@prismatic-io/lux 0.0.1

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 (636) hide show
  1. package/LICENSE +21 -0
  2. package/lib/answerers/claude-code/index.d.ts +17 -0
  3. package/lib/answerers/claude-code/index.d.ts.map +1 -0
  4. package/lib/answerers/claude-code/index.js +280 -0
  5. package/lib/answerers/claude-code/index.js.map +1 -0
  6. package/lib/answerers/persona/index.d.ts +40 -0
  7. package/lib/answerers/persona/index.d.ts.map +1 -0
  8. package/lib/answerers/persona/index.js +265 -0
  9. package/lib/answerers/persona/index.js.map +1 -0
  10. package/lib/answerers/scripted/index.d.ts +8 -0
  11. package/lib/answerers/scripted/index.d.ts.map +1 -0
  12. package/lib/answerers/scripted/index.js +163 -0
  13. package/lib/answerers/scripted/index.js.map +1 -0
  14. package/lib/answerers/shared.d.ts +13 -0
  15. package/lib/answerers/shared.d.ts.map +1 -0
  16. package/lib/answerers/shared.js +17 -0
  17. package/lib/answerers/shared.js.map +1 -0
  18. package/lib/answerers/terminal/index.d.ts +9 -0
  19. package/lib/answerers/terminal/index.d.ts.map +1 -0
  20. package/lib/answerers/terminal/index.js +97 -0
  21. package/lib/answerers/terminal/index.js.map +1 -0
  22. package/lib/assertions/core/command-exits-zero.d.ts +21 -0
  23. package/lib/assertions/core/command-exits-zero.d.ts.map +1 -0
  24. package/lib/assertions/core/command-exits-zero.js +87 -0
  25. package/lib/assertions/core/command-exits-zero.js.map +1 -0
  26. package/lib/assertions/core/event-checks.d.ts +57 -0
  27. package/lib/assertions/core/event-checks.d.ts.map +1 -0
  28. package/lib/assertions/core/event-checks.js +123 -0
  29. package/lib/assertions/core/event-checks.js.map +1 -0
  30. package/lib/assertions/core/file-checks.d.ts +103 -0
  31. package/lib/assertions/core/file-checks.d.ts.map +1 -0
  32. package/lib/assertions/core/file-checks.js +288 -0
  33. package/lib/assertions/core/file-checks.js.map +1 -0
  34. package/lib/assertions/core/helpers.d.ts +66 -0
  35. package/lib/assertions/core/helpers.d.ts.map +1 -0
  36. package/lib/assertions/core/helpers.js +170 -0
  37. package/lib/assertions/core/helpers.js.map +1 -0
  38. package/lib/assertions/core/index.d.ts +15 -0
  39. package/lib/assertions/core/index.d.ts.map +1 -0
  40. package/lib/assertions/core/index.js +56 -0
  41. package/lib/assertions/core/index.js.map +1 -0
  42. package/lib/assertions/core/json-pointer-equals.d.ts +19 -0
  43. package/lib/assertions/core/json-pointer-equals.d.ts.map +1 -0
  44. package/lib/assertions/core/json-pointer-equals.js +71 -0
  45. package/lib/assertions/core/json-pointer-equals.js.map +1 -0
  46. package/lib/assertions/core/output-checks.d.ts +46 -0
  47. package/lib/assertions/core/output-checks.d.ts.map +1 -0
  48. package/lib/assertions/core/output-checks.js +121 -0
  49. package/lib/assertions/core/output-checks.js.map +1 -0
  50. package/lib/assertions/core/predicate.d.ts +14 -0
  51. package/lib/assertions/core/predicate.d.ts.map +1 -0
  52. package/lib/assertions/core/predicate.js +55 -0
  53. package/lib/assertions/core/predicate.js.map +1 -0
  54. package/lib/assertions/core/resource-checks.d.ts +71 -0
  55. package/lib/assertions/core/resource-checks.d.ts.map +1 -0
  56. package/lib/assertions/core/resource-checks.js +159 -0
  57. package/lib/assertions/core/resource-checks.js.map +1 -0
  58. package/lib/assertions/core/safe-regex.d.ts +23 -0
  59. package/lib/assertions/core/safe-regex.d.ts.map +1 -0
  60. package/lib/assertions/core/safe-regex.js +102 -0
  61. package/lib/assertions/core/safe-regex.js.map +1 -0
  62. package/lib/assertions/core/tool-called-with.d.ts +21 -0
  63. package/lib/assertions/core/tool-called-with.d.ts.map +1 -0
  64. package/lib/assertions/core/tool-called-with.js +69 -0
  65. package/lib/assertions/core/tool-called-with.js.map +1 -0
  66. package/lib/assertions/core/tool-calls.d.ts +22 -0
  67. package/lib/assertions/core/tool-calls.d.ts.map +1 -0
  68. package/lib/assertions/core/tool-calls.js +17 -0
  69. package/lib/assertions/core/tool-calls.js.map +1 -0
  70. package/lib/assertions/core/tool-invocation.d.ts +36 -0
  71. package/lib/assertions/core/tool-invocation.d.ts.map +1 -0
  72. package/lib/assertions/core/tool-invocation.js +170 -0
  73. package/lib/assertions/core/tool-invocation.js.map +1 -0
  74. package/lib/assertions/core/tool-result-checks.d.ts +59 -0
  75. package/lib/assertions/core/tool-result-checks.d.ts.map +1 -0
  76. package/lib/assertions/core/tool-result-checks.js +193 -0
  77. package/lib/assertions/core/tool-result-checks.js.map +1 -0
  78. package/lib/assertions/mcp/discovery-checks.d.ts +36 -0
  79. package/lib/assertions/mcp/discovery-checks.d.ts.map +1 -0
  80. package/lib/assertions/mcp/discovery-checks.js +122 -0
  81. package/lib/assertions/mcp/discovery-checks.js.map +1 -0
  82. package/lib/assertions/mcp/index.d.ts +10 -0
  83. package/lib/assertions/mcp/index.d.ts.map +1 -0
  84. package/lib/assertions/mcp/index.js +25 -0
  85. package/lib/assertions/mcp/index.js.map +1 -0
  86. package/lib/assertions/rubric/index.d.ts +104 -0
  87. package/lib/assertions/rubric/index.d.ts.map +1 -0
  88. package/lib/assertions/rubric/index.js +191 -0
  89. package/lib/assertions/rubric/index.js.map +1 -0
  90. package/lib/assertions/rubric/internal.d.ts +14 -0
  91. package/lib/assertions/rubric/internal.d.ts.map +1 -0
  92. package/lib/assertions/rubric/internal.js +335 -0
  93. package/lib/assertions/rubric/internal.js.map +1 -0
  94. package/lib/authoring.d.ts +3 -0
  95. package/lib/authoring.d.ts.map +1 -0
  96. package/lib/authoring.js +3 -0
  97. package/lib/authoring.js.map +1 -0
  98. package/lib/cli/bin.d.ts +3 -0
  99. package/lib/cli/bin.d.ts.map +1 -0
  100. package/lib/cli/bin.js +15 -0
  101. package/lib/cli/bin.js.map +1 -0
  102. package/lib/cli/command-runtime.d.ts +5 -0
  103. package/lib/cli/command-runtime.d.ts.map +1 -0
  104. package/lib/cli/command-runtime.js +30 -0
  105. package/lib/cli/command-runtime.js.map +1 -0
  106. package/lib/cli/init.d.ts +21 -0
  107. package/lib/cli/init.d.ts.map +1 -0
  108. package/lib/cli/init.js +340 -0
  109. package/lib/cli/init.js.map +1 -0
  110. package/lib/cli/program.d.ts +22 -0
  111. package/lib/cli/program.d.ts.map +1 -0
  112. package/lib/cli/program.js +757 -0
  113. package/lib/cli/program.js.map +1 -0
  114. package/lib/cli/render/format.d.ts +14 -0
  115. package/lib/cli/render/format.d.ts.map +1 -0
  116. package/lib/cli/render/format.js +71 -0
  117. package/lib/cli/render/format.js.map +1 -0
  118. package/lib/cli/render/reporter.d.ts +73 -0
  119. package/lib/cli/render/reporter.d.ts.map +1 -0
  120. package/lib/cli/render/reporter.js +519 -0
  121. package/lib/cli/render/reporter.js.map +1 -0
  122. package/lib/cli/render/tty.d.ts +77 -0
  123. package/lib/cli/render/tty.d.ts.map +1 -0
  124. package/lib/cli/render/tty.js +283 -0
  125. package/lib/cli/render/tty.js.map +1 -0
  126. package/lib/cli/run-options.d.ts +27 -0
  127. package/lib/cli/run-options.d.ts.map +1 -0
  128. package/lib/cli/run-options.js +58 -0
  129. package/lib/cli/run-options.js.map +1 -0
  130. package/lib/cli/skills.d.ts +38 -0
  131. package/lib/cli/skills.d.ts.map +1 -0
  132. package/lib/cli/skills.js +84 -0
  133. package/lib/cli/skills.js.map +1 -0
  134. package/lib/core/annotation.d.ts +56 -0
  135. package/lib/core/annotation.d.ts.map +1 -0
  136. package/lib/core/annotation.js +92 -0
  137. package/lib/core/annotation.js.map +1 -0
  138. package/lib/core/answerer.d.ts +68 -0
  139. package/lib/core/answerer.d.ts.map +1 -0
  140. package/lib/core/answerer.js +17 -0
  141. package/lib/core/answerer.js.map +1 -0
  142. package/lib/core/artifact-evidence.d.ts +50 -0
  143. package/lib/core/artifact-evidence.d.ts.map +1 -0
  144. package/lib/core/artifact-evidence.js +185 -0
  145. package/lib/core/artifact-evidence.js.map +1 -0
  146. package/lib/core/artifact-snapshot.d.ts +37 -0
  147. package/lib/core/artifact-snapshot.d.ts.map +1 -0
  148. package/lib/core/artifact-snapshot.js +339 -0
  149. package/lib/core/artifact-snapshot.js.map +1 -0
  150. package/lib/core/assertion.d.ts +129 -0
  151. package/lib/core/assertion.d.ts.map +1 -0
  152. package/lib/core/assertion.js +75 -0
  153. package/lib/core/assertion.js.map +1 -0
  154. package/lib/core/case.d.ts +79 -0
  155. package/lib/core/case.d.ts.map +1 -0
  156. package/lib/core/case.js +90 -0
  157. package/lib/core/case.js.map +1 -0
  158. package/lib/core/conversation-turn.d.ts +23 -0
  159. package/lib/core/conversation-turn.d.ts.map +1 -0
  160. package/lib/core/conversation-turn.js +31 -0
  161. package/lib/core/conversation-turn.js.map +1 -0
  162. package/lib/core/driver.d.ts +64 -0
  163. package/lib/core/driver.d.ts.map +1 -0
  164. package/lib/core/driver.js +18 -0
  165. package/lib/core/driver.js.map +1 -0
  166. package/lib/core/emitter.d.ts +20 -0
  167. package/lib/core/emitter.d.ts.map +1 -0
  168. package/lib/core/emitter.js +34 -0
  169. package/lib/core/emitter.js.map +1 -0
  170. package/lib/core/events.d.ts +202 -0
  171. package/lib/core/events.d.ts.map +1 -0
  172. package/lib/core/events.js +107 -0
  173. package/lib/core/events.js.map +1 -0
  174. package/lib/core/exec.d.ts +82 -0
  175. package/lib/core/exec.d.ts.map +1 -0
  176. package/lib/core/exec.js +168 -0
  177. package/lib/core/exec.js.map +1 -0
