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 +1,1127 @@
1
+ import { z } from "zod";
2
+ export declare const ProfileEvaluationDiscoverySchema: z.ZodObject<{
3
+ profileRef: z.ZodObject<{
4
+ source: z.ZodEnum<{
5
+ local: "local";
6
+ github: "github";
7
+ openpond_git: "openpond_git";
8
+ }>;
9
+ repositoryId: z.ZodString;
10
+ profileId: z.ZodString;
11
+ }, z.core.$strict>;
12
+ sourceRevision: z.ZodString;
13
+ harnessRelease: z.ZodObject<{
14
+ id: z.ZodString;
15
+ contentHash: z.ZodString;
16
+ }, z.core.$strict>;
17
+ catalogHash: z.ZodNullable<z.ZodString>;
18
+ definitions: z.ZodArray<z.ZodObject<{
19
+ id: z.ZodString;
20
+ label: z.ZodString;
21
+ description: z.ZodString;
22
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
23
+ kind: z.ZodLiteral<"profile">;
24
+ }, z.core.$strict>, z.ZodObject<{
25
+ kind: z.ZodLiteral<"workflow">;
26
+ workflowId: z.ZodString;
27
+ }, z.core.$strict>, z.ZodObject<{
28
+ kind: z.ZodLiteral<"skill">;
29
+ skillPath: z.ZodString;
30
+ }, z.core.$strict>, z.ZodObject<{
31
+ kind: z.ZodLiteral<"agent_action">;
32
+ actionId: z.ZodString;
33
+ }, z.core.$strict>], "kind">;
34
+ tasksetRelease: z.ZodObject<{
35
+ id: z.ZodString;
36
+ contentHash: z.ZodString;
37
+ }, z.core.$strict>;
38
+ split: z.ZodEnum<{
39
+ train: "train";
40
+ validation: "validation";
41
+ test: "test";
42
+ frozen_eval: "frozen_eval";
43
+ }>;
44
+ taskIds: z.ZodArray<z.ZodString>;
45
+ seeds: z.ZodArray<z.ZodString>;
46
+ criterion: z.ZodObject<{
47
+ minimumPassRate: z.ZodNumber;
48
+ requireComplete: z.ZodBoolean;
49
+ }, z.core.$strict>;
50
+ }, z.core.$strict>>;
51
+ suites: z.ZodArray<z.ZodObject<{
52
+ id: z.ZodString;
53
+ label: z.ZodString;
54
+ scope: z.ZodEnum<{
55
+ profile: "profile";
56
+ component: "component";
57
+ }>;
58
+ definitionIds: z.ZodArray<z.ZodString>;
59
+ }, z.core.$strict>>;
60
+ runs: z.ZodArray<z.ZodObject<{
61
+ profileRef: z.ZodObject<{
62
+ source: z.ZodEnum<{
63
+ local: "local";
64
+ github: "github";
65
+ openpond_git: "openpond_git";
66
+ }>;
67
+ repositoryId: z.ZodString;
68
+ profileId: z.ZodString;
69
+ }, z.core.$strict>;
70
+ manifest: z.ZodObject<{
71
+ schemaVersion: z.ZodLiteral<"openpond.tasksetRunManifest.v1">;
72
+ id: z.ZodString;
73
+ tasksetRelease: z.ZodObject<{
74
+ id: z.ZodString;
75
+ contentHash: z.ZodString;
76
+ }, z.core.$strict>;
77
+ packageHash: z.ZodString;
78
+ execution: z.ZodDiscriminatedUnion<[z.ZodObject<{
79
+ kind: z.ZodLiteral<"taskset">;
80
+ environmentRelease: z.ZodObject<{
81
+ id: z.ZodString;
82
+ contentHash: z.ZodString;
83
+ }, z.core.$strict>;
84
+ verifierSetRelease: z.ZodObject<{
85
+ id: z.ZodString;
86
+ contentHash: z.ZodString;
87
+ }, z.core.$strict>;
88
+ policyHash: z.ZodString;
89
+ }, z.core.$strict>, z.ZodObject<{
90
+ kind: z.ZodLiteral<"harness">;
91
+ harnessRelease: z.ZodObject<{
92
+ id: z.ZodString;
93
+ contentHash: z.ZodString;
94
+ }, z.core.$strict>;
95
+ }, z.core.$strict>], "kind">;
96
+ profileEvaluation: z.ZodOptional<z.ZodObject<{
97
+ profileId: z.ZodString;
98
+ sourceRevision: z.ZodString;
99
+ harnessRelease: z.ZodObject<{
100
+ id: z.ZodString;
101
+ contentHash: z.ZodString;
102
+ }, z.core.$strict>;
103
+ catalogHash: z.ZodString;
104
+ definitionId: z.ZodString;
105
+ definitionHash: z.ZodString;
106
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
107
+ kind: z.ZodLiteral<"profile">;
108
+ }, z.core.$strict>, z.ZodObject<{
109
+ kind: z.ZodLiteral<"workflow">;
110
+ workflowId: z.ZodString;
111
+ }, z.core.$strict>, z.ZodObject<{
112
+ kind: z.ZodLiteral<"skill">;
113
+ skillPath: z.ZodString;
114
+ }, z.core.$strict>, z.ZodObject<{
115
+ kind: z.ZodLiteral<"agent_action">;
116
+ actionId: z.ZodString;
117
+ }, z.core.$strict>], "kind">;
118
+ environmentHash: z.ZodString;
119
+ externalDatasetBinding: z.ZodOptional<z.ZodObject<{
120
+ schemaVersion: z.ZodLiteral<"openpond.profileExternalDatasetBinding.v1">;
121
+ dataset: z.ZodObject<{
122
+ id: z.ZodString;
123
+ contentHash: z.ZodString;
124
+ revision: z.ZodNumber;
125
+ }, z.core.$strict>;
126
+ packageHash: z.ZodString;
127
+ profileId: z.ZodString;
