@openpond/evals 0.11.16 → 0.11.17

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 +9 -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 +1,878 @@
1
+ import { type HumanReview, type HumanReviewCommand, type HumanReviewRequest } from "./contracts.js";
2
+ import type { HumanReviewAuthority, HumanReviewContext, HumanReviewRepository } from "./repository.js";
3
+ export declare function createHumanReviewService(repository: HumanReviewRepository, authority: HumanReviewAuthority, now?: () => string): {
4
+ command: (context: HumanReviewContext, raw: HumanReviewCommand) => Promise<HumanReview>;
5
+ get: (context: HumanReviewContext, id: string, revision?: number) => Promise<{
6
+ schemaVersion: "openpond.humanReview.v1";
7
+ id: string;
8
+ revision: number;
9
+ scope: string;
10
+ projectId: string;
11
+ createdBy: string;
12
+ kind: "author" | "execute" | "grade";
13
+ title: string;
14
+ authorPublication: {
15
+ proposalHash: string;
16
+ dataset: {
17
+ id: string;
18
+ revision: number;
19
+ contentHash: string;
20
+ };
21
+ operationId: string;
22
+ publishedBy: string;
23
+ publishedAt: string;
24
+ } | null;
25
+ evidence: {
26
+ dataset: {
27
+ id: string;
28
+ revision: number;
29
+ contentHash: string;
30
+ };
31
+ grader: {
32
+ id: string;
33
+ revision: number;
34
+ contentHash: string;
35
+ };
36
+ graderBinding: {
37
+ id: string;
38
+ version: string;
39
+ contentHash: string;
40
+ };
41
+ rubric: {
42
+ id: string;
43
+ revision: number;
44
+ contentHash: string;
45
+ };
46
+ attempts: {
47
+ experimentId: string;
48
+ attemptId: string;
49
+ taskId: string;
50
+ targetId: string;
51
+ output: {
52
+ id: string;
53
+ revision: number;
54
+ contentHash: string;
55
+ };
56
+ trace: {
57
+ id: string;
58
+ revision: number;
59
+ contentHash: string;
60
+ } | null;
61
+ familyKey: string;
62
+ split: "train" | "validation" | "test" | "frozen_eval";
63
+ localSource?: {
64
+ kind: "owner_attested_local";
65
+ ownerActorId: string;
66
+ teamId: string;
67
+ executionHash: string;
68
+ configurationHash: string;
69
+ retainedAttemptHash: string;
70
+ traceHash: string;
71
+ publicationCutoffHash: string;
72
+ } | {
73
+ ownerActorId: string;
74
+ teamId: string;
75
+ executionHash: string;
76
+ configurationHash: string;
77
+ retainedAttemptHash: string;
78
+ traceHash: string;
79
+ kind: "native_local";
80
+ } | undefined;
81
+ recordedSource?: {
82
+ id: string;
83
+ snapshotHash: string;
84
+ boundaryId: string;
85
+ boundaryRevisionHash: string;
86
+ } | undefined;
87
+ }[];
88
+ publication?: {
89
+ id: string;
90
+ revision: number;
91
+ contentHash: string;
92
+ } | undefined;
93
+ taskIds?: string[] | undefined;
94
+ };
95
+ form: {
96
+ schemaVersion: "openpond.humanForm.v1";
97
+ mode: "individual" | "pairwise";
98
+ instructions: string;
99
+ criteria: ({
100
+ id: string;
101
+ label: string;
102
+ instructions: string;
103
+ required: boolean;
104
+ allowAbstain: boolean;
105
+ kind: "score";
106
+ minimum: number;
107
+ maximum: number;
108
+ step: number;
109
+ requireRationale?: boolean | undefined;
110
+ anchors?: {
111
+ value: string | number | boolean;
112
+ description: string;
113
+ }[] | undefined;
114
+ } | {
115
+ id: string;
116
+ label: string;
117
+ instructions: string;
118
+ required: boolean;
119
+ allowAbstain: boolean;
120
+ kind: "category";
121
+ options: string[];