  178. package/lib/core/experiment.d.ts +285 -0
  179. package/lib/core/experiment.d.ts.map +1 -0
  180. package/lib/core/experiment.js +389 -0
  181. package/lib/core/experiment.js.map +1 -0
  182. package/lib/core/glob.d.ts +3 -0
  183. package/lib/core/glob.d.ts.map +1 -0
  184. package/lib/core/glob.js +32 -0
  185. package/lib/core/glob.js.map +1 -0
  186. package/lib/core/hash.d.ts +8 -0
  187. package/lib/core/hash.d.ts.map +1 -0
  188. package/lib/core/hash.js +214 -0
  189. package/lib/core/hash.js.map +1 -0
  190. package/lib/core/index.d.ts +29 -0
  191. package/lib/core/index.d.ts.map +1 -0
  192. package/lib/core/index.js +29 -0
  193. package/lib/core/index.js.map +1 -0
  194. package/lib/core/lifecycle-fixtures.d.ts +20 -0
  195. package/lib/core/lifecycle-fixtures.d.ts.map +1 -0
  196. package/lib/core/lifecycle-fixtures.js +109 -0
  197. package/lib/core/lifecycle-fixtures.js.map +1 -0
  198. package/lib/core/lifecycle-hooks.d.ts +52 -0
  199. package/lib/core/lifecycle-hooks.d.ts.map +1 -0
  200. package/lib/core/lifecycle-hooks.js +42 -0
  201. package/lib/core/lifecycle-hooks.js.map +1 -0
  202. package/lib/core/mcp-events.d.ts +44 -0
  203. package/lib/core/mcp-events.d.ts.map +1 -0
  204. package/lib/core/mcp-events.js +23 -0
  205. package/lib/core/mcp-events.js.map +1 -0
  206. package/lib/core/model-cli.d.ts +57 -0
  207. package/lib/core/model-cli.d.ts.map +1 -0
  208. package/lib/core/model-cli.js +237 -0
  209. package/lib/core/model-cli.js.map +1 -0
  210. package/lib/core/observer.d.ts +31 -0
  211. package/lib/core/observer.d.ts.map +1 -0
  212. package/lib/core/observer.js +25 -0
  213. package/lib/core/observer.js.map +1 -0
  214. package/lib/core/platform-process.d.ts +33 -0
  215. package/lib/core/platform-process.d.ts.map +1 -0
  216. package/lib/core/platform-process.js +214 -0
  217. package/lib/core/platform-process.js.map +1 -0
  218. package/lib/core/registry.d.ts +22 -0
  219. package/lib/core/registry.d.ts.map +1 -0
  220. package/lib/core/registry.js +45 -0
  221. package/lib/core/registry.js.map +1 -0
  222. package/lib/core/rubric.d.ts +18 -0
  223. package/lib/core/rubric.d.ts.map +1 -0
  224. package/lib/core/rubric.js +49 -0
  225. package/lib/core/rubric.js.map +1 -0
  226. package/lib/core/run.d.ts +419 -0
  227. package/lib/core/run.d.ts.map +1 -0
  228. package/lib/core/run.js +126 -0
  229. package/lib/core/run.js.map +1 -0
  230. package/lib/core/statistics.d.ts +39 -0
  231. package/lib/core/statistics.d.ts.map +1 -0
  232. package/lib/core/statistics.js +163 -0
  233. package/lib/core/statistics.js.map +1 -0
  234. package/lib/core/subject.d.ts +4 -0
  235. package/lib/core/subject.d.ts.map +1 -0
  236. package/lib/core/subject.js +46 -0
  237. package/lib/core/subject.js.map +1 -0
  238. package/lib/core/tool-events.d.ts +69 -0
  239. package/lib/core/tool-events.d.ts.map +1 -0
  240. package/lib/core/tool-events.js +242 -0
  241. package/lib/core/tool-events.js.map +1 -0
  242. package/lib/core/usage.d.ts +78 -0
  243. package/lib/core/usage.d.ts.map +1 -0
  244. package/lib/core/usage.js +104 -0
  245. package/lib/core/usage.js.map +1 -0
  246. package/lib/core/wire.d.ts +83 -0
  247. package/lib/core/wire.d.ts.map +1 -0
  248. package/lib/core/wire.js +101 -0
  249. package/lib/core/wire.js.map +1 -0
  250. package/lib/core/zod.d.ts +10 -0
  251. package/lib/core/zod.d.ts.map +1 -0
  252. package/lib/core/zod.js +18 -0
  253. package/lib/core/zod.js.map +1 -0
  254. package/lib/drivers/claude-code/index.d.ts +55 -0
  255. package/lib/drivers/claude-code/index.d.ts.map +1 -0
  256. package/lib/drivers/claude-code/index.js +779 -0
  257. package/lib/drivers/claude-code/index.js.map +1 -0
  258. package/lib/drivers/claude-code/parse-events.d.ts +46 -0
  259. package/lib/drivers/claude-code/parse-events.d.ts.map +1 -0
  260. package/lib/drivers/claude-code/parse-events.js +387 -0
  261. package/lib/drivers/claude-code/parse-events.js.map +1 -0
  262. package/lib/drivers/codex/app-events.d.ts +16 -0
  263. package/lib/drivers/codex/app-events.d.ts.map +1 -0
  264. package/lib/drivers/codex/app-events.js +184 -0
  265. package/lib/drivers/codex/app-events.js.map +1 -0
  266. package/lib/drivers/codex/config.d.ts +55 -0
  267. package/lib/drivers/codex/config.d.ts.map +1 -0
  268. package/lib/drivers/codex/config.js +125 -0
  269. package/lib/drivers/codex/config.js.map +1 -0
  270. package/lib/drivers/codex/index.d.ts +8 -0
  271. package/lib/drivers/codex/index.d.ts.map +1 -0
  272. package/lib/drivers/codex/index.js +904 -0
  273. package/lib/drivers/codex/index.js.map +1 -0
  274. package/lib/drivers/mcp/index.d.ts +52 -0
  275. package/lib/drivers/mcp/index.d.ts.map +1 -0
  276. package/lib/drivers/mcp/index.js +282 -0
  277. package/lib/drivers/mcp/index.js.map +1 -0
  278. package/lib/drivers/mcp/transport.d.ts +50 -0
  279. package/lib/drivers/mcp/transport.d.ts.map +1 -0
  280. package/lib/drivers/mcp/transport.js +190 -0
  281. package/lib/drivers/mcp/transport.js.map +1 -0
  282. package/lib/drivers/shared/artifacts.d.ts +4 -0
  283. package/lib/drivers/shared/artifacts.d.ts.map +1 -0
  284. package/lib/drivers/shared/artifacts.js +81 -0
  285. package/lib/drivers/shared/artifacts.js.map +1 -0
  286. package/lib/drivers/shared/secure-copy.d.ts +2 -0
  287. package/lib/drivers/shared/secure-copy.d.ts.map +1 -0
  288. package/lib/drivers/shared/secure-copy.js +35 -0
  289. package/lib/drivers/shared/secure-copy.js.map +1 -0
  290. package/lib/drivers/subprocess/index.d.ts +24 -0
  291. package/lib/drivers/subprocess/index.d.ts.map +1 -0
  292. package/lib/drivers/subprocess/index.js +451 -0
  293. package/lib/drivers/subprocess/index.js.map +1 -0
  294. package/lib/index.d.ts +40 -0
  295. package/lib/index.d.ts.map +1 -0
  296. package/lib/index.js +48 -0
  297. package/lib/index.js.map +1 -0
  298. package/lib/optimization/reflective.d.ts +27 -0
  299. package/lib/optimization/reflective.d.ts.map +1 -0
  300. package/lib/optimization/reflective.js +257 -0
  301. package/lib/optimization/reflective.js.map +1 -0
  302. package/lib/orchestrator/annotation-loader.d.ts +65 -0
  303. package/lib/orchestrator/annotation-loader.d.ts.map +1 -0
  304. package/lib/orchestrator/annotation-loader.js +73 -0
  305. package/lib/orchestrator/annotation-loader.js.map +1 -0
  306. package/lib/orchestrator/annotation-store.d.ts +3 -0
  307. package/lib/orchestrator/annotation-store.d.ts.map +1 -0
  308. package/lib/orchestrator/annotation-store.js +100 -0
  309. package/lib/orchestrator/annotation-store.js.map +1 -0
  310. package/lib/orchestrator/authored-dependencies.d.ts +3 -0
  311. package/lib/orchestrator/authored-dependencies.d.ts.map +1 -0
  312. package/lib/orchestrator/authored-dependencies.js +146 -0
  313. package/lib/orchestrator/authored-dependencies.js.map +1 -0
  314. package/lib/orchestrator/campaign-lifecycle.d.ts +105 -0
  315. package/lib/orchestrator/campaign-lifecycle.d.ts.map +1 -0
  316. package/lib/orchestrator/campaign-lifecycle.js +141 -0
  317. package/lib/orchestrator/campaign-lifecycle.js.map +1 -0
  318. package/lib/orchestrator/candidate-integrity.d.ts +4 -0
  319. package/lib/orchestrator/candidate-integrity.d.ts.map +1 -0
  320. package/lib/orchestrator/candidate-integrity.js +14 -0
  321. package/lib/orchestrator/candidate-integrity.js.map +1 -0
  322. package/lib/orchestrator/candidates.d.ts +13 -0
  323. package/lib/orchestrator/candidates.d.ts.map +1 -0
  324. package/lib/orchestrator/candidates.js +577 -0
  325. package/lib/orchestrator/candidates.js.map +1 -0
  326. package/lib/orchestrator/compare.d.ts +80 -0
  327. package/lib/orchestrator/compare.d.ts.map +1 -0
  328. package/lib/orchestrator/compare.js +270 -0
  329. package/lib/orchestrator/compare.js.map +1 -0
  330. package/lib/orchestrator/comparison-identity.d.ts +10 -0
  331. package/lib/orchestrator/comparison-identity.d.ts.map +1 -0
  332. package/lib/orchestrator/comparison-identity.js +39 -0
  333. package/lib/orchestrator/comparison-identity.js.map +1 -0
  334. package/lib/orchestrator/config.d.ts +121 -0
  335. package/lib/orchestrator/config.d.ts.map +1 -0
  336. package/lib/orchestrator/config.js +176 -0
  337. package/lib/orchestrator/config.js.map +1 -0
  338. package/lib/orchestrator/content-bound-json.d.ts +5 -0
  339. package/lib/orchestrator/content-bound-json.d.ts.map +1 -0
  340. package/lib/orchestrator/content-bound-json.js +59 -0
  341. package/lib/orchestrator/content-bound-json.js.map +1 -0
  342. package/lib/orchestrator/discover.d.ts +45 -0
  343. package/lib/orchestrator/discover.d.ts.map +1 -0
  344. package/lib/orchestrator/discover.js +98 -0
  345. package/lib/orchestrator/discover.js.map +1 -0
  346. package/lib/orchestrator/doctor.d.ts +141 -0
  347. package/lib/orchestrator/doctor.d.ts.map +1 -0
  348. package/lib/orchestrator/doctor.js +525 -0
  349. package/lib/orchestrator/doctor.js.map +1 -0
  350. package/lib/orchestrator/driver-session.d.ts +177 -0
  351. package/lib/orchestrator/driver-session.d.ts.map +1 -0
  352. package/lib/orchestrator/driver-session.js +396 -0
  353. package/lib/orchestrator/driver-session.js.map +1 -0
  354. package/lib/orchestrator/evaluator-identity.d.ts +7 -0
  355. package/lib/orchestrator/evaluator-identity.d.ts.map +1 -0
  356. package/lib/orchestrator/evaluator-identity.js +42 -0