128
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
129
+ kind: z.ZodLiteral<"profile">;
130
+ }, z.core.$strict>, z.ZodObject<{
131
+ kind: z.ZodLiteral<"workflow">;
132
+ workflowId: z.ZodString;
133
+ }, z.core.$strict>, z.ZodObject<{
134
+ kind: z.ZodLiteral<"skill">;
135
+ skillPath: z.ZodString;
136
+ }, z.core.$strict>, z.ZodObject<{
137
+ kind: z.ZodLiteral<"agent_action">;
138
+ actionId: z.ZodString;
139
+ }, z.core.$strict>], "kind">;
140
+ declaredProfileCatalogHash: z.ZodNullable<z.ZodString>;
141
+ protectedProfileClosureHash: z.ZodString;
142
+ recordedOrigin: z.ZodOptional<z.ZodObject<{
143
+ execution: z.ZodObject<{
144
+ id: z.ZodString;
145
+ contentHash: z.ZodString;
146
+ }, z.core.$strict>;
147
+ originalEvidence: z.ZodObject<{
148
+ id: z.ZodString;
149
+ contentHash: z.ZodString;
150
+ }, z.core.$strict>;
151
+ members: z.ZodArray<z.ZodObject<{
152
+ caseId: z.ZodString;
153
+ seed: z.ZodString;
154
+ fixtureId: z.ZodString;
155
+ sourceId: z.ZodString;
156
+ snapshotHash: z.ZodString;
157
+ boundaryRevisionHash: z.ZodString;
158
+ }, z.core.$strict>>;
159
+ }, z.core.$strict>>;
160
+ gradingSource: z.ZodOptional<z.ZodObject<{
161
+ pass: z.ZodObject<{
162
+ id: z.ZodString;
163
+ contentHash: z.ZodString;
164
+ }, z.core.$strict>;
165
+ evidence: z.ZodObject<{
166
+ id: z.ZodString;
167
+ contentHash: z.ZodString;
168
+ }, z.core.$strict>;
169
+ }, z.core.$strict>>;
170
+ split: z.ZodEnum<{
171
+ train: "train";
172
+ validation: "validation";
173
+ test: "test";
174
+ frozen_eval: "frozen_eval";
175
+ }>;
176
+ population: z.ZodArray<z.ZodObject<{
177
+ taskId: z.ZodString;
178
+ seed: z.ZodString;
179
+ fixtureId: z.ZodNullable<z.ZodString>;
180
+ }, z.core.$strict>>;
181
+ evaluators: z.ZodArray<z.ZodObject<{
182
+ release: z.ZodObject<{
183
+ id: z.ZodString;
184
+ contentHash: z.ZodString;
185
+ revision: z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>;
186
+ }, z.core.$strict>;
187
+ feedbackKey: z.ZodString;
188
+ configurationHash: z.ZodString;
189
+ output: z.ZodEnum<{
190
+ boolean: "boolean";
191
+ score: "score";
192
+ category: "category";
193
+ }>;
194
+ categories: z.ZodArray<z.ZodString>;
195
+ }, z.core.$strict>>;
196
+ fieldMappingsHash: z.ZodString;
197
+ fieldMappings: z.ZodArray<z.ZodObject<{
198
+ graderId: z.ZodString;
199
+ fields: z.ZodArray<z.ZodObject<{
200
+ destination: z.ZodEnum<{
201
+ output: "output";
202
+ input: "input";
203
+ expectedOutput: "expectedOutput";
204
+ evaluatorContext: "evaluatorContext";
205
+ }>;
206
+ path: z.ZodArray<z.ZodString>;
207
+ source: z.ZodEnum<{
208
+ output: "output";
209
+ input: "input";
210
+ expectedOutput: "expectedOutput";
211
+ evaluatorContext: "evaluatorContext";
212
+ }>;
213
+ sourcePath: z.ZodArray<z.ZodString>;
214
+ }, z.core.$strict>>;
215
+ }, z.core.$strict>>;
216
+ environmentHash: z.ZodString;
217
+ privateDatasetClosureHash: z.ZodString;
218
+ criterion: z.ZodObject<{
219
+ minimumPassRate: z.ZodNumber;
220
+ requireComplete: z.ZodBoolean;
221
+ }, z.core.$strict>;
222
+ contentHash: z.ZodString;
223
+ }, z.core.$strict>>;
224
+ sourceCandidate: z.ZodOptional<z.ZodObject<{
225
+ candidateId: z.ZodString;
226
+ frozenRevision: z.ZodNumber;
227
+ frozenHash: z.ZodString;
228
+ harnessRelease: z.ZodObject<{
229
+ id: z.ZodString;
230
+ contentHash: z.ZodString;
231
+ }, z.core.$strict>;
232
+ protectedClosureHash: z.ZodString;
233
+ }, z.core.$strict>>;
234
+ }, z.core.$strict>>;
235
+ policy: z.ZodDiscriminatedUnion<[z.ZodObject<{
236
+ kind: z.ZodLiteral<"model">;
237
+ model: z.ZodObject<{
238
+ provider: z.ZodString;
239
+ model: z.ZodString;
240
+ revision: z.ZodDefault<z.ZodNullable<z.ZodString>>;
241
+ artifactHash: z.ZodDefault<z.ZodNullable<z.ZodString>>;
242
+ tokenizerRevision: z.ZodDefault<z.ZodNullable<z.ZodString>>;
243
+ chatTemplateHash: z.ZodDefault<z.ZodNullable<z.ZodString>>;
244
+ }, z.core.$strict>;
245
+ configurationHash: z.ZodString;
246
+ }, z.core.$strict>, z.ZodObject<{
247
+ kind: z.ZodLiteral<"fixture">;
248
+ }, z.core.$strict>], "kind">;
249
+ gradingRole: z.ZodLiteral<"evaluation">;
250
