@tangle-network/agent-eval 0.95.0 → 0.96.0

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 (968) hide show
  1. package/CHANGELOG.md +17 -0
  2. package/dist/adapters/http.d.ts +17 -10
  3. package/dist/adapters/langchain.d.ts +14 -7
  4. package/dist/adapters/otel.d.ts +25 -13
  5. package/dist/{rl/adversarial.d.ts → adversarial-DIVcDoI_.d.ts} +7 -6
  6. package/dist/analyst/index.d.ts +236 -28
  7. package/dist/analyst/index.js +1 -1
  8. package/dist/{trace-analyst/analyst.d.ts → analyst-C8HHvfJp.d.ts} +14 -12
  9. package/dist/{contract/analyze-runs.d.ts → analyze-runs-DtT6F_6T.d.ts} +10 -9
  10. package/dist/authenticity/index.d.ts +16 -15
  11. package/dist/{baseline.d.ts → baseline-Bbid3WoO.d.ts} +44 -8
  12. package/dist/belief-state/index.d.ts +605 -14
  13. package/dist/benchmarks/index.d.ts +5 -23
  14. package/dist/builder-eval/index.d.ts +250 -5
  15. package/dist/calibration-BPmzuVPk.d.ts +101 -0
  16. package/dist/campaign/index.d.ts +1452 -38
  17. package/dist/{chunk-AQ5WQAIV.js → chunk-3NHEO6ZC.js} +2 -2
  18. package/dist/chunk-3NHEO6ZC.js.map +1 -0
  19. package/dist/cli.d.ts +0 -2
  20. package/dist/{contract/intake/code-agent-session.d.ts → code-agent-session-CPHRCb4-.d.ts} +17 -15
  21. package/dist/contract/index.d.ts +824 -107
  22. package/dist/contract/index.js +145 -1
  23. package/dist/contract/index.js.map +1 -1
  24. package/dist/control-Doncu-B_.d.ts +259 -0
  25. package/dist/{control-runtime.d.ts → control-runtime-Acf9CGhw.d.ts} +26 -23
  26. package/dist/control.d.ts +10 -11
  27. package/dist/corpus-D4YW9UoJ.d.ts +560 -0
  28. package/dist/{cost-ledger.d.ts → cost-ledger-DuSqlw5B.d.ts} +11 -10
  29. package/dist/{counterfactual.d.ts → counterfactual-DlOz8PBx.d.ts} +13 -12
  30. package/dist/{dataset.d.ts → dataset-BbGkaN2I.d.ts} +14 -11
  31. package/dist/{analyst/registry.d.ts → default-registry-GyE8X5SP.d.ts} +37 -8
  32. package/dist/diagnose.d.ts +252 -1
  33. package/dist/{trace/emitter.d.ts → emitter-C2rqGH_l.d.ts} +12 -9
  34. package/dist/{errors.d.ts → errors-CzMUYo7b.d.ts} +11 -10
  35. package/dist/failure-cluster-DH9Flgcf.d.ts +76 -0
  36. package/dist/{feedback-trajectory.d.ts → feedback-trajectory-BxY0cKfs.d.ts} +38 -36
  37. package/dist/fuzz.d.ts +547 -1
  38. package/dist/gepa-Dprxvz8r.d.ts +414 -0
  39. package/dist/governance/index.d.ts +135 -5
  40. package/dist/harness-optimizer-mOl9XX_O.d.ts +106 -0
  41. package/dist/hosted/index.d.ts +239 -10
  42. package/dist/index-_Y4oNOOb.d.ts +159 -0
  43. package/dist/index.d.ts +5660 -277
  44. package/dist/index.js +1 -1
  45. package/dist/{contract/insight-report.d.ts → insight-report-BnRjTibG.d.ts} +19 -16
  46. package/dist/{trace/integrity.d.ts → integrity-D2t12mMw.d.ts} +14 -11
  47. package/dist/{judge-calibration.d.ts → judge-calibration-0p2QcWNE.d.ts} +17 -16
  48. package/dist/{analyst/kind-factory.d.ts → kind-factory-X3eDYbKn.d.ts} +61 -10
  49. package/dist/knowledge/index.d.ts +103 -3
  50. package/dist/{llm-client.d.ts → llm-client-Bj7g0rqu.d.ts} +23 -21
  51. package/dist/matrix/index.d.ts +30 -12
  52. package/dist/meta-eval/index.d.ts +182 -6
  53. package/dist/{multi-layer-verifier.d.ts → multi-layer-verifier-DUZXrPDA.d.ts} +15 -12
  54. package/dist/multishot/index.d.ts +290 -7
  55. package/dist/{rl/off-policy.d.ts → off-policy-DiwuKKg7.d.ts} +9 -8
  56. package/dist/openapi.json +1 -1
  57. package/dist/{meta-eval/outcome-store.d.ts → outcome-store-rnXLEqSn.d.ts} +8 -7
  58. package/dist/{pareto.d.ts → pareto-E-pembql.d.ts} +10 -9
  59. package/dist/perf/index.d.ts +119 -13
  60. package/dist/pipelines/index.d.ts +173 -8
  61. package/dist/pre-registration-nfUdc9EQ.d.ts +483 -0
  62. package/dist/prm/index.d.ts +104 -5
  63. package/dist/provenance-CIxfBnkl.d.ts +426 -0
  64. package/dist/query-B7GGjRox.d.ts +32 -0
