@velum-labs/routekit-eval-setup 1.3.2 → 1.5.0

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 (342) hide show
  1. package/dist/adapters/authoring-responses-request.d.ts +5 -0
  2. package/dist/adapters/authoring-responses-request.js +38 -0
  3. package/dist/adapters/evaluation-evidence-freshness.d.ts +7 -0
  4. package/dist/adapters/evaluation-evidence-freshness.js +76 -0
  5. package/dist/adapters/git-task-history.d.ts +67 -0
  6. package/dist/adapters/git-task-history.js +171 -0
  7. package/dist/adapters/integrated-repository-history.d.ts +21 -0
  8. package/dist/adapters/integrated-repository-history.js +175 -0
  9. package/dist/adapters/repository-command-diagnostic.d.ts +8 -0
  10. package/dist/adapters/repository-command-diagnostic.js +46 -0
  11. package/dist/adapters/repository-command-evidence.d.ts +13 -0
  12. package/dist/adapters/repository-command-evidence.js +102 -0
  13. package/dist/adapters/repository-command-runner.d.ts +238 -0
  14. package/dist/adapters/repository-command-runner.js +1483 -0
  15. package/dist/adapters/repository-import-context.d.ts +47 -0
  16. package/dist/adapters/repository-import-context.js +469 -0
  17. package/dist/adapters/repository-node-test-reporter.d.ts +3 -0
  18. package/dist/adapters/repository-node-test-reporter.js +27 -0
  19. package/dist/adapters/repository-review-evidence.d.ts +39 -0
  20. package/dist/adapters/repository-review-evidence.js +632 -0
  21. package/dist/adapters/repository-seed-selection.d.ts +7 -0
  22. package/dist/adapters/repository-seed-selection.js +79 -0
  23. package/dist/adapters/repository-solution-edits.d.ts +49 -0
  24. package/dist/adapters/repository-solution-edits.js +136 -0
  25. package/dist/adapters/repository-vitest-phase-adapter.d.ts +8 -0
  26. package/dist/adapters/repository-vitest-phase-adapter.js +310 -0
  27. package/dist/adapters/repository-vitest-reporter.d.ts +24 -0
  28. package/dist/adapters/repository-vitest-reporter.js +314 -0
  29. package/dist/adapters/strict-authoring-schema.d.ts +5 -0
  30. package/dist/adapters/strict-authoring-schema.js +158 -0
  31. package/dist/adapters/test-discovery.d.ts +30 -0
  32. package/dist/adapters/test-discovery.js +124 -0
  33. package/dist/adapters/typescript-repository-index.d.ts +51 -0
  34. package/dist/adapters/typescript-repository-index.js +226 -0
  35. package/dist/agentic-capabilities-protocol.d.ts +1373 -0
  36. package/dist/agentic-capabilities-protocol.js +786 -0
  37. package/dist/case-checkpoint-store.d.ts +29 -0
  38. package/dist/case-checkpoint-store.js +133 -0
  39. package/dist/case-pipeline-protocol-v2.d.ts +184 -0
  40. package/dist/case-pipeline-protocol-v2.js +193 -0
  41. package/dist/case-pipeline-protocol.d.ts +2626 -0
  42. package/dist/case-pipeline-protocol.js +371 -0
  43. package/dist/effect-api.d.ts +74 -10
  44. package/dist/effect-api.js +56 -6
  45. package/dist/errors.d.ts +31 -0
  46. package/dist/errors.js +10 -0
  47. package/dist/eval-capability-execution-envelope.d.ts +64 -0
  48. package/dist/eval-capability-execution-envelope.js +98 -0
  49. package/dist/eval-capability-policy.d.ts +90 -0
  50. package/dist/eval-capability-policy.js +107 -0
  51. package/dist/eval-event-log.d.ts +140 -0
  52. package/dist/eval-event-log.js +220 -0
  53. package/dist/evaluation-authoring-policy.d.ts +18 -0
  54. package/dist/evaluation-authoring-policy.js +19 -0
  55. package/dist/evaluation-authoring-validation.d.ts +22 -0
  56. package/dist/evaluation-authoring-validation.js +72 -0
  57. package/dist/evaluation-evidence.d.ts +20 -0
  58. package/dist/evaluation-evidence.js +319 -0
  59. package/dist/evaluation-grader-calibration-protocol.d.ts +108 -0
  60. package/dist/evaluation-grader-calibration-protocol.js +80 -0
  61. package/dist/evaluation-grader-calibration.d.ts +18 -0
  62. package/dist/evaluation-grader-calibration.js +334 -0
  63. package/dist/evaluation-grading-policy.d.ts +24 -0
  64. package/dist/evaluation-grading-policy.js +54 -0
  65. package/dist/evaluation-proposal-policy.d.ts +4 -0
  66. package/dist/evaluation-proposal-policy.js +91 -0
  67. package/dist/evaluation-source-retrieval.d.ts +68 -0
  68. package/dist/evaluation-source-retrieval.js +513 -0
  69. package/dist/evaluation-structure-policy.d.ts +29 -0
  70. package/dist/evaluation-structure-policy.js +138 -0
  71. package/dist/index.d.ts +124 -17
  72. package/dist/index.js +69 -11
  73. package/dist/inspection.js +2 -3
  74. package/dist/project-artifacts.d.ts +7 -2
  75. package/dist/project-artifacts.js +49 -136
  76. package/dist/project-authoring.d.ts +66 -5
  77. package/dist/project-authoring.js +783 -109
  78. package/dist/project-contracts.d.ts +419 -84
  79. package/dist/project-contracts.js +160 -52
  80. package/dist/project-store.js +2 -1
  81. package/dist/project-workflow.d.ts +5 -4
  82. package/dist/project-workflow.js +154 -35
  83. package/dist/repository-adversary-protocol.d.ts +64 -0
  84. package/dist/repository-adversary-protocol.js +105 -0
  85. package/dist/repository-behavior-protocol.d.ts +188 -0
  86. package/dist/repository-behavior-protocol.js +202 -0
  87. package/dist/repository-benchmark-protocol.d.ts +487 -0
  88. package/dist/repository-benchmark-protocol.js +96 -0
  89. package/dist/repository-execution-protocol.d.ts +150 -0
  90. package/dist/repository-execution-protocol.js +38 -0
  91. package/dist/repository-fixture-instructions.d.ts +3 -0
  92. package/dist/repository-fixture-instructions.js +91 -0
  93. package/dist/repository-fixture-protocol.d.ts +79 -0
  94. package/dist/repository-fixture-protocol.js +79 -0
  95. package/dist/repository-foundry-plan-protocol.d.ts +118 -0
  96. package/dist/repository-foundry-plan-protocol.js +296 -0
  97. package/dist/repository-foundry-progress-protocol.d.ts +52 -0
  98. package/dist/repository-foundry-progress-protocol.js +52 -0
  99. package/dist/repository-improvement-protocol.d.ts +100 -0
  100. package/dist/repository-improvement-protocol.js +106 -0
  101. package/dist/repository-language-model-protocol.d.ts +43 -0
  102. package/dist/repository-language-model-protocol.js +146 -0
  103. package/dist/repository-oracle-coverage-protocol.d.ts +18 -0
  104. package/dist/repository-oracle-coverage-protocol.js +39 -0
  105. package/dist/repository-oracle-execution-binding.d.ts +27 -0
  106. package/dist/repository-oracle-execution-binding.js +59 -0
  107. package/dist/repository-oracle-protocol.d.ts +230 -0
  108. package/dist/repository-oracle-protocol.js +156 -0
  109. package/dist/repository-oracle-scope-policy.d.ts +22 -0
  110. package/dist/repository-oracle-scope-policy.js +92 -0
  111. package/dist/repository-quality-policy.d.ts +15 -0
  112. package/dist/repository-quality-policy.js +357 -0
  113. package/dist/repository-routing-benchmark-protocol.d.ts +176 -0
  114. package/dist/repository-routing-benchmark-protocol.js +103 -0
  115. package/dist/repository-routing-model-protocol.d.ts +36 -0
  116. package/dist/repository-routing-model-protocol.js +89 -0
  117. package/dist/repository-routing-plan-protocol.d.ts +112 -0
  118. package/dist/repository-routing-plan-protocol.js +58 -0
  119. package/dist/repository-routing-quality-policy.d.ts +9 -0
  120. package/dist/repository-routing-quality-policy.js +191 -0
  121. package/dist/repository-seed-qualification-progress-protocol.d.ts +205 -0
  122. package/dist/repository-seed-qualification-progress-protocol.js +28 -0