+ metricPolicy: z.ZodObject<{
251
+ schemaVersion: z.ZodLiteral<"openpond.tasksetMetricPolicy.v1">;
252
+ primaryMetric: z.ZodString;
253
+ aggregation: z.ZodEnum<{
254
+ mean_score: "mean_score";
255
+ pass_rate: "pass_rate";
256
+ weighted_mean: "weighted_mean";
257
+ custom: "custom";
258
+ }>;
259
+ missingReward: z.ZodEnum<{
260
+ zero: "zero";
261
+ exclude: "exclude";
262
+ }>;
263
+ taskWeights: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodNumber>>;
264
+ customAggregator: z.ZodNullable<z.ZodObject<{
265
+ module: z.ZodString;
266
+ exportName: z.ZodString;
267
+ contentHash: z.ZodString;
268
+ timeoutMs: z.ZodNumber;
269
+ networkPolicy: z.ZodLiteral<"none">;
270
+ }, z.core.$strict>>;
271
+ }, z.core.$strict>;
272
+ population: z.ZodArray<z.ZodObject<{
273
+ receiptId: z.ZodString;
274
+ taskId: z.ZodString;
275
+ seed: z.ZodString;
276
+ fixtureId: z.ZodNullable<z.ZodString>;
277
+ }, z.core.$strict>>;
278
+ runtimeTarget: z.ZodObject<{
279
+ adapterId: z.ZodString;
280
+ placement: z.ZodEnum<{
281
+ local: "local";
282
+ remote: "remote";
283
+ colocated: "colocated";
284
+ provider_native: "provider_native";
285
+ }>;
286
+ runtimeVersion: z.ZodString;
287
+ capabilityReceipt: z.ZodString;
288
+ }, z.core.$strict>;
289
+ limits: z.ZodObject<{
290
+ maxTurns: z.ZodNumber;
291
+ timeoutMs: z.ZodNumber;
292
+ maxOutputBytes: z.ZodNumber;
293
+ maximumSpendUsd: z.ZodNullable<z.ZodNumber>;
294
+ }, z.core.$strict>;
295
+ createdAt: z.ZodString;
296
+ metadata: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
297
+ contentHash: z.ZodString;
298
+ }, z.core.$strict>;
299
+ metric: z.ZodObject<{
300
+ schemaVersion: z.ZodLiteral<"openpond.tasksetMetricResult.v1">;
301
+ runManifest: z.ZodObject<{
302
+ id: z.ZodString;
303
+ contentHash: z.ZodString;
304
+ }, z.core.$strict>;
305
+ tasksetRelease: z.ZodObject<{
306
+ id: z.ZodString;
307
+ contentHash: z.ZodString;
308
+ }, z.core.$strict>;
309
+ policy: z.ZodObject<{
310
+ schemaVersion: z.ZodLiteral<"openpond.tasksetMetricPolicy.v1">;
311
+ primaryMetric: z.ZodString;
312
+ aggregation: z.ZodEnum<{
313
+ mean_score: "mean_score";
314
+ pass_rate: "pass_rate";
315
+ weighted_mean: "weighted_mean";
316
+ custom: "custom";
317
+ }>;
318
+ missingReward: z.ZodEnum<{
319
+ zero: "zero";
320
+ exclude: "exclude";
321
+ }>;
322
+ taskWeights: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodNumber>>;
323
+ customAggregator: z.ZodNullable<z.ZodObject<{
324
+ module: z.ZodString;
325
+ exportName: z.ZodString;
326
+ contentHash: z.ZodString;
327
+ timeoutMs: z.ZodNumber;
328
+ networkPolicy: z.ZodLiteral<"none">;
329
+ }, z.core.$strict>>;
330
+ }, z.core.$strict>;
331
+ policyHash: z.ZodString;
332
+ receiptRefs: z.ZodArray<z.ZodObject<{
333
+ id: z.ZodString;
334
+ contentHash: z.ZodString;
335
+ }, z.core.$strict>>;
336
+ includedCount: z.ZodNumber;
337
+ missingRewardCount: z.ZodNumber;
338
+ excludedCount: z.ZodNumber;
339
+ value: z.ZodNullable<z.ZodNumber>;
340
+ contentHash: z.ZodString;
341
+ }, z.core.$strict>;
342
+ gradeRefs: z.ZodArray<z.ZodObject<{
343
+ id: z.ZodString;
344
+ contentHash: z.ZodString;
345
+ }, z.core.$strict>>;
346
+ receiptRefs: z.ZodArray<z.ZodObject<{
347
+ id: z.ZodString;
348
+ contentHash: z.ZodString;
349
+ }, z.core.$strict>>;
350
+ passRate: z.ZodNumber;
351
+ passed: z.ZodBoolean;
352
+ completedAt: z.ZodISODateTime;
353
+ contentHash: z.ZodString;
354
+ }, z.core.$strict>>;
355
+ comparisons: z.ZodArray<z.ZodObject<{
356
+ schemaVersion: z.ZodLiteral<"openpond.profileEvaluationComparison.v1">;
357
+ id: z.ZodString;
358
+ tasksetRelease: z.ZodObject<{
359
+ id: z.ZodString;
360
+ contentHash: z.ZodString;
361
+ }, z.core.$strict>;
362
+ populationHash: z.ZodString;
363
+ environmentHash: z.ZodString;
364
+ metricPolicyHash: z.ZodString;
365
+ limitsHash: z.ZodString;
366
+ members: z.ZodArray<z.ZodObject<{
367
+ runManifest: z.ZodObject<{
368
+ id: z.ZodString;
369
+ contentHash: z.ZodString;
370
+ }, z.core.$strict>;
371
+ metricResultHash: z.ZodString;
372
+ source: z.ZodObject<{
373
+ profileId: z.ZodString;
374
+ sourceRevision: z.ZodString;
375