  65. package/dist/{trace/raw-provider-sink.d.ts → raw-provider-sink-C46HDghv.d.ts} +14 -13
  66. package/dist/{red-team.d.ts → red-team-BWdoyleI.d.ts} +16 -14
  67. package/dist/{trace/redact.d.ts → redact-B40YG2M_.d.ts} +8 -7
  68. package/dist/release-report-pidWUMZ2.d.ts +233 -0
  69. package/dist/reporting.d.ts +16 -15
  70. package/dist/{eval-campaign.d.ts → researcher-Jr8ME1dZ.d.ts} +156 -34
  71. package/dist/rl.d.ts +1193 -1
  72. package/dist/{prm/rubric.d.ts → rubric-Cc6UHvUb.d.ts} +13 -10
  73. package/dist/{meta-eval/rubric-predictive-validity.d.ts → rubric-predictive-validity-C2hDKM8Z.d.ts} +11 -8
  74. package/dist/run-critic-CmMf05uV.d.ts +56 -0
  75. package/dist/{run-record.d.ts → run-record-CP2ObebC.d.ts} +117 -15
  76. package/dist/runtime-trajectory-BOUUjI0y.d.ts +49 -0
  77. package/dist/{trace/schema.d.ts → schema-m0gsnbt3.d.ts} +30 -29
  78. package/dist/semantic-concept-judge-DSBB2Cfp.d.ts +624 -0
  79. package/dist/{sequential.d.ts → sequential-5iSVfzl2.d.ts} +10 -9
  80. package/dist/{series-convergence.d.ts → series-convergence-D5OWMBg6.d.ts} +5 -4
  81. package/dist/sink-fetch-B1Yg4Til.d.ts +101 -0
  82. package/dist/{statistics.d.ts → statistics-CCJpTGOS.d.ts} +48 -46
  83. package/dist/{trace/store.d.ts → store-BcFXE6LG.d.ts} +12 -21
  84. package/dist/{trace-analyst/types.d.ts → store-C1YxJDEK.d.ts} +74 -18
  85. package/dist/storyboard/index.d.ts +81 -16
  86. package/dist/{summary-report.d.ts → summary-report-CInXwsza.d.ts} +160 -23
  87. package/dist/telemetry/{sink-file.d.ts → file.d.ts} +7 -5
  88. package/dist/telemetry/index.d.ts +35 -17
  89. package/dist/{sandbox-harness.d.ts → test-graded-scenario-DeODGLra.d.ts} +60 -15
  90. package/dist/{locked-jsonl-appender.d.ts → testing-C21CHsq2.d.ts} +4 -3
  91. package/dist/testing.d.ts +1 -5
  92. package/dist/traces.d.ts +976 -4
  93. package/dist/{trajectory.d.ts → trajectory-2TkpSEVh.d.ts} +9 -8
  94. package/dist/{analyst/types.d.ts → types-B5x54y6n.d.ts} +112 -19
  95. package/dist/{campaign/types.d.ts → types-BMahhhio.d.ts} +47 -44
  96. package/dist/{matrix/types.d.ts → types-BUxNaJ8c.d.ts} +11 -9
  97. package/dist/{types.d.ts → types-C7DGg5ex.d.ts} +33 -31
  98. package/dist/{verdict.d.ts → verdict-C9MlYujm.d.ts} +3 -2
  99. package/dist/wire/index.d.ts +570 -13
  100. package/dist/workflow/index.d.ts +496 -22
  101. package/package.json +2 -2
  102. package/dist/action-policy.d.ts +0 -24
  103. package/dist/action-policy.d.ts.map +0 -1
  104. package/dist/action-policy.test.d.ts +0 -2
  105. package/dist/action-policy.test.d.ts.map +0 -1
  106. package/dist/active-learning.d.ts +0 -41
  107. package/dist/active-learning.d.ts.map +0 -1
  108. package/dist/adapters/http.d.ts.map +0 -1
  109. package/dist/adapters/langchain.d.ts.map +0 -1
  110. package/dist/adapters/otel.d.ts.map +0 -1
  111. package/dist/agent-profile-cell.d.ts +0 -101
  112. package/dist/agent-profile-cell.d.ts.map +0 -1
  113. package/dist/agent-profile.d.ts +0 -27
  114. package/dist/agent-profile.d.ts.map +0 -1
  115. package/dist/agent-profile.test.d.ts +0 -2
  116. package/dist/agent-profile.test.d.ts.map +0 -1
  117. package/dist/analyst/adapters.d.ts +0 -62
  118. package/dist/analyst/adapters.d.ts.map +0 -1
  119. package/dist/analyst/analyst.test.d.ts +0 -2
  120. package/dist/analyst/analyst.test.d.ts.map +0 -1
  121. package/dist/analyst/ax-service.d.ts +0 -27
  122. package/dist/analyst/ax-service.d.ts.map +0 -1
  123. package/dist/analyst/behavioral-analyst.d.ts +0 -28
  124. package/dist/analyst/behavioral-analyst.d.ts.map +0 -1
  125. package/dist/analyst/chat-client.d.ts +0 -91
  126. package/dist/analyst/chat-client.d.ts.map +0 -1
  127. package/dist/analyst/default-registry.d.ts +0 -27
  128. package/dist/analyst/default-registry.d.ts.map +0 -1
  129. package/dist/analyst/default-registry.test.d.ts +0 -2
