openpond-sdk 0.9.9 → 0.11.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (253) hide show
  1. package/README.md +18 -17
  2. package/dist/advanced-refiner-evaluations.js +42 -0
  3. package/dist/advanced-refiner-evaluations.js.map +7 -0
  4. package/dist/candidate-evaluations.js +25 -0
  5. package/dist/candidate-evaluations.js.map +7 -0
  6. package/dist/{chunk-BGXWHDYM.js → chunk-2JKV5OJO.js} +95 -59
  7. package/dist/chunk-2JKV5OJO.js.map +7 -0
  8. package/dist/{chunk-4HZORGK3.js → chunk-2NSNBED7.js} +18 -5
  9. package/dist/chunk-2NSNBED7.js.map +7 -0
  10. package/dist/chunk-5RZAPZWI.js +445 -0
  11. package/dist/chunk-5RZAPZWI.js.map +7 -0
  12. package/dist/{chunk-JCPLAIZL.js → chunk-5U25KAJM.js} +106 -76
  13. package/dist/chunk-5U25KAJM.js.map +7 -0
  14. package/dist/chunk-7Q7I6FAT.js +139 -0
  15. package/dist/chunk-7Q7I6FAT.js.map +7 -0
  16. package/dist/chunk-A64VYVXN.js +64 -0
  17. package/dist/chunk-A64VYVXN.js.map +7 -0
  18. package/dist/{chunk-DT7FRNIZ.js → chunk-CSWVE3E6.js} +2 -2
  19. package/dist/chunk-DJWHN7NU.js +394 -0
  20. package/dist/chunk-DJWHN7NU.js.map +7 -0
  21. package/dist/{chunk-KYCQ7LZ4.js → chunk-DSGGM2ME.js} +2 -2
  22. package/dist/chunk-DSGGM2ME.js.map +7 -0
  23. package/dist/chunk-GDJRGRHA.js +144 -0
  24. package/dist/chunk-GDJRGRHA.js.map +7 -0
  25. package/dist/chunk-HOAYWW33.js +587 -0
  26. package/dist/chunk-HOAYWW33.js.map +7 -0
  27. package/dist/{chunk-Y4NGWGTA.js → chunk-I2MBVOVX.js} +265 -264
  28. package/dist/chunk-I2MBVOVX.js.map +7 -0
  29. package/dist/{chunk-6D4JIOOC.js → chunk-IXBSFJCX.js} +5 -5
  30. package/dist/{chunk-WOPM32SE.js → chunk-JA3KFK3U.js} +7 -1
  31. package/dist/chunk-JA3KFK3U.js.map +7 -0
  32. package/dist/chunk-JMQ4HBWP.js +160 -0
  33. package/dist/chunk-JMQ4HBWP.js.map +7 -0
  34. package/dist/chunk-KQWCXW3A.js +294 -0
  35. package/dist/chunk-KQWCXW3A.js.map +7 -0
  36. package/dist/{chunk-3HV3G6DG.js → chunk-NHCGOFQW.js} +3 -3
  37. package/dist/{chunk-XMKEFQ62.js → chunk-PJ2VLNJF.js} +20 -2
  38. package/dist/chunk-PJ2VLNJF.js.map +7 -0
  39. package/dist/{chunk-32H2EJRA.js → chunk-R3M6BIOK.js} +2 -2
  40. package/dist/chunk-SLS3GFPG.js +125 -0
  41. package/dist/chunk-SLS3GFPG.js.map +7 -0
  42. package/dist/chunk-SVU4RZ4K.js +70 -0
  43. package/dist/chunk-SVU4RZ4K.js.map +7 -0
  44. package/dist/chunk-T4APHDES.js +74 -0
  45. package/dist/chunk-T4APHDES.js.map +7 -0
  46. package/dist/chunk-VC5OV5AS.js +174 -0
  47. package/dist/chunk-VC5OV5AS.js.map +7 -0
  48. package/dist/{chunk-YYJ4T4U4.js → chunk-WX22WN33.js} +13 -7
  49. package/dist/chunk-WX22WN33.js.map +7 -0
  50. package/dist/chunk-XEUOJ5WF.js +31 -0
  51. package/dist/chunk-XEUOJ5WF.js.map +7 -0
  52. package/dist/{chunk-LGEF4S5I.js → chunk-XJDWHTH5.js} +390 -362
  53. package/dist/chunk-XJDWHTH5.js.map +7 -0
  54. package/dist/chunk-YCQVVFE5.js +37 -0
  55. package/dist/chunk-YCQVVFE5.js.map +7 -0
  56. package/dist/chunk-ZVJU5CJ5.js +267 -0
  57. package/dist/chunk-ZVJU5CJ5.js.map +7 -0
  58. package/dist/connected-evidence.js +81 -0
  59. package/dist/connected-evidence.js.map +7 -0
  60. package/dist/dataset-workspaces.js +45 -80
  61. package/dist/dataset-workspaces.js.map +4 -4
  62. package/dist/experiment-evaluation-schedules.js +48 -0
  63. package/dist/experiment-evaluation-schedules.js.map +7 -0
  64. package/dist/experiment-improvements.js +52 -0
  65. package/dist/experiment-improvements.js.map +7 -0
  66. package/dist/experiments.js +63 -124
  67. package/dist/experiments.js.map +4 -4
  68. package/dist/human-review.js +13 -0
  69. package/dist/human-review.js.map +7 -0
  70. package/dist/index.js +208 -17
  71. package/dist/index.js.map +2 -2
  72. package/dist/learning.js +1 -1
  73. package/dist/model-starter-attempts.js +2 -2
  74. package/dist/model-starter-catalog.js +9 -4
  75. package/dist/model-starter-catalog.js.map +1 -1
  76. package/dist/model-starters.js +15 -6
  77. package/dist/model-taskset-runs.js +3 -3
  78. package/dist/post-training.js +47 -0
  79. package/dist/post-training.js.map +7 -0
  80. package/dist/taskset-drafts.js +13 -6
  81. package/dist/taskset-drafts.js.map +2 -2
  82. package/dist/taskset-packages.js +31 -559
  83. package/dist/taskset-packages.js.map +4 -4
  84. package/dist/training-bundle.js +59 -18
  85. package/dist/training-bundle.js.map +2 -2
  86. package/dist/training-projects.js +3 -1
  87. package/dist/training-projects.js.map +2 -2
  88. package/dist/training.js +7 -2
  89. package/dist/types/packages/sdk/src/advanced-refiner-evaluations.d.ts +2673 -0
  90. package/dist/types/packages/sdk/src/advanced-refiner-evaluations.d.ts.map +1 -0
  91. package/dist/types/packages/sdk/src/candidate-evaluation-client.d.ts +1079 -0
  92. package/dist/types/packages/sdk/src/candidate-evaluation-client.d.ts.map +1 -0
  93. package/dist/types/packages/sdk/src/candidate-evaluations.d.ts +845 -0
  94. package/dist/types/packages/sdk/src/candidate-evaluations.d.ts.map +1 -0
  95. package/dist/types/packages/sdk/src/claude-code-runtime.d.ts +51 -0
  96. package/dist/types/packages/sdk/src/claude-code-runtime.d.ts.map +1 -0
  97. package/dist/types/packages/sdk/src/connected-dataset-publication.d.ts +15 -0
  98. package/dist/types/packages/sdk/src/connected-dataset-publication.d.ts.map +1 -0
  99. package/dist/types/packages/sdk/src/connected-dataset-source.d.ts +48 -0
  100. package/dist/types/packages/sdk/src/connected-dataset-source.d.ts.map +1 -0
  101. package/dist/types/packages/sdk/src/connected-evidence-client.d.ts +2460 -0
  102. package/dist/types/packages/sdk/src/connected-evidence-client.d.ts.map +1 -0
  103. package/dist/types/packages/sdk/src/connected-evidence-contracts.d.ts +215 -0
  104. package/dist/types/packages/sdk/src/connected-evidence-contracts.d.ts.map +1 -0
  105. package/dist/types/packages/sdk/src/connected-evidence-http.d.ts +6 -0
  106. package/dist/types/packages/sdk/src/connected-evidence-http.d.ts.map +1 -0
  107. package/dist/types/packages/sdk/src/connected-evidence.d.ts +6 -0
  108. package/dist/types/packages/sdk/src/connected-evidence.d.ts.map +1 -0
  109. package/dist/types/packages/sdk/src/connected-recorded-execution.d.ts +690 -0
  110. package/dist/types/packages/sdk/src/connected-recorded-execution.d.ts.map +1 -0
  111. package/dist/types/packages/sdk/src/dataset-population.d.ts +1 -0
  112. package/dist/types/packages/sdk/src/dataset-population.d.ts.map +1 -1
  113. package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts +771 -120
  114. package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts.map +1 -1
  115. package/dist/types/packages/sdk/src/dataset-workspace-contracts.d.ts +278 -38
  116. package/dist/types/packages/sdk/src/dataset-workspace-contracts.d.ts.map +1 -1
  117. package/dist/types/packages/sdk/src/dataset-workspace-operations.d.ts +280 -40
  118. package/dist/types/packages/sdk/src/dataset-workspace-operations.d.ts.map +1 -1