122
+ requireRationale?: boolean | undefined;
123
+ anchors?: {
124
+ value: string | number | boolean;
125
+ description: string;
126
+ }[] | undefined;
127
+ } | {
128
+ id: string;
129
+ label: string;
130
+ instructions: string;
131
+ required: boolean;
132
+ allowAbstain: boolean;
133
+ kind: "boolean";
134
+ requireRationale?: boolean | undefined;
135
+ anchors?: {
136
+ value: string | number | boolean;
137
+ description: string;
138
+ }[] | undefined;
139
+ } | {
140
+ id: string;
141
+ label: string;
142
+ instructions: string;
143
+ required: boolean;
144
+ allowAbstain: boolean;
145
+ kind: "text";
146
+ maximumLength: number;
147
+ requireRationale?: boolean | undefined;
148
+ anchors?: {
149
+ value: string | number | boolean;
150
+ description: string;
151
+ }[] | undefined;
152
+ minimumLength?: number | undefined;
153
+ } | {
154
+ id: string;
155
+ label: string;
156
+ instructions: string;
157
+ required: boolean;
158
+ allowAbstain: boolean;
159
+ kind: "preference";
160
+ allowTie: boolean;
161
+ requireRationale?: boolean | undefined;
162
+ anchors?: {
163
+ value: string | number | boolean;
164
+ description: string;
165
+ }[] | undefined;
166
+ requireStrength?: boolean | undefined;
167
+ })[];
168
+ };
169
+ policy: {
170
+ approval: "none" | "manager";
171
+ minimumRaters: number;
172
+ managerIds: string[];
173
+ reviewerIds: string[];
174
+ queueClaim: boolean;
175
+ teamVisible: boolean;
176
+ distinctManager?: boolean | undefined;
177
+ independentRaters?: boolean | undefined;
178
+ blindPair?: boolean | undefined;
179
+ };
180
+ status: "queued" | "assigned" | "in_progress" | "submitted" | "returned" | "accepted" | "rejected" | "cancelled";
181
+ assigneeId: string | null;
182
+ generation: number;
183
+ drafts: {
184
+ actorId: string;
185
+ generation: number;
186
+ answers: {
187
+ criterionId: string;
188
+ value: string | number | boolean | null;
189
+ abstain: boolean;
190
+ note: string;
191
+ strength?: "slight" | "moderate" | "strong" | undefined;
192
+ }[];
193
+ note: string;
194
+ updatedAt: string;
195
+ taskProposal?: {
196
+ schemaVersion: "openpond.humanTaskProposal.v1";
197
+ tasks: {
198
+ taskId: string;
199
+ input: Record<string, unknown>;
200
+ policyVisibleContext: Record<string, unknown>;
201
+ expectedOutput?: Record<string, unknown> | null | undefined;
202
+ tags?: string[] | undefined;
203
+ }[];
204
+ contentHash: string;
205
+ } | undefined;
206
+ }[];
207
+ submissions: {
208
+ id: string;
209
+ revision: number;
210
+ actorId: string;
211
+ generation: number;
212
+ evidenceHash: string;
213
+ formHash: string;
214
+ answers: {
215
+ criterionId: string;
216
+ value: string | number | boolean | null;
217
+ abstain: boolean;
218
+ note: string;
219
+ strength?: "slight" | "moderate" | "strong" | undefined;
220
+ }[];
221
+ note: string;
222
+ submittedAt: string;
223
+ supersedes: string | null;
224
+ contentHash: string;
225
+ taskProposal?: {
226
+ schemaVersion: "openpond.humanTaskProposal.v1";
227
+ tasks: {
228
+ taskId: string;
229
+ input: Record<string, unknown>;
230
+ policyVisibleContext: Record<string, unknown>;
231
+ expectedOutput?: Record<string, unknown> | null | undefined;
232
+ tags?: string[] | undefined;
233
+ }[];
234
+ contentHash: string;
235
+ } | undefined;
236
+ }[];
237
+ decisions: {
238
+ id: string;
239
+ revision: number;