+ harnessRelease: z.ZodObject<{
376
+ id: z.ZodString;
377
+ contentHash: z.ZodString;
378
+ }, z.core.$strict>;
379
+ catalogHash: z.ZodString;
380
+ definitionId: z.ZodString;
381
+ definitionHash: z.ZodString;
382
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
383
+ kind: z.ZodLiteral<"profile">;
384
+ }, z.core.$strict>, z.ZodObject<{
385
+ kind: z.ZodLiteral<"workflow">;
386
+ workflowId: z.ZodString;
387
+ }, z.core.$strict>, z.ZodObject<{
388
+ kind: z.ZodLiteral<"skill">;
389
+ skillPath: z.ZodString;
390
+ }, z.core.$strict>, z.ZodObject<{
391
+ kind: z.ZodLiteral<"agent_action">;
392
+ actionId: z.ZodString;
393
+ }, z.core.$strict>], "kind">;
394
+ environmentHash: z.ZodString;
395
+ externalDatasetBinding: z.ZodOptional<z.ZodObject<{
396
+ schemaVersion: z.ZodLiteral<"openpond.profileExternalDatasetBinding.v1">;
397
+ dataset: z.ZodObject<{
398
+ id: z.ZodString;
399
+ contentHash: z.ZodString;
400
+ revision: z.ZodNumber;
401
+ }, z.core.$strict>;
402
+ packageHash: z.ZodString;
403
+ profileId: z.ZodString;
404
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
405
+ kind: z.ZodLiteral<"profile">;
406
+ }, z.core.$strict>, z.ZodObject<{
407
+ kind: z.ZodLiteral<"workflow">;
408
+ workflowId: z.ZodString;
409
+ }, z.core.$strict>, z.ZodObject<{
410
+ kind: z.ZodLiteral<"skill">;
411
+ skillPath: z.ZodString;
412
+ }, z.core.$strict>, z.ZodObject<{
413
+ kind: z.ZodLiteral<"agent_action">;
414
+ actionId: z.ZodString;
415
+ }, z.core.$strict>], "kind">;
416
+ declaredProfileCatalogHash: z.ZodNullable<z.ZodString>;
417
+ protectedProfileClosureHash: z.ZodString;
418
+ recordedOrigin: z.ZodOptional<z.ZodObject<{
419
+ execution: z.ZodObject<{
420
+ id: z.ZodString;
421
+ contentHash: z.ZodString;
422
+ }, z.core.$strict>;
423
+ originalEvidence: z.ZodObject<{
424
+ id: z.ZodString;
425
+ contentHash: z.ZodString;
426
+ }, z.core.$strict>;
427
+ members: z.ZodArray<z.ZodObject<{
428
+ caseId: z.ZodString;
429
+ seed: z.ZodString;
430
+ fixtureId: z.ZodString;
431
+ sourceId: z.ZodString;
432
+ snapshotHash: z.ZodString;
433
+ boundaryRevisionHash: z.ZodString;
434
+ }, z.core.$strict>>;
435
+ }, z.core.$strict>>;
436
+ gradingSource: z.ZodOptional<z.ZodObject<{
437
+ pass: z.ZodObject<{
438
+ id: z.ZodString;
439
+ contentHash: z.ZodString;
440
+ }, z.core.$strict>;
441
+ evidence: z.ZodObject<{
442
+ id: z.ZodString;
443
+ contentHash: z.ZodString;
444
+ }, z.core.$strict>;
445
+ }, z.core.$strict>>;
446
+ split: z.ZodEnum<{
447
+ train: "train";
448
+ validation: "validation";
449
+ test: "test";
450
+ frozen_eval: "frozen_eval";
451
+ }>;
452
+ population: z.ZodArray<z.ZodObject<{
453
+ taskId: z.ZodString;
454
+ seed: z.ZodString;
455
+ fixtureId: z.ZodNullable<z.ZodString>;
456
+ }, z.core.$strict>>;
457
+ evaluators: z.ZodArray<z.ZodObject<{
458
+ release: z.ZodObject<{
459
+ id: z.ZodString;
460
+ contentHash: z.ZodString;
461
+ revision: z.ZodUnion<readonly [z.ZodNumber, z.ZodString]>;
462
+ }, z.core.$strict>;
463
+ feedbackKey: z.ZodString;
464
+ configurationHash: z.ZodString;
465
+ output: z.ZodEnum<{
466
+ boolean: "boolean";
467
+ score: "score";
468
+ category: "category";
469
+ }>;
470
+ categories: z.ZodArray<z.ZodString>;
471
+ }, z.core.$strict>>;
472
+ fieldMappingsHash: z.ZodString;
473
+ fieldMappings: z.ZodArray<z.ZodObject<{
474
+ graderId: z.ZodString;
475
+ fields: z.ZodArray<z.ZodObject<{
476
+ destination: z.ZodEnum<{
477
+ output: "output";
478
+ input: "input";
479
+ expectedOutput: "expectedOutput";
480
+ evaluatorContext: "evaluatorContext";
481
+ }>;
482
+ path: z.ZodArray<z.ZodString>;
483
+ source: z.ZodEnum<{
484
+ output: "output";
485
+ input: "input";
486
+ expectedOutput: "expectedOutput";
487
+ evaluatorContext: "evaluatorContext";
488
+ }>;
489
+ sourcePath: z.ZodArray<z.ZodString>;
490
+ }, z.core.$strict>>;
491
+ }, z.core.$strict>>;
492
+ environmentHash: z.ZodString;
493
+ privateDatasetClosureHash: z.ZodString;
494
+ criterion: z.ZodObject<{
495
+ minimumPassRate: z.ZodNumber;
496
+ requireComplete: z.ZodBoolean;
497
+ }, z.core.$strict>;
498
+ contentHash: z.ZodString;
499