  357. package/lib/orchestrator/evaluator-identity.js.map +1 -0
  358. package/lib/orchestrator/experiment-budget.d.ts +11 -0
  359. package/lib/orchestrator/experiment-budget.d.ts.map +1 -0
  360. package/lib/orchestrator/experiment-budget.js +51 -0
  361. package/lib/orchestrator/experiment-budget.js.map +1 -0
  362. package/lib/orchestrator/experiment-context.d.ts +43 -0
  363. package/lib/orchestrator/experiment-context.d.ts.map +1 -0
  364. package/lib/orchestrator/experiment-context.js +9 -0
  365. package/lib/orchestrator/experiment-context.js.map +1 -0
  366. package/lib/orchestrator/experiment-contracts.d.ts +290 -0
  367. package/lib/orchestrator/experiment-contracts.d.ts.map +1 -0
  368. package/lib/orchestrator/experiment-contracts.js +38 -0
  369. package/lib/orchestrator/experiment-contracts.js.map +1 -0
  370. package/lib/orchestrator/experiment-corpus.d.ts +21 -0
  371. package/lib/orchestrator/experiment-corpus.d.ts.map +1 -0
  372. package/lib/orchestrator/experiment-corpus.js +276 -0
  373. package/lib/orchestrator/experiment-corpus.js.map +1 -0
  374. package/lib/orchestrator/experiment-evaluation.d.ts +44 -0
  375. package/lib/orchestrator/experiment-evaluation.d.ts.map +1 -0
  376. package/lib/orchestrator/experiment-evaluation.js +542 -0
  377. package/lib/orchestrator/experiment-evaluation.js.map +1 -0
  378. package/lib/orchestrator/experiment-finish.d.ts +6 -0
  379. package/lib/orchestrator/experiment-finish.d.ts.map +1 -0
  380. package/lib/orchestrator/experiment-finish.js +19 -0
  381. package/lib/orchestrator/experiment-finish.js.map +1 -0
  382. package/lib/orchestrator/experiment-identity.d.ts +6 -0
  383. package/lib/orchestrator/experiment-identity.d.ts.map +1 -0
  384. package/lib/orchestrator/experiment-identity.js +257 -0
  385. package/lib/orchestrator/experiment-identity.js.map +1 -0
  386. package/lib/orchestrator/experiment-loader.d.ts +175 -0
  387. package/lib/orchestrator/experiment-loader.d.ts.map +1 -0
  388. package/lib/orchestrator/experiment-loader.js +77 -0
  389. package/lib/orchestrator/experiment-loader.js.map +1 -0
  390. package/lib/orchestrator/experiment-report.d.ts +450 -0
  391. package/lib/orchestrator/experiment-report.d.ts.map +1 -0
  392. package/lib/orchestrator/experiment-report.js +570 -0
  393. package/lib/orchestrator/experiment-report.js.map +1 -0
  394. package/lib/orchestrator/experiment-runs.d.ts +379 -0
  395. package/lib/orchestrator/experiment-runs.d.ts.map +1 -0
  396. package/lib/orchestrator/experiment-runs.js +796 -0
  397. package/lib/orchestrator/experiment-runs.js.map +1 -0
  398. package/lib/orchestrator/experiment-runtime-identity.d.ts +41 -0
  399. package/lib/orchestrator/experiment-runtime-identity.d.ts.map +1 -0
  400. package/lib/orchestrator/experiment-runtime-identity.js +411 -0
  401. package/lib/orchestrator/experiment-runtime-identity.js.map +1 -0
  402. package/lib/orchestrator/experiment-selection.d.ts +25 -0
  403. package/lib/orchestrator/experiment-selection.d.ts.map +1 -0
  404. package/lib/orchestrator/experiment-selection.js +286 -0
  405. package/lib/orchestrator/experiment-selection.js.map +1 -0
  406. package/lib/orchestrator/experiment.d.ts +11 -0
  407. package/lib/orchestrator/experiment.d.ts.map +1 -0
  408. package/lib/orchestrator/experiment.js +639 -0
  409. package/lib/orchestrator/experiment.js.map +1 -0
  410. package/lib/orchestrator/fixture-identity.d.ts +7 -0
  411. package/lib/orchestrator/fixture-identity.d.ts.map +1 -0
  412. package/lib/orchestrator/fixture-identity.js +55 -0
  413. package/lib/orchestrator/fixture-identity.js.map +1 -0
  414. package/lib/orchestrator/fs-read.d.ts +19 -0
  415. package/lib/orchestrator/fs-read.d.ts.map +1 -0
  416. package/lib/orchestrator/fs-read.js +113 -0
  417. package/lib/orchestrator/fs-read.js.map +1 -0
  418. package/lib/orchestrator/grader-experiment.d.ts +190 -0
  419. package/lib/orchestrator/grader-experiment.d.ts.map +1 -0
  420. package/lib/orchestrator/grader-experiment.js +579 -0
  421. package/lib/orchestrator/grader-experiment.js.map +1 -0
  422. package/lib/orchestrator/grading.d.ts +31 -0
  423. package/lib/orchestrator/grading.d.ts.map +1 -0
  424. package/lib/orchestrator/grading.js +186 -0
  425. package/lib/orchestrator/grading.js.map +1 -0
  426. package/lib/orchestrator/index.d.ts +22 -0
  427. package/lib/orchestrator/index.d.ts.map +1 -0
  428. package/lib/orchestrator/index.js +30 -0
  429. package/lib/orchestrator/index.js.map +1 -0
  430. package/lib/orchestrator/lifecycle-hook-execution.d.ts +4 -0
  431. package/lib/orchestrator/lifecycle-hook-execution.d.ts.map +1 -0
  432. package/lib/orchestrator/lifecycle-hook-execution.js +67 -0
  433. package/lib/orchestrator/lifecycle-hook-execution.js.map +1 -0
  434. package/lib/orchestrator/list-runs.d.ts +33 -0
  435. package/lib/orchestrator/list-runs.d.ts.map +1 -0
  436. package/lib/orchestrator/list-runs.js +308 -0
  437. package/lib/orchestrator/list-runs.js.map +1 -0
  438. package/lib/orchestrator/loader.d.ts +3 -0
  439. package/lib/orchestrator/loader.d.ts.map +1 -0
  440. package/lib/orchestrator/loader.js +17 -0
  441. package/lib/orchestrator/loader.js.map +1 -0
  442. package/lib/orchestrator/optimization-lifecycle.d.ts +12 -0
  443. package/lib/orchestrator/optimization-lifecycle.d.ts.map +1 -0
  444. package/lib/orchestrator/optimization-lifecycle.js +92 -0
  445. package/lib/orchestrator/optimization-lifecycle.js.map +1 -0
  446. package/lib/orchestrator/orchestrator.d.ts +144 -0
  447. package/lib/orchestrator/orchestrator.d.ts.map +1 -0
  448. package/lib/orchestrator/orchestrator.js +632 -0
  449. package/lib/orchestrator/orchestrator.js.map +1 -0
  450. package/lib/orchestrator/owned-lock.d.ts +17 -0
  451. package/lib/orchestrator/owned-lock.d.ts.map +1 -0
  452. package/lib/orchestrator/owned-lock.js +149 -0
  453. package/lib/orchestrator/owned-lock.js.map +1 -0
  454. package/lib/orchestrator/package-version.d.ts +2 -0
  455. package/lib/orchestrator/package-version.d.ts.map +1 -0
  456. package/lib/orchestrator/package-version.js +4 -0
  457. package/lib/orchestrator/package-version.js.map +1 -0
  458. package/lib/orchestrator/provenance.d.ts +5 -0
  459. package/lib/orchestrator/provenance.d.ts.map +1 -0
  460. package/lib/orchestrator/provenance.js +137 -0
  461. package/lib/orchestrator/provenance.js.map +1 -0
  462. package/lib/orchestrator/regrade.d.ts +80 -0
  463. package/lib/orchestrator/regrade.d.ts.map +1 -0
  464. package/lib/orchestrator/regrade.js +433 -0
  465. package/lib/orchestrator/regrade.js.map +1 -0
  466. package/lib/orchestrator/rubric-loader.d.ts +7 -0
  467. package/lib/orchestrator/rubric-loader.d.ts.map +1 -0
  468. package/lib/orchestrator/rubric-loader.js +65 -0
  469. package/lib/orchestrator/rubric-loader.js.map +1 -0
  470. package/lib/orchestrator/run-dir.d.ts +40 -0
  471. package/lib/orchestrator/run-dir.d.ts.map +1 -0
  472. package/lib/orchestrator/run-dir.js +161 -0
  473. package/lib/orchestrator/run-dir.js.map +1 -0
  474. package/lib/orchestrator/run-execution.d.ts +30 -0
  475. package/lib/orchestrator/run-execution.d.ts.map +1 -0
  476. package/lib/orchestrator/run-execution.js +531 -0
  477. package/lib/orchestrator/run-execution.js.map +1 -0
  478. package/lib/orchestrator/run-lifecycle.d.ts +115 -0
  479. package/lib/orchestrator/run-lifecycle.d.ts.map +1 -0
  480. package/lib/orchestrator/run-lifecycle.js +136 -0
  481. package/lib/orchestrator/run-lifecycle.js.map +1 -0
  482. package/lib/orchestrator/suite-compare.d.ts +76 -0
  483. package/lib/orchestrator/suite-compare.d.ts.map +1 -0
  484. package/lib/orchestrator/suite-compare.js +159 -0
  485. package/lib/orchestrator/suite-compare.js.map +1 -0
  486. package/lib/orchestrator/suite-lifecycle.d.ts +157 -0
  487. package/lib/orchestrator/suite-lifecycle.d.ts.map +1 -0
  488. package/lib/orchestrator/suite-lifecycle.js +253 -0
  489. package/lib/orchestrator/suite-lifecycle.js.map +1 -0
  490. package/lib/orchestrator/suite-runs.d.ts +91 -0
  491. package/lib/orchestrator/suite-runs.d.ts.map +1 -0
  492. package/lib/orchestrator/suite-runs.js +157 -0
  493. package/lib/orchestrator/suite-runs.js.map +1 -0
  494. package/lib/orchestrator/suite-summary.d.ts +109 -0
  495. package/lib/orchestrator/suite-summary.d.ts.map +1 -0
  496. package/lib/orchestrator/suite-summary.js +213 -0
  497. package/lib/orchestrator/suite-summary.js.map +1 -0
  498. package/lib/orchestrator/trace-limits.d.ts +20 -0
  499. package/lib/orchestrator/trace-limits.d.ts.map +1 -0
  500. package/lib/orchestrator/trace-limits.js +28 -0
  501. package/lib/orchestrator/trace-limits.js.map +1 -0
  502. package/lib/orchestrator/validate-plugin-config.d.ts +10 -0
  503. package/lib/orchestrator/validate-plugin-config.d.ts.map +1 -0
  504. package/lib/orchestrator/validate-plugin-config.js +26 -0
  505. package/lib/orchestrator/validate-plugin-config.js.map +1 -0
  506. package/package.json +84 -0
  507. package/skills/lux/SKILL.md +85 -0
  508. package/skills/lux/references/cli.md +102 -0
  509. package/skills/lux/references/eval-authoring.md +105 -0
  510. package/skills/lux-answerer/SKILL.md +169 -0
  511. package/src/answerers/claude-code/index.ts +342 -0
  512. package/src/answerers/persona/index.ts +327 -0
  513. package/src/answerers/scripted/index.ts +186 -0
  514. package/src/answerers/shared.ts +23 -0