  130. package/dist/analyst/default-registry.test.d.ts.map +0 -1
  131. package/dist/analyst/finding-signature.d.ts +0 -48
  132. package/dist/analyst/finding-signature.d.ts.map +0 -1
  133. package/dist/analyst/finding-subject.d.ts +0 -146
  134. package/dist/analyst/finding-subject.d.ts.map +0 -1
  135. package/dist/analyst/finding-subject.test.d.ts +0 -2
  136. package/dist/analyst/finding-subject.test.d.ts.map +0 -1
  137. package/dist/analyst/findings-store.d.ts +0 -75
  138. package/dist/analyst/findings-store.d.ts.map +0 -1
  139. package/dist/analyst/index.d.ts.map +0 -1
  140. package/dist/analyst/kind-factory.d.ts.map +0 -1
  141. package/dist/analyst/kinds/failure-mode.d.ts +0 -19
  142. package/dist/analyst/kinds/failure-mode.d.ts.map +0 -1
  143. package/dist/analyst/kinds/improvement.d.ts +0 -23
  144. package/dist/analyst/kinds/improvement.d.ts.map +0 -1
  145. package/dist/analyst/kinds/index.d.ts +0 -22
  146. package/dist/analyst/kinds/index.d.ts.map +0 -1
  147. package/dist/analyst/kinds/kinds.test.d.ts +0 -2
  148. package/dist/analyst/kinds/kinds.test.d.ts.map +0 -1
  149. package/dist/analyst/kinds/knowledge-gap.d.ts +0 -28
  150. package/dist/analyst/kinds/knowledge-gap.d.ts.map +0 -1
  151. package/dist/analyst/kinds/knowledge-poisoning.d.ts +0 -22
  152. package/dist/analyst/kinds/knowledge-poisoning.d.ts.map +0 -1
  153. package/dist/analyst/kinds/skill-usage.d.ts +0 -84
  154. package/dist/analyst/kinds/skill-usage.d.ts.map +0 -1
  155. package/dist/analyst/kinds/skill-usage.test.d.ts +0 -2
  156. package/dist/analyst/kinds/skill-usage.test.d.ts.map +0 -1
  157. package/dist/analyst/parse-tolerant.d.ts +0 -26
  158. package/dist/analyst/parse-tolerant.d.ts.map +0 -1
  159. package/dist/analyst/parse-tolerant.test.d.ts +0 -2
  160. package/dist/analyst/parse-tolerant.test.d.ts.map +0 -1
  161. package/dist/analyst/registry.budget.test.d.ts +0 -2
  162. package/dist/analyst/registry.budget.test.d.ts.map +0 -1
  163. package/dist/analyst/registry.d.ts.map +0 -1
  164. package/dist/analyst/steer-firewall.d.ts +0 -35
  165. package/dist/analyst/steer-firewall.d.ts.map +0 -1
  166. package/dist/analyst/steer-firewall.test.d.ts +0 -2
  167. package/dist/analyst/steer-firewall.test.d.ts.map +0 -1
  168. package/dist/analyst/structure-findings.d.ts +0 -37
  169. package/dist/analyst/structure-findings.d.ts.map +0 -1
  170. package/dist/analyst/structure-findings.test.d.ts +0 -2
  171. package/dist/analyst/structure-findings.test.d.ts.map +0 -1
  172. package/dist/analyst/tool-groups.d.ts +0 -34
  173. package/dist/analyst/tool-groups.d.ts.map +0 -1
  174. package/dist/analyst/types.d.ts.map +0 -1
  175. package/dist/anti-slop.d.ts +0 -59
  176. package/dist/anti-slop.d.ts.map +0 -1
  177. package/dist/artifact-validator.d.ts +0 -74
  178. package/dist/artifact-validator.d.ts.map +0 -1
  179. package/dist/attestation.d.ts +0 -63
  180. package/dist/attestation.d.ts.map +0 -1
  181. package/dist/attestation.test.d.ts +0 -2
  182. package/dist/attestation.test.d.ts.map +0 -1
  183. package/dist/authenticity/index.d.ts.map +0 -1
  184. package/dist/authenticity/index.test.d.ts +0 -2
  185. package/dist/authenticity/index.test.d.ts.map +0 -1
  186. package/dist/auto-pr.d.ts +0 -120
  187. package/dist/auto-pr.d.ts.map +0 -1
  188. package/dist/baseline.d.ts.map +0 -1
  189. package/dist/behavior-dsl.d.ts +0 -73
  190. package/dist/behavior-dsl.d.ts.map +0 -1
  191. package/dist/belief-state/calibration.d.ts +0 -10
  192. package/dist/belief-state/calibration.d.ts.map +0 -1
  193. package/dist/belief-state/calibration.test.d.ts +0 -2
  194. package/dist/belief-state/calibration.test.d.ts.map +0 -1
  195. package/dist/belief-state/code-agent-corpus.d.ts +0 -66
  196. package/dist/belief-state/code-agent-corpus.d.ts.map +0 -1