240
+ actorId: string;
241
+ outcome: "accept" | "reject" | "request_changes";
242
+ submissions: {
243
+ id: string;
244
+ contentHash: string;
245
+ }[];
246
+ answers: {
247
+ criterionId: string;
248
+ value: string | number | boolean | null;
249
+ abstain: boolean;
250
+ note: string;
251
+ strength?: "slight" | "moderate" | "strong" | undefined;
252
+ }[];
253
+ note: string;
254
+ decidedAt: string;
255
+ contentHash: string;
256
+ taskProposalHash?: string | undefined;
257
+ }[];
258
+ audit: {
259
+ actorId: string;
260
+ action: "claim" | "reassign" | "cancel";
261
+ priorAssigneeId: string | null;
262
+ assigneeId: string | null;
263
+ generation: number;
264
+ note: string;
265
+ occurredAt: string;
266
+ }[];
267
+ execution: {
268
+ executionId: string | null;
269
+ operationId: string;
270
+ boundBy: string;
271
+ configurationHash: string;
272
+ generation: number;
273
+ result: {
274
+ id: string;
275
+ revision: number;
276
+ contentHash: string;
277
+ } | null;
278
+ } | null;
279
+ createdAt: string;
280
+ updatedAt: string;
281
+ contentHash: string;
282
+ dueAt?: string | undefined;
283
+ }>;
284
+ inbox: (context: HumanReviewContext, query: Extract<HumanReviewRequest, {
285
+ endpoint: "inbox";
286
+ }>) => Promise<{
287
+ items: {
288
+ schemaVersion: "openpond.humanReview.v1";
289
+ id: string;
290
+ revision: number;
291
+ scope: string;
292
+ projectId: string;
293
+ createdBy: string;
294
+ kind: "author" | "execute" | "grade";
295
+ title: string;
296
+ authorPublication: {
297
+ proposalHash: string;
298
+ dataset: {
299
+ id: string;
300
+ revision: number;
301
+ contentHash: string;
302
+ };
303
+ operationId: string;
304
+ publishedBy: string;
305
+ publishedAt: string;
306
+ } | null;
307
+ evidence: {
308
+ dataset: {
309
+ id: string;
310
+ revision: number;
311
+ contentHash: string;
312
+ };
313
+ grader: {
314
+ id: string;
315
+ revision: number;
316
+ contentHash: string;
317
+ };
318
+ graderBinding: {
319
+ id: string;
320
+ version: string;
321
+ contentHash: string;
322
+ };
323
+ rubric: {
324
+ id: string;
325
+ revision: number;
326
+ contentHash: string;
327
+ };
328
+ attempts: {
329
+ experimentId: string;
330
+ attemptId: string;
331
+ taskId: string;
332
+ targetId: string;
333
+ output: {
334
+ id: string;
335
+ revision: number;
336
+ contentHash: string;
337
+ };
338
+ trace: {
339
+ id: string;
340
+ revision: number;
341
+ contentHash: string;
342
+ } | null;
343
+ familyKey: string;
344
+ split: "train" | "validation" | "test" | "frozen_eval";
345
+ localSource?: {
346
+ kind: "owner_attested_local";
347
+ ownerActorId: string;
348
+ teamId: string;
349
+ executionHash: string;
350
+ configurationHash: string;
351
+ retainedAttemptHash: string;
352
+ traceHash: string;
353
+ publicationCutoffHash: string;
354
+ } | {
355
+ ownerActorId: string;
356
+ teamId: string;
357
+ executionHash: string;
358
+ configurationHash: string;
359
+ retainedAttemptHash: string;
360
+ traceHash: string;
361
+ kind: "native_local";
362
+ } | undefined;
363
+ recordedSource?: {
364
+ id: string;
365
+ snapshotHash: string;
366
+ boundaryId: string;
367
+ boundaryRevisionHash: string;
368
+ } | undefined;
369
+ }[];
370
+ publication?: {
371
+ id: string;
372
+ revision: number;
373
+ contentHash: string;
374
+ } | undefined;
375
+ taskIds?: string[] | undefined;
376
+ };