  119. package/dist/types/packages/sdk/src/evaluation-coordination-http.d.ts +17 -0
  120. package/dist/types/packages/sdk/src/evaluation-coordination-http.d.ts.map +1 -0
  121. package/dist/types/packages/sdk/src/experiment-client.d.ts +4353 -490
  122. package/dist/types/packages/sdk/src/experiment-client.d.ts.map +1 -1
  123. package/dist/types/packages/sdk/src/experiment-contracts.d.ts +1067 -2
  124. package/dist/types/packages/sdk/src/experiment-contracts.d.ts.map +1 -1
  125. package/dist/types/packages/sdk/src/experiment-evaluation-schedules.d.ts +2780 -0
  126. package/dist/types/packages/sdk/src/experiment-evaluation-schedules.d.ts.map +1 -0
  127. package/dist/types/packages/sdk/src/experiment-execution-selection.d.ts +58 -0
  128. package/dist/types/packages/sdk/src/experiment-execution-selection.d.ts.map +1 -0
  129. package/dist/types/packages/sdk/src/experiment-grading-contracts.d.ts +8 -0
  130. package/dist/types/packages/sdk/src/experiment-grading-contracts.d.ts.map +1 -1
  131. package/dist/types/packages/sdk/src/experiment-harness-catalog.d.ts +24 -0
  132. package/dist/types/packages/sdk/src/experiment-harness-catalog.d.ts.map +1 -0
  133. package/dist/types/packages/sdk/src/experiment-improvement-client.d.ts +5724 -0
  134. package/dist/types/packages/sdk/src/experiment-improvement-client.d.ts.map +1 -0
  135. package/dist/types/packages/sdk/src/experiment-improvement-contracts.d.ts +13024 -0
  136. package/dist/types/packages/sdk/src/experiment-improvement-contracts.d.ts.map +1 -0
  137. package/dist/types/packages/sdk/src/experiment-improvement-hosted-adoption.d.ts +1361 -0
  138. package/dist/types/packages/sdk/src/experiment-improvement-hosted-adoption.d.ts.map +1 -0
  139. package/dist/types/packages/sdk/src/experiment-improvements.d.ts +4 -0
  140. package/dist/types/packages/sdk/src/experiment-improvements.d.ts.map +1 -0
  141. package/dist/types/packages/sdk/src/experiment-run-contracts.d.ts +383 -0
  142. package/dist/types/packages/sdk/src/experiment-run-contracts.d.ts.map +1 -0
  143. package/dist/types/packages/sdk/src/experiment-run-details.d.ts +2111 -0
  144. package/dist/types/packages/sdk/src/experiment-run-details.d.ts.map +1 -0
  145. package/dist/types/packages/sdk/src/experiment-scoring-contracts.d.ts +7 -2
  146. package/dist/types/packages/sdk/src/experiment-scoring-contracts.d.ts.map +1 -1
  147. package/dist/types/packages/sdk/src/experiments.d.ts +7 -1
  148. package/dist/types/packages/sdk/src/experiments.d.ts.map +1 -1
  149. package/dist/types/packages/sdk/src/grader-inspection.d.ts +307 -5
  150. package/dist/types/packages/sdk/src/grader-inspection.d.ts.map +1 -1
  151. package/dist/types/packages/sdk/src/harness-experiment-policy.d.ts +238 -0
  152. package/dist/types/packages/sdk/src/harness-experiment-policy.d.ts.map +1 -1
  153. package/dist/types/packages/sdk/src/human-review-client.d.ts +1267 -0
  154. package/dist/types/packages/sdk/src/human-review-client.d.ts.map +1 -0
  155. package/dist/types/packages/sdk/src/human-review.d.ts +3 -0
  156. package/dist/types/packages/sdk/src/human-review.d.ts.map +1 -0
  157. package/dist/types/packages/sdk/src/index.d.ts +10 -0
  158. package/dist/types/packages/sdk/src/index.d.ts.map +1 -1
  159. package/dist/types/packages/sdk/src/managed-training-preparation.d.ts +181 -0
  160. package/dist/types/packages/sdk/src/managed-training-preparation.d.ts.map +1 -1
  161. package/dist/types/packages/sdk/src/model-batch-review-contracts.d.ts +20 -4
  162. package/dist/types/packages/sdk/src/model-batch-review-contracts.d.ts.map +1 -1
  163. package/dist/types/packages/sdk/src/model-batch-review-preparation.d.ts +79 -5
  164. package/dist/types/packages/sdk/src/model-batch-review-preparation.d.ts.map +1 -1
  165. package/dist/types/packages/sdk/src/model-batch-review.d.ts +15 -13
  166. package/dist/types/packages/sdk/src/model-batch-review.d.ts.map +1 -1
  167. package/dist/types/packages/sdk/src/model-learning-policy.d.ts +359 -1
  168. package/dist/types/packages/sdk/src/model-learning-policy.d.ts.map +1 -1
  169. package/dist/types/packages/sdk/src/model-starter-attempts.d.ts +444 -19
  170. package/dist/types/packages/sdk/src/model-starter-attempts.d.ts.map +1 -1
  171. package/dist/types/packages/sdk/src/model-starter-catalog.d.ts +234 -12
  172. package/dist/types/packages/sdk/src/model-starter-catalog.d.ts.map +1 -1
  173. package/dist/types/packages/sdk/src/model-starter-execution.d.ts +79 -1
  174. package/dist/types/packages/sdk/src/model-starter-execution.d.ts.map +1 -1
  175. package/dist/types/packages/sdk/src/model-starters.d.ts +467 -13
  176. package/dist/types/packages/sdk/src/model-starters.d.ts.map +1 -1
  177. package/dist/types/packages/sdk/src/model-taskset-derivation.d.ts +312 -4
  178. package/dist/types/packages/sdk/src/model-taskset-derivation.d.ts.map +1 -1
  179. package/dist/types/packages/sdk/src/model-taskset-resources.d.ts +78 -1
  180. package/dist/types/packages/sdk/src/model-taskset-resources.d.ts.map +1 -1
  181. package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts +1326 -4
  182. package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts.map +1 -1
  183. package/dist/types/packages/sdk/src/model-taskset-runs.d.ts +1329 -5
  184. package/dist/types/packages/sdk/src/model-taskset-runs.d.ts.map +1 -1
  185. package/dist/types/packages/sdk/src/post-training.d.ts +3518 -0
  186. package/dist/types/packages/sdk/src/post-training.d.ts.map +1 -0
  187. package/dist/types/packages/sdk/src/profile-evaluation-discovery.d.ts +1127 -0
  188. package/dist/types/packages/sdk/src/profile-evaluation-discovery.d.ts.map +1 -0
  189. package/dist/types/packages/sdk/src/profile-external-dataset-package.d.ts +1369 -0
  190. package/dist/types/packages/sdk/src/profile-external-dataset-package.d.ts.map +1 -0
  191. package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts +142 -22
  192. package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts.map +1 -1
  193. package/dist/types/packages/sdk/src/taskset-authored-files.d.ts +111 -18
  194. package/dist/types/packages/sdk/src/taskset-authored-files.d.ts.map +1 -1
  195. package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts +278 -38
  196. package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts.map +1 -1
  197. package/dist/types/packages/sdk/src/taskset-draft-client.d.ts +661 -103
  198. package/dist/types/packages/sdk/src/taskset-draft-client.d.ts.map +1 -1
  199. package/dist/types/packages/sdk/src/taskset-draft-core.d.ts +217 -20
  200. package/dist/types/packages/sdk/src/taskset-draft-core.d.ts.map +1 -1
  201. package/dist/types/packages/sdk/src/taskset-draft-dataset-artifacts.d.ts +12 -12
  202. package/dist/types/packages/sdk/src/taskset-draft-dataset-sources.d.ts +17 -17
  203. package/dist/types/packages/sdk/src/taskset-draft-document.d.ts +140 -20
  204. package/dist/types/packages/sdk/src/taskset-draft-document.d.ts.map +1 -1