  123. package/dist/repository-semantic-calibration-protocol.d.ts +768 -0
  124. package/dist/repository-semantic-calibration-protocol.js +276 -0
  125. package/dist/repository-semantic-calibration.d.ts +163 -0
  126. package/dist/repository-semantic-calibration.js +581 -0
  127. package/dist/repository-specification-contract-facts-protocol.d.ts +224 -0
  128. package/dist/repository-specification-contract-facts-protocol.js +276 -0
  129. package/dist/repository-specification-critique-protocol.d.ts +189 -0
  130. package/dist/repository-specification-critique-protocol.js +103 -0
  131. package/dist/repository-task-family-protocol.d.ts +24 -0
  132. package/dist/repository-task-family-protocol.js +37 -0
  133. package/dist/repository-task-seed-protocol.d.ts +384 -0
  134. package/dist/repository-task-seed-protocol.js +236 -0
  135. package/dist/repository-trajectory-protocol.d.ts +20 -0
  136. package/dist/repository-trajectory-protocol.js +42 -0
  137. package/dist/service.js +1 -1
  138. package/dist/services/adversary/service.d.ts +64 -0
  139. package/dist/services/adversary/service.js +330 -0
  140. package/dist/services/benchmark-compiler/service.d.ts +450 -0
  141. package/dist/services/benchmark-compiler/service.js +9 -0
  142. package/dist/services/budgeted-model/service.d.ts +118 -0
  143. package/dist/services/budgeted-model/service.js +460 -0
  144. package/dist/services/case-authoring/service.d.ts +163 -0
  145. package/dist/services/case-authoring/service.js +1456 -0
  146. package/dist/services/case-finalization/service.d.ts +283 -0
  147. package/dist/services/case-finalization/service.js +370 -0
  148. package/dist/services/case-generation/service.d.ts +619 -0
  149. package/dist/services/case-generation/service.js +2628 -0
  150. package/dist/services/case-pipeline/service.d.ts +31 -0
  151. package/dist/services/case-pipeline/service.js +485 -0
  152. package/dist/services/case-pipeline-v2/service.d.ts +70 -0
  153. package/dist/services/case-pipeline-v2/service.js +477 -0
  154. package/dist/services/command-observability/service.d.ts +13 -0
  155. package/dist/services/command-observability/service.js +3 -0
  156. package/dist/services/dimension-labeling/service.d.ts +77 -0
  157. package/dist/services/dimension-labeling/service.js +188 -0
  158. package/dist/services/eval-candidate/service.d.ts +208 -0
  159. package/dist/services/eval-candidate/service.js +64 -0
  160. package/dist/services/eval-capabilities/service.d.ts +183 -0
  161. package/dist/services/eval-capabilities/service.js +1433 -0
  162. package/dist/services/eval-environment/service.d.ts +173 -0
  163. package/dist/services/eval-environment/service.js +127 -0
  164. package/dist/services/evidence-reconstruction/service.d.ts +36 -0
  165. package/dist/services/evidence-reconstruction/service.js +145 -0
  166. package/dist/services/fixture-builder/service.d.ts +62 -0
  167. package/dist/services/fixture-builder/service.js +36 -0
  168. package/dist/services/fixture-validation/service.d.ts +75 -0
  169. package/dist/services/fixture-validation/service.js +295 -0
  170. package/dist/services/foundry/service.d.ts +831 -0
  171. package/dist/services/foundry/service.js +442 -0
  172. package/dist/services/foundry-progress/service.d.ts +62 -0
  173. package/dist/services/foundry-progress/service.js +149 -0
  174. package/dist/services/foundry-v2/service.d.ts +54 -0
  175. package/dist/services/foundry-v2/service.js +28 -0
  176. package/dist/services/grounded-authoring/service.d.ts +126 -0
  177. package/dist/services/grounded-authoring/service.js +822 -0
  178. package/dist/services/historical-case/service.d.ts +722 -0
  179. package/dist/services/historical-case/service.js +177 -0
  180. package/dist/services/improvement-loop/service.d.ts +59 -0
  181. package/dist/services/improvement-loop/service.js +176 -0
  182. package/dist/services/language-model/service.d.ts +52 -0
  183. package/dist/services/language-model/service.js +194 -0
  184. package/dist/services/oracle-builder/service.d.ts +146 -0
  185. package/dist/services/oracle-builder/service.js +513 -0
  186. package/dist/services/oracle-coverage/service.d.ts +28 -0
  187. package/dist/services/oracle-coverage/service.js +50 -0
  188. package/dist/services/oracle-coverage-witness/service.d.ts +130 -0
  189. package/dist/services/oracle-coverage-witness/service.js +538 -0
  190. package/dist/services/pipeline-challenge/service.d.ts +551 -0
  191. package/dist/services/pipeline-challenge/service.js +427 -0
  192. package/dist/services/pipeline-controls/service.d.ts +130 -0
  193. package/dist/services/pipeline-controls/service.js +483 -0
  194. package/dist/services/pipeline-oracle/service.d.ts +8 -0
  195. package/dist/services/pipeline-oracle/service.js +256 -0
  196. package/dist/services/pipeline-seed/service.d.ts +298 -0
  197. package/dist/services/pipeline-seed/service.js +428 -0
  198. package/dist/services/pipeline-spec/service.d.ts +103 -0
  199. package/dist/services/pipeline-spec/service.js +619 -0
  200. package/dist/services/pipeline-tournament/service.d.ts +258 -0
  201. package/dist/services/pipeline-tournament/service.js +476 -0
  202. package/dist/services/quality-gate/service.d.ts +233 -0
  203. package/dist/services/quality-gate/service.js +136 -0
  204. package/dist/services/repository-bundle/service.d.ts +33 -0
  205. package/dist/services/repository-bundle/service.js +114 -0
  206. package/dist/services/repository-model/service.d.ts +105 -0
  207. package/dist/services/repository-model/service.js +250 -0
  208. package/dist/services/repository-public-artifact/service.d.ts +133 -0
  209. package/dist/services/repository-public-artifact/service.js +330 -0
  210. package/dist/services/routing-benchmark/service.d.ts +362 -0
  211. package/dist/services/routing-benchmark/service.js +96 -0
  212. package/dist/services/specification-critic/service.d.ts +92 -0
  213. package/dist/services/specification-critic/service.js +172 -0
  214. package/dist/services/task-family/service.d.ts +40 -0
  215. package/dist/services/task-family/service.js +55 -0
  216. package/dist/services/task-seed/service.d.ts +906 -0
  217. package/dist/services/task-seed/service.js +1406 -0
  218. package/dist/services/task-specification/service.d.ts +27 -0
  219. package/dist/services/task-specification/service.js +40 -0
  220. package/dist/services/trajectory-policy/service.d.ts +110 -0
  221. package/dist/services/trajectory-policy/service.js +216 -0
  222. package/dist/test/agentic-capabilities-protocol.test.d.ts +1 -0
  223. package/dist/test/agentic-capabilities-protocol.test.js +570 -0
  224. package/dist/test/agentic-capabilities.test.d.ts +1 -0
  225. package/dist/test/agentic-capabilities.test.js +1461 -0
  226. package/dist/test/agentic-environment.test.d.ts +1 -0
  227. package/dist/test/agentic-environment.test.js +213 -0
  228. package/dist/test/case-pipeline-foundation.test.d.ts +1 -0
  229. package/dist/test/case-pipeline-foundation.test.js +535 -0
  230. package/dist/test/case-pipeline-protocol-v2.test.d.ts +1 -0
  231. package/dist/test/case-pipeline-protocol-v2.test.js +124 -0
  232. package/dist/test/case-pipeline-v2.test.d.ts +1 -0
  233. package/dist/test/case-pipeline-v2.test.js +286 -0
  234. package/dist/test/case-pipeline.test.d.ts +1 -0
  235. package/dist/test/case-pipeline.test.js +851 -0
  236. package/dist/test/eval-capability-policy.test.d.ts +1 -0
  237. package/dist/test/eval-capability-policy.test.js +50 -0