+ }, z.core.$strict>>;
500
+ sourceCandidate: z.ZodOptional<z.ZodObject<{
501
+ candidateId: z.ZodString;
502
+ frozenRevision: z.ZodNumber;
503
+ frozenHash: z.ZodString;
504
+ harnessRelease: z.ZodObject<{
505
+ id: z.ZodString;
506
+ contentHash: z.ZodString;
507
+ }, z.core.$strict>;
508
+ protectedClosureHash: z.ZodString;
509
+ }, z.core.$strict>>;
510
+ }, z.core.$strict>;
511
+ policy: z.ZodDiscriminatedUnion<[z.ZodObject<{
512
+ kind: z.ZodLiteral<"model">;
513
+ model: z.ZodObject<{
514
+ provider: z.ZodString;
515
+ model: z.ZodString;
516
+ revision: z.ZodDefault<z.ZodNullable<z.ZodString>>;
517
+ artifactHash: z.ZodDefault<z.ZodNullable<z.ZodString>>;
518
+ tokenizerRevision: z.ZodDefault<z.ZodNullable<z.ZodString>>;
519
+ chatTemplateHash: z.ZodDefault<z.ZodNullable<z.ZodString>>;
520
+ }, z.core.$strict>;
521
+ configurationHash: z.ZodString;
522
+ }, z.core.$strict>, z.ZodObject<{
523
+ kind: z.ZodLiteral<"fixture">;
524
+ }, z.core.$strict>], "kind">;
525
+ score: z.ZodNullable<z.ZodNumber>;
526
+ }, z.core.$strict>>;
527
+ createdAt: z.ZodString;
528
+ contentHash: z.ZodString;
529
+ }, z.core.$strict>>;
530
+ suiteRuns: z.ZodArray<z.ZodObject<{
531
+ schemaVersion: z.ZodLiteral<"openpond.profileEvaluationSuiteRun.v1">;
532
+ id: z.ZodString;
533
+ suiteId: z.ZodString;
534
+ suiteHash: z.ZodString;
535
+ catalogHash: z.ZodString;
536
+ profileId: z.ZodString;
537
+ sourceRevision: z.ZodString;
538
+ harnessRelease: z.ZodObject<{
539
+ id: z.ZodString;
540
+ contentHash: z.ZodString;
541
+ }, z.core.$strict>;
542
+ members: z.ZodArray<z.ZodObject<{
543
+ definitionId: z.ZodString;
544
+ runManifest: z.ZodObject<{
545
+ id: z.ZodString;
546
+ contentHash: z.ZodString;
547
+ }, z.core.$strict>;
548
+ runHash: z.ZodString;
549
+ passed: z.ZodBoolean;
550
+ score: z.ZodNullable<z.ZodNumber>;
551
+ }, z.core.$strict>>;
552
+ passed: z.ZodBoolean;
553
+ createdAt: z.ZodString;
554
+ completedAt: z.ZodString;
555
+ contentHash: z.ZodString;
556
+ }, z.core.$strict>>;
557
+ reports: z.ZodArray<z.ZodObject<{
558
+ schemaVersion: z.ZodLiteral<"openpond.profileEvaluationReport.v1">;
559
+ id: z.ZodString;
560
+ createdAt: z.ZodString;
561
+ testedSources: z.ZodArray<z.ZodObject<{
562
+ profileId: z.ZodString;
563
+ sourceRevision: z.ZodString;
564
+ harnessRelease: z.ZodObject<{
565
+ id: z.ZodString;
566
+ contentHash: z.ZodString;
567
+ }, z.core.$strict>;
568
+ }, z.core.$strict>>;
569
+ tasksetReleases: z.ZodArray<z.ZodObject<{
570
+ id: z.ZodString;
571
+ contentHash: z.ZodString;
572
+ }, z.core.$strict>>;
573
+ metricPolicyHashes: z.ZodArray<z.ZodString>;
574
+ modelConfigurationHashes: z.ZodArray<z.ZodString>;
575
+ evidence: z.ZodArray<z.ZodObject<{
576
+ kind: z.ZodEnum<{
577
+ run: "run";
578
+ suite: "suite";
579
+ comparison: "comparison";
580
+ }>;
581
+ id: z.ZodString;
582
+ contentHash: z.ZodString;
583
+ }, z.core.$strict>>;
584
+ summary: z.ZodObject<{
585
+ totalChecks: z.ZodNumber;
586
+ passedChecks: z.ZodNumber;
587
+ failedChecks: z.ZodNumber;
588
+ unscoredChecks: z.ZodNumber;
589
+ meanScore: z.ZodNullable<z.ZodNumber>;
590
+ }, z.core.$strict>;
591
+ contentHash: z.ZodString;
592
+ }, z.core.$strict>>;
593
+ }, z.core.$strict>;
594
+ export type ProfileEvaluationDiscovery = z.infer<typeof ProfileEvaluationDiscoverySchema>;
595
+ export declare const ProfileEvaluationDiscoveryRequestSchema: z.ZodObject<{
596
+ teamId: z.ZodString;
597
+ profileRepositoryId: z.ZodString;
598
+ trainingProjectId: z.ZodOptional<z.ZodString>;
599
+ }, z.core.$strict>;
600
+ /** Reads the existing source owner's receipts. No resource is copied and no
601
+ * incomplete execution is turned into a synthetic rerunnable configuration. */
602
+ export declare class OpenPondProfileEvaluationDiscoveryClient {
603
+ private readonly options;
604
+ constructor(options: {
605
+ baseUrl: string;
606
+ apiKey: string;
607
+ teamId: string;
608
+ fetch?: typeof fetch;
609
+ });
610
+ read(input: {
611
+ profileRepositoryId: string;
612
+ trainingProjectId?: string;
613
+ }, signal?: AbortSignal): Promise<{
614
+ profileRef: {
615
+ source: "local" | "github" | "openpond_git";