  205. package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts +638 -42
  206. package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts.map +1 -1
  207. package/dist/types/packages/sdk/src/taskset-draft-source.d.ts +249 -36
  208. package/dist/types/packages/sdk/src/taskset-draft-source.d.ts.map +1 -1
  209. package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts +278 -38
  210. package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts.map +1 -1
  211. package/dist/types/packages/sdk/src/taskset-drafts.d.ts +1 -0
  212. package/dist/types/packages/sdk/src/taskset-drafts.d.ts.map +1 -1
  213. package/dist/types/packages/sdk/src/taskset-package-client.d.ts +2805 -41
  214. package/dist/types/packages/sdk/src/taskset-package-client.d.ts.map +1 -1
  215. package/dist/types/packages/sdk/src/taskset-package-contracts.d.ts +660 -12
  216. package/dist/types/packages/sdk/src/taskset-package-contracts.d.ts.map +1 -1
  217. package/dist/types/packages/sdk/src/taskset-package-grader-projection.d.ts +77 -3
  218. package/dist/types/packages/sdk/src/taskset-package-grader-projection.d.ts.map +1 -1
  219. package/dist/types/packages/sdk/src/taskset-package-learning.d.ts +29 -11
  220. package/dist/types/packages/sdk/src/taskset-package-learning.d.ts.map +1 -1
  221. package/dist/types/packages/sdk/src/taskset-package-projection.d.ts +430 -39
  222. package/dist/types/packages/sdk/src/taskset-package-projection.d.ts.map +1 -1
  223. package/dist/types/packages/sdk/src/training-acceptance.d.ts +10 -10
  224. package/dist/types/packages/sdk/src/training-candidate-decisions.d.ts +2 -2
  225. package/dist/types/packages/sdk/src/training-evaluation-results.d.ts +1 -1
  226. package/dist/types/packages/sdk/src/training-learning-batch.d.ts +376 -42
  227. package/dist/types/packages/sdk/src/training-learning-batch.d.ts.map +1 -1
  228. package/dist/types/packages/sdk/src/training-preparations.d.ts +3930 -0
  229. package/dist/types/packages/sdk/src/training-preparations.d.ts.map +1 -1
  230. package/dist/types/packages/sdk/src/training-project-client.d.ts +828 -8
  231. package/dist/types/packages/sdk/src/training-project-client.d.ts.map +1 -1
  232. package/dist/types/packages/sdk/src/training-project-contracts.d.ts +1226 -11
  233. package/dist/types/packages/sdk/src/training-project-contracts.d.ts.map +1 -1
  234. package/dist/types/packages/sdk/src/training.d.ts +2655 -24
  235. package/dist/types/packages/sdk/src/training.d.ts.map +1 -1
  236. package/package.json +30 -2
  237. package/dist/chunk-4HZORGK3.js.map +0 -7
  238. package/dist/chunk-4JRSH2TT.js +0 -256
  239. package/dist/chunk-4JRSH2TT.js.map +0 -7
  240. package/dist/chunk-BGXWHDYM.js.map +0 -7
  241. package/dist/chunk-JCPLAIZL.js.map +0 -7
  242. package/dist/chunk-KYCQ7LZ4.js.map +0 -7
  243. package/dist/chunk-LGEF4S5I.js.map +0 -7
  244. package/dist/chunk-WOPM32SE.js.map +0 -7
  245. package/dist/chunk-XMKEFQ62.js.map +0 -7
  246. package/dist/chunk-Y4NGWGTA.js.map +0 -7
  247. package/dist/chunk-YYJ4T4U4.js.map +0 -7
  248. package/dist/types/packages/sdk/src/experiment-history.d.ts +0 -266
  249. package/dist/types/packages/sdk/src/experiment-history.d.ts.map +0 -1
  250. /package/dist/{chunk-DT7FRNIZ.js.map → chunk-CSWVE3E6.js.map} +0 -0
  251. /package/dist/{chunk-6D4JIOOC.js.map → chunk-IXBSFJCX.js.map} +0 -0
  252. /package/dist/{chunk-3HV3G6DG.js.map → chunk-NHCGOFQW.js.map} +0 -0
  253. /package/dist/{chunk-32H2EJRA.js.map → chunk-R3M6BIOK.js.map} +0 -0
@@ -0,0 +1,1369 @@
1
+ import { z } from "zod";
2
+ import { type ProfileExternalDatasetBinding } from "@openpond/evals";
3
+ import { type TasksetPackage } from "./taskset-packages.js";
4
+ export type ExternalDatasetGraderPin = {
5
+ id: string;
6
+ version: string;
7
+ contentHash: string;
8
+ feedbackKey: string;
9
+ release: {
10
+ id: string;
11
+ revision: number | string;
12
+ contentHash: string;
13
+ } | null;
14
+ };
15
+ /** Recompute admission from actual authorized releases. None of the binding's
16
+ * self-declared hashes grants Dataset or Profile access. */
17
+ export declare function verifyExternalDatasetPackage(input: {
18
+ binding: ProfileExternalDatasetBinding;
19
+ packageValue: TasksetPackage;
20
+ graders: ExternalDatasetGraderPin[];
21
+ actualGrading?: {
22
+ pass: unknown;
23
+ evidence: unknown;
24
+ };
25
+ }): {
26
+ schemaVersion: "openpond.tasksetPackage.v1";
27
+ taskset: {
28
+ schemaVersion: "openpond.tasksetRelease.v2";
29
+ id: string;
30
+ revision: number;
31
+ policy: {
32
+ policyVisibleFields: string[];
33
+ privilegedFields: string[];
34
+ hiddenGraderRefs: string[];
35
+ connectedAppScopes: string[];
36
+ };
37
+ environment: {
38
+ protocolVersion: "openpond.environment.v1";
39
+ kind: "text" | "agent" | "work" | "custom_program";
40
+ entrypoint: string;
41
+ stateful: boolean;
42
+ deterministicSeeds: boolean;
43
+ lifecycle: ("create" | "reset" | "step" | "collect" | "destroy")[];
44
+ networkPolicy: "none" | "declared_read_only" | "declared_scoped";
45
+ defaultTimeoutMs: number;
46
+ };
47
+ tools: {
48
+ name: string;
49
+ description: string;
50
+ inputSchema: Record<string, unknown>;
51
+ inputSchemaHash: string;
52
+ sideEffect: "read" | "write";
53
+ timeoutMs: number;
54
+ }[];
55
+ capabilities: {
56
+ id: string;
57
+ required: boolean;
58
+ scopes: string[];
59
+ portability: "portable" | "host_adapter" | "local_only" | "hosted_only";
60
+ }[];
61
+ tasks: {
62
+ id: string;
63
+ clusterKey: string;
64
+ split: "train" | "validation" | "test" | "frozen_eval";
65
+ input: Record<string, unknown>;
66
+ expectedOutput: Record<string, unknown> | null;
67
+ policyVisibleContext: Record<string, unknown>;
68
+ privilegedContextRef: string | null;
69
+ artifactRefs: {
70
+ id: string;
71
+ path: string;
72
+ contentHash: string;
73
+ sizeBytes: number;
74
+ mediaType: string;
75
+ visibility: "policy" | "verifier" | "host_private";
76
+ }[];
77
+ tags: string[];
78
+ requiredOutputs?: {
79
+ path: string;
80
+ mediaType: string;
81
+ schemaRef: {
82
+ id: string;
83
+ path: string;
84
+ contentHash: string;
85
+ sizeBytes: number;
86
+ mediaType: string;
87
+ visibility: "policy" | "verifier" | "host_private";
88
+ } | null;
89
+ maxBytes: number | null;
90
+ metadata: Record<string, unknown>;
91
+ }[] | undefined;
92
+ }[];
93
+ graders: ({
94
+ id: string;
95
+ version: string;
96
+ weight: number;
97
+ hardGate: boolean;
98
+ rewardEligible: boolean;
99
+ privileged: boolean;
100
+ kind: "content" | "state" | "artifact" | "schema" | "runtime_event";
101
+ config: Record<string, unknown>;
102
+ } | {
103
+ id: string;
104
+ version: string;
105
+ weight: number;
106
+ hardGate: boolean;
107
+ rewardEligible: boolean;
108
+ privileged: boolean;
109
+ kind: "model_judge";
110
+ rubricRef: {
111
+ id: string;
112
+ path: string;
113
+ contentHash: string;
114
+ sizeBytes: number;