  197. package/dist/belief-state/code-agent-corpus.test.d.ts +0 -2
  198. package/dist/belief-state/code-agent-corpus.test.d.ts.map +0 -1
  199. package/dist/belief-state/code-agent-evidence.d.ts +0 -22
  200. package/dist/belief-state/code-agent-evidence.d.ts.map +0 -1
  201. package/dist/belief-state/code-agent-evidence.test.d.ts +0 -2
  202. package/dist/belief-state/code-agent-evidence.test.d.ts.map +0 -1
  203. package/dist/belief-state/extract.d.ts +0 -7
  204. package/dist/belief-state/extract.d.ts.map +0 -1
  205. package/dist/belief-state/extract.test.d.ts +0 -2
  206. package/dist/belief-state/extract.test.d.ts.map +0 -1
  207. package/dist/belief-state/index.d.ts.map +0 -1
  208. package/dist/belief-state/ope.d.ts +0 -17
  209. package/dist/belief-state/ope.d.ts.map +0 -1
  210. package/dist/belief-state/ope.test.d.ts +0 -2
  211. package/dist/belief-state/ope.test.d.ts.map +0 -1
  212. package/dist/belief-state/phase0-measurement.d.ts +0 -55
  213. package/dist/belief-state/phase0-measurement.d.ts.map +0 -1
  214. package/dist/belief-state/report.d.ts +0 -17
  215. package/dist/belief-state/report.d.ts.map +0 -1
  216. package/dist/belief-state/report.test.d.ts +0 -2
  217. package/dist/belief-state/report.test.d.ts.map +0 -1
  218. package/dist/belief-state/research-evidence.d.ts +0 -23
  219. package/dist/belief-state/research-evidence.d.ts.map +0 -1
  220. package/dist/belief-state/research-evidence.test.d.ts +0 -2
  221. package/dist/belief-state/research-evidence.test.d.ts.map +0 -1
  222. package/dist/belief-state/runtime-benchmark-corpus.d.ts +0 -32
  223. package/dist/belief-state/runtime-benchmark-corpus.d.ts.map +0 -1
  224. package/dist/belief-state/runtime-hooks.d.ts +0 -87
  225. package/dist/belief-state/runtime-hooks.d.ts.map +0 -1
  226. package/dist/belief-state/runtime-hooks.test.d.ts +0 -2
  227. package/dist/belief-state/runtime-hooks.test.d.ts.map +0 -1
  228. package/dist/belief-state/selective.d.ts +0 -15
  229. package/dist/belief-state/selective.d.ts.map +0 -1
  230. package/dist/belief-state/selective.test.d.ts +0 -2
  231. package/dist/belief-state/selective.test.d.ts.map +0 -1
  232. package/dist/belief-state/shadow-probe.d.ts +0 -80
  233. package/dist/belief-state/shadow-probe.d.ts.map +0 -1
  234. package/dist/belief-state/shadow-probe.test.d.ts +0 -2
  235. package/dist/belief-state/shadow-probe.test.d.ts.map +0 -1
  236. package/dist/belief-state/types.d.ts +0 -195
  237. package/dist/belief-state/types.d.ts.map +0 -1
  238. package/dist/belief-state/types.test.d.ts +0 -2
  239. package/dist/belief-state/types.test.d.ts.map +0 -1
  240. package/dist/benchmark.d.ts +0 -14
  241. package/dist/benchmark.d.ts.map +0 -1
  242. package/dist/benchmarks/index.d.ts.map +0 -1
  243. package/dist/benchmarks/routing/dataset.d.ts +0 -34
  244. package/dist/benchmarks/routing/dataset.d.ts.map +0 -1
  245. package/dist/benchmarks/routing/index.d.ts +0 -34
  246. package/dist/benchmarks/routing/index.d.ts.map +0 -1
  247. package/dist/benchmarks/types.d.ts +0 -49
  248. package/dist/benchmarks/types.d.ts.map +0 -1
  249. package/dist/bisector.d.ts +0 -81
  250. package/dist/bisector.d.ts.map +0 -1
  251. package/dist/budget-guard.d.ts +0 -31
  252. package/dist/budget-guard.d.ts.map +0 -1
  253. package/dist/builder-eval/builder-session.d.ts +0 -111
  254. package/dist/builder-eval/builder-session.d.ts.map +0 -1
  255. package/dist/builder-eval/correlation.d.ts +0 -32
  256. package/dist/builder-eval/correlation.d.ts.map +0 -1
  257. package/dist/builder-eval/index.d.ts.map +0 -1
  258. package/dist/builder-eval/project-registry.d.ts +0 -51
  259. package/dist/builder-eval/project-registry.d.ts.map +0 -1
  260. package/dist/builder-eval/three-layer-eval.d.ts +0 -55
  261. package/dist/builder-eval/three-layer-eval.d.ts.map +0 -1