  515. package/src/answerers/terminal/index.ts +117 -0
  516. package/src/assertions/core/command-exits-zero.ts +94 -0
  517. package/src/assertions/core/event-checks.ts +138 -0
  518. package/src/assertions/core/file-checks.ts +336 -0
  519. package/src/assertions/core/helpers.ts +203 -0
  520. package/src/assertions/core/index.ts +80 -0
  521. package/src/assertions/core/json-pointer-equals.ts +84 -0
  522. package/src/assertions/core/output-checks.ts +134 -0
  523. package/src/assertions/core/predicate.ts +67 -0
  524. package/src/assertions/core/resource-checks.ts +180 -0
  525. package/src/assertions/core/safe-regex.ts +196 -0
  526. package/src/assertions/core/tool-called-with.ts +68 -0
  527. package/src/assertions/core/tool-calls.ts +28 -0
  528. package/src/assertions/core/tool-invocation.ts +198 -0
  529. package/src/assertions/core/tool-result-checks.ts +211 -0
  530. package/src/assertions/mcp/discovery-checks.ts +139 -0
  531. package/src/assertions/mcp/index.ts +37 -0
  532. package/src/assertions/rubric/index.ts +258 -0
  533. package/src/assertions/rubric/internal.ts +367 -0
  534. package/src/authoring.ts +2 -0
  535. package/src/cli/bin.ts +14 -0
  536. package/src/cli/command-runtime.ts +34 -0
  537. package/src/cli/init.ts +410 -0
  538. package/src/cli/program.ts +992 -0
  539. package/src/cli/render/format.ts +69 -0
  540. package/src/cli/render/reporter.ts +643 -0
  541. package/src/cli/render/tty.ts +314 -0
  542. package/src/cli/run-options.ts +82 -0
  543. package/src/cli/skills.ts +129 -0
  544. package/src/core/annotation.ts +104 -0
  545. package/src/core/answerer.ts +75 -0
  546. package/src/core/artifact-evidence.ts +249 -0
  547. package/src/core/artifact-snapshot.ts +429 -0
  548. package/src/core/assertion.ts +156 -0
  549. package/src/core/case.ts +103 -0
  550. package/src/core/conversation-turn.ts +47 -0
  551. package/src/core/driver.ts +75 -0
  552. package/src/core/emitter.ts +38 -0
  553. package/src/core/events.ts +135 -0
  554. package/src/core/exec.ts +263 -0
  555. package/src/core/experiment.ts +467 -0
  556. package/src/core/glob.ts +31 -0
  557. package/src/core/hash.ts +237 -0
  558. package/src/core/index.ts +28 -0
  559. package/src/core/lifecycle-fixtures.ts +145 -0
  560. package/src/core/lifecycle-hooks.ts +101 -0
  561. package/src/core/mcp-events.ts +57 -0
  562. package/src/core/model-cli.ts +271 -0
  563. package/src/core/observer.ts +50 -0
  564. package/src/core/platform-process.ts +284 -0
  565. package/src/core/registry.ts +57 -0
  566. package/src/core/rubric.ts +52 -0
  567. package/src/core/run.ts +144 -0
  568. package/src/core/statistics.ts +194 -0
  569. package/src/core/subject.ts +50 -0
  570. package/src/core/tool-events.ts +292 -0
  571. package/src/core/usage.ts +123 -0
  572. package/src/core/wire.ts +135 -0
  573. package/src/core/zod.ts +20 -0
  574. package/src/drivers/claude-code/index.ts +866 -0
  575. package/src/drivers/claude-code/parse-events.ts +464 -0
  576. package/src/drivers/codex/app-events.ts +218 -0
  577. package/src/drivers/codex/config.ts +128 -0
  578. package/src/drivers/codex/index.ts +995 -0
  579. package/src/drivers/mcp/index.ts +351 -0
  580. package/src/drivers/mcp/transport.ts +222 -0
  581. package/src/drivers/shared/artifacts.ts +86 -0
  582. package/src/drivers/shared/secure-copy.ts +40 -0
  583. package/src/drivers/subprocess/index.ts +490 -0
  584. package/src/index.ts +594 -0
  585. package/src/optimization/reflective.ts +319 -0
  586. package/src/orchestrator/annotation-loader.ts +97 -0
  587. package/src/orchestrator/annotation-store.ts +112 -0
  588. package/src/orchestrator/authored-dependencies.ts +193 -0
  589. package/src/orchestrator/campaign-lifecycle.ts +181 -0
  590. package/src/orchestrator/candidate-integrity.ts +18 -0
  591. package/src/orchestrator/candidates.ts +683 -0
  592. package/src/orchestrator/compare.ts +358 -0
  593. package/src/orchestrator/comparison-identity.ts +45 -0
  594. package/src/orchestrator/config.ts +205 -0
  595. package/src/orchestrator/content-bound-json.ts +71 -0
  596. package/src/orchestrator/discover.ts +132 -0
  597. package/src/orchestrator/doctor.ts +781 -0
  598. package/src/orchestrator/driver-session.ts +507 -0
  599. package/src/orchestrator/evaluator-identity.ts +45 -0
  600. package/src/orchestrator/experiment-budget.ts +61 -0
  601. package/src/orchestrator/experiment-context.ts +41 -0
  602. package/src/orchestrator/experiment-contracts.ts +43 -0
  603. package/src/orchestrator/experiment-corpus.ts +416 -0
  604. package/src/orchestrator/experiment-evaluation.ts +749 -0
  605. package/src/orchestrator/experiment-finish.ts +33 -0
  606. package/src/orchestrator/experiment-identity.ts +356 -0
  607. package/src/orchestrator/experiment-loader.ts +105 -0
  608. package/src/orchestrator/experiment-report.ts +788 -0
  609. package/src/orchestrator/experiment-runs.ts +960 -0
  610. package/src/orchestrator/experiment-runtime-identity.ts +465 -0
  611. package/src/orchestrator/experiment-selection.ts +467 -0
  612. package/src/orchestrator/experiment.ts +991 -0
  613. package/src/orchestrator/fixture-identity.ts +67 -0
  614. package/src/orchestrator/fs-read.ts +126 -0
  615. package/src/orchestrator/grader-experiment.ts +856 -0
  616. package/src/orchestrator/grading.ts +260 -0
  617. package/src/orchestrator/index.ts +186 -0
  618. package/src/orchestrator/lifecycle-hook-execution.ts +90 -0
  619. package/src/orchestrator/list-runs.ts +384 -0
  620. package/src/orchestrator/loader.ts +19 -0
  621. package/src/orchestrator/optimization-lifecycle.ts +126 -0
  622. package/src/orchestrator/orchestrator.ts +919 -0
  623. package/src/orchestrator/owned-lock.ts +147 -0
  624. package/src/orchestrator/package-version.ts +5 -0
  625. package/src/orchestrator/provenance.ts +140 -0
  626. package/src/orchestrator/regrade.ts +580 -0
  627. package/src/orchestrator/rubric-loader.ts +74 -0
  628. package/src/orchestrator/run-dir.ts +209 -0
  629. package/src/orchestrator/run-execution.ts +694 -0
  630. package/src/orchestrator/run-lifecycle.ts +174 -0
  631. package/src/orchestrator/suite-compare.ts +195 -0
  632. package/src/orchestrator/suite-lifecycle.ts +335 -0
  633. package/src/orchestrator/suite-runs.ts +202 -0
  634. package/src/orchestrator/suite-summary.ts +276 -0
  635. package/src/orchestrator/trace-limits.ts +65 -0
  636. package/src/orchestrator/validate-plugin-config.ts +34 -0
@@ -0,0 +1,276 @@
1
+ import { readFile } from "node:fs/promises";
2
+ import { dirname, join } from "node:path";
3
+ import { RubricConfigSchema } from "../assertions/rubric/index.js";
4
+ import { bindSubjectRoot, contentHash, valueHash, } from "../core/index.js";
5
+ import { assertDisjointAnnotationSplits, selectGraderAnnotations } from "./annotation-loader.js";
6
+ import { discoverCases, selectCases } from "./discover.js";
7
+ import { profileOrThrow } from "./experiment-context.js";
8
+ import { claudeConfigRoot, claudeUserBehaviorHash, cliIdentity, configuredClaudeSourceHashes, referencedAnswererCli, referencedCli, } from "./experiment-runtime-identity.js";
9
+ import { fixtureIdentity } from "./fixture-identity.js";
10
+ import { loadEvalCase } from "./loader.js";
11
+ import { resolveAnswererReference, resolveDriverReference } from "./orchestrator.js";
12
+ import { loadFrozenRegradableRun, releaseFrozenRegradableRun } from "./regrade.js";
13
+ import { loadRubricDefinition } from "./rubric-loader.js";
14
+ export const selectorsBySplit = (campaign) => [
15
+ ...(campaign.splits.train ? [{ name: "train", selector: campaign.splits.train }] : []),
16
+ { name: "validation", selector: campaign.splits.validation },
17
+ ...(campaign.splits.test ? [{ name: "test", selector: campaign.splits.test }] : []),
18
+ ];
19
+ const fixtureEvidenceHash = async (execution, casePath, fixtures) => {
20
+ if (!fixtures)
21
+ return null;
22
+ return (await fixtureIdentity(fixtures, execution.config.fixturesRoot ?? dirname(casePath))).hash;
23
+ };
24
+ const agentCaseEvidenceHash = async (execution, casePath) => {
25
+ const authored = await loadEvalCase(casePath);
26
+ const { id: _caseId, meta: _meta, fixtures, assertions = [], ...behavior } = authored;
27
+ const normalizedAssertions = assertions.map(({ id: _assertionId, name: _assertionName, ...assertion }) => assertion);
28
+ return valueHash({
29
+ case: { ...behavior, assertions: normalizedAssertions },
30
+ fixtures: await fixtureEvidenceHash(execution, casePath, fixtures),
31
+ });
32
+ };
33
+ export const caseRubricCliReferences = (execution, authored) => {
34
+ const rubrics = (authored.assertions ?? []).flatMap((assertion, index) => assertion.type === "rubric" ? [{ index, config: RubricConfigSchema.parse(assertion) }] : []);
35
+ return execution.loaded.campaign.profiles.flatMap((profileId) => {
36
+ const profile = profileOrThrow(execution.config, profileId);
37
+ const harness = profile.harness ?? execution.config.harness;
38
+ return rubrics.map(({ index, config }) => {
39
+ const provider = config.provider ?? harness?.provider ?? "claude-code";
40
+ const inheritsHarness = config.provider === undefined || config.provider === harness?.provider;
41
+ return {
42
+ assertionId: config.id ?? `assertion-${index}`,
43
+ profileId,
44
+ provider,
45
+ command: config.command ??
46
+ (inheritsHarness ? harness?.command : undefined) ??