115
+ mediaType: string;
116
+ visibility: "policy" | "verifier" | "host_private";
117
+ };
118
+ calibrationStatus: "passed" | "pending" | "failed";
119
+ model?: {
120
+ providerId: string;
121
+ modelId: string;
122
+ revision: string | null;
123
+ } | undefined;
124
+ temperature?: number | undefined;
125
+ } | {
126
+ id: string;
127
+ version: string;
128
+ weight: number;
129
+ hardGate: boolean;
130
+ rewardEligible: boolean;
131
+ privileged: boolean;
132
+ kind: "custom_verifier";
133
+ verifierRef: {
134
+ id: string;
135
+ path: string;
136
+ contentHash: string;
137
+ sizeBytes: number;
138
+ mediaType: string;
139
+ visibility: "policy" | "verifier" | "host_private";
140
+ };
141
+ timeoutMs: number;
142
+ networkPolicy: "none";
143
+ runtime?: "isolated_javascript" | "sandbox_process" | undefined;
144
+ exportName?: string | undefined;
145
+ } | {
146
+ id: string;
147
+ version: string;
148
+ weight: number;
149
+ hardGate: boolean;
150
+ rewardEligible: boolean;
151
+ privileged: boolean;
152
+ kind: "human";
153
+ rubricRef: {
154
+ id: string;
155
+ path: string;
156
+ contentHash: string;
157
+ sizeBytes: number;
158
+ mediaType: string;
159
+ visibility: "policy" | "verifier" | "host_private";
160
+ };
161
+ reviewerRole: string;
162
+ form?: {
163
+ schemaVersion: "openpond.humanForm.v1";
164
+ mode: "individual" | "pairwise";
165
+ instructions: string;
166
+ criteria: ({
167
+ id: string;
168
+ label: string;
169
+ instructions: string;
170
+ required: boolean;
171
+ allowAbstain: boolean;
172
+ kind: "score";
173
+ minimum: number;
174
+ maximum: number;
175
+ step: number;
176
+ requireRationale?: boolean | undefined;
177
+ anchors?: {
178
+ value: string | number | boolean;
179
+ description: string;
180
+ }[] | undefined;
181
+ } | {
182
+ id: string;
183
+ label: string;
184
+ instructions: string;
185
+ required: boolean;
186
+ allowAbstain: boolean;
187
+ kind: "category";
188
+ options: string[];
189
+ requireRationale?: boolean | undefined;
190
+ anchors?: {
191
+ value: string | number | boolean;
192
+ description: string;
193
+ }[] | undefined;
194
+ } | {
195
+ id: string;
196
+ label: string;
197
+ instructions: string;
198
+ required: boolean;
199
+ allowAbstain: boolean;
200
+ kind: "boolean";
201
+ requireRationale?: boolean | undefined;
202
+ anchors?: {
203
+ value: string | number | boolean;
204
+ description: string;
205
+ }[] | undefined;
206
+ } | {
207
+ id: string;
208
+ label: string;
209
+ instructions: string;
210
+ required: boolean;
211
+ allowAbstain: boolean;
212
+ kind: "text";
213
+ maximumLength: number;
214
+ requireRationale?: boolean | undefined;
215
+ anchors?: {
216
+ value: string | number | boolean;
217
+ description: string;
218
+ }[] | undefined;
219
+ minimumLength?: number | undefined;
220
+ } | {
221
+ id: string;
222
+ label: string;
223
+ instructions: string;
224
+ required: boolean;
225
+ allowAbstain: boolean;
226
+ kind: "preference";
227
+ allowTie: boolean;
228
+ requireRationale?: boolean | undefined;
229
+ anchors?: {
230
+ value: string | number | boolean;
231
+ description: string;
232
+ }[] | undefined;
233
+ requireStrength?: boolean | undefined;
234
+ })[];
235
+ } | undefined;
236
+ })[];
237
+ metadata: Record<string, unknown>;
238
+ contentHash: string;
239
+ environmentRelease?: {
240
+ id: string;
241
+ contentHash: string;
242
+ } | undefined;
243
+ verifierSetRelease?: {
244
+ id: string;
245
+ contentHash: string;
246
+ } | undefined;
247
+ metrics?: {
248
+ schemaVersion: "openpond.tasksetMetricPolicy.v1";
249
+ primaryMetric: string;
250
+ aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
251
+ missingReward: "zero" | "exclude";
252
+ customAggregator: {
253
+ module: string;
254
+ exportName: string;
255
+ contentHash: string;
256
+ timeoutMs: number;
257
+ networkPolicy: "none";
258
+ } | null;
259
+ taskWeights?: Record<string, number> | undefined;
260
+ } | undefined;
261
+ };
262
+ environment: {
263
+ schemaVersion: "openpond.environmentRelease.v1";
264
+ id: string;
265
+ revision: number;
266
+ contract: {
267
+ protocolVersion: "openpond.environment.v1";
268
+ kind: "text" | "agent" | "work" | "custom_program";
269
+ entrypoint: string;
270
+ stateful: boolean;
271
+ deterministicSeeds: boolean;
272
+ lifecycle: ("create" | "reset" | "step" | "collect" | "destroy")[];
273
+ networkPolicy: "none" | "declared_read_only" | "declared_scoped";
274
+ defaultTimeoutMs: number;
275
+ };
276
+ actionSchemaRef: {
277
+ id: string;
278
+ path: string;
279
+ contentHash: string;
280
+ sizeBytes: number;
281
+ mediaType: string;
282
+ visibility: "policy" | "verifier" | "host_private";
283
+ } | null;
284
+ observationSchemaRef: {
285
+ id: string;
286
+ path: string;
287
+ contentHash: string;
288
+ sizeBytes: number;
289
+ mediaType: string;
290
+ visibility: "policy" | "verifier" | "host_private";
291
+ } | null;
292
+ stateSchemaRef: {
293
+ id: string;
294
+ path: string;
295
+ contentHash: string;
296
+ sizeBytes: number;
297
+ mediaType: string;
298
+ visibility: "policy" | "verifier" | "host_private";
299
+ } | null;
300
+ artifactCollection: {
301
+ maxArtifacts: number;
302
+ maxTotalBytes: number;
303
+ };
304
+ adapterConformanceHashes: Record<string, string>;
305
+ metadata: Record<string, unknown>;
306
+ contentHash: string;
307
+ };
308
+ verifierSet: {
309
+ schemaVersion: "openpond.verifierSetRelease.v1";
310
+ id: string;
311
+ revision: number;
312
+ graders: ({
313
+ id: string;
314
+ version: string;
315
+ weight: number;
316
+ hardGate: boolean;
317
+ rewardEligible: boolean;
318
+ privileged: boolean;
319
+ kind: "content" | "state" | "artifact" | "schema" | "runtime_event";
320
+ config: Record<string, unknown>;
321
+ } | {
322
+ id: string;
323
+ version: string;
324
+ weight: number;
325
+ hardGate: boolean;
326
+ rewardEligible: boolean;
327
+ privileged: boolean;
328
+ kind: "model_judge";
329
+ rubricRef: {
330
+ id: string;
331
+ path: string;
332
+ contentHash: string;
333
+ sizeBytes: number;
334
+ mediaType: string;
335
+ visibility: "policy" | "verifier" | "host_private";
336
+ };
337
+ calibrationStatus: "passed" | "pending" | "failed";
338
+ model?: {
339
+ providerId: string;
340
+ modelId: string;
341
+ revision: string | null;
342
+ } | undefined;
343
+ temperature?: number | undefined;
344
+ } | {
345
+ id: string;
346
+ version: string;
347
+ weight: number;
348
+ hardGate: boolean;
349
+ rewardEligible: boolean;
350
+ privileged: boolean;
351
+ kind: "custom_verifier";
352
+ verifierRef: {
353
+ id: string;
354
+ path: string;
355
+ contentHash: string;
356
+ sizeBytes: number;
357
+ mediaType: string;
358
+ visibility: "policy" | "verifier" | "host_private";
359
+ };
360
+ timeoutMs: number;
361
+ networkPolicy: "none";
362
+ runtime?: "isolated_javascript" | "sandbox_process" | undefined;
363
+ exportName?: string | undefined;
364
+ } | {
365