  262. package/dist/campaign/analyst-surface.d.ts +0 -108
  263. package/dist/campaign/analyst-surface.d.ts.map +0 -1
  264. package/dist/campaign/analyst-surface.test.d.ts +0 -2
  265. package/dist/campaign/analyst-surface.test.d.ts.map +0 -1
  266. package/dist/campaign/auto-pr.d.ts +0 -46
  267. package/dist/campaign/auto-pr.d.ts.map +0 -1
  268. package/dist/campaign/distillation/agreement-judge.d.ts +0 -69
  269. package/dist/campaign/distillation/agreement-judge.d.ts.map +0 -1
  270. package/dist/campaign/distillation/cli.d.ts +0 -35
  271. package/dist/campaign/distillation/cli.d.ts.map +0 -1
  272. package/dist/campaign/distillation/distillation.test.d.ts +0 -2
  273. package/dist/campaign/distillation/distillation.test.d.ts.map +0 -1
  274. package/dist/campaign/distillation/gold-scenarios.d.ts +0 -54
  275. package/dist/campaign/distillation/gold-scenarios.d.ts.map +0 -1
  276. package/dist/campaign/distillation/run-distillation.d.ts +0 -119
  277. package/dist/campaign/distillation/run-distillation.d.ts.map +0 -1
  278. package/dist/campaign/gates/compose.d.ts +0 -12
  279. package/dist/campaign/gates/compose.d.ts.map +0 -1
  280. package/dist/campaign/gates/default-production-gate.d.ts +0 -58
  281. package/dist/campaign/gates/default-production-gate.d.ts.map +0 -1
  282. package/dist/campaign/gates/heldout-gate.d.ts +0 -12
  283. package/dist/campaign/gates/heldout-gate.d.ts.map +0 -1
  284. package/dist/campaign/gates/promotion-policy.d.ts +0 -125
  285. package/dist/campaign/gates/promotion-policy.d.ts.map +0 -1
  286. package/dist/campaign/gates/promotion-policy.test.d.ts +0 -2
  287. package/dist/campaign/gates/promotion-policy.test.d.ts.map +0 -1
  288. package/dist/campaign/gates/sequential.d.ts +0 -146
  289. package/dist/campaign/gates/sequential.d.ts.map +0 -1
  290. package/dist/campaign/gates/sequential.test.d.ts +0 -2
  291. package/dist/campaign/gates/sequential.test.d.ts.map +0 -1
  292. package/dist/campaign/gates/statistical-heldout.d.ts +0 -99
  293. package/dist/campaign/gates/statistical-heldout.d.ts.map +0 -1
  294. package/dist/campaign/gates/statistical-heldout.test.d.ts +0 -2
  295. package/dist/campaign/gates/statistical-heldout.test.d.ts.map +0 -1
  296. package/dist/campaign/index.d.ts.map +0 -1
  297. package/dist/campaign/labeled-store/fs-adapter.d.ts +0 -59
  298. package/dist/campaign/labeled-store/fs-adapter.d.ts.map +0 -1
  299. package/dist/campaign/presets/compare-proposers.d.ts +0 -146
  300. package/dist/campaign/presets/compare-proposers.d.ts.map +0 -1
  301. package/dist/campaign/presets/playback.d.ts +0 -120
  302. package/dist/campaign/presets/playback.d.ts.map +0 -1
  303. package/dist/campaign/presets/playback.test.d.ts +0 -2
  304. package/dist/campaign/presets/playback.test.d.ts.map +0 -1
  305. package/dist/campaign/presets/run-eval.d.ts +0 -14
  306. package/dist/campaign/presets/run-eval.d.ts.map +0 -1
  307. package/dist/campaign/presets/run-improvement-loop.d.ts +0 -63
  308. package/dist/campaign/presets/run-improvement-loop.d.ts.map +0 -1
  309. package/dist/campaign/presets/run-improvement-loop.test.d.ts +0 -2
  310. package/dist/campaign/presets/run-improvement-loop.test.d.ts.map +0 -1
  311. package/dist/campaign/presets/run-optimization.d.ts +0 -92
  312. package/dist/campaign/presets/run-optimization.d.ts.map +0 -1
  313. package/dist/campaign/presets/run-profile-matrix.d.ts +0 -151
  314. package/dist/campaign/presets/run-profile-matrix.d.ts.map +0 -1
  315. package/dist/campaign/presets/run-skill-opt.d.ts +0 -96
  316. package/dist/campaign/presets/run-skill-opt.d.ts.map +0 -1
  317. package/dist/campaign/proposers/_findings-text.d.ts +0 -22
  318. package/dist/campaign/proposers/_findings-text.d.ts.map +0 -1
  319. package/dist/campaign/proposers/ace.d.ts +0 -33