616
+ repositoryId: string;
617
+ profileId: string;
618
+ };
619
+ sourceRevision: string;
620
+ harnessRelease: {
621
+ id: string;
622
+ contentHash: string;
623
+ };
624
+ catalogHash: string | null;
625
+ definitions: {
626
+ id: string;
627
+ label: string;
628
+ description: string;
629
+ target: {
630
+ kind: "profile";
631
+ } | {
632
+ kind: "workflow";
633
+ workflowId: string;
634
+ } | {
635
+ kind: "skill";
636
+ skillPath: string;
637
+ } | {
638
+ kind: "agent_action";
639
+ actionId: string;
640
+ };
641
+ tasksetRelease: {
642
+ id: string;
643
+ contentHash: string;
644
+ };
645
+ split: "train" | "validation" | "test" | "frozen_eval";
646
+ taskIds: string[];
647
+ seeds: string[];
648
+ criterion: {
649
+ minimumPassRate: number;
650
+ requireComplete: boolean;
651
+ };
652
+ }[];
653
+ suites: {
654
+ id: string;
655
+ label: string;
656
+ scope: "component" | "profile";
657
+ definitionIds: string[];
658
+ }[];
659
+ runs: {
660
+ profileRef: {
661
+ source: "local" | "github" | "openpond_git";
662
+ repositoryId: string;
663
+ profileId: string;
664
+ };
665
+ manifest: {
666
+ schemaVersion: "openpond.tasksetRunManifest.v1";
667
+ id: string;
668
+ tasksetRelease: {
669
+ id: string;
670
+ contentHash: string;
671
+ };
672
+ packageHash: string;
673
+ execution: {
674
+ kind: "taskset";
675
+ environmentRelease: {
676
+ id: string;
677
+ contentHash: string;
678
+ };
679
+ verifierSetRelease: {
680
+ id: string;
681
+ contentHash: string;
682
+ };
683
+ policyHash: string;
684
+ } | {
685
+ kind: "harness";
686
+ harnessRelease: {
687
+ id: string;
688
+ contentHash: string;
689
+ };
690
+ };
691
+ policy: {
692
+ kind: "model";
693
+ model: {
694
+ provider: string;
695
+ model: string;
696
+ revision: string | null;
697
+ artifactHash: string | null;
698
+ tokenizerRevision: string | null;
699
+ chatTemplateHash: string | null;
700
+ };
701
+ configurationHash: string;
702
+ } | {
703
+ kind: "fixture";
704
+ };
705
+ gradingRole: "evaluation";
706
+ metricPolicy: {
707
+ schemaVersion: "openpond.tasksetMetricPolicy.v1";
708
+ primaryMetric: string;
709
+ aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
710
+ missingReward: "zero" | "exclude";
711
+ customAggregator: {
712
+ module: string;
713
+ exportName: string;
714
+ contentHash: string;
715
+ timeoutMs: number;
716
+ networkPolicy: "none";
717
+ } | null;
718
+ taskWeights?: Record<string, number> | undefined;
719
+ };
720
+ population: {
721
+ receiptId: string;
722
+ taskId: string;
723
+ seed: string;
724
+ fixtureId: string | null;
725
+ }[];
726
+ runtimeTarget: {
727
+ adapterId: string;
728
+ placement: "local" | "remote" | "colocated" | "provider_native";
729
+ runtimeVersion: string;
730
+ capabilityReceipt: string;
731
+ };
732
+ limits: {
733
+ maxTurns: number;
734
+ timeoutMs: number;
735
+ maxOutputBytes: number;
736
+ maximumSpendUsd: number | null;
737
+ };
738
+ createdAt: string;
739
+ metadata: Record<string, unknown>;
740
+ contentHash: string;
741
+ profileEvaluation?: {
742
+ profileId: string;
743
+ sourceRevision: string;
744
+ harnessRelease: {
745
+ id: string;
746
+ contentHash: string;
747
+ };
748
+ catalogHash: string;
749
+ definitionId: string;
750
+ definitionHash: string;
751
+ target: {
752
+ kind: "profile";
753
+ } | {
754
+ kind: "workflow";
755
+ workflowId: string;
756
+ } | {
757
+ kind: "skill";
758
+ skillPath: string;
759
+ } | {
760
+ kind: "agent_action";
761
+ actionId: string;
762
+ };
763
+ environmentHash: string;
764
+ externalDatasetBinding?: {
765
+ schemaVersion: "openpond.profileExternalDatasetBinding.v1";
766
+ dataset: {
767
+ id: string;
768
+ contentHash: string;
769
+ revision: number;
770
+ };
771
+ packageHash: string;
772
+ profileId: string;
773
+ target: {
774
+ kind: "profile";
775
+ } | {
776
+ kind: "workflow";
777
+ workflowId: string;
778
+ } | {
779
+ kind: "skill";
780
+ skillPath: string;
781
+ } | {
782
+ kind: "agent_action";
783
+ actionId: string;
784
+ };
785
+ declaredProfileCatalogHash: string | null;
786
+ protectedProfileClosureHash: string;
787