+ id: string;
366
+ version: string;
367
+ weight: number;
368
+ hardGate: boolean;
369
+ rewardEligible: boolean;
370
+ privileged: boolean;
371
+ kind: "human";
372
+ rubricRef: {
373
+ id: string;
374
+ path: string;
375
+ contentHash: string;
376
+ sizeBytes: number;
377
+ mediaType: string;
378
+ visibility: "policy" | "verifier" | "host_private";
379
+ };
380
+ reviewerRole: string;
381
+ form?: {
382
+ schemaVersion: "openpond.humanForm.v1";
383
+ mode: "individual" | "pairwise";
384
+ instructions: string;
385
+ criteria: ({
386
+ id: string;
387
+ label: string;
388
+ instructions: string;
389
+ required: boolean;
390
+ allowAbstain: boolean;
391
+ kind: "score";
392
+ minimum: number;
393
+ maximum: number;
394
+ step: number;
395
+ requireRationale?: boolean | undefined;
396
+ anchors?: {
397
+ value: string | number | boolean;
398
+ description: string;
399
+ }[] | undefined;
400
+ } | {
401
+ id: string;
402
+ label: string;
403
+ instructions: string;
404
+ required: boolean;
405
+ allowAbstain: boolean;
406
+ kind: "category";
407
+ options: string[];
408
+ requireRationale?: boolean | undefined;
409
+ anchors?: {
410
+ value: string | number | boolean;
411
+ description: string;
412
+ }[] | undefined;
413
+ } | {
414
+ id: string;
415
+ label: string;
416
+ instructions: string;
417
+ required: boolean;
418
+ allowAbstain: boolean;
419
+ kind: "boolean";
420
+ requireRationale?: boolean | undefined;
421
+ anchors?: {
422
+ value: string | number | boolean;
423
+ description: string;
424
+ }[] | undefined;
425
+ } | {
426
+ id: string;
427
+ label: string;
428
+ instructions: string;
429
+ required: boolean;
430
+ allowAbstain: boolean;
431
+ kind: "text";
432
+ maximumLength: number;
433
+ requireRationale?: boolean | undefined;
434
+ anchors?: {
435
+ value: string | number | boolean;
436
+ description: string;
437
+ }[] | undefined;
438
+ minimumLength?: number | undefined;
439
+ } | {
440
+ id: string;
441
+ label: string;
442
+ instructions: string;
443
+ required: boolean;
444
+ allowAbstain: boolean;
445
+ kind: "preference";
446
+ allowTie: boolean;
447
+ requireRationale?: boolean | undefined;
448
+ anchors?: {
449
+ value: string | number | boolean;
450
+ description: string;
451
+ }[] | undefined;
452
+ requireStrength?: boolean | undefined;
453
+ })[];
454
+ } | undefined;
455
+ })[];
456
+ isolation: {
457
+ processBoundary: "same_process" | "isolated_process" | "container";
458
+ networkPolicy: "none";
459
+ defaultTimeoutMs: number;
460
+ };
461
+ calibrationReceiptRefs: {
462
+ id: string;
463
+ contentHash: string;
464
+ }[];
465
+ metadata: Record<string, unknown>;
466
+ contentHash: string;
467
+ };
468
+ files: {
469
+ asset: {
470
+ id: string;
471
+ path: string;
472
+ contentHash: string;
473
+ sizeBytes: number;
474
+ mediaType: string;
475
+ visibility: "policy" | "verifier" | "host_private";
476
+ };
477
+ base64: string;
478
+ }[];
479
+ contentHash: string;
480
+ modelResources?: {
481
+ taskDefinition: {
482
+ schemaVersion: "openpond.taskDefinition.v1";
483
+ id: string;
484
+ revision: number;
485
+ name: string;
486
+ description: string;
487
+ instructions: string;
488
+ category: "custom" | "structured" | "question_answering" | "coding" | "tool_workflow";
489
+ familyNamespace: string;
490
+ inputSchema: Record<string, unknown>;
491
+ outputSchema: Record<string, unknown>;
492
+ rewardBinding: {
493
+ id: string;
494
+ contentHash: string;
495
+ revision: number;
496
+ };
497
+ harness: {
498
+ id: string;
499
+ contentHash: string;
500
+ } | null;
501
+ execution: {
502
+ policy: {
503
+ policyVisibleFields: string[];
504
+ privilegedFields: string[];
505
+ hiddenGraderRefs: string[];
506
+ connectedAppScopes: string[];
507
+ };
508
+ environment: {
509
+ protocolVersion: "openpond.environment.v1";
510
+ kind: "text" | "agent" | "work" | "custom_program";
511
+ entrypoint: string;
512
+ stateful: boolean;
513
+ deterministicSeeds: boolean;
514
+ lifecycle: ("create" | "reset" | "step" | "collect" | "destroy")[];
515
+ networkPolicy: "none" | "declared_read_only" | "declared_scoped";
516
+ defaultTimeoutMs: number;
517
+ };
518
+ tools: {
519
+ name: string;
520
+ description: string;
521
+ inputSchema: Record<string, unknown>;
522
+ inputSchemaHash: string;
523
+ sideEffect: "read" | "write";
524
+ timeoutMs: number;
525
+ }[];
526
+ capabilities: {
527
+ id: string;
528
+ required: boolean;
529
+ scopes: string[];
530
+ portability: "portable" | "host_adapter" | "local_only" | "hosted_only";
531
+ }[];
532
+ environmentRelease?: {
533
+ id: string;
534
+ contentHash: string;
535
+ } | undefined;
536
+ verifierSetRelease?: {
537
+ id: string;
538
+ contentHash: string;
539
+ } | undefined;
540
+ };
541
+ contentHash: string;
542
+ requiredOutputs?: {
543
+ path: string;
544
+ mediaType: string;
545
+ schemaRef: {
546
+ id: string;
547
+ path: string;
548
+ contentHash: string;
549
+ sizeBytes: number;
550
+ mediaType: string;
551
+ visibility: "policy" | "verifier" | "host_private";
552
+ } | null;
553
+ maxBytes: number | null;
554
+ metadata: Record<string, unknown>;
555
+ }[] | undefined;
556
+ contextRequirements?: string[] | undefined;
557
+ };
558
+ assets: {
559
+ schemaVersion: "openpond.learningTextAsset.v1";
560
+ id: string;
561
+ revision: 1;
562
+ asset: {
563
+ id: string;
564
+ path: string;
565
+ contentHash: string;
566
+ sizeBytes: number;
567
+ mediaType: string;
568
+ visibility: "policy" | "verifier" | "host_private";
569
+ };
570
+ text: string;
571
+ contentHash: string;
572
+ }[];
573
+ rewards: {
574
+ schemaVersion: "openpond.rewardRelease.v1";
575
+ id: string;
576
+ revision: number;
577
+ name: string;
578
+ description: string;
579
+ implementation: {
580
+ kind: "content" | "state" | "artifact" | "schema" | "runtime_event";
581
+ config: Record<string, unknown>;
582
+ } | {
583
+ kind: "model_judge";
584
+ rubricRef: {
585
+ id: string;
586
+ path: string;
587
+ contentHash: string;
588
+ sizeBytes: number;
589
+ mediaType: string;
590
+ visibility: "policy" | "verifier" | "host_private";
591
+ };
592
+ calibrationStatus: "passed" | "pending" | "failed";
593
+ model?: {
594
+ providerId: string;
595
+ modelId: string;
596
+ revision: string | null;
597
+ } | undefined;
598
+ temperature?: number | undefined;
599
+ } | {
600
+ timeoutMs: number;
601
+ networkPolicy: "none";
602
+ kind: "custom_verifier";
603
+ verifierRef: {
604
+ id: string;
605
+ path: string;
606
+ contentHash: string;
607
+ sizeBytes: number;
608
+ mediaType: string;
609
+ visibility: "policy" | "verifier" | "host_private";
610
+ };
611
+ exportName?: string | undefined;
612
+ runtime?: "isolated_javascript" | "sandbox_process" | undefined;
613
+ } | {
614
+ kind: "human";
615
+ rubricRef: {
616
+ id: string;
617
+ path: string;
618
+ contentHash: string;
619
+ sizeBytes: number;
620
+ mediaType: string;
621
+ visibility: "policy" | "verifier" | "host_private";