  320. package/dist/campaign/proposers/ace.d.ts.map +0 -1
  321. package/dist/campaign/proposers/ace.test.d.ts +0 -2
  322. package/dist/campaign/proposers/ace.test.d.ts.map +0 -1
  323. package/dist/campaign/proposers/analysis-edit.d.ts +0 -32
  324. package/dist/campaign/proposers/analysis-edit.d.ts.map +0 -1
  325. package/dist/campaign/proposers/evolutionary.d.ts +0 -20
  326. package/dist/campaign/proposers/evolutionary.d.ts.map +0 -1
  327. package/dist/campaign/proposers/fapo.d.ts +0 -120
  328. package/dist/campaign/proposers/fapo.d.ts.map +0 -1
  329. package/dist/campaign/proposers/gepa.d.ts +0 -86
  330. package/dist/campaign/proposers/gepa.d.ts.map +0 -1
  331. package/dist/campaign/proposers/halo.d.ts +0 -44
  332. package/dist/campaign/proposers/halo.d.ts.map +0 -1
  333. package/dist/campaign/proposers/halo.test.d.ts +0 -2
  334. package/dist/campaign/proposers/halo.test.d.ts.map +0 -1
  335. package/dist/campaign/proposers/memory.d.ts +0 -47
  336. package/dist/campaign/proposers/memory.d.ts.map +0 -1
  337. package/dist/campaign/proposers/memory.test.d.ts +0 -2
  338. package/dist/campaign/proposers/memory.test.d.ts.map +0 -1
  339. package/dist/campaign/proposers/skill-opt.d.ts +0 -88
  340. package/dist/campaign/proposers/skill-opt.d.ts.map +0 -1
  341. package/dist/campaign/proposers/trace-analyst.d.ts +0 -48
  342. package/dist/campaign/proposers/trace-analyst.d.ts.map +0 -1
  343. package/dist/campaign/proposers/trace-analyst.test.d.ts +0 -2
  344. package/dist/campaign/proposers/trace-analyst.test.d.ts.map +0 -1
  345. package/dist/campaign/provenance.d.ts +0 -185
  346. package/dist/campaign/provenance.d.ts.map +0 -1
  347. package/dist/campaign/run-campaign.d.ts +0 -90
  348. package/dist/campaign/run-campaign.d.ts.map +0 -1
  349. package/dist/campaign/score-utils.d.ts +0 -26
  350. package/dist/campaign/score-utils.d.ts.map +0 -1
  351. package/dist/campaign/skill-patch.d.ts +0 -62
  352. package/dist/campaign/skill-patch.d.ts.map +0 -1
  353. package/dist/campaign/storage.d.ts +0 -38
  354. package/dist/campaign/storage.d.ts.map +0 -1
  355. package/dist/campaign/types.d.ts.map +0 -1
  356. package/dist/campaign/worktree/index.d.ts +0 -53
  357. package/dist/campaign/worktree/index.d.ts.map +0 -1
  358. package/dist/canary.d.ts +0 -101
  359. package/dist/canary.d.ts.map +0 -1
  360. package/dist/causal-attribution.d.ts +0 -45
  361. package/dist/causal-attribution.d.ts.map +0 -1
  362. package/dist/chunk-AQ5WQAIV.js.map +0 -1
  363. package/dist/ci-gate.d.ts +0 -44
  364. package/dist/ci-gate.d.ts.map +0 -1
  365. package/dist/cli.d.ts.map +0 -1
  366. package/dist/client.d.ts +0 -77
  367. package/dist/client.d.ts.map +0 -1
  368. package/dist/client.test.d.ts +0 -2
  369. package/dist/client.test.d.ts.map +0 -1
  370. package/dist/command-runner.d.ts +0 -74
  371. package/dist/command-runner.d.ts.map +0 -1
  372. package/dist/command-runner.test.d.ts +0 -2
  373. package/dist/command-runner.test.d.ts.map +0 -1
  374. package/dist/completion-verifier.d.ts +0 -147
  375. package/dist/completion-verifier.d.ts.map +0 -1
  376. package/dist/completion-verifier.test.d.ts +0 -9
  377. package/dist/completion-verifier.test.d.ts.map +0 -1
  378. package/dist/concurrency.d.ts +0 -23
  379. package/dist/concurrency.d.ts.map +0 -1
  380. package/dist/contamination-guard.d.ts +0 -81
  381. package/dist/contamination-guard.d.ts.map +0 -1
  382. package/dist/contract/analyze-runs.d.ts.map +0 -1
  383. package/dist/contract/define-agent-eval.d.ts +0 -52
  384. package/dist/contract/define-agent-eval.d.ts.map +0 -1
  385. package/dist/contract/diff.d.ts +0 -114
  386. package/dist/contract/diff.d.ts.map +0 -1
  387. package/dist/contract/index.d.ts.map +0 -1
  388. package/dist/contract/insight-report.d.ts.map +0 -1