47
+ (provider === "codex" ? "codex" : "claude"),
48
+ };
49
+ });
50
+ });
51
+ };
52
+ const caseRubricCliIdentities = (execution, authored) => Promise.all(caseRubricCliReferences(execution, authored).map(async ({ command, ...reference }) => ({
53
+ ...reference,
54
+ cliIdentity: await cliIdentity(command, execution.config.rootDir),
55
+ })));
56
+ export const agentCorpusHash = async (execution, cases, splits) => {
57
+ const selected = [
58
+ ...new Map(splits
59
+ .flatMap((split) => selectCases(cases, split.selector))
60
+ .map((evalCase) => [evalCase.path, evalCase])).values(),
61
+ ];
62
+ const caseBehaviorHashes = new Map();
63
+ const configuredSourceHashes = new Map();
64
+ const definitions = await Promise.all(selected.map(async (evalCase) => {
65
+ const authored = await loadEvalCase(evalCase.path);
66
+ const bound = bindSubjectRoot(authored, execution.loaded.subjectRoot);
67
+ const profileCliIdentities = await Promise.all(execution.loaded.campaign.profiles.map(async (profileId) => {
68
+ const profile = profileOrThrow(execution.config, profileId);
69
+ const harness = profile.harness ?? execution.config.harness;
70
+ const driver = resolveDriverReference(bound, profile.driver ?? execution.config.defaultDriver);
71
+ const answerer = resolveAnswererReference(bound, profile.answerer ?? execution.config.defaultAnswerer, harness);
72
+ const driverCli = referencedCli(driver);
73
+ const answererCli = referencedAnswererCli(answerer);
74
+ const configuredRoot = driverCli?.name === "claude-code" ? claudeConfigRoot(driver) : null;
75
+ let driverClaudeBehaviorHash = null;
76
+ if (configuredRoot) {
77
+ let hash = caseBehaviorHashes.get(configuredRoot);
78
+ if (!hash) {
79
+ hash = claudeUserBehaviorHash(configuredRoot);
80
+ caseBehaviorHashes.set(configuredRoot, hash);
81
+ }
82
+ driverClaudeBehaviorHash = await hash;
83
+ }
84
+ return {
85
+ profileId,
86
+ driver,
87
+ driverCliIdentity: driverCli
88
+ ? await cliIdentity(driverCli.command, execution.config.rootDir)
89
+ : null,
90
+ driverClaudeBehaviorHash,
91
+ configuredDriverSources: await configuredClaudeSourceHashes(driver, execution.config.rootDir, configuredSourceHashes),
92
+ answerer,
93
+ answererCliIdentity: answererCli
94
+ ? await cliIdentity(answererCli.command, execution.config.rootDir)
95
+ : null,
96
+ };
97
+ }));
98
+ return {
99
+ path: evalCase.path,
100
+ sourceHash: contentHash(await readFile(evalCase.path)),
101
+ case: authored,
102
+ fixturesHash: await fixtureEvidenceHash(execution, evalCase.path, authored.fixtures),
103
+ profileCliIdentities,
104
+ caseRubricCliIdentities: await caseRubricCliIdentities(execution, bound),
105
+ };
106
+ }));
107
+ return valueHash(definitions);
108
+ };
109
+ export const graderCorpusHash = async (execution, splits) => {
110
+ const loaded = execution.loaded.graderAnnotations;
111
+ if (!loaded)
112
+ throw new Error("grader experiment is missing its annotation dataset");
113
+ const selected = [
114
+ ...new Map(splits
115
+ .flatMap((split) => selectGraderAnnotations(loaded.annotations, split.selector))
116
+ .map((annotation) => [annotation.id, annotation])).values(),
117
+ ];
118
+ const evidence = await Promise.all(selected.map(async (annotation) => {
119
+ const run = await loadFrozenRegradableRun(annotation.runDir);
120
+ try {
121
+ const rubric = await loadRubricDefinition(join(execution.loaded.subjectRoot, annotation.rubricPath));
122
+ if (rubric.id !== run.case.id) {
123
+ throw new Error(`rubric '${annotation.rubricPath}' id '${rubric.id}' does not match persisted case '${run.case.id}'`);
124
+ }
125
+ const artifacts = await Promise.all(run.artifacts.map(async (artifact) => {
126
+ const root = artifact.root ?? join(annotation.runDir, "artifacts");
127
+ try {
128
+ return {
129
+ ...artifact,
130
+ root: undefined,
131
+ contentHash: contentHash(await readFile(join(root, artifact.path))),
132
+ };
133
+ }
134
+ catch (error) {
135
+ if (error.code === "ENOENT") {
136
+ return { ...artifact, root: undefined, contentHash: null };
137
+ }
138
+ throw error;
139
+ }
140
+ }));
141
+ return {
142
+ annotationId: annotation.id,
143
+ runDir: annotation.runDir,
144
+ evidenceHash: valueHash({
145
+ caseId: run.case.id,
146
+ prompt: run.case.prompt,
147
+ events: run.events.flatMap((event) => {
148
+ if (event.type === "lifecycle" || event.type === "ready")
149
+ return [];
150
+ const { ts: _timestamp, ...visible } = event;
151
+ return [visible];
152
+ }),
153
+ artifacts,
154
+ }),
155
+ rubric,
156
+ };
157
+ }
158
+ finally {
159
+ await releaseFrozenRegradableRun(run);
160
+ }
161
+ }));
162
+ const splitByAnnotation = new Map(splits.flatMap((split) => selectGraderAnnotations(loaded.annotations, split.selector).map((annotation) => [annotation.id, split.name])));
163
+ const ownerByEvidence = new Map();
164
+ for (const item of evidence) {
165
+ const split = splitByAnnotation.get(item.annotationId);
166
+ if (!split)
167
+ continue;
168
+ const owner = ownerByEvidence.get(item.evidenceHash);
169
+ if (owner) {
170
+ const scope = owner.split === split ? `duplicate evidence within ${split}` : "split leakage";
171
+ throw new Error(`grader ${scope}: annotations '${owner.annotationId}' (${owner.split}) and '${item.annotationId}' (${split}) contain identical frozen evidence`);
172
+ }
173
+ ownerByEvidence.set(item.evidenceHash, { annotationId: item.annotationId, split });
174
+ }
175
+ return valueHash({ annotationsHash: loaded.hash, evidence });
176
+ };
177
+ export const prepareCorpus = async (execution) => {
178
+ const campaign = execution.loaded.campaign;
179
+ const splits = selectorsBySplit(campaign);
180
+ if (campaign.evaluation.kind === "grader") {
181
+ const loaded = execution.loaded.graderAnnotations;
182
+ if (!loaded)
183
+ throw new Error("grader experiment is missing its annotation dataset");
184
+ const selectedSplits = splits.map((split) => ({
185
+ name: split.name,
186
+ annotations: selectGraderAnnotations(loaded.annotations, split.selector),
187
+ }));
188
+ for (const split of selectedSplits) {
189
+ if (split.annotations.length === 0) {
190
+ throw new Error(`experiment '${campaign.id}' ${split.name} selector matched no grader annotations`);
191
+ }
192
+ }
193
+ assertDisjointAnnotationSplits(selectedSplits);
194
+ return { kind: "grader", hash: await graderCorpusHash(execution, splits) };
195
+ }
196
+ const discovery = await discoverCases(execution.config.casesRoot ?? join(execution.config.rootDir, "cases"));
197
+ if (discovery.problems.length > 0) {
198
+ throw new Error(`experiment case discovery failed: ${discovery.problems.map((problem) => `${problem.path}: ${problem.reason}`).join("; ")}`);
199
+ }
200
+ const owners = new Map();
201
+ const caseIds = new Map();
202
+ const evidenceOwners = new Map();
203
+ for (const split of splits) {
204
+ const selected = selectCases(discovery.cases, split.selector);
205
+ if (selected.length === 0) {
206
+ throw new Error(`experiment '${campaign.id}' ${split.name} selector matched no cases`);
207
+ }
208
+ for (const evalCase of selected) {
209
+ const existingPath = caseIds.get(evalCase.id);
210
+ if (existingPath && existingPath !== evalCase.path) {
211
+ throw new Error(`experiment case id '${evalCase.id}' is duplicated by ${existingPath} and ${evalCase.path}`);
212
+ }
213
+ caseIds.set(evalCase.id, evalCase.path);
214
+ const owner = owners.get(evalCase.path);
215
+ if (owner) {
216
+ throw new Error(`eval case '${evalCase.id}' appears in both ${owner} and ${split.name} splits`);
217
+ }
218
+ owners.set(evalCase.path, split.name);
219
+ const evidenceHash = await agentCaseEvidenceHash(execution, evalCase.path);