+ split: "train" | "validation" | "test" | "frozen_eval";
788
+ population: {
789
+ taskId: string;
790
+ seed: string;
791
+ fixtureId: string | null;
792
+ }[];
793
+ evaluators: {
794
+ release: {
795
+ id: string;
796
+ contentHash: string;
797
+ revision: string | number;
798
+ };
799
+ feedbackKey: string;
800
+ configurationHash: string;
801
+ output: "boolean" | "score" | "category";
802
+ categories: string[];
803
+ }[];
804
+ fieldMappingsHash: string;
805
+ fieldMappings: {
806
+ graderId: string;
807
+ fields: {
808
+ destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
809
+ path: string[];
810
+ source: "input" | "output" | "expectedOutput" | "evaluatorContext";
811
+ sourcePath: string[];
812
+ }[];
813
+ }[];
814
+ environmentHash: string;
815
+ privateDatasetClosureHash: string;
816
+ criterion: {
817
+ minimumPassRate: number;
818
+ requireComplete: boolean;
819
+ };
820
+ contentHash: string;
821
+ recordedOrigin?: {
822
+ execution: {
823
+ id: string;
824
+ contentHash: string;
825
+ };
826
+ originalEvidence: {
827
+ id: string;
828
+ contentHash: string;
829
+ };
830
+ members: {
831
+ caseId: string;
832
+ seed: string;
833
+ fixtureId: string;
834
+ sourceId: string;
835
+ snapshotHash: string;
836
+ boundaryRevisionHash: string;
837
+ }[];
838
+ } | undefined;
839
+ gradingSource?: {
840
+ pass: {
841
+ id: string;
842
+ contentHash: string;
843
+ };
844
+ evidence: {
845
+ id: string;
846
+ contentHash: string;
847
+ };
848
+ } | undefined;
849
+ } | undefined;
850
+ sourceCandidate?: {
851
+ candidateId: string;
852
+ frozenRevision: number;
853
+ frozenHash: string;
854
+ harnessRelease: {
855
+ id: string;
856
+ contentHash: string;
857
+ };
858
+ protectedClosureHash: string;
859
+ } | undefined;
860
+ } | undefined;
861
+ };
862
+ metric: {
863
+ schemaVersion: "openpond.tasksetMetricResult.v1";
864
+ runManifest: {
865
+ id: string;
866
+ contentHash: string;
867
+ };
868
+ tasksetRelease: {
869
+ id: string;
870
+ contentHash: string;
871
+ };
872
+ policy: {
873
+ schemaVersion: "openpond.tasksetMetricPolicy.v1";
874
+ primaryMetric: string;
875
+ aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
876
+ missingReward: "zero" | "exclude";
877
+ customAggregator: {
878
+ module: string;
879
+ exportName: string;
880
+ contentHash: string;
881
+ timeoutMs: number;
882
+ networkPolicy: "none";
883
+ } | null;
884
+ taskWeights?: Record<string, number> | undefined;
885
+ };
886
+ policyHash: string;
887
+ receiptRefs: {
888
+ id: string;
889
+ contentHash: string;
890
+ }[];
891
+ includedCount: number;
892
+ missingRewardCount: number;
893
+ excludedCount: number;
894
+ value: number | null;
895
+ contentHash: string;
896
+ };
897
+ gradeRefs: {
898
+ id: string;
899
+ contentHash: string;
900
+ }[];
901
+ receiptRefs: {
902
+ id: string;
903
+ contentHash: string;
904
+ }[];
905
+ passRate: number;
906
+ passed: boolean;
907
+ completedAt: string;
908
+ contentHash: string;
909
+ }[];
910
+ comparisons: {
911
+ schemaVersion: "openpond.profileEvaluationComparison.v1";
912
+ id: string;
913
+ tasksetRelease: {
914
+ id: string;
915
+ contentHash: string;
916
+ };
917
+ populationHash: string;
918
+ environmentHash: string;
919
+ metricPolicyHash: string;
920
+ limitsHash: string;
921
+ members: {
922
+ runManifest: {
923
+ id: string;
924
+ contentHash: string;
925
+ };
926
+ metricResultHash: string;
927
+ source: {
928
+ profileId: string;
929
+ sourceRevision: string;
930
+ harnessRelease: {
931
+ id: string;
932
+ contentHash: string;
933
+ };
934
+ catalogHash: string;
935
+ definitionId: string;
936
+ definitionHash: string;
937
+ target: {
938
+ kind: "profile";
939
+ } | {
940
+ kind: "workflow";
941
+ workflowId: string;
942
+ } | {
943
+ kind: "skill";
944
+ skillPath: string;
945
+ } | {
946
+ kind: "agent_action";
947
+ actionId: string;
948
+ };
949
+ environmentHash: string;
950
+ externalDatasetBinding?: {
951
+ schemaVersion: "openpond.profileExternalDatasetBinding.v1";
952
+ dataset: {
953
+ id: string;
954
+ contentHash: string;
955
+ revision: number;
956
+ };
957
+ packageHash: string;
958
+ profileId: string;
959