  238. package/dist/test/eval-event-log.test.d.ts +1 -0
  239. package/dist/test/eval-event-log.test.js +125 -0
  240. package/dist/test/evaluation-evidence-freshness.test.d.ts +1 -0
  241. package/dist/test/evaluation-evidence-freshness.test.js +44 -0
  242. package/dist/test/evaluation-evidence.test.d.ts +1 -0
  243. package/dist/test/evaluation-evidence.test.js +230 -0
  244. package/dist/test/evaluation-grader-calibration.test.d.ts +1 -0
  245. package/dist/test/evaluation-grader-calibration.test.js +373 -0
  246. package/dist/test/evaluation-proposal-digest.test.d.ts +1 -0
  247. package/dist/test/evaluation-proposal-digest.test.js +187 -0
  248. package/dist/test/evaluation-source-retrieval.test.d.ts +1 -0
  249. package/dist/test/evaluation-source-retrieval.test.js +237 -0
  250. package/dist/test/evaluation-structure-policy.test.d.ts +1 -0
  251. package/dist/test/evaluation-structure-policy.test.js +196 -0
  252. package/dist/test/fixtures/repository-resource-panel.d.ts +39 -0
  253. package/dist/test/fixtures/repository-resource-panel.js +111 -0
  254. package/dist/test/fixtures/vitest-boundary-panel.d.ts +84 -0
  255. package/dist/test/fixtures/vitest-boundary-panel.js +120 -0
  256. package/dist/test/fixtures/vitest-phase-panel.d.ts +135 -0
  257. package/dist/test/fixtures/vitest-phase-panel.js +213 -0
  258. package/dist/test/fixtures/vitest-reporter-results.d.ts +76 -0
  259. package/dist/test/fixtures/vitest-reporter-results.js +94 -0
  260. package/dist/test/grounded-authoring.test.d.ts +1 -0
  261. package/dist/test/grounded-authoring.test.js +565 -0
  262. package/dist/test/integrated-repository-history.test.d.ts +1 -0
  263. package/dist/test/integrated-repository-history.test.js +227 -0
  264. package/dist/test/project-authoring.test.js +593 -43
  265. package/dist/test/project-workflow.test.js +419 -40
  266. package/dist/test/repository-authoring-artifacts.test.d.ts +1 -0
  267. package/dist/test/repository-authoring-artifacts.test.js +185 -0
  268. package/dist/test/repository-bundle.test.d.ts +1 -0
  269. package/dist/test/repository-bundle.test.js +52 -0
  270. package/dist/test/repository-case-generation.test.d.ts +1 -0
  271. package/dist/test/repository-case-generation.test.js +3465 -0
  272. package/dist/test/repository-command-diagnostic.test.d.ts +1 -0
  273. package/dist/test/repository-command-diagnostic.test.js +55 -0
  274. package/dist/test/repository-command-signals.test.d.ts +1 -0
  275. package/dist/test/repository-command-signals.test.js +124 -0
  276. package/dist/test/repository-fixture-scope-coverage.test.d.ts +1 -0
  277. package/dist/test/repository-fixture-scope-coverage.test.js +127 -0
  278. package/dist/test/repository-fixture-validation.test.d.ts +1 -0
  279. package/dist/test/repository-fixture-validation.test.js +362 -0
  280. package/dist/test/repository-foundry-progress.test.d.ts +1 -0
  281. package/dist/test/repository-foundry-progress.test.js +110 -0
  282. package/dist/test/repository-foundry-quality.test.d.ts +1 -0
  283. package/dist/test/repository-foundry-quality.test.js +1138 -0
  284. package/dist/test/repository-import-context.test.d.ts +1 -0
  285. package/dist/test/repository-import-context.test.js +354 -0
  286. package/dist/test/repository-model-authoring.test.d.ts +1 -0
  287. package/dist/test/repository-model-authoring.test.js +544 -0
  288. package/dist/test/repository-model.test.d.ts +1 -0
  289. package/dist/test/repository-model.test.js +2195 -0
  290. package/dist/test/repository-node-test-reporter.test.d.ts +1 -0
  291. package/dist/test/repository-node-test-reporter.test.js +104 -0
  292. package/dist/test/repository-oracle-concurrency.test.d.ts +1 -0
  293. package/dist/test/repository-oracle-concurrency.test.js +542 -0
  294. package/dist/test/repository-oracle-coverage-witness.test.d.ts +1 -0
  295. package/dist/test/repository-oracle-coverage-witness.test.js +511 -0
  296. package/dist/test/repository-oracle-coverage.test.d.ts +1 -0
  297. package/dist/test/repository-oracle-coverage.test.js +168 -0
  298. package/dist/test/repository-oracle-evidence.test.d.ts +1 -0
  299. package/dist/test/repository-oracle-evidence.test.js +185 -0
  300. package/dist/test/repository-oracle-plan.test.d.ts +1 -0
  301. package/dist/test/repository-oracle-plan.test.js +176 -0
  302. package/dist/test/repository-overlay-isolation.test.d.ts +1 -0
  303. package/dist/test/repository-overlay-isolation.test.js +85 -0
  304. package/dist/test/repository-preparation-cache.test.d.ts +1 -0
  305. package/dist/test/repository-preparation-cache.test.js +414 -0
  306. package/dist/test/repository-public-artifact.test.d.ts +1 -0
  307. package/dist/test/repository-public-artifact.test.js +273 -0
  308. package/dist/test/repository-qualification-diagnostics.test.d.ts +1 -0
  309. package/dist/test/repository-qualification-diagnostics.test.js +524 -0
  310. package/dist/test/repository-reference-authoring.test.d.ts +1 -0
  311. package/dist/test/repository-reference-authoring.test.js +1633 -0
  312. package/dist/test/repository-review-evidence-v2.test.d.ts +1 -0
  313. package/dist/test/repository-review-evidence-v2.test.js +183 -0
  314. package/dist/test/repository-review-evidence.test.d.ts +1 -0
  315. package/dist/test/repository-review-evidence.test.js +124 -0
  316. package/dist/test/repository-seed-exclusions.test.d.ts +1 -0
  317. package/dist/test/repository-seed-exclusions.test.js +96 -0
  318. package/dist/test/repository-seed-selection.test.d.ts +1 -0
  319. package/dist/test/repository-seed-selection.test.js +504 -0
  320. package/dist/test/repository-semantic-calibration.test.d.ts +1 -0
  321. package/dist/test/repository-semantic-calibration.test.js +688 -0
  322. package/dist/test/repository-solution-edits.test.d.ts +1 -0
  323. package/dist/test/repository-solution-edits.test.js +377 -0
  324. package/dist/test/repository-specification-budget.test.d.ts +1 -0
  325. package/dist/test/repository-specification-budget.test.js +171 -0
  326. package/dist/test/repository-specification-contract-checkpoint.test.d.ts +1 -0
  327. package/dist/test/repository-specification-contract-checkpoint.test.js +228 -0
  328. package/dist/test/repository-specification-contract-facts.test.d.ts +1 -0
  329. package/dist/test/repository-specification-contract-facts.test.js +177 -0
  330. package/dist/test/repository-trajectory-authoring.test.d.ts +1 -0
  331. package/dist/test/repository-trajectory-authoring.test.js +176 -0
  332. package/dist/test/repository-valid-control-plan.test.d.ts +1 -0
  333. package/dist/test/repository-valid-control-plan.test.js +45 -0
  334. package/dist/test/repository-vitest-phase.test.d.ts +1 -0
  335. package/dist/test/repository-vitest-phase.test.js +848 -0
  336. package/dist/test/repository-vitest-reporter.test.d.ts +1 -0
  337. package/dist/test/repository-vitest-reporter.test.js +158 -0
  338. package/dist/test/repository-workspace-build.test.d.ts +1 -0
  339. package/dist/test/repository-workspace-build.test.js +160 -0
  340. package/dist/test/strict-authoring-schema.test.d.ts +1 -0
  341. package/dist/test/strict-authoring-schema.test.js +169 -0
  342. package/package.json +48 -6
@@ -1,5 +1,6 @@
1
1
  import { type EvalComparisonResult as EvalComparisonResultType } from "@velum-labs/routekit-eval-contracts";
2
- import { Schema } from "effect";
2
+ import { type CostServiceShape } from "@velum-labs/routekit-registry";
3
+ import { Effect, Schema } from "effect";