622
+ };
623
+ reviewerRole: string;
624
+ form?: {
625
+ schemaVersion: "openpond.humanForm.v1";
626
+ mode: "individual" | "pairwise";
627
+ instructions: string;
628
+ criteria: ({
629
+ id: string;
630
+ label: string;
631
+ instructions: string;
632
+ required: boolean;
633
+ allowAbstain: boolean;
634
+ kind: "score";
635
+ minimum: number;
636
+ maximum: number;
637
+ step: number;
638
+ requireRationale?: boolean | undefined;
639
+ anchors?: {
640
+ value: string | number | boolean;
641
+ description: string;
642
+ }[] | undefined;
643
+ } | {
644
+ id: string;
645
+ label: string;
646
+ instructions: string;
647
+ required: boolean;
648
+ allowAbstain: boolean;
649
+ kind: "category";
650
+ options: string[];
651
+ requireRationale?: boolean | undefined;
652
+ anchors?: {
653
+ value: string | number | boolean;
654
+ description: string;
655
+ }[] | undefined;
656
+ } | {
657
+ id: string;
658
+ label: string;
659
+ instructions: string;
660
+ required: boolean;
661
+ allowAbstain: boolean;
662
+ kind: "boolean";
663
+ requireRationale?: boolean | undefined;
664
+ anchors?: {
665
+ value: string | number | boolean;
666
+ description: string;
667
+ }[] | undefined;
668
+ } | {
669
+ id: string;
670
+ label: string;
671
+ instructions: string;
672
+ required: boolean;
673
+ allowAbstain: boolean;
674
+ kind: "text";
675
+ maximumLength: number;
676
+ requireRationale?: boolean | undefined;
677
+ anchors?: {
678
+ value: string | number | boolean;
679
+ description: string;
680
+ }[] | undefined;
681
+ minimumLength?: number | undefined;
682
+ } | {
683
+ id: string;
684
+ label: string;
685
+ instructions: string;
686
+ required: boolean;
687
+ allowAbstain: boolean;
688
+ kind: "preference";
689
+ allowTie: boolean;
690
+ requireRationale?: boolean | undefined;
691
+ anchors?: {
692
+ value: string | number | boolean;
693
+ description: string;
694
+ }[] | undefined;
695
+ requireStrength?: boolean | undefined;
696
+ })[];
697
+ } | undefined;
698
+ } | {
699
+ kind: "learned_model";
700
+ modelVersion: {
701
+ id: string;
702
+ contentHash: string;
703
+ };
704
+ inputContract: {
705
+ id: string;
706
+ path: string;
707
+ contentHash: string;
708
+ sizeBytes: number;
709
+ mediaType: string;
710
+ visibility: "policy" | "verifier" | "host_private";
711
+ };
712
+ };
713
+ rawScore: {
714
+ minimum: number;
715
+ maximum: number;
716
+ };
717
+ assets: {
718
+ id: string;
719
+ path: string;
720
+ contentHash: string;
721
+ sizeBytes: number;
722
+ mediaType: string;
723
+ visibility: "policy" | "verifier" | "host_private";
724
+ }[];
725
+ contentHash: string;
726
+ feedbackKey?: string | undefined;
727
+ fixtureSetRef?: {
728
+ id: string;
729
+ path: string;
730
+ contentHash: string;
731
+ sizeBytes: number;
732
+ mediaType: string;
733
+ visibility: "policy" | "verifier" | "host_private";
734
+ } | undefined;
735
+ calibrationCheckRef?: {
736
+ id: string;
737
+ contentHash: string;
738
+ revision: number;
739
+ } | undefined;
740
+ }[];
741
+ rewardBinding: {
742
+ schemaVersion: "openpond.rewardBinding.v1";
743
+ id: string;
744
+ revision: number;
745
+ sources: {
746
+ graderId: string;
747
+ reward: {
748
+ id: string;
749
+ contentHash: string;
750
+ revision: number;
751
+ };
752
+ role: "training" | "evaluation";
753
+ normalization: {
754
+ kind: "identity";
755
+ } | {
756
+ kind: "linear";
757
+ minimum: number;
758
+ maximum: number;
759
+ direction: "higher" | "lower";
760
+ };
761
+ weight: number;
762
+ required: boolean;
763
+ hardGate: boolean;
764
+ privileged: boolean;
765
+ fixtureRefs: {
766
+ id: string;
767
+ contentHash: string;
768
+ }[];
769
+ }[];
770
+ aggregation: "weighted_mean";
771
+ unscorable: "exclude_optional_require_all_required";
772
+ contentHash: string;
773
+ name?: string | undefined;
774
+ description?: string | undefined;
775
+ recipeRef?: {
776
+ id: string;
777
+ contentHash: string;
778
+ revision: number;
779
+ } | undefined;
780
+ };
781
+ execution?: {
782
+ environment: {
783
+ schemaVersion: "openpond.environmentRelease.v1";
784
+ id: string;
785
+ revision: number;
786
+ contract: {
787
+ protocolVersion: "openpond.environment.v1";
788
+ kind: "text" | "agent" | "work" | "custom_program";
789
+ entrypoint: string;
790
+ stateful: boolean;
791
+ deterministicSeeds: boolean;
792
+ lifecycle: ("create" | "reset" | "step" | "collect" | "destroy")[];
793
+ networkPolicy: "none" | "declared_read_only" | "declared_scoped";
794
+ defaultTimeoutMs: number;
795
+ };
796
+ actionSchemaRef: {
797
+ id: string;
798
+ path: string;
799
+ contentHash: string;
800
+ sizeBytes: number;
801
+ mediaType: string;
802
+ visibility: "policy" | "verifier" | "host_private";
803
+ } | null;
804
+ observationSchemaRef: {
805
+ id: string;
806
+ path: string;
807
+ contentHash: string;
808
+ sizeBytes: number;
809
+ mediaType: string;
810
+ visibility: "policy" | "verifier" | "host_private";
811
+ } | null;
812
+ stateSchemaRef: {
813
+ id: string;
814
+ path: string;
815
+ contentHash: string;
816
+ sizeBytes: number;
817
+ mediaType: string;
818
+ visibility: "policy" | "verifier" | "host_private";
819
+ } | null;
820
+ artifactCollection: {
821
+ maxArtifacts: number;
822
+ maxTotalBytes: number;
823
+ };
824
+ adapterConformanceHashes: Record<string, string>;
825
+ metadata: Record<string, unknown>;
826
+ contentHash: string;
827
+ };
828
+ verifierSet: {
829
+ schemaVersion: "openpond.verifierSetRelease.v1";
830
+ id: string;
831
+ revision: number;
832
+ graders: ({
833
+ id: string;
834
+ version: string;
835
+ weight: number;
836
+ hardGate: boolean;
837
+ rewardEligible: boolean;
838
+ privileged: boolean;
839
+ kind: "content" | "state" | "artifact" | "schema" | "runtime_event";
840
+ config: Record<string, unknown>;
841
+ } | {
842
+ id: string;
843
+ version: string;
844
+ weight: number;
845
+ hardGate: boolean;
846
+ rewardEligible: boolean;
847
+ privileged: boolean;
848
+ kind: "model_judge";
849
+ rubricRef: {
850
+ id: string;
851
+ path: string;
852
+ contentHash: string;
853
+ sizeBytes: number;
854
+ mediaType: string;
855
+ visibility: "policy" | "verifier" | "host_private";
856
+ };
857
+ calibrationStatus: "passed" | "pending" | "failed";
858
+ model?: {
859
+ providerId: string;
860
+ modelId: string;
861
+ revision: string | null;
862
+ } | undefined;
863
+ temperature?: number | undefined;
864
+ } | {
865
+ id: string;
866
+ version: string;
867
+ weight: number;
868
+ hardGate: boolean;
869
+ rewardEligible: boolean;
870
+ privileged: boolean;
871
+ kind: "custom_verifier";
872
+ verifierRef: {
873
+ id: string;
874
+ path: string;
875
+ contentHash: string;
876
+ sizeBytes: number;
877
+ mediaType: string;
878
+ visibility: "policy" | "verifier" | "host_private";
879
+ };
880
+ timeoutMs: number;
881
+ networkPolicy: "none";
882
+ runtime?: "isolated_javascript" | "sandbox_process" | undefined;
883