  389. package/dist/contract/insight-types-fwd.d.ts +0 -7
  390. package/dist/contract/insight-types-fwd.d.ts.map +0 -1
  391. package/dist/contract/intake/agent-trace.d.ts +0 -97
  392. package/dist/contract/intake/agent-trace.d.ts.map +0 -1
  393. package/dist/contract/intake/code-agent-session.d.ts.map +0 -1
  394. package/dist/contract/intake/feedback-table.d.ts +0 -87
  395. package/dist/contract/intake/feedback-table.d.ts.map +0 -1
  396. package/dist/contract/intake/index.d.ts +0 -22
  397. package/dist/contract/intake/index.d.ts.map +0 -1
  398. package/dist/contract/intake/otel-spans.d.ts +0 -33
  399. package/dist/contract/intake/otel-spans.d.ts.map +0 -1
  400. package/dist/contract/self-improve.d.ts +0 -284
  401. package/dist/contract/self-improve.d.ts.map +0 -1
  402. package/dist/control-runtime.d.ts.map +0 -1
  403. package/dist/control-runtime.test.d.ts +0 -2
  404. package/dist/control-runtime.test.d.ts.map +0 -1
  405. package/dist/control.d.ts.map +0 -1
  406. package/dist/convergence.d.ts +0 -29
  407. package/dist/convergence.d.ts.map +0 -1
  408. package/dist/cost-ledger.d.ts.map +0 -1
  409. package/dist/cost-ledger.test.d.ts +0 -2
  410. package/dist/cost-ledger.test.d.ts.map +0 -1
  411. package/dist/cost-report.d.ts +0 -43
  412. package/dist/cost-report.d.ts.map +0 -1
  413. package/dist/cost-report.test.d.ts +0 -2
  414. package/dist/cost-report.test.d.ts.map +0 -1
  415. package/dist/cost-tracker.d.ts +0 -76
  416. package/dist/cost-tracker.d.ts.map +0 -1
  417. package/dist/counterfactual.d.ts.map +0 -1
  418. package/dist/cross-trace-diff.d.ts +0 -56
  419. package/dist/cross-trace-diff.d.ts.map +0 -1
  420. package/dist/dataset.d.ts.map +0 -1
  421. package/dist/deploy-gate-layer.d.ts +0 -125
  422. package/dist/deploy-gate-layer.d.ts.map +0 -1
  423. package/dist/deploy-gate-layer.test.d.ts +0 -2
  424. package/dist/deploy-gate-layer.test.d.ts.map +0 -1
  425. package/dist/description-length-gate.d.ts +0 -119
  426. package/dist/description-length-gate.d.ts.map +0 -1
  427. package/dist/detectors/edge.test.d.ts +0 -2
  428. package/dist/detectors/edge.test.d.ts.map +0 -1
  429. package/dist/detectors/index.d.ts +0 -81
  430. package/dist/detectors/index.d.ts.map +0 -1
  431. package/dist/detectors/index.test.d.ts +0 -2
  432. package/dist/detectors/index.test.d.ts.map +0 -1
  433. package/dist/diagnose/causal-sweep.d.ts +0 -100
  434. package/dist/diagnose/causal-sweep.d.ts.map +0 -1
  435. package/dist/diagnose/index.d.ts +0 -36
  436. package/dist/diagnose/index.d.ts.map +0 -1
  437. package/dist/diagnose/remediation.d.ts +0 -68
  438. package/dist/diagnose/remediation.d.ts.map +0 -1
  439. package/dist/diagnose/repair.d.ts +0 -77
  440. package/dist/diagnose/repair.d.ts.map +0 -1
  441. package/dist/discover-personas.d.ts +0 -35
  442. package/dist/discover-personas.d.ts.map +0 -1
  443. package/dist/driver.d.ts +0 -95
  444. package/dist/driver.d.ts.map +0 -1
  445. package/dist/driver.test.d.ts +0 -8
  446. package/dist/driver.test.d.ts.map +0 -1
  447. package/dist/dual-agent-bench.d.ts +0 -81
  448. package/dist/dual-agent-bench.d.ts.map +0 -1
  449. package/dist/error-count-extractor.d.ts +0 -47
  450. package/dist/error-count-extractor.d.ts.map +0 -1
  451. package/dist/error-count-extractor.test.d.ts +0 -2
  452. package/dist/error-count-extractor.test.d.ts.map +0 -1
  453. package/dist/errors.d.ts.map +0 -1
  454. package/dist/eval-campaign.d.ts.map +0 -1
  455. package/dist/eval-campaign.test.d.ts +0 -2
  456. package/dist/eval-campaign.test.d.ts.map +0 -1
  457. package/dist/eval-tools.d.ts +0 -55
  458. package/dist/eval-tools.d.ts.map +0 -1
  459. package/dist/eval-trace-store.d.ts +0 -107
  460. package/dist/eval-trace-store.d.ts.map +0 -1
  461. package/dist/eval-trace-store.test.d.ts +0 -2
  462. package/dist/eval-trace-store.test.d.ts.map +0 -1