3
4
  export declare const EVAL_PROJECT_VERSION: 1;
4
5
  export declare const EVAL_PROJECT_V2_VERSION: 2;
5
6
  export declare const EvalProjectQuestion: Schema.Struct<{
@@ -4151,7 +4152,71 @@ export declare const EvalProjectDocument: Schema.Union<readonly [Schema.Union<re
4151
4152
  }>;
4152
4153
  }>]>]>;
4153
4154
  export type EvalProjectDocument = typeof EvalProjectDocument.Type;
4155
+ export declare const EvalEvidenceBlock: Schema.Struct<{
4156
+ readonly id: Schema.String;
4157
+ readonly path: Schema.String;
4158
+ readonly sourceDigest: Schema.String;
4159
+ readonly contentDigest: Schema.String;
4160
+ readonly ordinal: Schema.Finite;
4161
+ readonly startLine: Schema.Finite;
4162
+ readonly endLine: Schema.Finite;
4163
+ readonly text: Schema.String;
4164
+ }>;
4165
+ export type EvalEvidenceBlock = typeof EvalEvidenceBlock.Type;
4166
+ export declare const EvalEvidenceSource: Schema.Struct<{
4167
+ readonly path: Schema.String;
4168
+ readonly sourceDigest: Schema.String;
4169
+ readonly sourceBytes: Schema.optionalKey<Schema.Finite>;
4170
+ readonly sourceBlockCount: Schema.optionalKey<Schema.Finite>;
4171
+ readonly blocks: Schema.$Array<Schema.Struct<{
4172
+ readonly id: Schema.String;
4173
+ readonly path: Schema.String;
4174
+ readonly sourceDigest: Schema.String;
4175
+ readonly contentDigest: Schema.String;
4176
+ readonly ordinal: Schema.Finite;
4177
+ readonly startLine: Schema.Finite;
4178
+ readonly endLine: Schema.Finite;
4179
+ readonly text: Schema.String;
4180
+ }>>;
4181
+ }>;
4182
+ export type EvalEvidenceSource = typeof EvalEvidenceSource.Type;
4183
+ export declare const EvalCriterion: Schema.Struct<{
4184
+ readonly id: Schema.String;
4185
+ readonly description: Schema.String;
4186
+ readonly importance: Schema.Literals<readonly ["critical", "supporting"]>;
4187
+ readonly gradingMode: Schema.Literal<"semantic">;
4188
+ }>;
4189
+ export type EvalCriterion = typeof EvalCriterion.Type;
4190
+ export declare const EvalPositiveControl: Schema.Struct<{
4191
+ readonly id: Schema.String;
4192
+ readonly response: Schema.String;
4193
+ }>;
4194
+ export type EvalPositiveControl = typeof EvalPositiveControl.Type;
4195
+ export declare const EvalNegativeControl: Schema.Struct<{
4196
+ readonly id: Schema.String;
4197
+ readonly response: Schema.String;
4198
+ readonly targetedCriterionIds: Schema.$Array<Schema.String>;
4199
+ readonly kind: Schema.Literals<readonly ["missing-criterion", "forbidden-behavior", "non-responsive"]>;
4200
+ }>;
4201
+ export type EvalNegativeControl = typeof EvalNegativeControl.Type;
4154
4202
  export declare const EvalDimensionCase: Schema.Struct<{
4203
+ readonly evidenceIds: Schema.optionalKey<Schema.$Array<Schema.String>>;
4204
+ readonly criteria: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4205
+ readonly id: Schema.String;
4206
+ readonly description: Schema.String;
4207
+ readonly importance: Schema.Literals<readonly ["critical", "supporting"]>;
4208
+ readonly gradingMode: Schema.Literal<"semantic">;
4209
+ }>>>;
4210
+ readonly positiveControls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4211
+ readonly id: Schema.String;
4212
+ readonly response: Schema.String;
4213
+ }>>>;
4214
+ readonly negativeControls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4215
+ readonly id: Schema.String;
4216
+ readonly response: Schema.String;
4217
+ readonly targetedCriterionIds: Schema.$Array<Schema.String>;
4218
+ readonly kind: Schema.Literals<readonly ["missing-criterion", "forbidden-behavior", "non-responsive"]>;
4219
+ }>>>;
4155
4220
  readonly id: Schema.String;
4156
4221
  readonly prompt: Schema.String;
4157
4222
  readonly context: Schema.optionalKey<Schema.String>;
@@ -4163,6 +4228,23 @@ export declare const EvalDimensionSuite: Schema.Struct<{
4163
4228
  readonly dimensionId: Schema.String;
4164
4229
  readonly maximumOutputTokens: Schema.Finite;
4165
4230
  readonly cases: Schema.$Array<Schema.Struct<{
4231
+ readonly evidenceIds: Schema.optionalKey<Schema.$Array<Schema.String>>;
4232
+ readonly criteria: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4233
+ readonly id: Schema.String;
4234
+ readonly description: Schema.String;
4235
+ readonly importance: Schema.Literals<readonly ["critical", "supporting"]>;
4236
+ readonly gradingMode: Schema.Literal<"semantic">;
4237
+ }>>>;
4238
+ readonly positiveControls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4239
+ readonly id: Schema.String;
4240
+ readonly response: Schema.String;
4241
+ }>>>;
4242
+ readonly negativeControls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4243
+ readonly id: Schema.String;
4244
+ readonly response: Schema.String;
4245
+ readonly targetedCriterionIds: Schema.$Array<Schema.String>;
4246
+ readonly kind: Schema.Literals<readonly ["missing-criterion", "forbidden-behavior", "non-responsive"]>;
4247
+ }>>>;
4166
4248
  readonly id: Schema.String;
4167
4249
  readonly prompt: Schema.String;
4168
4250
  readonly context: Schema.optionalKey<Schema.String>;
@@ -4198,10 +4280,6 @@ export declare const EvalDecompositionBenchmark: Schema.Struct<{
4198
4280
  }>;
4199
4281
  export type EvalDecompositionBenchmark = typeof EvalDecompositionBenchmark.Type;
4200
4282
  export declare const EvalCompositionCase: Schema.Struct<{
4201
- readonly id: Schema.String;
4202
- readonly prompt: Schema.String;
4203
- readonly context: Schema.optionalKey<Schema.String>;
4204
- readonly rubric: Schema.String;
4205
4283
  readonly decomposition: Schema.Struct<{
4206
4284
  readonly weights: Schema.$Array<Schema.Struct<{
4207
4285
  readonly dimensionId: Schema.String;
@@ -4216,6 +4294,27 @@ export declare const EvalCompositionCase: Schema.Struct<{
4216
4294
  readonly inputTokens: Schema.optionalKey<Schema.Finite>;
4217
4295
  readonly maxOutputTokens: Schema.optionalKey<Schema.Finite>;
4218
4296
  }>;
4297
+ readonly evidenceIds: Schema.optionalKey<Schema.$Array<Schema.String>>;
4298
+ readonly criteria: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4299
+ readonly id: Schema.String;
4300
+ readonly description: Schema.String;
4301
+ readonly importance: Schema.Literals<readonly ["critical", "supporting"]>;
4302
+ readonly gradingMode: Schema.Literal<"semantic">;
4303
+ }>>>;
4304
+ readonly positiveControls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4305
+ readonly id: Schema.String;
4306
+ readonly response: Schema.String;
4307
+ }>>>;
4308
+ readonly negativeControls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4309
+ readonly id: Schema.String;
4310
+ readonly response: Schema.String;
4311
+ readonly targetedCriterionIds: Schema.$Array<Schema.String>;
4312
+ readonly kind: Schema.Literals<readonly ["missing-criterion", "forbidden-behavior", "non-responsive"]>;
4313
+ }>>>;
4314
+ readonly id: Schema.String;
4315
+ readonly prompt: Schema.String;
4316
+ readonly context: Schema.optionalKey<Schema.String>;
4317
+ readonly rubric: Schema.String;
4219
4318
  }>;
4220
4319
  export type EvalCompositionCase = typeof EvalCompositionCase.Type;
4221
4320
  export declare const EvalCompositionSuite: Schema.Struct<{
@@ -4223,10 +4322,6 @@ export declare const EvalCompositionSuite: Schema.Struct<{
4223
4322
  readonly minimumWinnerScoreGap: Schema.Finite;
4224
4323
  readonly minimumWinnerAgreement: Schema.Finite;
4225
4324
  readonly cases: Schema.$Array<Schema.Struct<{
4226
- readonly id: Schema.String;
4227
- readonly prompt: Schema.String;
4228