220
+ const evidenceOwner = evidenceOwners.get(evidenceHash);
221
+ if (evidenceOwner) {
222
+ const scope = evidenceOwner.split === split.name
223
+ ? `duplicate evidence within ${split.name}`
224
+ : "split leakage";
225
+ throw new Error(`agent ${scope}: cases '${evidenceOwner.caseId}' (${evidenceOwner.split}) and '${evalCase.id}' (${split.name}) contain identical behavioral evidence`);
226
+ }
227
+ evidenceOwners.set(evidenceHash, {
228
+ caseId: evalCase.id,
229
+ path: evalCase.path,
230
+ split: split.name,
231
+ });
232
+ }
233
+ }
234
+ return {
235
+ kind: "agent",
236
+ cases: discovery.cases,
237
+ hash: await agentCorpusHash(execution, discovery.cases, splits),
238
+ };
239
+ };
240
+ export const assertGraderSubject = async (execution, baseline) => {
241
+ if (execution.loaded.campaign.evaluation.kind !== "grader")
242
+ return;
243
+ const loadedAnnotations = execution.loaded.graderAnnotations?.annotations ?? [];
244
+ const annotations = [
245
+ ...new Map(selectorsBySplit(execution.loaded.campaign)
246
+ .flatMap((split) => selectGraderAnnotations(loadedAnnotations, split.selector))
247
+ .map((annotation) => [annotation.id, annotation])).values(),
248
+ ];
249
+ const rubricPaths = [...new Set(annotations.map((annotation) => annotation.rubricPath))].sort();
250
+ const componentPaths = Object.keys(baseline.components).sort();
251
+ if (rubricPaths.length !== componentPaths.length ||
252
+ rubricPaths.some((path, index) => path !== componentPaths[index])) {
253
+ throw new Error("grader experiment mutable components must exactly match the annotation rubric paths");
254
+ }
255
+ for (const rubricPath of rubricPaths) {
256
+ const rubric = await loadRubricDefinition(join(execution.loaded.subjectRoot, rubricPath));
257
+ const baselineIds = new Set(rubric.assertions.map((assertion) => assertion.id));
258
+ const relevant = annotations.filter((annotation) => annotation.rubricPath === rubricPath);
259
+ const allowedAdditions = new Set(execution.loaded.campaign.evaluation.allowedAssertionAdditions[rubricPath] ?? []);
260
+ const allLabeledIds = new Set(relevant.flatMap((annotation) => Object.keys(annotation.expected.assertions)));
261
+ const undeclaredAdditions = [...allLabeledIds].filter((id) => !baselineIds.has(id) && !allowedAdditions.has(id));
262
+ if (undeclaredAdditions.length > 0) {
263
+ throw new Error(`annotation assertion IDs absent from '${rubricPath}' must be declared in evaluation.allowedAssertionAdditions: ${undeclaredAdditions.join(", ")}`);
264
+ }
265
+ const incompleteAdditions = [...allowedAdditions].filter((id) => baselineIds.has(id) ||
266
+ !relevant.every((annotation) => Object.hasOwn(annotation.expected.assertions, id)));
267
+ if (incompleteAdditions.length > 0) {
268
+ throw new Error(`allowed assertion additions for '${rubricPath}' must be absent from baseline and labeled on every frozen run: ${incompleteAdditions.join(", ")}`);
269
+ }
270
+ }
271
+ const unknownRubricPaths = Object.keys(execution.loaded.campaign.evaluation.allowedAssertionAdditions).filter((path) => !rubricPaths.includes(path));
272
+ if (unknownRubricPaths.length > 0) {
273
+ throw new Error(`allowed assertion additions reference rubric paths absent from annotations: ${unknownRubricPaths.join(", ")}`);
274
+ }
275
+ };
276
+ //# sourceMappingURL=experiment-corpus.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"experiment-corpus.js","sourceRoot":"","sources":["../../src/orchestrator/experiment-corpus.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,QAAQ,EAAE,MAAM,kBAAkB,CAAC;AAC5C,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAC1C,OAAO,EAAE,kBAAkB,EAAE,MAAM,+BAA+B,CAAC;AACnE,OAAO,EACL,eAAe,EAEf,WAAW,EAGX,SAAS,GACV,MAAM,kBAAkB,CAAC;AAC1B,OAAO,EAAE,8BAA8B,EAAE,uBAAuB,EAAE,MAAM,wBAAwB,CAAC;AACjG,OAAO,EAAE,aAAa,EAAE,WAAW,EAAE,MAAM,eAAe,CAAC;AAE3D,OAAO,EAAE,cAAc,EAAE,MAAM,yBAAyB,CAAC;AACzD,OAAO,EACL,gBAAgB,EAChB,sBAAsB,EACtB,WAAW,EACX,4BAA4B,EAC5B,qBAAqB,EACrB,aAAa,GACd,MAAM,kCAAkC,CAAC;AAC1C,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AACxD,OAAO,EAAE,YAAY,EAAE,MAAM,aAAa,CAAC;AAC3C,OAAO,EAAE,wBAAwB,EAAE,sBAAsB,EAAE,MAAM,mBAAmB,CAAC;AACrF,OAAO,EAAE,uBAAuB,EAAE,0BAA0B,EAAE,MAAM,cAAc,CAAC;AACnF,OAAO,EAAE,oBAAoB,EAAE,MAAM,oBAAoB,CAAC;AAI1D,MAAM,CAAC,MAAM,gBAAgB,GAAG,CAC9B,QAA4B,EACoC,EAAE,CAAC;IACnE,GAAG,CAAC,QAAQ,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,OAAgB,EAAE,QAAQ,EAAE,QAAQ,CAAC,MAAM,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IAC/F,EAAE,IAAI,EAAE,YAAY,EAAE,QAAQ,EAAE,QAAQ,CAAC,MAAM,CAAC,UAAU,EAAE;IAC5D,GAAG,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,MAAe,EAAE,QAAQ,EAAE,QAAQ,CAAC,MAAM,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;CAC7F,CAAC;AAEF,MAAM,mBAAmB,GAAG,KAAK,EAC/B,SAAmC,EACnC,QAAgB,EAChB,QAA8D,EACtC,EAAE;IAC1B,IAAI,CAAC,QAAQ;QAAE,OAAO,IAAI,CAAC;IAC3B,OAAO,CAAC,MAAM,eAAe,CAAC,QAAQ,EAAE,SAAS,CAAC,MAAM,CAAC,YAAY,IAAI,OAAO,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC;AACpG,CAAC,CAAC;AAEF,MAAM,qBAAqB,GAAG,KAAK,EACjC,SAAmC,EACnC,QAAgB,EACC,EAAE;IACnB,MAAM,QAAQ,GAAG,MAAM,YAAY,CAAC,QAAQ,CAAC,CAAC;IAC9C,MAAM,EAAE,EAAE,EAAE,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,QAAQ,EAAE,UAAU,GAAG,EAAE,EAAE,GAAG,QAAQ,EAAE,GAAG,QAAQ,CAAC;IACtF,MAAM,oBAAoB,GAAG,UAAU,CAAC,GAAG,CACzC,CAAC,EAAE,EAAE,EAAE,YAAY,EAAE,IAAI,EAAE,cAAc,EAAE,GAAG,SAAS,EAAE,EAAE,EAAE,CAAC,SAAS,CACxE,CAAC;IACF,OAAO,SAAS,CAAC;QACf,IAAI,EAAE,EAAE,GAAG,QAAQ,EAAE,UAAU,EAAE,oBAAoB,EAAE;QACvD,QAAQ,EAAE,MAAM,mBAAmB,CAAC,SAAS,EAAE,QAAQ,EAAE,QAAQ,CAAC;KACnE,CAAC,CAAC;AACL,CAAC,CAAC;AAEF,MAAM,CAAC,MAAM,uBAAuB,GAAG,CACrC,SAAmC,EACnC,QAAkD,EAMjD,EAAE;IACH,MAAM,OAAO,GAAG,CAAC,QAAQ,CAAC,UAAU,IAAI,EAAE,CAAC,CAAC,OAAO,CAAC,CAAC,SAAS,EAAE,KAAK,EAAE,EAAE,CACvE,SAAS,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,kBAAkB,CAAC,KAAK,CAAC,SAAS,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAC5F,CAAC;IACF,OAAO,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC,SAAS,EAAE,EAAE;QAC9D,MAAM,OAAO,GAAG,cAAc,CAAC,SAAS,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;QAC5D,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,IAAI,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC;QAC5D,OAAO,OAAO,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,EAAE,EAAE;YACvC,MAAM,QAAQ,GAAG,MAAM,CAAC,QAAQ,IAAI,OAAO,EAAE,QAAQ,IAAI,aAAa,CAAC;YACvE,MAAM,eAAe,GACnB,MAAM,CAAC,QAAQ,KAAK,SAAS,IAAI,MAAM,CAAC,QAAQ,KAAK,OAAO,EAAE,QAAQ,CAAC;YACzE,OAAO;gBACL,WAAW,EAAE,MAAM,CAAC,EAAE,IAAI,aAAa,KAAK,EAAE;gBAC9C,SAAS;gBACT,QAAQ;gBACR,OAAO,EACL,MAAM,CAAC,OAAO;oBACd,CAAC,eAAe,CAAC,CAAC,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC,CAAC,SAAS,CAAC;oBAChD,CAAC,QAAQ,KAAK,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,QAAQ,CAAC;aAC9C,CAAC;QACJ,CAAC,CAAC,CAAC;IACL,CAAC,CAAC,CAAC;AACL,CAAC,CAAC;AAEF,MAAM,uBAAuB,GAAG,CAC9B,SAAmC,EACnC,QAAkD,EAQlD,EAAE,CACF,OAAO,CAAC,GAAG,CACT,uBAAuB,CAAC,SAAS,EAAE,QAAQ,CAAC,CAAC,GAAG,CAAC,KAAK,EAAE,EAAE,OAAO,EAAE,GAAG,SAAS,EAAE,EAAE,EAAE,CAAC,CAAC;IACrF,GAAG,SAAS;IACZ,WAAW,EAAE,MAAM,WAAW,CAAC,OAAO,EAAE,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC;CAClE,CAAC,CAAC,CACJ,CAAC;AAEJ,MAAM,CAAC,MAAM,eAAe,GAAG,KAAK,EAClC,SAAmC,EACnC,KAAyD,EACzD,MAA2C,EAC1B,EAAE;IACnB,MAAM,QAAQ,GAAG;QACf,GAAG,IAAI,GAAG,CACR,MAAM;aACH,OAAO,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,WAAW,CAAC,KAAK,EAAE,KAAK,CAAC,QAAQ,CAAC,CAAC;aACtD,GAAG,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC,QAAQ,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC,CAChD,CAAC,MAAM,EAAE;KACX,CAAC;IACF,MAAM,kBAAkB,GAAG,IAAI,GAAG,EAAkC,CAAC;IACrE,MAAM,sBAAsB,GAAG,IAAI,GAAG,EAA2B,CAAC;IAClE,MAAM,WAAW,GAAG,MAAM,OAAO,CAAC,GAAG,CACnC,QAAQ,CAAC,GAAG,CAAC,KAAK,EAAE,QAAQ,EAAE,EAAE;QAC9B,MAAM,QAAQ,GAAG,MAAM,YAAY,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;QACnD,MAAM,KAAK,GAAG,eAAe,CAAC,QAAQ,EAAE,SAAS,CAAC,MAAM,CAAC,WAAW,CAAC,CAAC;QACtE,MAAM,oBAAoB,GAAG,MAAM,OAAO,CAAC,GAAG,CAC5C,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,QAAQ,CAAC,GAAG,CAAC,KAAK,EAAE,SAAS,EAAE,EAAE;YACzD,MAAM,OAAO,GAAG,cAAc,CAAC,SAAS,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;YAC5D,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,IAAI,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC;YAC5D,MAAM,MAAM,GAAG,sBAAsB,CACnC,KAAK,EACL,OAAO,CAAC,MAAM,IAAI,SAAS,CAAC,MAAM,CAAC,aAAa,CACjD,CAAC;YACF,MAAM,QAAQ,GAAG,wBAAwB,CACvC,KAAK,EACL,OAAO,CAAC,QAAQ,IAAI,SAAS,CAAC,MAAM,CAAC,eAAe,EACpD,OAAO,CACR,CAAC;YACF,MAAM,SAAS,GAAG,aAAa,CAAC,MAAM,CAAC,CAAC;YACxC,MAAM,WAAW,GAAG,qBAAqB,CAAC,QAAQ,CAAC,CAAC;YACpD,MAAM,cAAc,GAClB,SAAS,EAAE,IAAI,KAAK,aAAa,CAAC,CAAC,CAAC,gBAAgB,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC;YACtE,IAAI,wBAAwB,GAAkB,IAAI,CAAC;YACnD,IAAI,cAAc,EAAE,CAAC;gBACnB,IAAI,IAAI,GAAG,kBAAkB,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC;gBAClD,IAAI,CAAC,IAAI,EAAE,CAAC;oBACV,IAAI,GAAG,sBAAsB,CAAC,cAAc,CAAC,CAAC;oBAC9C,kBAAkB,CAAC,GAAG,CAAC,cAAc,EAAE,IAAI,CAAC,CAAC;gBAC/C,CAAC;gBACD,wBAAwB,GAAG,MAAM,IAAI,CAAC;YACxC,CAAC;YACD,OAAO;gBACL,SAAS;gBACT,MAAM;gBACN,iBAAiB,EAAE,SAAS;oBAC1B,CAAC,CAAC,MAAM,WAAW,CAAC,SAAS,CAAC,OAAO,EAAE,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC;oBAChE,CAAC,CAAC,IAAI;gBACR,wBAAwB;gBACxB,uBAAuB,EAAE,MAAM,4BAA4B,CACzD,MAAM,EACN,SAAS,CAAC,MAAM,CAAC,OAAO,EACxB,sBAAsB,CACvB;gBACD,QAAQ;gBACR,mBAAmB,EAAE,WAAW;oBAC9B,CAAC,CAAC,MAAM,WAAW,CAAC,WAAW,CAAC,OAAO,EAAE,SAAS,CAAC,MAAM,CAAC,OAAO,CAAC;oBAClE,CAAC,CAAC,IAAI;aACT,CAAC;QACJ,CAAC,CAAC,CACH,CAAC;QACF,OAAO;YACL,IAAI,EAAE,QAAQ,CAAC,IAAI;YACnB,UAAU,EAAE,WAAW,CAAC,MAAM,QAAQ,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;YACtD,IAAI,EAAE,QAAQ;YACd,YAAY,EAAE,MAAM,mBAAmB,CAAC,SAAS,EAAE,QAAQ,CAAC,IAAI,EAAE,QAAQ,CAAC,QAAQ,CAAC;YACpF,oBAAoB;YACpB,uBAAuB,EAAE,MAAM,uBAAuB,CAAC,SAAS,EAAE,KAAK,CAAC;SACzE,CAAC;IACJ,CAAC,CAAC,CACH,CAAC;IACF,OAAO,SAAS,CAAC,WAAW,CAAC,CAAC;AAChC,CAAC,CAAC;AAEF,MAAM,CAAC,MAAM,gBAAgB,GAAG,KAAK,EACnC,SAAmC,EACnC,MAA2C,EAC1B,EAAE;IACnB,MAAM,MAAM,GAAG,SAAS,CAAC,MAAM,CAAC,iBAAiB,CAAC;IAClD,IAAI,CAAC,MAAM;QAAE,MAAM,IAAI,KAAK,CAAC,qDAAqD,CAAC,CAAC;IACpF,MAAM,QAAQ,GAAG;QACf,GAAG,IAAI,GAAG,CACR,MAAM;aACH,OAAO,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,uBAAuB,CAAC,MAAM,CAAC,WAAW,EAAE,KAAK,CAAC,QAAQ,CAAC,CAAC;aAC/E,GAAG,CAAC,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC,UAAU,CAAC,EAAE,EAAE,UAAU,CAAC,CAAC,CACpD,CAAC,MAAM,EAAE;KACX,CAAC;IACF,MAAM,QAAQ,GAAG,MAAM,OAAO,CAAC,GAAG,CAChC,QAAQ,CAAC,GAAG,CAAC,KAAK,EAAE,UAAU,EAAE,EAAE;QAChC,MAAM,GAAG,GAAG,MAAM,uBAAuB,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC;QAC7D,IAAI,CAAC;YACH,MAAM,MAAM,GAAG,MAAM,oBAAoB,CACvC,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,WAAW,EAAE,UAAU,CAAC,UAAU,CAAC,CAC1D,CAAC;YACF,IAAI,MAAM,CAAC,EAAE,KAAK,GAAG,CAAC,IAAI,CAAC,EAAE,EAAE,CAAC;gBAC9B,MAAM,IAAI,KAAK,CACb,WAAW,UAAU,CAAC,UAAU,SAAS,MAAM,CAAC,EAAE,oCAAoC,GAAG,CAAC,IAAI,CAAC,EAAE,GAAG,CACrG,CAAC;YACJ,CAAC;YACD,MAAM,SAAS,GAAG,MAAM,OAAO,CAAC,GAAG,CACjC,GAAG,CAAC,SAAS,CAAC,GAAG,CAAC,KAAK,EAAE,QAAQ,EAAE,EAAE;gBACnC,MAAM,IAAI,GAAG,QAAQ,CAAC,IAAI,IAAI,IAAI,CAAC,UAAU,CAAC,MAAM,EAAE,WAAW,CAAC,CAAC;gBACnE,IAAI,CAAC;oBACH,OAAO;wBACL,GAAG,QAAQ;wBACX,IAAI,EAAE,SAAS;wBACf,WAAW,EAAE,WAAW,CAAC,MAAM,QAAQ,CAAC,IAAI,CAAC,IAAI,EAAE,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC;qBACpE,CAAC;gBACJ,CAAC;gBAAC,OAAO,KAAK,EAAE,CAAC;oBACf,IAAK,KAA+B,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;wBACvD,OAAO,EAAE,GAAG,QAAQ,EAAE,IAAI,EAAE,SAAS,EAAE,WAAW,EAAE,IAAI,EAAE,CAAC;oBAC7D,CAAC;oBACD,MAAM,KAAK,CAAC;gBACd,CAAC;YACH,CAAC,CAAC,CACH,CAAC;YACF,OAAO;gBACL,YAAY,EAAE,UAAU,CAAC,EAAE;gBAC3B,MAAM,EAAE,UAAU,CAAC,MAAM;gBACzB,YAAY,EAAE,SAAS,CAAC;oBACtB,MAAM,EAAE,GAAG,CAAC,IAAI,CAAC,EAAE;oBACnB,MAAM,EAAE,GAAG,CAAC,IAAI,CAAC,MAAM;oBACvB,MAAM,EAAE,GAAG,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,KAAK,EAAE,EAAE;wBACnC,IAAI,KAAK,CAAC,IAAI,KAAK,WAAW,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO;4BAAE,OAAO,EAAE,CAAC;wBACpE,MAAM,EAAE,EAAE,EAAE,UAAU,EAAE,GAAG,OAAO,EAAE,GAAG,KAAK,CAAC;wBAC7C,OAAO,CAAC,OAAO,CAAC,CAAC;oBACnB,CAAC,CAAC;oBACF,SAAS;iBACV,CAAC;gBACF,MAAM;aACP,CAAC;QACJ,CAAC;gBAAS,CAAC;YACT,MAAM,0BAA0B,CAAC,GAAG,CAAC,CAAC;QACxC,CAAC;IACH,CAAC,CAAC,CACH,CAAC;IACF,MAAM,iBAAiB,GAAG,IAAI,GAAG,CAC/B,MAAM,CAAC,OAAO,CAAC,CAAC,KAAK,EAAE,EAAE,CACvB,uBAAuB,CAAC,MAAM,CAAC,WAAW,EAAE,KAAK,CAAC,QAAQ,CAAC,CAAC,GAAG,CAC7D,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC,UAAU,CAAC,EAAE,EAAE,KAAK,CAAC,IAAI,CAAU,CACrD,CACF,CACF,CAAC;IACF,MAAM,eAAe,GAAG,IAAI,GAAG,EAA4D,CAAC;IAC5F,KAAK,MAAM,IAAI,IAAI,QAAQ,EAAE,CAAC;QAC5B,MAAM,KAAK,GAAG,iBAAiB,CAAC,GAAG,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;QACvD,IAAI,CAAC,KAAK;YAAE,SAAS;QACrB,MAAM,KAAK,GAAG,eAAe,CAAC,GAAG,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;QACrD,IAAI,KAAK,EAAE,CAAC;YACV,MAAM,KAAK,GAAG,KAAK,CAAC,KAAK,KAAK,KAAK,CAAC,CAAC,CAAC,6BAA6B,KAAK,EAAE,CAAC,CAAC,CAAC,eAAe,CAAC;YAC7F,MAAM,IAAI,KAAK,CACb,UAAU,KAAK,kBAAkB,KAAK,CAAC,YAAY,MAAM,KAAK,CAAC,KAAK,UAAU,IAAI,CAAC,YAAY,MAAM,KAAK,qCAAqC,CAChJ,CAAC;QACJ,CAAC;QACD,eAAe,CAAC,GAAG,CAAC,IAAI,CAAC,YAAY,EAAE,EAAE,YAAY,EAAE,IAAI,CAAC,YAAY,EAAE,KAAK,EAAE,CAAC,CAAC;IACrF,CAAC;IACD,OAAO,SAAS,CAAC,EAAE,eAAe,EAAE,MAAM,CAAC,IAAI,EAAE,QAAQ,EAAE,CAAC,CAAC;AAC/D,CAAC,CAAC;AAEF,MAAM,CAAC,MAAM,aAAa,GAAG,KAAK,EAChC,SAAmC,EACR,EAAE;IAC7B,MAAM,QAAQ,GAAG,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC;IAC3C,MAAM,MAAM,GAAG,gBAAgB,CAAC,QAAQ,CAAC,CAAC;IAC1C,IAAI,QAAQ,CAAC,UAAU,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;QAC1C,MAAM,MAAM,GAAG,SAAS,CAAC,MAAM,CAAC,iBAAiB,CAAC;QAClD,IAAI,CAAC,MAAM;YAAE,MAAM,IAAI,KAAK,CAAC,qDAAqD,CAAC,CAAC;QACpF,MAAM,cAAc,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC;YAC5C,IAAI,EAAE,KAAK,CAAC,IAAI;YAChB,WAAW,EAAE,uBAAuB,CAAC,MAAM,CAAC,WAAW,EAAE,KAAK,CAAC,QAAQ,CAAC;SACzE,CAAC,CAAC,CAAC;QACJ,KAAK,MAAM,KAAK,IAAI,cAAc,EAAE,CAAC;YACnC,IAAI,KAAK,CAAC,WAAW,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;gBACnC,MAAM,IAAI,KAAK,CACb,eAAe,QAAQ,CAAC,EAAE,KAAK,KAAK,CAAC,IAAI,yCAAyC,CACnF,CAAC;YACJ,CAAC;QACH,CAAC;QACD,8BAA8B,CAAC,cAAc,CAAC,CAAC;QAC/C,OAAO,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,MAAM,gBAAgB,CAAC,SAAS,EAAE,MAAM,CAAC,EAAE,CAAC;IAC7E,CAAC;IACD,MAAM,SAAS,GAAG,MAAM,aAAa,CACnC,SAAS,CAAC,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,OAAO,EAAE,OAAO,CAAC,CACtE,CAAC;IACF,IAAI,SAAS,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAClC,MAAM,IAAI,KAAK,CACb,qCAAqC,SAAS,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,GAAG,OAAO,CAAC,IAAI,KAAK,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAC5H,CAAC;IACJ,CAAC;IACD,MAAM,MAAM,GAAG,IAAI,GAAG,EAAkB,CAAC;IACzC,MAAM,OAAO,GAAG,IAAI,GAAG,EAAkB,CAAC;IAC1C,MAAM,cAAc,GAAG,IAAI,GAAG,EAG3B,CAAC;IACJ,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;QAC3B,MAAM,QAAQ,GAAG,WAAW,CAAC,SAAS,CAAC,KAAK,EAAE,KAAK,CAAC,QAAQ,CAAC,CAAC;QAC9D,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YAC1B,MAAM,IAAI,KAAK,CAAC,eAAe,QAAQ,CAAC,EAAE,KAAK,KAAK,CAAC,IAAI,4BAA4B,CAAC,CAAC;QACzF,CAAC;QACD,KAAK,MAAM,QAAQ,IAAI,QAAQ,EAAE,CAAC;YAChC,MAAM,YAAY,GAAG,OAAO,CAAC,GAAG,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC;YAC9C,IAAI,YAAY,IAAI,YAAY,KAAK,QAAQ,CAAC,IAAI,EAAE,CAAC;gBACnD,MAAM,IAAI,KAAK,CACb,uBAAuB,QAAQ,CAAC,EAAE,sBAAsB,YAAY,QAAQ,QAAQ,CAAC,IAAI,EAAE,CAC5F,CAAC;YACJ,CAAC;YACD,OAAO,CAAC,GAAG,CAAC,QAAQ,CAAC,EAAE,EAAE,QAAQ,CAAC,IAAI,CAAC,CAAC;YACxC,MAAM,KAAK,GAAG,MAAM,CAAC,GAAG,