377
+ form: {
378
+ schemaVersion: "openpond.humanForm.v1";
379
+ mode: "individual" | "pairwise";
380
+ instructions: string;
381
+ criteria: ({
382
+ id: string;
383
+ label: string;
384
+ instructions: string;
385
+ required: boolean;
386
+ allowAbstain: boolean;
387
+ kind: "score";
388
+ minimum: number;
389
+ maximum: number;
390
+ step: number;
391
+ requireRationale?: boolean | undefined;
392
+ anchors?: {
393
+ value: string | number | boolean;
394
+ description: string;
395
+ }[] | undefined;
396
+ } | {
397
+ id: string;
398
+ label: string;
399
+ instructions: string;
400
+ required: boolean;
401
+ allowAbstain: boolean;
402
+ kind: "category";
403
+ options: string[];
404
+ requireRationale?: boolean | undefined;
405
+ anchors?: {
406
+ value: string | number | boolean;
407
+ description: string;
408
+ }[] | undefined;
409
+ } | {
410
+ id: string;
411
+ label: string;
412
+ instructions: string;
413
+ required: boolean;
414
+ allowAbstain: boolean;
415
+ kind: "boolean";
416
+ requireRationale?: boolean | undefined;
417
+ anchors?: {
418
+ value: string | number | boolean;
419
+ description: string;
420
+ }[] | undefined;
421
+ } | {
422
+ id: string;
423
+ label: string;
424
+ instructions: string;
425
+ required: boolean;
426
+ allowAbstain: boolean;
427
+ kind: "text";
428
+ maximumLength: number;
429
+ requireRationale?: boolean | undefined;
430
+ anchors?: {
431
+ value: string | number | boolean;
432
+ description: string;
433
+ }[] | undefined;
434
+ minimumLength?: number | undefined;
435
+ } | {
436
+ id: string;
437
+ label: string;
438
+ instructions: string;
439
+ required: boolean;
440
+ allowAbstain: boolean;
441
+ kind: "preference";
442
+ allowTie: boolean;
443
+ requireRationale?: boolean | undefined;
444
+ anchors?: {
445
+ value: string | number | boolean;
446
+ description: string;
447
+ }[] | undefined;
448
+ requireStrength?: boolean | undefined;
449
+ })[];
450
+ };
451
+ policy: {
452
+ approval: "none" | "manager";
453
+ minimumRaters: number;
454
+ managerIds: string[];
455
+ reviewerIds: string[];
456
+ queueClaim: boolean;
457
+ teamVisible: boolean;
458
+ distinctManager?: boolean | undefined;
459
+ independentRaters?: boolean | undefined;
460
+ blindPair?: boolean | undefined;
461
+ };
462
+ status: "queued" | "assigned" | "in_progress" | "submitted" | "returned" | "accepted" | "rejected" | "cancelled";
463
+ assigneeId: string | null;
464
+ generation: number;
465
+ drafts: {
466
+ actorId: string;
467
+ generation: number;
468
+ answers: {
469
+ criterionId: string;
470
+ value: string | number | boolean | null;
471
+ abstain: boolean;
472
+ note: string;
473
+ strength?: "slight" | "moderate" | "strong" | undefined;
474
+ }[];
475
+ note: string;
476
+ updatedAt: string;
477
+ taskProposal?: {
478
+ schemaVersion: "openpond.humanTaskProposal.v1";
479
+ tasks: {
480
+ taskId: string;
481
+ input: Record<string, unknown>;
482
+ policyVisibleContext: Record<string, unknown>;
483
+ expectedOutput?: Record<string, unknown> | null | undefined;
484
+ tags?: string[] | undefined;
485
+ }[];
486
+ contentHash: string;
487
+ } | undefined;
488
+ }[];
489
+ submissions: {
490
+ id: string;
491
+ revision: number;
492
+ actorId: string;
493
+ generation: number;
494
+ evidenceHash: string;
495
+ formHash: string;
496
+ answers: {
497
+ criterionId: string;
498