- readonly context: Schema.optionalKey<Schema.String>;
4229
- readonly rubric: Schema.String;
4230
4325
  readonly decomposition: Schema.Struct<{
4231
4326
  readonly weights: Schema.$Array<Schema.Struct<{
4232
4327
  readonly dimensionId: Schema.String;
@@ -4241,6 +4336,27 @@ export declare const EvalCompositionSuite: Schema.Struct<{
4241
4336
  readonly inputTokens: Schema.optionalKey<Schema.Finite>;
4242
4337
  readonly maxOutputTokens: Schema.optionalKey<Schema.Finite>;
4243
4338
  }>;
4339
+ readonly evidenceIds: Schema.optionalKey<Schema.$Array<Schema.String>>;
4340
+ readonly criteria: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4341
+ readonly id: Schema.String;
4342
+ readonly description: Schema.String;
4343
+ readonly importance: Schema.Literals<readonly ["critical", "supporting"]>;
4344
+ readonly gradingMode: Schema.Literal<"semantic">;
4345
+ }>>>;
4346
+ readonly positiveControls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4347
+ readonly id: Schema.String;
4348
+ readonly response: Schema.String;
4349
+ }>>>;
4350
+ readonly negativeControls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4351
+ readonly id: Schema.String;
4352
+ readonly response: Schema.String;
4353
+ readonly targetedCriterionIds: Schema.$Array<Schema.String>;
4354
+ readonly kind: Schema.Literals<readonly ["missing-criterion", "forbidden-behavior", "non-responsive"]>;
4355
+ }>>>;
4356
+ readonly id: Schema.String;
4357
+ readonly prompt: Schema.String;
4358
+ readonly context: Schema.optionalKey<Schema.String>;
4359
+ readonly rubric: Schema.String;
4244
4360
  }>>;
4245
4361
  }>;
4246
4362
  export type EvalCompositionSuite = typeof EvalCompositionSuite.Type;
@@ -4251,11 +4367,44 @@ export declare const EvalEvaluationProposal: Schema.Struct<{
4251
4367
  readonly candidateModels: Schema.$Array<Schema.String>;
4252
4368
  readonly candidateReasoningEfforts: Schema.optionalKey<Schema.$Record<Schema.String, Schema.NonEmptyString>>;
4253
4369
  readonly judgeModel: Schema.String;
4370
+ readonly evidenceSources: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4371
+ readonly path: Schema.String;
4372
+ readonly sourceDigest: Schema.String;
4373
+ readonly sourceBytes: Schema.optionalKey<Schema.Finite>;
4374
+ readonly sourceBlockCount: Schema.optionalKey<Schema.Finite>;
4375
+ readonly blocks: Schema.$Array<Schema.Struct<{
4376
+ readonly id: Schema.String;
4377
+ readonly path: Schema.String;
4378
+ readonly sourceDigest: Schema.String;
4379
+ readonly contentDigest: Schema.String;
4380
+ readonly ordinal: Schema.Finite;
4381
+ readonly startLine: Schema.Finite;
4382
+ readonly endLine: Schema.Finite;
4383
+ readonly text: Schema.String;
4384
+ }>>;
4385
+ }>>>;
4254
4386
  readonly suites: Schema.$Array<Schema.Struct<{
4255
4387
  readonly version: Schema.Literal<1>;
4256
4388
  readonly dimensionId: Schema.String;
4257
4389
  readonly maximumOutputTokens: Schema.Finite;
4258
4390
  readonly cases: Schema.$Array<Schema.Struct<{
4391
+ readonly evidenceIds: Schema.optionalKey<Schema.$Array<Schema.String>>;
4392
+ readonly criteria: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4393
+ readonly id: Schema.String;
4394
+ readonly description: Schema.String;
4395
+ readonly importance: Schema.Literals<readonly ["critical", "supporting"]>;
4396
+ readonly gradingMode: Schema.Literal<"semantic">;
4397
+ }>>>;
4398
+ readonly positiveControls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4399
+ readonly id: Schema.String;
4400
+ readonly response: Schema.String;
4401
+ }>>>;
4402
+ readonly negativeControls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4403
+ readonly id: Schema.String;
4404
+ readonly response: Schema.String;
4405
+ readonly targetedCriterionIds: Schema.$Array<Schema.String>;
4406
+ readonly kind: Schema.Literals<readonly ["missing-criterion", "forbidden-behavior", "non-responsive"]>;
4407
+ }>>>;
4259
4408
  readonly id: Schema.String;
4260
4409
  readonly prompt: Schema.String;
4261
4410
  readonly context: Schema.optionalKey<Schema.String>;
@@ -4281,10 +4430,6 @@ export declare const EvalEvaluationProposal: Schema.Struct<{
4281
4430
  readonly minimumWinnerScoreGap: Schema.Finite;
4282
4431
  readonly minimumWinnerAgreement: Schema.Finite;
4283
4432
  readonly cases: Schema.$Array<Schema.Struct<{
4284
- readonly id: Schema.String;
4285
- readonly prompt: Schema.String;
4286
- readonly context: Schema.optionalKey<Schema.String>;
4287
- readonly rubric: Schema.String;
4288
4433
  readonly decomposition: Schema.Struct<{
4289
4434
  readonly weights: Schema.$Array<Schema.Struct<{
4290
4435
  readonly dimensionId: Schema.String;
@@ -4299,6 +4444,27 @@ export declare const EvalEvaluationProposal: Schema.Struct<{
4299
4444
  readonly inputTokens: Schema.optionalKey<Schema.Finite>;
4300
4445
  readonly maxOutputTokens: Schema.optionalKey<Schema.Finite>;
4301
4446
  }>;
4447
+ readonly evidenceIds: Schema.optionalKey<Schema.$Array<Schema.String>>;
4448
+ readonly criteria: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4449
+ readonly id: Schema.String;
4450
+ readonly description: Schema.String;
4451
+ readonly importance: Schema.Literals<readonly ["critical", "supporting"]>;
4452
+ readonly gradingMode: Schema.Literal<"semantic">;
4453
+ }>>>;
4454
+ readonly positiveControls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4455
+ readonly id: Schema.String;
4456
+ readonly response: Schema.String;
4457
+ }>>>;
4458
+ readonly negativeControls: Schema.optionalKey<Schema.$Array<Schema.Struct<{
4459
+ readonly id: Schema.String;
4460
+ readonly response: Schema.String;
4461
+ readonly targetedCriterionIds: Schema.$Array<Schema.String>;
4462
+ readonly kind: Schema.Literals<readonly ["missing-criterion", "forbidden-behavior", "non-responsive"]>;
4463
+ }>>>;
4464
+ readonly id: Schema.String;
4465
+ readonly prompt: Schema.String;
4466
+ readonly context: Schema.optionalKey<Schema.String>;
4467
+ readonly rubric: Schema.String;
4302
4468
  }>>;
4303
4469
  }>;
4304
4470
  }>;
@@ -4308,10 +4474,23 @@ export declare const EvalArtifactApproval: Schema.Struct<{
4308
4474
  readonly kind: Schema.Literals<readonly ["routing-basis", "evaluations"]>;
4309
4475
  readonly digest: Schema.String;
4310
4476
  readonly approvedAt: Schema.String;
4477
+ readonly structurePolicyVersion: Schema.optionalKey<Schema.Finite>;
4478
+ readonly structureReportDigest: Schema.optionalKey<Schema.String>;
4311
4479
  }>;
4312
4480
  export type EvalArtifactApproval = typeof EvalArtifactApproval.Type;
4313
4481
  export declare const EvalPlanScope: Schema.Literals<readonly ["pilot", "full"]>;
4314
4482
  export type EvalPlanScope = typeof EvalPlanScope.Type;
4483
+ export declare const EvalCostEstimate: Schema.Struct<{
4484
+ readonly callCount: Schema.Finite;
4485
+ readonly maximumInputTokens: Schema.Finite;
4486
+ readonly maximumOutputTokens: Schema.Finite;
4487
+ readonly maximumCostUsd: Schema.optionalKey<Schema.Finite>;
4488
+ readonly knownPricedMaximumUsd: Schema.Finite;
4489
+ readonly unpricedCalls: Schema.Finite;
4490
+ readonly subscriptionUnitCalls: Schema.Finite;
4491
+ readonly unknowns: Schema.$Array<Schema.String>;
4492
+ }>;
4493
+ export type EvalCostEstimate = typeof EvalCostEstimate.Type;
4315
4494
  export declare const EvalExecutionPlan: Schema.Struct<{
4316
4495
  readonly version: Schema.Literal<1>;
4317
4496
  readonly planId: Schema.String;
@@ -4319,6 +4498,7 @@ export declare const EvalExecutionPlan: Schema.Struct<{