+ target: {
960
+ kind: "profile";
961
+ } | {
962
+ kind: "workflow";
963
+ workflowId: string;
964
+ } | {
965
+ kind: "skill";
966
+ skillPath: string;
967
+ } | {
968
+ kind: "agent_action";
969
+ actionId: string;
970
+ };
971
+ declaredProfileCatalogHash: string | null;
972
+ protectedProfileClosureHash: string;
973
+ split: "train" | "validation" | "test" | "frozen_eval";
974
+ population: {
975
+ taskId: string;
976
+ seed: string;
977
+ fixtureId: string | null;
978
+ }[];
979
+ evaluators: {
980
+ release: {
981
+ id: string;
982
+ contentHash: string;
983
+ revision: string | number;
984
+ };
985
+ feedbackKey: string;
986
+ configurationHash: string;
987
+ output: "boolean" | "score" | "category";
988
+ categories: string[];
989
+ }[];
990
+ fieldMappingsHash: string;
991
+ fieldMappings: {
992
+ graderId: string;
993
+ fields: {
994
+ destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
995
+ path: string[];
996
+ source: "input" | "output" | "expectedOutput" | "evaluatorContext";
997
+ sourcePath: string[];
998
+ }[];
999
+ }[];
1000
+ environmentHash: string;
1001
+ privateDatasetClosureHash: string;
1002
+ criterion: {
1003
+ minimumPassRate: number;
1004
+ requireComplete: boolean;
1005
+ };
1006
+ contentHash: string;
1007
+ recordedOrigin?: {
1008
+ execution: {
1009
+ id: string;
1010
+ contentHash: string;
1011
+ };
1012
+ originalEvidence: {
1013
+ id: string;
1014
+ contentHash: string;
1015
+ };
1016
+ members: {
1017
+ caseId: string;
1018
+ seed: string;
1019
+ fixtureId: string;
1020
+ sourceId: string;
1021
+ snapshotHash: string;
1022
+ boundaryRevisionHash: string;
1023
+ }[];
1024
+ } | undefined;
1025
+ gradingSource?: {
1026
+ pass: {
1027
+ id: string;
1028
+ contentHash: string;
1029
+ };
1030
+ evidence: {
1031
+ id: string;
1032
+ contentHash: string;
1033
+ };
1034
+ } | undefined;
1035
+ } | undefined;
1036
+ sourceCandidate?: {
1037
+ candidateId: string;
1038
+ frozenRevision: number;
1039
+ frozenHash: string;
1040
+ harnessRelease: {
1041
+ id: string;
1042
+ contentHash: string;
1043
+ };
1044
+ protectedClosureHash: string;
1045
+ } | undefined;
1046
+ };
1047
+ policy: {
1048
+ kind: "model";
1049
+ model: {
1050
+ provider: string;
1051
+ model: string;
1052
+ revision: string | null;
1053
+ artifactHash: string | null;
1054
+ tokenizerRevision: string | null;
1055
+ chatTemplateHash: string | null;
1056
+ };
1057
+ configurationHash: string;
1058
+ } | {
1059
+ kind: "fixture";
1060
+ };
1061
+ score: number | null;
1062
+ }[];
1063
+ createdAt: string;
1064
+ contentHash: string;
1065
+ }[];
1066
+ suiteRuns: {
1067
+ schemaVersion: "openpond.profileEvaluationSuiteRun.v1";
1068
+ id: string;
1069
+ suiteId: string;
1070
+ suiteHash: string;
1071
+ catalogHash: string;
1072
+ profileId: string;
1073
+ sourceRevision: string;
1074
+ harnessRelease: {
1075
+ id: string;
1076
+ contentHash: string;
1077
+ };
1078
+ members: {
1079
+ definitionId: string;
1080
+ runManifest: {
1081
+ id: string;
1082
+ contentHash: string;
1083
+ };
1084
+ runHash: string;
1085
+ passed: boolean;
1086
+ score: number | null;
1087
+ }[];
1088
+ passed: boolean;
1089
+ createdAt: string;
1090
+ completedAt: string;
1091
+ contentHash: string;
1092
+ }[];
1093
+ reports: {
1094
+ schemaVersion: "openpond.profileEvaluationReport.v1";
1095
+ id: string;
1096
+ createdAt: string;
1097
+ testedSources: {
1098
+ profileId: string;
1099
+ sourceRevision: string;
1100
+ harnessRelease: {
1101
+ id: string;
1102
+ contentHash: string;
1103
+ };
1104
+ }[];
1105
+ tasksetReleases: {
1106
+ id: string;
1107
+ contentHash: string;
1108
+ }[];
1109
+ metricPolicyHashes: string[];
1110
+ modelConfigurationHashes: string[];
1111
+ evidence: {
1112
+ kind: "run" | "comparison" | "suite";
1113
+ id: string;
1114
+ contentHash: string;
1115
+ }[];
1116
+ summary: {
1117
+ totalChecks: number;
1118
+ passedChecks: number;
1119
+ failedChecks: number;
1120
+ unscoredChecks: number;
1121
+ meanScore: number | null;
1122
+ };
1123
+ contentHash: string;
1124
+ }[];
1125
+ }>;
1126
+ }
1127
+ //# sourceMappingURL=profile-evaluation-discovery.d.ts.map