+ value: string | number | boolean | null;
499
+ abstain: boolean;
500
+ note: string;
501
+ strength?: "slight" | "moderate" | "strong" | undefined;
502
+ }[];
503
+ note: string;
504
+ submittedAt: string;
505
+ supersedes: string | null;
506
+ contentHash: string;
507
+ taskProposal?: {
508
+ schemaVersion: "openpond.humanTaskProposal.v1";
509
+ tasks: {
510
+ taskId: string;
511
+ input: Record<string, unknown>;
512
+ policyVisibleContext: Record<string, unknown>;
513
+ expectedOutput?: Record<string, unknown> | null | undefined;
514
+ tags?: string[] | undefined;
515
+ }[];
516
+ contentHash: string;
517
+ } | undefined;
518
+ }[];
519
+ decisions: {
520
+ id: string;
521
+ revision: number;
522
+ actorId: string;
523
+ outcome: "accept" | "reject" | "request_changes";
524
+ submissions: {
525
+ id: string;
526
+ contentHash: string;
527
+ }[];
528
+ answers: {
529
+ criterionId: string;
530
+ value: string | number | boolean | null;
531
+ abstain: boolean;
532
+ note: string;
533
+ strength?: "slight" | "moderate" | "strong" | undefined;
534
+ }[];
535
+ note: string;
536
+ decidedAt: string;
537
+ contentHash: string;
538
+ taskProposalHash?: string | undefined;
539
+ }[];
540
+ audit: {
541
+ actorId: string;
542
+ action: "claim" | "reassign" | "cancel";
543
+ priorAssigneeId: string | null;
544
+ assigneeId: string | null;
545
+ generation: number;
546
+ note: string;
547
+ occurredAt: string;
548
+ }[];
549
+ execution: {
550
+ executionId: string | null;
551
+ operationId: string;
552
+ boundBy: string;
553
+ configurationHash: string;
554
+ generation: number;
555
+ result: {
556
+ id: string;
557
+ revision: number;
558
+ contentHash: string;
559
+ } | null;
560
+ } | null;
561
+ createdAt: string;
562
+ updatedAt: string;
563
+ contentHash: string;
564
+ dueAt?: string | undefined;
565
+ }[];
566
+ nextCursor: string | null;
567
+ actionCount: number;
568
+ approvalAvailable: boolean;
569
+ }>;
570
+ view: (context: HumanReviewContext, record: HumanReview) => Promise<{
571
+ evidence: {
572
+ dataset: {
573
+ id: string;
574
+ revision: number;
575
+ contentHash: string;
576
+ };
577
+ grader: {
578
+ id: string;
579
+ revision: number;
580
+ contentHash: string;
581
+ };
582
+ graderBinding: {
583
+ id: string;
584
+ version: string;
585
+ contentHash: string;
586
+ };
587
+ rubric: {
588
+ id: string;
589
+ revision: number;
590
+ contentHash: string;
591
+ };
592
+ attempts: {
593
+ experimentId: string;
594
+ attemptId: string;
595
+ taskId: string;
596
+ targetId: string;
597
+ output: {
598
+ id: string;
599
+ revision: number;
600
+ contentHash: string;
601
+ };
602
+ trace: {
603
+ id: string;
604
+ revision: number;
605
+ contentHash: string;
606
+ } | null;
607
+ familyKey: string;
608
+ split: "train" | "validation" | "test" | "frozen_eval";
609
+ localSource?: {
610
+ kind: "owner_attested_local";
611
+ ownerActorId: string;
612
+ teamId: string;
613
+ executionHash: string;
614
+ configurationHash: string;
615
+ retainedAttemptHash: string;
616
+ traceHash: string;
617
+ publicationCutoffHash: string;
618
+ } | {
619
+ ownerActorId: string;
620
+ teamId: string;
621
+ executionHash: string;
622
+ configurationHash: string;
623
+ retainedAttemptHash: string;
624
+ traceHash: string;
625
+ kind: "native_local";
626
+ } | undefined;
627
+ recordedSource?: {
628
+ id: string;
629
+ snapshotHash: string;