4319
4498
  readonly projectRevision: Schema.Finite;
4320
4499
  readonly createdAt: Schema.String;
4321
4500
  readonly scope: Schema.Literals<readonly ["pilot", "full"]>;
4501
+ readonly repetitions: Schema.withDecodingDefaultKey<Schema.Finite, never>;
4322
4502
  readonly basisDigest: Schema.String;
4323
4503
  readonly evaluationDigest: Schema.String;
4324
4504
  readonly candidateModels: Schema.$Array<Schema.String>;
@@ -4341,6 +4521,16 @@ export declare const EvalExecutionPlan: Schema.Struct<{
4341
4521
  readonly expectedCandidateCalls: Schema.Finite;
4342
4522
  readonly expectedJudgeCalls: Schema.Finite;
4343
4523
  readonly expectedCallCount: Schema.Finite;
4524
+ readonly costEstimate: Schema.optionalKey<Schema.Struct<{
4525
+ readonly callCount: Schema.Finite;
4526
+ readonly maximumInputTokens: Schema.Finite;
4527
+ readonly maximumOutputTokens: Schema.Finite;
4528
+ readonly maximumCostUsd: Schema.optionalKey<Schema.Finite>;
4529
+ readonly knownPricedMaximumUsd: Schema.Finite;
4530
+ readonly unpricedCalls: Schema.Finite;
4531
+ readonly subscriptionUnitCalls: Schema.Finite;
4532
+ readonly unknowns: Schema.$Array<Schema.String>;
4533
+ }>>;
4344
4534
  }>;
4345
4535
  export type EvalExecutionPlan = typeof EvalExecutionPlan.Type;
4346
4536
  export type EvalScopedExpectedCalls = {
@@ -4354,6 +4544,8 @@ export type EvalScopedExpectedCalls = {
4354
4544
  readonly expectedCallCount: number;
4355
4545
  };
4356
4546
  export declare function scopedExpectedCalls(plan: Pick<EvalExecutionPlan, "candidateModels" | "selectedCaseIds" | "selectedDecompositionCaseIds" | "selectedCompositionCaseIds">): EvalScopedExpectedCalls;
4547
+ export declare const EVAL_MAXIMUM_INPUT_TOKENS_PER_CALL: number;
4548
+ export declare function estimateEvalPlanCost(plan: Pick<EvalExecutionPlan, "candidateModels" | "candidateReasoningEfforts" | "classifierModel" | "judgeModel" | "maximumOutputTokens" | "expectedDimensionCandidateCalls" | "expectedClassifierCalls" | "expectedCompositionCandidateCalls">, costService: CostServiceShape): Effect.Effect<EvalCostEstimate>;
4357
4549
  export declare const EvalRunTarget: Schema.Struct<{
4358
4550
  readonly kind: Schema.Literals<readonly ["configured", "external"]>;
4359
4551
  readonly identity: Schema.String;
@@ -4375,6 +4567,7 @@ export declare const EvalRunLedger: Schema.Struct<{
4375
4567
  readonly unknownTokenMeasurements: Schema.Finite;
4376
4568
  readonly knownPricedSubtotalUsd: Schema.Finite;
4377
4569
  readonly unpricedCalls: Schema.Finite;
4570
+ readonly subscriptionUnitCalls: Schema.withDecodingDefaultKey<Schema.Finite, never>;
4378
4571
  }>;
4379
4572
  export type EvalRunLedger = typeof EvalRunLedger.Type;
4380
4573
  export declare const EvalClassifierObservation: Schema.Struct<{
@@ -4397,10 +4590,13 @@ export declare const EvalClassifierObservation: Schema.Struct<{
4397
4590
  export type EvalClassifierObservation = typeof EvalClassifierObservation.Type;
4398
4591
  export declare const EvalCompositionCaseResult: Schema.Struct<{
4399
4592
  readonly caseId: Schema.String;
4400
- readonly predictedWinner: Schema.String;
4401
- readonly observedWinner: Schema.String;
4593
+ readonly predictedWinner: Schema.NullOr<Schema.String>;
4594
+ readonly predictedTiedModels: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.$Array<Schema.String>>, never>;
4595
+ readonly observedWinner: Schema.NullOr<Schema.String>;
4596
+ readonly observedTiedModels: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.$Array<Schema.String>>, never>;
4402
4597
  readonly predictedScoreGap: Schema.Finite;
4403
4598
  readonly observedScoreGap: Schema.Finite;
4599
+ readonly comparable: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.Boolean>, never>;
4404
4600
  readonly passed: Schema.Boolean;
4405
4601
  }>;
4406
4602
  export type EvalCompositionCaseResult = typeof EvalCompositionCaseResult.Type;
@@ -4431,19 +4627,25 @@ export declare const EvalRunQualification: Schema.Struct<{
4431
4627
  readonly expectedCases: Schema.Finite;
4432
4628
  readonly comparableCases: Schema.Finite;
4433
4629
  readonly agreeingCases: Schema.Finite;
4434
- readonly winnerAgreement: Schema.Finite;
4630
+ readonly winnerAgreement: Schema.optionalKey<Schema.Finite>;
4631
+ readonly conclusion: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.Literals<readonly ["passed", "failed", "inconclusive"]>>, never>;
4632
+ readonly reason: Schema.optionalKey<Schema.String>;
4633
+ readonly publishable: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.Boolean>, never>;
4435
4634
  readonly cases: Schema.$Array<Schema.Struct<{
4436
4635
  readonly caseId: Schema.String;
4437
- readonly predictedWinner: Schema.String;
4438
- readonly observedWinner: Schema.String;
4636
+ readonly predictedWinner: Schema.NullOr<Schema.String>;
4637
+ readonly predictedTiedModels: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.$Array<Schema.String>>, never>;
4638
+ readonly observedWinner: Schema.NullOr<Schema.String>;
4639
+ readonly observedTiedModels: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.$Array<Schema.String>>, never>;
4439
4640
  readonly predictedScoreGap: Schema.Finite;
4440
4641
  readonly observedScoreGap: Schema.Finite;
4642
+ readonly comparable: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.Boolean>, never>;
4441
4643
  readonly passed: Schema.Boolean;
4442
4644
  }>>;
4443
4645
  }>;
4444
4646
  }>;
4445
4647
  export type EvalRunQualification = typeof EvalRunQualification.Type;
4446
- export declare function summarizeEvalRunLedger(comparisons: readonly EvalComparisonResultType[], expectedCalls: number, classifierObservations?: readonly EvalClassifierObservation[]): EvalRunLedger;
4648
+ export declare function summarizeEvalRunLedger(comparisons: readonly EvalComparisonResultType[], expectedCalls: number, costService: CostServiceShape, classifierObservations?: readonly EvalClassifierObservation[], classifierModel?: string): Effect.Effect<EvalRunLedger>;
4447
4649
  export declare const EvalRunFailureError: Schema.Struct<{
4448
4650
  readonly name: Schema.String;
4449
4651
  readonly message: Schema.String;
@@ -4597,13 +4799,19 @@ export declare const EvalRunReport: Schema.Union<readonly [Schema.Struct<{
4597
4799
  readonly expectedCases: Schema.Finite;
4598
4800
  readonly comparableCases: Schema.Finite;
4599
4801
  readonly agreeingCases: Schema.Finite;
4600
- readonly winnerAgreement: Schema.Finite;
4802
+ readonly winnerAgreement: Schema.optionalKey<Schema.Finite>;
4803
+ readonly conclusion: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.Literals<readonly ["passed", "failed", "inconclusive"]>>, never>;
4804
+ readonly reason: Schema.optionalKey<Schema.String>;
4805
+ readonly publishable: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.Boolean>, never>;
4601
4806
  readonly cases: Schema.$Array<Schema.Struct<{
4602
4807
  readonly caseId: Schema.String;
4603
- readonly predictedWinner: Schema.String;
4604
- readonly observedWinner: Schema.String;
4808
+ readonly predictedWinner: Schema.NullOr<Schema.String>;
4809
+ readonly predictedTiedModels: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.$Array<Schema.String>>, never>;
4810
+ readonly observedWinner: Schema.NullOr<Schema.String>;
4811
+ readonly observedTiedModels: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.$Array<Schema.String>>, never>;
4605
4812
  readonly predictedScoreGap: Schema.Finite;
4606
4813