+ exportName?: string | undefined;
884
+ } | {
885
+ id: string;
886
+ version: string;
887
+ weight: number;
888
+ hardGate: boolean;
889
+ rewardEligible: boolean;
890
+ privileged: boolean;
891
+ kind: "human";
892
+ rubricRef: {
893
+ id: string;
894
+ path: string;
895
+ contentHash: string;
896
+ sizeBytes: number;
897
+ mediaType: string;
898
+ visibility: "policy" | "verifier" | "host_private";
899
+ };
900
+ reviewerRole: string;
901
+ form?: {
902
+ schemaVersion: "openpond.humanForm.v1";
903
+ mode: "individual" | "pairwise";
904
+ instructions: string;
905
+ criteria: ({
906
+ id: string;
907
+ label: string;
908
+ instructions: string;
909
+ required: boolean;
910
+ allowAbstain: boolean;
911
+ kind: "score";
912
+ minimum: number;
913
+ maximum: number;
914
+ step: number;
915
+ requireRationale?: boolean | undefined;
916
+ anchors?: {
917
+ value: string | number | boolean;
918
+ description: string;
919
+ }[] | undefined;
920
+ } | {
921
+ id: string;
922
+ label: string;
923
+ instructions: string;
924
+ required: boolean;
925
+ allowAbstain: boolean;
926
+ kind: "category";
927
+ options: string[];
928
+ requireRationale?: boolean | undefined;
929
+ anchors?: {
930
+ value: string | number | boolean;
931
+ description: string;
932
+ }[] | undefined;
933
+ } | {
934
+ id: string;
935
+ label: string;
936
+ instructions: string;
937
+ required: boolean;
938
+ allowAbstain: boolean;
939
+ kind: "boolean";
940
+ requireRationale?: boolean | undefined;
941
+ anchors?: {
942
+ value: string | number | boolean;
943
+ description: string;
944
+ }[] | undefined;
945
+ } | {
946
+ id: string;
947
+ label: string;
948
+ instructions: string;
949
+ required: boolean;
950
+ allowAbstain: boolean;
951
+ kind: "text";
952
+ maximumLength: number;
953
+ requireRationale?: boolean | undefined;
954
+ anchors?: {
955
+ value: string | number | boolean;
956
+ description: string;
957
+ }[] | undefined;
958
+ minimumLength?: number | undefined;
959
+ } | {
960
+ id: string;
961
+ label: string;
962
+ instructions: string;
963
+ required: boolean;
964
+ allowAbstain: boolean;
965
+ kind: "preference";
966
+ allowTie: boolean;
967
+ requireRationale?: boolean | undefined;
968
+ anchors?: {
969
+ value: string | number | boolean;
970
+ description: string;
971
+ }[] | undefined;
972
+ requireStrength?: boolean | undefined;
973
+ })[];
974
+ } | undefined;
975
+ })[];
976
+ isolation: {
977
+ processBoundary: "same_process" | "isolated_process" | "container";
978
+ networkPolicy: "none";
979
+ defaultTimeoutMs: number;
980
+ };
981
+ calibrationReceiptRefs: {
982
+ id: string;
983
+ contentHash: string;
984
+ }[];
985
+ metadata: Record<string, unknown>;
986
+ contentHash: string;
987
+ };
988
+ javascript: {
989
+ schemaVersion: "openpond.javascriptEnvironment.v1";
990
+ id: string;
991
+ revision: number;
992
+ module: {
993
+ id: string;
994
+ path: string;
995
+ contentHash: string;
996
+ sizeBytes: number;
997
+ mediaType: string;
998
+ visibility: "policy" | "verifier" | "host_private";
999
+ };
1000
+ tools: {
1001
+ name: string;
1002
+ description: string;
1003
+ inputSchema: Record<string, unknown>;
1004
+ inputSchemaHash: string;
1005
+ sideEffect: "read" | "write";
1006
+ timeoutMs: number;
1007
+ }[];
1008
+ maxSteps: number;
1009
+ maxStateBytes: number;
1010
+ maxObservationBytes: number;
1011
+ operationTimeoutMs: number;
1012
+ contentHash: string;
1013
+ executionServices?: ({
1014
+ kind: "sqlite.v1";
1015
+ sql: {
1016
+ scope: "input" | "state" | "arguments" | "initialState";
1017
+ path: string[];
1018
+ };
1019
+ snapshot: {
1020
+ scope: "input" | "state" | "arguments" | "initialState";
1021
+ path: string[];
1022
+ };
1023
+ maxRows: number;
1024
+ maxResultBytes: number;
1025
+ id: string;
1026
+ operation: "step" | "collect";
1027
+ timeoutMs: number;
1028
+ toolName?: string | undefined;
1029
+ } | {
1030
+ kind: "javascript.v1";
1031
+ source: {
1032
+ scope: "input" | "state" | "arguments" | "initialState";
1033
+ path: string[];
1034
+ };
1035
+ cases: {
1036
+ scope: "input" | "state" | "arguments" | "initialState";
1037
+ path: string[];
1038
+ };
1039
+ exportName: string;
1040
+ maxCases: number;
1041
+ maxResultBytes: number;
1042
+ id: string;
1043
+ operation: "step" | "collect";
1044
+ timeoutMs: number;
1045
+ toolName?: string | undefined;
1046
+ })[] | undefined;
1047
+ };
1048
+ } | undefined;
1049
+ instructionMode?: "per_task" | undefined;
1050
+ } | undefined;
1051
+ learningResources?: {
1052
+ batch: {
1053
+ schemaVersion: "openpond.taskBatch.v1";
1054
+ id: string;
1055
+ revision: 1;
1056
+ taskDefinition: {
1057
+ id: string;
1058
+ contentHash: string;
1059
+ revision: number;
1060
+ };
1061
+ rewardBinding: {
1062
+ id: string;
1063
+ contentHash: string;
1064
+ revision: number;
1065
+ };
1066
+ purpose: "evaluation" | "supervised_training" | "reward_training";
1067
+ examples: {
1068
+ evidence: {
1069
+ id: string;
1070
+ contentHash: string;
1071
+ revision: number;
1072
+ };
1073
+ decision: {
1074
+ id: string;
1075
+ contentHash: string;
1076
+ revision: number;
1077
+ };
1078
+ familyKey: string;
1079
+ inputHash: string;
1080
+ split: "train" | "validation" | "test" | "frozen_eval";
1081
+ }[];
1082
+ sealedAt: string;
1083
+ sealedBy: string;
1084
+ contentHash: string;
1085
+ };
1086
+ evidence: {
1087
+ schemaVersion: "openpond.taskEvidence.v1";
1088
+ id: string;
1089
+ revision: number;
1090
+ source: {
1091
+ id: string;
1092
+ contentHash: string;
1093
+ revision: number;
1094
+ };
1095
+ submission: {
1096
+ schemaVersion: "openpond.taskExample.v1";
1097
+ sourceId: string;
1098
+ idempotencyKey: string;
1099
+ taskDefinition: {
1100
+ id: string;
1101
+ contentHash: string;
1102
+ revision: number;
1103
+ };
1104
+ exampleId: string;
1105
+ attemptId: string;
1106
+ occurredAt: string;
1107
+ familyKey: string | null;
1108
+ split: "train" | "validation" | "test" | "frozen_eval";
1109
+ input: Record<string, z.core.util.JSONType>;
1110
+ observedOutput: Record<string, z.core.util.JSONType> | null;
1111
+ expected: Record<string, z.core.util.JSONType> | null;
1112
+ evaluatorContext: Record<string, z.core.util.JSONType> | null;
1113
+ assets: {
1114
+ id: string;
1115
+ path: string;
1116
+ contentHash: string;
1117
+ sizeBytes: number;
1118
+ mediaType: string;
1119
+ visibility: "policy" | "verifier" | "host_private";
1120
+ }[];
1121
+ provenance: {
1122
+ sourceRecordRef: string | null;
1123
+ mappingHash: string | null;
1124
+ };
1125
+ requiredOutputs?: {
1126
+ path: string;
1127
+ mediaType: string;
1128
+ schemaRef: {
1129
+ id: string;
1130
+ path: string;
1131
+ contentHash: string;
1132
+ sizeBytes: number;
1133
+ mediaType: string;
1134
+ visibility: "policy" | "verifier" | "host_private";
1135
+ } | null;
1136
+ maxBytes: number | null;
1137
+ metadata: Record<string, unknown>;
1138
+ }[] | undefined;
1139
+ };
1140
+ supersedes: {
