash: string;
1836
+ };
1837
+ agentSnapshot: {
1838
+ id: string;
1839
+ contentHash: string;
1840
+ };
1841
+ sourcePackageHash: string;
1842
+ };
1843
+ model: {
1844
+ modelId: string;
1845
+ configurationHash: string;
1846
+ };
1847
+ } | {
1848
+ kind: "fixture";
1849
+ configurationHash: string;
1850
+ };
1851
+ evaluators: {
1852
+ release: {
1853
+ id: string;
1854
+ contentHash: string;
1855
+ revision: string | number;
1856
+ };
1857
+ feedbackKey: string;
1858
+ output: "boolean" | "score" | "category";
1859
+ categories: string[];
1860
+ configurationHash?: string | undefined;
1861
+ }[];
1862
+ population: {
1863
+ caseId: string;
1864
+ seed: string;
1865
+ fixtureId: string | null;
1866
+ }[];
1867
+ createdAt: string;
1868
+ contentHash: string;
1869
+ execution?: {
1870
+ packageHash: string;
1871
+ runtimeTargetHash: string;
1872
+ metricPolicyHash: string;
1873
+ compatibility?: {
1874
+ protocol: "openpond.evaluation-execution.v1";
1875
+ targetKind: "agent" | "model" | "fixture" | "harness";
1876
+ } | undefined;
1877
+ } | undefined;
1878
+ lineage?: {
1879
+ definition: {
1880
+ id: string;
1881
+ contentHash: string;
1882
+ revision: number;
1883
+ } | null;
1884
+ execution: {
1885
+ id: string;
1886
+ contentHash: string;
1887
+ };
1888
+ scoringPassId: string | null;
1889
+ humanProjection?: {
1890
+ originalManifest: {
1891
+ id: string;
1892
+ contentHash: string;
1893
+ };
1894
+ selectionHash: string;
1895
+ rulesHash: string;
1896
+ } | undefined;
1897
+ } | undefined;
1898
+ };
1899
+ result: {
1900
+ schemaVersion: "openpond.experimentResult.v1";
1901
+ manifest: {
1902
+ id: string;
1903
+ contentHash: string;
1904
+ };
1905
+ status: "cancelled" | "failed" | "completed";
1906
+ cases: {
1907
+ identity: {
1908
+ caseId: string;
1909
+ seed: string;
1910
+ fixtureId: string | null;
1911
+ };
1912
+ status: "cancelled" | "failed" | "completed" | "unavailable";
1913
+ output: unknown;
1914
+ error: {
1915
+ code: string;
1916
+ message: string;
1917
+ } | null;
1918
+ feedback: {
1919
+ feedbackKey: string;
1920
+ evaluator: {
1921
+ id: string;
1922
+ contentHash: string;
1923
+ revision: string | number;
1924
+ };
1925
+ status: "pending" | "failed" | "scored" | "unavailable";
1926
+ value: string | number | boolean | null;
1927
+ reasoning: string | null;
1928
+ evidenceRefs: {
1929
+ id: string;
1930
+ contentHash: string;
1931
+ mediaType: string | null;
1932
+ sizeBytes: number | null;
1933
+ }[];
1934
+ passed?: boolean | null | undefined;
1935
+ }[];
1936
+ usage: {
1937
+ inputTokens: number | null;
1938
+ outputTokens: number | null;
1939
+ totalTokens: number | null;
1940
+ costUsd: number | null;
1941
+ latencyMs: number | null;
1942
+ };
1943
+ traceRef: {
1944
+ id: string;
1945
+ contentHash: string;
1946
+ mediaType: string | null;
1947
+ sizeBytes: number | null;
1948
+ } | null;
1949
+ startedAt: string | null;
1950
+ completedAt: string | null;
1951
+ }[];
1952
+ completedAt: string;
1953
+ contentHash: string;
1954
+ };
1955
+ };
1956
+ contentHash: string;
1957
+ };
1958
+ export {};
1959
+ //# sourceMappingURL=comparison.d.ts.map