  readonly observedScoreGap: Schema.Finite;
4814
+ readonly comparable: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.Boolean>, never>;
4607
4815
  readonly passed: Schema.Boolean;
4608
4816
  }>>;
4609
4817
  }>;
@@ -4617,65 +4825,66 @@ export declare const EvalRunReport: Schema.Union<readonly [Schema.Struct<{
4617
4825
  readonly unknownTokenMeasurements: Schema.Finite;
4618
4826
  readonly knownPricedSubtotalUsd: Schema.Finite;
4619
4827
  readonly unpricedCalls: Schema.Finite;
4828
+ readonly subscriptionUnitCalls: Schema.withDecodingDefaultKey<Schema.Finite, never>;
4620
4829
  }>;
4830
+ readonly events: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.$Array<Schema.Union<readonly [Schema.Struct<{
4831
+ readonly name: Schema.Literal<"plan">;
4832
+ readonly at: Schema.String;
4833
+ readonly runId: Schema.String;
4834
+ readonly plannedCalls: Schema.Finite;
4835
+ }>, Schema.Struct<{
4836
+ readonly at: Schema.String;
4837
+ readonly runId: Schema.String;
4838
+ readonly dimensionId: Schema.String;
4839
+ readonly caseId: Schema.String;
4840
+ readonly sequence: Schema.Finite;
4841
+ readonly name: Schema.Literal<"candidate#">;
4842
+ }>, Schema.Struct<{
4843
+ readonly at: Schema.String;
4844
+ readonly runId: Schema.String;
4845
+ readonly dimensionId: Schema.String;
4846
+ readonly caseId: Schema.String;
4847
+ readonly sequence: Schema.Finite;
4848
+ readonly name: Schema.Literal<"judge#">;
4849
+ }>, Schema.Struct<{
4850
+ readonly name: Schema.Literal<"failure-detail">;
4851
+ readonly at: Schema.String;
4852
+ readonly runId: Schema.String;
4853
+ readonly phase: Schema.Literals<readonly ["target", "classifier", "suite-inspection", "comparison", "activation", "accounting", "completion"]>;
4854
+ readonly reason: Schema.String;
4855
+ readonly retryable: Schema.Boolean;
4856
+ readonly callId: Schema.optionalKey<Schema.String>;
4857
+ readonly reservationId: Schema.optionalKey<Schema.String>;
4858
+ readonly accounting: Schema.Struct<{
4859
+ readonly expectedCalls: Schema.Finite;
4860
+ readonly observedCalls: Schema.Finite;
4861
+ readonly knownInputTokens: Schema.Finite;
4862
+ readonly knownOutputTokens: Schema.Finite;
4863
+ readonly unknownTokenMeasurements: Schema.Finite;
4864
+ readonly knownPricedSubtotalUsd: Schema.Finite;
4865
+ readonly unpricedCalls: Schema.Finite;
4866
+ readonly subscriptionUnitCalls: Schema.Finite;
4867
+ }>;
4868
+ readonly cleanup: Schema.Struct<{
4869
+ readonly sessionOpened: Schema.Boolean;
4870
+ readonly sessionClosed: Schema.Boolean;
4871
+ readonly detail: Schema.optionalKey<Schema.String>;
4872
+ }>;
4873
+ }>, Schema.Struct<{
4874
+ readonly at: Schema.String;
4875
+ readonly operationId: Schema.String;
4876
+ readonly name: Schema.Literal<"basis">;
4877
+ }>, Schema.Struct<{
4878
+ readonly at: Schema.String;
4879
+ readonly operationId: Schema.String;
4880
+ readonly name: Schema.Literal<"dimensions">;
4881
+ }>, Schema.Struct<{
4882
+ readonly at: Schema.String;
4883
+ readonly operationId: Schema.String;
4884
+ readonly name: Schema.Literal<"activation">;
4885
+ }>]>>>, never>;
4621
4886
  }>, Schema.Struct<{
4622
4887
  readonly status: Schema.Literal<"completed">;
4623
- readonly activation: Schema.Struct<{
4624
- readonly version: Schema.Literal<2>;
4625
- readonly generatedAt: Schema.String;
4626
- readonly basisDigest: Schema.String;
4627
- readonly evidenceDigest: Schema.String;
4628
- readonly classifierModel: Schema.String;
4629
- readonly objective: Schema.Union<readonly [Schema.Struct<{
4630
- readonly kind: Schema.Literal<"highest-quality">;
4631
- }>, Schema.Struct<{
4632
- readonly kind: Schema.Literal<"lowest-cost">;
4633
- readonly minimumQuality: Schema.Finite;
4634
- }>, Schema.Struct<{
4635
- readonly kind: Schema.Literal<"lowest-latency">;
4636
- readonly minimumQuality: Schema.Finite;
4637
- }>, Schema.Struct<{
4638
- readonly kind: Schema.Literal<"balanced">;
4639
- readonly minimumQuality: Schema.Finite;
4640
- readonly weights: Schema.Struct<{
4641
- readonly quality: Schema.Finite;
4642
- readonly cost: Schema.Finite;
4643
- readonly latency: Schema.Finite;
4644
- }>;
4645
- }>, Schema.Struct<{
4646
- readonly kind: Schema.Literal<"pareto">;
4647
- readonly minimumQuality: Schema.Finite;
4648
- readonly preference: Schema.Literals<readonly ["quality", "cost", "latency"]>;
4649
- }>]>;
4650
- readonly maximumUnknownWeight: Schema.Finite;
4651
- readonly constraints: Schema.optionalKey<Schema.Struct<{
4652
- readonly minimumDimensionQuality: Schema.optionalKey<Schema.$Record<Schema.String, Schema.Finite>>;
4653
- readonly maximumFailureRate: Schema.optionalKey<Schema.Finite>;
4654
- }>>;
4655
- readonly dimensions: Schema.$Array<Schema.Struct<{
4656
- readonly id: Schema.String;
4657
- readonly description: Schema.String;
4658
- readonly includes: Schema.$Array<Schema.String>;
4659
- readonly excludes: Schema.$Array<Schema.String>;
4660
- }>>;
4661
- readonly candidateModels: Schema.$Array<Schema.String>;
4662
- readonly evidence: Schema.$Array<Schema.Struct<{
4663
- readonly model: Schema.String;
4664
- readonly dimensionId: Schema.String;
4665
- readonly suiteDigest: Schema.String;
4666
- readonly evidenceDigest: Schema.String;
4667
- readonly quality: Schema.Struct<{
4668
- readonly passRate: Schema.Finite;
4669
- readonly lowerConfidenceBound: Schema.Finite;
4670
- readonly sampleCount: Schema.Finite;
4671
- }>;
4672
- readonly failureRate: Schema.Finite;
4673
- readonly averageJudgeScore: Schema.optionalKey<Schema.Finite>;
4674
- readonly p95DurationMs: Schema.optionalKey<Schema.Finite>;
4675
- readonly averageCostUsd: Schema.optionalKey<Schema.Finite>;
4676
- readonly unpricedCalls: Schema.Finite;
4677
- }>>;
4678
- }>;
4679
4888
  readonly version: Schema.Literal<1>;
4680
4889
  readonly runId: Schema.String;
4681
4890
  readonly planId: Schema.String;
@@ -4757,13 +4966,19 @@ export declare const EvalRunReport: Schema.Union<readonly [Schema.Struct<{
4757
4966
  readonly expectedCases: Schema.Finite;
4758
4967
  readonly comparableCases: Schema.Finite;
4759
4968
  readonly agreeingCases: Schema.Finite;
4760
- readonly winnerAgreement: Schema.Finite;
4969
+ readonly winnerAgreement: Schema.optionalKey<Schema.Finite>;
4970
+ readonly conclusion: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.Literals<readonly ["passed", "failed", "inconclusive"]>>, never>;
4971
+ readonly reason: Schema.optionalKey<Schema.String>;
4972
+ readonly publishable: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.Boolean>, never>;
4761
4973
  readonly cases: Schema.$Array<Schema.Struct<{
4762
4974
  readonly caseId: Schema.String;
4763
- readonly predictedWinner: Schema.String;
4764
- readonly observedWinner: Schema.String;
4975
+ readonly predictedWinner: Schema.NullOr<Schema.String>;
4976
+ readonly predictedTiedModels: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.$Array<Schema.String>>, never>;
4977
+ readonly observedWinner: Schema.NullOr<Schema.String>;
4978
+ readonly observedTiedModels: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.$Array<Schema.String>>, never>;
4765
4979
  readonly predictedScoreGap: Schema.Finite;
4766
4980
  readonly observedScoreGap: Schema.Finite;
4981
+ readonly comparable: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.Boolean>, never>;
4767
4982
  readonly passed: Schema.Boolean;
4768
4983
  }>>;
4769
4984
  }>;
@@ -4777,7 +4992,64 @@ export declare const EvalRunReport: Schema.Union<readonly [Schema.Struct<{