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;YACxC,IAAI,KAAK,EAAE,CAAC;gBACV,MAAM,IAAI,KAAK,CACb,cAAc,QAAQ,CAAC,EAAE,qBAAqB,KAAK,QAAQ,KAAK,CAAC,IAAI,SAAS,CAC/E,CAAC;YACJ,CAAC;YACD,MAAM,CAAC,GAAG,CAAC,QAAQ,CAAC,IAAI,EAAE,KAAK,CAAC,IAAI,CAAC,CAAC;YACtC,MAAM,YAAY,GAAG,MAAM,qBAAqB,CAAC,SAAS,EAAE,QAAQ,CAAC,IAAI,CAAC,CAAC;YAC3E,MAAM,aAAa,GAAG,cAAc,CAAC,GAAG,CAAC,YAAY,CAAC,CAAC;YACvD,IAAI,aAAa,EAAE,CAAC;gBAClB,MAAM,KAAK,GACT,aAAa,CAAC,KAAK,KAAK,KAAK,CAAC,IAAI;oBAChC,CAAC,CAAC,6BAA6B,KAAK,CAAC,IAAI,EAAE;oBAC3C,CAAC,CAAC,eAAe,CAAC;gBACtB,MAAM,IAAI,KAAK,CACb,SAAS,KAAK,YAAY,aAAa,CAAC,MAAM,MAAM,aAAa,CAAC,KAAK,UAAU,QAAQ,CAAC,EAAE,MAAM,KAAK,CAAC,IAAI,yCAAyC,CACtJ,CAAC;YACJ,CAAC;YACD,cAAc,CAAC,GAAG,CAAC,YAAY,EAAE;gBAC/B,MAAM,EAAE,QAAQ,CAAC,EAAE;gBACnB,IAAI,EAAE,QAAQ,CAAC,IAAI;gBACnB,KAAK,EAAE,KAAK,CAAC,IAAI;aAClB,CAAC,CAAC;QACL,CAAC;IACH,CAAC;IACD,OAAO;QACL,IAAI,EAAE,OAAO;QACb,KAAK,EAAE,SAAS,CAAC,KAAK;QACtB,IAAI,EAAE,MAAM,eAAe,CAAC,SAAS,EAAE,SAAS,CAAC,KAAK,EAAE,MAAM,CAAC;KAChE,CAAC;AACJ,CAAC,CAAC;AAEF,MAAM,CAAC,MAAM,mBAAmB,GAAG,KAAK,EACtC,SAAmC,EACnC,QAAmB,EACJ,EAAE;IACjB,IAAI,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,UAAU,CAAC,IAAI,KAAK,QAAQ;QAAE,OAAO;IACnE,MAAM,iBAAiB,GAAG,SAAS,CAAC,MAAM,CAAC,iBAAiB,EAAE,WAAW,IAAI,EAAE,CAAC;IAChF,MAAM,WAAW,GAAG;QAClB,GAAG,IAAI,GAAG,CACR,gBAAgB,CAAC,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC;aACxC,OAAO,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,uBAAuB,CAAC,iBAAiB,EAAE,KAAK,CAAC,QAAQ,CAAC,CAAC;aAC9E,GAAG,CAAC,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC,UAAU,CAAC,EAAE,EAAE,UAAU,CAAC,CAAC,CACpD,CAAC,MAAM,EAAE;KACX,CAAC;IACF,MAAM,WAAW,GAAG,CAAC,GAAG,IAAI,GAAG,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,UAAU,EAAE,EAAE,CAAC,UAAU,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC;IAChG,MAAM,cAAc,GAAG,MAAM,CAAC,IAAI,CAAC,QAAQ,CAAC,UAAU,CAAC,CAAC,IAAI,EAAE,CAAC;IAC/D,IACE,WAAW,CAAC,MAAM,KAAK,cAAc,CAAC,MAAM;QAC5C,WAAW,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC,IAAI,KAAK,cAAc,CAAC,KAAK,CAAC,CAAC,EACjE,CAAC;QACD,MAAM,IAAI,KAAK,CACb,qFAAqF,CACtF,CAAC;IACJ,CAAC;IACD,KAAK,MAAM,UAAU,IAAI,WAAW,EAAE,CAAC;QACrC,MAAM,MAAM,GAAG,MAAM,oBAAoB,CAAC,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,WAAW,EAAE,UAAU,CAAC,CAAC,CAAC;QAC1F,MAAM,WAAW,GAAG,IAAI,GAAG,CAAC,MAAM,CAAC,UAAU,CAAC,GAAG,CAAC,CAAC,SAAS,EAAE,EAAE,CAAC,SAAS,CAAC,EAAE,CAAC,CAAC,CAAC;QAChF,MAAM,QAAQ,GAAG,WAAW,CAAC,MAAM,CAAC,CAAC,UAAU,EAAE,EAAE,CAAC,UAAU,CAAC,UAAU,KAAK,UAAU,CAAC,CAAC;QAC1F,MAAM,gBAAgB,GAAG,IAAI,GAAG,CAC9B,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,UAAU,CAAC,yBAAyB,CAAC,UAAU,CAAC,IAAI,EAAE,CACjF,CAAC;QACF,MAAM,aAAa,GAAG,IAAI,GAAG,CAC3B,QAAQ,CAAC,OAAO,CAAC,CAAC,UAAU,EAAE,EAAE,CAAC,MAAM,CAAC,IAAI,CAAC,UAAU,CAAC,QAAQ,CAAC,UAAU,CAAC,CAAC,CAC9E,CAAC;QACF,MAAM,mBAAmB,GAAG,CAAC,GAAG,aAAa,CAAC,CAAC,MAAM,CACnD,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,WAAW,CAAC,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,gBAAgB,CAAC,GAAG,CAAC,EAAE,CAAC,CAC1D,CAAC;QACF,IAAI,mBAAmB,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACnC,MAAM,IAAI,KAAK,CACb,yCAAyC,UAAU,+DAA+D,mBAAmB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CACnJ,CAAC;QACJ,CAAC;QACD,MAAM,mBAAmB,GAAG,CAAC,GAAG,gBAAgB,CAAC,CAAC,MAAM,CACtD,CAAC,EAAE,EAAE,EAAE,CACL,WAAW,CAAC,GAAG,CAAC,EAAE,CAAC;YACnB,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,UAAU,EAAE,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,UAAU,CAAC,QAAQ,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC,CACrF,CAAC;QACF,IAAI,mBAAmB,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACnC,MAAM,IAAI,KAAK,CACb,oCAAoC,UAAU,mEAAmE,mBAAmB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAClJ,CAAC;QACJ,CAAC;IACH,CAAC;IACD,MAAM,kBAAkB,GAAG,MAAM,CAAC,IAAI,CACpC,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,UAAU,CAAC,yBAAyB,CAC/D,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC,WAAW,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC;IAChD,IAAI,kBAAkB,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAClC,MAAM,IAAI,KAAK,CACb,+EAA+E,kBAAkB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAC/G,CAAC;IACJ,CAAC;AACH,CAAC,CAAC"}
@@ -0,0 +1,44 @@
1
+ import { z } from "zod";
2
+ import { type Candidate, type ExperimentSelector } from "../core/index.js";
3
+ import { type discoverCases } from "./discover.js";
4
+ import type { CampaignExecutionOptions } from "./experiment-context.js";
5
+ import type { ExperimentEvaluation, OpenedExperimentRun } from "./experiment-runs.js";
6
+ type ExperimentSplit = "train" | "validation" | "test";
7
+ declare const EvaluationAccountingSchema: z.ZodObject<{
8
+ metricCalls: z.ZodOptional<z.ZodNumber>;
9
+ judgeCalls: z.ZodOptional<z.ZodNumber>;
10
+ unpricedModelCalls: z.ZodOptional<z.ZodNumber>;
11
+ usage: z.ZodOptional<z.ZodObject<{
12
+ input: z.ZodNumber;
13
+ output: z.ZodNumber;
14
+ cacheRead: z.ZodNumber;
15
+ cacheCreation: z.ZodNumber;
16
+ }, z.core.$strip>>;
17
+ costUsd: z.ZodOptional<z.ZodNumber>;
18
+ }, z.core.$strict>;
19
+ export type EvaluationAccounting = z.infer<typeof EvaluationAccountingSchema>;
20
+ export declare const invokeWithVerifiedExecutionIdentity: <T>(verifyExecutionIdentity: () => Promise<void>, invocation: () => Promise<T>) => Promise<T>;
21
+ export declare const evaluateAgentCandidate: (options: {
22
+ execution: CampaignExecutionOptions;
23
+ experiment: OpenedExperimentRun;
24
+ candidate: Candidate;
25
+ allCases: Awaited<ReturnType<typeof discoverCases>>["cases"];
26
+ profileId: string;
27
+ split: ExperimentSplit;
28
+ selector: ExperimentSelector;
29
+ repetitions: number;
30
+ verifyExecutionIdentity: () => Promise<void>;
31
+ }) => Promise<ExperimentEvaluation>;
32
+ export declare const evaluateOfflineGraderCandidate: (options: {
33
+ execution: CampaignExecutionOptions;
34
+ experiment: OpenedExperimentRun;
35
+ candidate: Candidate;
36
+ materializedRoot: string;
37
+ profileId: string;
38
+ split: ExperimentSplit;
39
+ selector: ExperimentSelector;
40
+ repetitions: number;
41
+ verifyExecutionIdentity: () => Promise<void>;
42
+ }) => Promise<ExperimentEvaluation>;
43
+ export {};
44
+ //# sourceMappingURL=experiment-evaluation.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"experiment-evaluation.d.ts","sourceRoot":"","sources":["../../src/orchestrator/experiment-evaluation.ts"],"names":[],"mappings":"AAGA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAKxB,OAAO,EAEL,KAAK,SAAS,EAGd,KAAK,kBAAkB,EAIxB,MAAM,kBAAkB,CAAC;AAG1B,OAAO,EAAE,KAAK,aAAa,EAAe,MAAM,eAAe,CAAC;AAEhE,OAAO,KAAK,EAAE,wBAAwB,EAAE,MAAM,yBAAyB,CAAC;AAExE,OAAO,KAAK,EAAE,oBAAoB,EAAE,mBAAmB,EAAE,MAAM,sBAAsB,CAAC;AAOtF,KAAK,eAAe,GAAG,OAAO,GAAG,YAAY,GAAG,MAAM,CAAC;AAuFvD,QAAA,MAAM,0BAA0B;;;;;;;;;;;kBAQrB,CAAC;AACZ,MAAM,MAAM,oBAAoB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,0BAA0B,CAAC,CAAC;AAkQ9E,eAAO,MAAM,mCAAmC,GAAU,CAAC,2BAChC,MAAM,OAAO,CAAC,IAAI,CAAC,cAChC,MAAM,OAAO,CAAC,CAAC,CAAC,KAC3B,OAAO,CAAC,CAAC,CAOX,CAAC;AAYF,eAAO,MAAM,sBAAsB,YAAmB;IACpD,SAAS,EAAE,wBAAwB,CAAC;IACpC,UAAU,EAAE,mBAAmB,CAAC;IAChC,SAAS,EAAE,SAAS,CAAC;IACrB,QAAQ,EAAE,OAAO,CAAC,UAAU,CAAC,OAAO,aAAa,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC;IAC7D,SAAS,EAAE,MAAM,CAAC;IAClB,KAAK,EAAE,eAAe,CAAC;IACvB,QAAQ,EAAE,kBAAkB,CAAC;IAC7B,WAAW,EAAE,MAAM,CAAC;IACpB,uBAAuB,EAAE,MAAM,OAAO,CAAC,IAAI,CAAC,CAAC;CAC9C,KAAG,OAAO,CAAC,oBAAoB,CAgP/B,CAAC;AAEF,eAAO,MAAM,8BAA8B,YAAmB;IAC5D,SAAS,EAAE,wBAAwB,CAAC;IACpC,UAAU,EAAE,mBAAmB,CAAC;IAChC,SAAS,EAAE,SAAS,CAAC;IACrB,gBAAgB,EAAE,MAAM,CAAC;IACzB,SAAS,EAAE,MAAM,CAAC;IAClB,KAAK,EAAE,eAAe,CAAC;IACvB,QAAQ,EAAE,kBAAkB,CAAC;IAC7B,WAAW,EAAE,MAAM,CAAC;IACpB,uBAAuB,EAAE,MAAM,OAAO,CAAC,IAAI,CAAC,CAAC;CAC9C,KAAG,OAAO,CAAC,oBAAoB,CA+E/B,CAAC"}