630
+ boundaryId: string;
631
+ boundaryRevisionHash: string;
632
+ } | undefined;
633
+ }[];
634
+ publication?: {
635
+ id: string;
636
+ revision: number;
637
+ contentHash: string;
638
+ } | undefined;
639
+ taskIds?: string[] | undefined;
640
+ } | null;
641
+ evidenceHash: string;
642
+ formHash: string;
643
+ drafts: {
644
+ taskProposal: {
645
+ schemaVersion: "openpond.humanTaskProposal.v1";
646
+ tasks: {
647
+ taskId: string;
648
+ input: Record<string, unknown>;
649
+ policyVisibleContext: Record<string, unknown>;
650
+ expectedOutput?: Record<string, unknown> | null | undefined;
651
+ tags?: string[] | undefined;
652
+ }[];
653
+ contentHash: string;
654
+ } | {
655
+ referenceRedacted: true;
656
+ tasks: {
657
+ taskId: string;
658
+ input: Record<string, unknown>;
659
+ policyVisibleContext: Record<string, unknown>;
660
+ tags?: string[] | undefined;
661
+ }[];
662
+ schemaVersion: "openpond.humanTaskProposal.v1";
663
+ contentHash: string;
664
+ } | undefined;
665
+ actorId: string;
666
+ generation: number;
667
+ answers: {
668
+ criterionId: string;
669
+ value: string | number | boolean | null;
670
+ abstain: boolean;
671
+ note: string;
672
+ strength?: "slight" | "moderate" | "strong" | undefined;
673
+ }[];
674
+ note: string;
675
+ updatedAt: string;
676
+ }[];
677
+ submissions: {
678
+ taskProposal: {
679
+ schemaVersion: "openpond.humanTaskProposal.v1";
680
+ tasks: {
681
+ taskId: string;
682
+ input: Record<string, unknown>;
683
+ policyVisibleContext: Record<string, unknown>;
684
+ expectedOutput?: Record<string, unknown> | null | undefined;
685
+ tags?: string[] | undefined;
686
+ }[];
687
+ contentHash: string;
688
+ } | {
689
+ referenceRedacted: true;
690
+ tasks: {
691
+ taskId: string;
692
+ input: Record<string, unknown>;
693
+ policyVisibleContext: Record<string, unknown>;
694
+ tags?: string[] | undefined;
695
+ }[];
696
+ schemaVersion: "openpond.humanTaskProposal.v1";
697
+ contentHash: string;
698
+ } | undefined;
699
+ id: string;
700
+ revision: number;
701
+ actorId: string;
702
+ generation: number;
703
+ evidenceHash: string;
704
+ formHash: string;
705
+ answers: {
706
+ criterionId: string;
707
+ value: string | number | boolean | null;
708
+ abstain: boolean;
709
+ note: string;
710
+ strength?: "slight" | "moderate" | "strong" | undefined;
711
+ }[];
712
+ note: string;
713
+ submittedAt: string;
714
+ supersedes: string | null;
715
+ contentHash: string;
716
+ }[];
717
+ decisions: {
718
+ id: string;
719
+ revision: number;
720
+ actorId: string;
721
+ outcome: "accept" | "reject" | "request_changes";
722
+ submissions: {
723
+ id: string;
724
+ contentHash: string;
725
+ }[];
726
+ answers: {
727
+ criterionId: string;
728
+ value: string | number | boolean | null;
729
+ abstain: boolean;
730
+ note: string;
731
+ strength?: "slight" | "moderate" | "strong" | undefined;
732
+ }[];
733
+ note: string;
734
+ decidedAt: string;
735
+ contentHash: string;
736
+ taskProposalHash?: string | undefined;
737
+ }[];
738
+ audit: {
739
+ actorId: string;
740
+ action: "claim" | "reassign" | "cancel";
741
+ priorAssigneeId: string | null;
742
+ assigneeId: string | null;
743
+ generation: number;
744
+ note: string;
745
+ occurredAt: string;
746
+ }[];
747
+ createdBy: string;
748
+ permissions: {
749
+ claim: boolean;
750
+ owner: boolean;
751
+ manager: boolean;
752