4777
4992
  readonly unknownTokenMeasurements: Schema.Finite;
4778
4993
  readonly knownPricedSubtotalUsd: Schema.Finite;
4779
4994
  readonly unpricedCalls: Schema.Finite;
4995
+ readonly subscriptionUnitCalls: Schema.withDecodingDefaultKey<Schema.Finite, never>;
4780
4996
  }>;
4997
+ readonly events: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.$Array<Schema.Union<readonly [Schema.Struct<{
4998
+ readonly name: Schema.Literal<"plan">;
4999
+ readonly at: Schema.String;
5000
+ readonly runId: Schema.String;
5001
+ readonly plannedCalls: Schema.Finite;
5002
+ }>, Schema.Struct<{
5003
+ readonly at: Schema.String;
5004
+ readonly runId: Schema.String;
5005
+ readonly dimensionId: Schema.String;
5006
+ readonly caseId: Schema.String;
5007
+ readonly sequence: Schema.Finite;
5008
+ readonly name: Schema.Literal<"candidate#">;
5009
+ }>, Schema.Struct<{
5010
+ readonly at: Schema.String;
5011
+ readonly runId: Schema.String;
5012
+ readonly dimensionId: Schema.String;
5013
+ readonly caseId: Schema.String;
5014
+ readonly sequence: Schema.Finite;
5015
+ readonly name: Schema.Literal<"judge#">;
5016
+ }>, Schema.Struct<{
5017
+ readonly name: Schema.Literal<"failure-detail">;
5018
+ readonly at: Schema.String;
5019
+ readonly runId: Schema.String;
5020
+ readonly phase: Schema.Literals<readonly ["target", "classifier", "suite-inspection", "comparison", "activation", "accounting", "completion"]>;
5021
+ readonly reason: Schema.String;
5022
+ readonly retryable: Schema.Boolean;
5023
+ readonly callId: Schema.optionalKey<Schema.String>;
5024
+ readonly reservationId: Schema.optionalKey<Schema.String>;
5025
+ readonly accounting: Schema.Struct<{
5026
+ readonly expectedCalls: Schema.Finite;
5027
+ readonly observedCalls: Schema.Finite;
5028
+ readonly knownInputTokens: Schema.Finite;
5029
+ readonly knownOutputTokens: Schema.Finite;
5030
+ readonly unknownTokenMeasurements: Schema.Finite;
5031
+ readonly knownPricedSubtotalUsd: Schema.Finite;
5032
+ readonly unpricedCalls: Schema.Finite;
5033
+ readonly subscriptionUnitCalls: Schema.Finite;
5034
+ }>;
5035
+ readonly cleanup: Schema.Struct<{
5036
+ readonly sessionOpened: Schema.Boolean;
5037
+ readonly sessionClosed: Schema.Boolean;
5038
+ readonly detail: Schema.optionalKey<Schema.String>;
5039
+ }>;
5040
+ }>, Schema.Struct<{
5041
+ readonly at: Schema.String;
5042
+ readonly operationId: Schema.String;
5043
+ readonly name: Schema.Literal<"basis">;
5044
+ }>, Schema.Struct<{
5045
+ readonly at: Schema.String;
5046
+ readonly operationId: Schema.String;
5047
+ readonly name: Schema.Literal<"dimensions">;
5048
+ }>, Schema.Struct<{
5049
+ readonly at: Schema.String;
5050
+ readonly operationId: Schema.String;
5051
+ readonly name: Schema.Literal<"activation">;
5052
+ }>]>>>, never>;
4781
5053
  }>, Schema.Struct<{
4782
5054
  readonly status: Schema.Literal<"failed">;
4783
5055
  readonly failure: Schema.String;
@@ -4867,13 +5139,19 @@ export declare const EvalRunReport: Schema.Union<readonly [Schema.Struct<{
4867
5139
  readonly expectedCases: Schema.Finite;
4868
5140
  readonly comparableCases: Schema.Finite;
4869
5141
  readonly agreeingCases: Schema.Finite;
4870
- readonly winnerAgreement: Schema.Finite;
5142
+ readonly winnerAgreement: Schema.optionalKey<Schema.Finite>;
5143
+ readonly conclusion: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.Literals<readonly ["passed", "failed", "inconclusive"]>>, never>;
5144
+ readonly reason: Schema.optionalKey<Schema.String>;
5145
+ readonly publishable: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.Boolean>, never>;
4871
5146
  readonly cases: Schema.$Array<Schema.Struct<{
4872
5147
  readonly caseId: Schema.String;
4873
- readonly predictedWinner: Schema.String;
4874
- readonly observedWinner: Schema.String;
5148
+ readonly predictedWinner: Schema.NullOr<Schema.String>;
5149
+ readonly predictedTiedModels: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.$Array<Schema.String>>, never>;
5150
+ readonly observedWinner: Schema.NullOr<Schema.String>;
5151
+ readonly observedTiedModels: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.$Array<Schema.String>>, never>;
4875
5152
  readonly predictedScoreGap: Schema.Finite;
4876
5153
  readonly observedScoreGap: Schema.Finite;
5154
+ readonly comparable: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.Boolean>, never>;
4877
5155
  readonly passed: Schema.Boolean;
4878
5156
  }>>;
4879
5157
  }>;
@@ -4887,7 +5165,64 @@ export declare const EvalRunReport: Schema.Union<readonly [Schema.Struct<{
4887
5165
  readonly unknownTokenMeasurements: Schema.Finite;
4888
5166
  readonly knownPricedSubtotalUsd: Schema.Finite;
4889
5167
  readonly unpricedCalls: Schema.Finite;
5168
+ readonly subscriptionUnitCalls: Schema.withDecodingDefaultKey<Schema.Finite, never>;
4890
5169
  }>;
5170
+ readonly events: Schema.withDecodingDefaultKey<Schema.optionalKey<Schema.$Array<Schema.Union<readonly [Schema.Struct<{
5171
+ readonly name: Schema.Literal<"plan">;
5172
+ readonly at: Schema.String;
5173
+ readonly runId: Schema.String;
5174
+ readonly plannedCalls: Schema.Finite;
5175
+ }>, Schema.Struct<{
5176
+ readonly at: Schema.String;
5177
+ readonly runId: Schema.String;
5178
+ readonly dimensionId: Schema.String;
5179
+ readonly caseId: Schema.String;
5180
+ readonly sequence: Schema.Finite;
5181
+ readonly name: Schema.Literal<"candidate#">;
5182
+ }>, Schema.Struct<{
5183
+ readonly at: Schema.String;
5184
+ readonly runId: Schema.String;
5185
+ readonly dimensionId: Schema.String;
5186
+ readonly caseId: Schema.String;
5187
+ readonly sequence: Schema.Finite;
5188
+ readonly name: Schema.Literal<"judge#">;
5189
+ }>, Schema.Struct<{
5190
+ readonly name: Schema.Literal<"failure-detail">;
5191
+ readonly at: Schema.String;
5192
+ readonly runId: Schema.String;
5193
+ readonly phase: Schema.Literals<readonly ["target", "classifier", "suite-inspection", "comparison", "activation", "accounting", "completion"]>;
5194
+ readonly reason: Schema.String;
5195
+ readonly retryable: Schema.Boolean;
5196
+ readonly callId: Schema.optionalKey<Schema.String>;
5197
+ readonly reservationId: Schema.optionalKey<Schema.String>;
5198
+ readonly accounting: Schema.Struct<{
5199
+ readonly expectedCalls: Schema.Finite;
5200
+ readonly observedCalls: Schema.Finite;
5201
+ readonly knownInputTokens: Schema.Finite;
5202
+ readonly knownOutputTokens: Schema.Finite;
5203
+ readonly unknownTokenMeasurements: Schema.Finite;
5204
+ readonly knownPricedSubtotalUsd: Schema.Finite;
5205
+ readonly unpricedCalls: Schema.Finite;
5206
+ readonly subscriptionUnitCalls: Schema.Finite;
5207
+ }>;
5208
+ readonly cleanup: Schema.Struct<{
5209
+ readonly sessionOpened: Schema.Boolean;
5210
+ readonly sessionClosed: Schema.Boolean;
5211
+ readonly detail: Schema.optionalKey<Schema.String>;
5212
+ }>;
5213
+ }>, Schema.Struct<{
5214
+ readonly at: Schema.String;
5215
+ readonly operationId: Schema.String;
5216
+ readonly name: Schema.Literal<"basis">;
5217
+ }>, Schema.Struct<{
5218
+ readonly at: Schema.String;
5219
+ readonly operationId: Schema.String;
5220
+ readonly name: Schema.Literal<"dimensions">;
5221
+ }>, Schema.Struct<{
5222
+ readonly at: Schema.String;
5223
+ readonly operationId: Schema.String;
5224
+ readonly name: Schema.Literal<"activation">;
5225
+ }>]>>>, never>;
4891
5226
  }>]>;
4892
5227
  export type EvalRunReport = typeof EvalRunReport.Type;
4893
5228
  export type EvalProjectArtifactsStatus = {