1141
+ id: string;
1142
+ contentHash: string;
1143
+ revision: number;
1144
+ } | null;
1145
+ correctionFeedbackId: string | null;
1146
+ receivedAt: string;
1147
+ contentHash: string;
1148
+ }[];
1149
+ decisions: {
1150
+ schemaVersion: "openpond.taskAdmissionDecision.v1";
1151
+ id: string;
1152
+ revision: number;
1153
+ evidence: {
1154
+ id: string;
1155
+ contentHash: string;
1156
+ revision: number;
1157
+ };
1158
+ supersedes: {
1159
+ id: string;
1160
+ contentHash: string;
1161
+ revision: number;
1162
+ } | null;
1163
+ actor: {
1164
+ kind: "human" | "qualified_policy";
1165
+ id: string;
1166
+ policy: {
1167
+ id: string;
1168
+ contentHash: string;
1169
+ revision: number;
1170
+ } | null;
1171
+ };
1172
+ evidenceValidity: "pending" | "valid" | "invalid";
1173
+ taskAdmissibility: "pending" | "approved" | "rejected";
1174
+ observedQuality: "passed" | "failed" | "unavailable" | "unscored";
1175
+ targetApproval: "pending" | "approved" | "rejected" | "not_required";
1176
+ approvedTarget: Record<string, z.core.util.JSONType> | null;
1177
+ grade: {
1178
+ schemaVersion: "openpond.rewardComposition.v1";
1179
+ binding: {
1180
+ id: string;
1181
+ contentHash: string;
1182
+ revision: number;
1183
+ };
1184
+ taskHash: string;
1185
+ outputHash: string;
1186
+ results: {
1187
+ graderId: string;
1188
+ reward: {
1189
+ id: string;
1190
+ contentHash: string;
1191
+ revision: number;
1192
+ };
1193
+ role: "training" | "evaluation";
1194
+ status: "scored" | "pending" | "failed" | "unavailable";
1195
+ rawScore: number | null;
1196
+ normalizedScore: number | null;
1197
+ passed: boolean | null;
1198
+ evidenceHashes: string[];
1199
+ message: string | null;
1200
+ graderEvidence?: {
1201
+ schemaVersion: "openpond.graderEvidence.v1";
1202
+ graderId: string;
1203
+ graderVersion: string;
1204
+ score: number | null;
1205
+ passed: boolean;
1206
+ rewardEligible: boolean;
1207
+ failureClass: "policy_failure" | "grader_failure" | "environment_failure" | "infrastructure_failure" | "timeout" | "cancelled" | null;
1208
+ feedback: string[];
1209
+ visibleEvidenceRefs: string[];
1210
+ privilegedEvidenceRefs: string[];
1211
+ contentHash: string;
1212
+ status?: "scored" | "pending" | "failed" | "unavailable" | undefined;
1213
+ modelJudgeReceipt?: {
1214
+ schemaVersion: "openpond.modelJudgeReceipt.v1";
1215
+ providerId: string;
1216
+ modelId: string;
1217
+ modelRevision: string | null;
1218
+ responseId: string | null;
1219
+ requestHash: string;
1220
+ responseHash: string;
1221
+ inputTokens: number | null;
1222
+ outputTokens: number | null;
1223
+ costUsd: number | null;
1224
+ } | undefined;
1225
+ } | undefined;
1226
+ }[];
1227
+ training: {
1228
+ status: "scored" | "pending" | "unscorable" | "not_configured";
1229
+ score: number | null;
1230
+ passed: boolean | null;
1231
+ };
1232
+ evaluation: {
1233
+ status: "scored" | "pending" | "unscorable" | "not_configured";
1234
+ score: number | null;
1235
+ passed: boolean | null;
1236
+ };
1237
+ contentHash: string;
1238
+ } | null;
1239
+ targetGrade: {
1240
+ schemaVersion: "openpond.rewardComposition.v1";
1241
+ binding: {
1242
+ id: string;
1243
+ contentHash: string;
1244
+ revision: number;
1245
+ };
1246
+ taskHash: string;
1247
+ outputHash: string;
1248
+ results: {
1249
+ graderId: string;
1250
+ reward: {
1251
+ id: string;
1252
+ contentHash: string;
1253
+ revision: number;
1254
+ };
1255
+ role: "training" | "evaluation";
1256
+ status: "scored" | "pending" | "failed" | "unavailable";
1257
+ rawScore: number | null;
1258
+ normalizedScore: number | null;
1259
+ passed: boolean | null;
1260
+ evidenceHashes: string[];
1261
+ message: string | null;
1262
+ graderEvidence?: {
1263
+ schemaVersion: "openpond.graderEvidence.v1";
1264
+ graderId: string;
1265
+ graderVersion: string;
1266
+ score: number | null;
1267
+ passed: boolean;
1268
+ rewardEligible: boolean;
1269
+ failureClass: "policy_failure" | "grader_failure" | "environment_failure" | "infrastructure_failure" | "timeout" | "cancelled" | null;
1270
+ feedback: string[];
1271
+ visibleEvidenceRefs: string[];
1272
+ privilegedEvidenceRefs: string[];
1273
+ contentHash: string;
1274
+ status?: "scored" | "pending" | "failed" | "unavailable" | undefined;
1275
+ modelJudgeReceipt?: {
1276
+ schemaVersion: "openpond.modelJudgeReceipt.v1";
1277
+ providerId: string;
1278
+ modelId: string;
1279
+ modelRevision: string | null;
1280
+ responseId: string | null;
1281
+ requestHash: string;
1282
+ responseHash: string;
1283
+ inputTokens: number | null;
1284
+ outputTokens: number | null;
1285
+ costUsd: number | null;
1286
+ } | undefined;
1287
+ } | undefined;
1288
+ }[];
1289
+ training: {
1290
+ status: "scored" | "pending" | "unscorable" | "not_configured";
1291
+ score: number | null;
1292
+ passed: boolean | null;
1293
+ };
1294
+ evaluation: {
1295
+ status: "scored" | "pending" | "unscorable" | "not_configured";
1296
+ score: number | null;
1297
+ passed: boolean | null;
1298
+ };
1299
+ contentHash: string;
1300
+ } | null;
1301
+ note: string;
1302
+ decidedAt: string;
1303
+ contentHash: string;
1304
+ }[];
1305
+ sources: {
1306
+ schemaVersion: "openpond.learningSource.v1";
1307
+ id: string;
1308
+ revision: number;
1309
+ name: string;
1310
+ kind: "provider" | "work" | "benchmark" | "direct" | "opentelemetry";
1311
+ taskDefinition: {
1312
+ id: string;
1313
+ contentHash: string;
1314
+ revision: number;
1315
+ };
1316
+ enabled: boolean;
1317
+ allowedSplits: ("train" | "validation" | "test" | "frozen_eval")[];
1318
+ mapping: {
1319
+ schemaVersion: "openpond.taskSourceMapping.v1";
1320
+ exampleId: string;
1321
+ attemptId: string;
1322
+ occurredAt: string;
1323
+ input: string;
1324
+ observedOutput: string | null;
1325
+ expected: string | null;
1326
+ evaluatorContext: string | null;
1327
+ familyKey: string;
1328
+ split: "train" | "validation" | "test" | "frozen_eval";
1329
+ } | null;
1330
+ adapterVersion: string | null;
1331
+ contentHash: string;
1332
+ reviewOrigin?: {
1333
+ modelId: string;
1334
+ packageHash: string;
1335
+ taskset: {
1336
+ id: string;
1337
+ contentHash: string;
1338
+ revision: number;
1339
+ };
1340
+ batch: {
1341
+ id: string;
1342
+ contentHash: string;
1343
+ revision: number;
1344
+ };
1345
+ rewardBinding: {
1346
+ id: string;
1347
+ contentHash: string;
1348
+ revision: number;
1349
+ };
1350
+ } | undefined;
1351
+ }[];
1352
+ assets: {
1353
+ schemaVersion: "openpond.learningTextAsset.v1";
1354
+ id: string;
1355
+ revision: 1;
1356
+ asset: {
1357
+ id: string;
1358
+ path: string;
1359
+ contentHash: string;
1360
+ sizeBytes: number;
1361
+ mediaType: string;
1362
+ visibility: "policy" | "verifier" | "host_private";
1363
+ };
1364
+ text: string;
1365
+ contentHash: string;
1366
+ }[];
1367
+ } | undefined;
1368
+ };
1369
+ //# sourceMappingURL=profile-external-dataset-package.d.ts.map