+ reviewer: boolean;
753
+ };
754
+ schemaVersion: "openpond.humanReview.v1";
755
+ id: string;
756
+ revision: number;
757
+ scope: string;
758
+ projectId: string;
759
+ kind: "author" | "execute" | "grade";
760
+ title: string;
761
+ authorPublication: {
762
+ proposalHash: string;
763
+ dataset: {
764
+ id: string;
765
+ revision: number;
766
+ contentHash: string;
767
+ };
768
+ operationId: string;
769
+ publishedBy: string;
770
+ publishedAt: string;
771
+ } | null;
772
+ form: {
773
+ schemaVersion: "openpond.humanForm.v1";
774
+ mode: "individual" | "pairwise";
775
+ instructions: string;
776
+ criteria: ({
777
+ id: string;
778
+ label: string;
779
+ instructions: string;
780
+ required: boolean;
781
+ allowAbstain: boolean;
782
+ kind: "score";
783
+ minimum: number;
784
+ maximum: number;
785
+ step: number;
786
+ requireRationale?: boolean | undefined;
787
+ anchors?: {
788
+ value: string | number | boolean;
789
+ description: string;
790
+ }[] | undefined;
791
+ } | {
792
+ id: string;
793
+ label: string;
794
+ instructions: string;
795
+ required: boolean;
796
+ allowAbstain: boolean;
797
+ kind: "category";
798
+ options: string[];
799
+ requireRationale?: boolean | undefined;
800
+ anchors?: {
801
+ value: string | number | boolean;
802
+ description: string;
803
+ }[] | undefined;
804
+ } | {
805
+ id: string;
806
+ label: string;
807
+ instructions: string;
808
+ required: boolean;
809
+ allowAbstain: boolean;
810
+ kind: "boolean";
811
+ requireRationale?: boolean | undefined;
812
+ anchors?: {
813
+ value: string | number | boolean;
814
+ description: string;
815
+ }[] | undefined;
816
+ } | {
817
+ id: string;
818
+ label: string;
819
+ instructions: string;
820
+ required: boolean;
821
+ allowAbstain: boolean;
822
+ kind: "text";
823
+ maximumLength: number;
824
+ requireRationale?: boolean | undefined;
825
+ anchors?: {
826
+ value: string | number | boolean;
827
+ description: string;
828
+ }[] | undefined;
829
+ minimumLength?: number | undefined;
830
+ } | {
831
+ id: string;
832
+ label: string;
833
+ instructions: string;
834
+ required: boolean;
835
+ allowAbstain: boolean;
836
+ kind: "preference";
837
+ allowTie: boolean;
838
+ requireRationale?: boolean | undefined;
839
+ anchors?: {
840
+ value: string | number | boolean;
841
+ description: string;
842
+ }[] | undefined;
843
+ requireStrength?: boolean | undefined;
844
+ })[];
845
+ };
846
+ policy: {
847
+ approval: "none" | "manager";
848
+ minimumRaters: number;
849
+ managerIds: string[];
850
+ reviewerIds: string[];
851
+ queueClaim: boolean;
852
+ teamVisible: boolean;
853
+ distinctManager?: boolean | undefined;
854
+ independentRaters?: boolean | undefined;
855
+ blindPair?: boolean | undefined;
856
+ };
857
+ status: "queued" | "assigned" | "in_progress" | "submitted" | "returned" | "accepted" | "rejected" | "cancelled";
858
+ assigneeId: string | null;
859
+ generation: number;
860
+ execution: {
861
+ executionId: string | null;
862
+ operationId: string;
863
+ boundBy: string;
864
+ configurationHash: string;
865
+ generation: number;
866
+ result: {
867
+ id: string;
868
+ revision: number;
869
+ contentHash: string;
870
+ } | null;
871
+ } | null;
872
+ createdAt: string;
873
+ updatedAt: string;
874
+ contentHash: string;
875
+ dueAt?: string | undefined;
876
+ }>;
877
+ };
878
+ //# sourceMappingURL=service.d.ts.map