  463. package/dist/executor.d.ts +0 -38
  464. package/dist/executor.d.ts.map +0 -1
  465. package/dist/executor.test.d.ts +0 -10
  466. package/dist/executor.test.d.ts.map +0 -1
  467. package/dist/experiment-tracker.d.ts +0 -178
  468. package/dist/experiment-tracker.d.ts.map +0 -1
  469. package/dist/experiment-tracker.test.d.ts +0 -2
  470. package/dist/experiment-tracker.test.d.ts.map +0 -1
  471. package/dist/failure-taxonomy.d.ts +0 -38
  472. package/dist/failure-taxonomy.d.ts.map +0 -1
  473. package/dist/feedback-trajectory.d.ts.map +0 -1
  474. package/dist/feedback-trajectory.test.d.ts +0 -2
  475. package/dist/feedback-trajectory.test.d.ts.map +0 -1
  476. package/dist/flow-layer.d.ts +0 -90
  477. package/dist/flow-layer.d.ts.map +0 -1
  478. package/dist/flow-layer.test.d.ts +0 -2
  479. package/dist/flow-layer.test.d.ts.map +0 -1
  480. package/dist/fuzz/capsule.d.ts +0 -46
  481. package/dist/fuzz/capsule.d.ts.map +0 -1
  482. package/dist/fuzz/cube.d.ts +0 -36
  483. package/dist/fuzz/cube.d.ts.map +0 -1
  484. package/dist/fuzz/explorer-cost.test.d.ts +0 -2
  485. package/dist/fuzz/explorer-cost.test.d.ts.map +0 -1
  486. package/dist/fuzz/explorer.d.ts +0 -64
  487. package/dist/fuzz/explorer.d.ts.map +0 -1
  488. package/dist/fuzz/fuzz-agent.d.ts +0 -16
  489. package/dist/fuzz/fuzz-agent.d.ts.map +0 -1
  490. package/dist/fuzz/fuzz-agent.test.d.ts +0 -2
  491. package/dist/fuzz/fuzz-agent.test.d.ts.map +0 -1
  492. package/dist/fuzz/gates.d.ts +0 -33
  493. package/dist/fuzz/gates.d.ts.map +0 -1
  494. package/dist/fuzz/index.d.ts +0 -26
  495. package/dist/fuzz/index.d.ts.map +0 -1
  496. package/dist/fuzz/policies.d.ts +0 -28
  497. package/dist/fuzz/policies.d.ts.map +0 -1
  498. package/dist/fuzz/tools.d.ts +0 -20
  499. package/dist/fuzz/tools.d.ts.map +0 -1
  500. package/dist/fuzz/types.d.ts +0 -307
  501. package/dist/fuzz/types.d.ts.map +0 -1
  502. package/dist/golden-matcher.d.ts +0 -71
  503. package/dist/golden-matcher.d.ts.map +0 -1
  504. package/dist/governance/eu-ai-act.d.ts +0 -37
  505. package/dist/governance/eu-ai-act.d.ts.map +0 -1
  506. package/dist/governance/index.d.ts.map +0 -1
  507. package/dist/governance/nist-ai-rmf.d.ts +0 -15
  508. package/dist/governance/nist-ai-rmf.d.ts.map +0 -1
  509. package/dist/governance/soc2.d.ts +0 -12
  510. package/dist/governance/soc2.d.ts.map +0 -1
  511. package/dist/governance/types.d.ts +0 -66
  512. package/dist/governance/types.d.ts.map +0 -1
  513. package/dist/harness-optimizer.d.ts +0 -82
  514. package/dist/harness-optimizer.d.ts.map +0 -1
  515. package/dist/held-out-gate.d.ts +0 -135
  516. package/dist/held-out-gate.d.ts.map +0 -1
  517. package/dist/hosted/client.d.ts +0 -73
  518. package/dist/hosted/client.d.ts.map +0 -1
  519. package/dist/hosted/from-env.test.d.ts +0 -8
  520. package/dist/hosted/from-env.test.d.ts.map +0 -1
  521. package/dist/hosted/index.d.ts.map +0 -1
  522. package/dist/hosted/types.d.ts +0 -159
  523. package/dist/hosted/types.d.ts.map +0 -1
  524. package/dist/index.d.ts.map +0 -1
  525. package/dist/integrity/backend-integrity.d.ts +0 -71
  526. package/dist/integrity/backend-integrity.d.ts.map +0 -1
  527. package/dist/integrity/preflight.d.ts +0 -72
  528. package/dist/integrity/preflight.d.ts.map +0 -1
  529. package/dist/integrity/preflight.test.d.ts +0 -2
  530. package/dist/integrity/preflight.test.d.ts.map +0 -1
  531. package/dist/integrity/single-backend.d.ts +0 -67
  532. package/dist/integrity/single-backend.d.ts.map +0 -1
  533. package/dist/intent-match-judge.d.ts +0 -69
  534. package/dist/intent-match-judge.d.ts.map +0 -1
  535. package/dist/intent-match-judge.test.d.ts +0 -2
  536. package/dist/intent-match-judge.test.d.ts.map +0 -1
  537. package/dist/judge-calibration.d.ts.map +0 -1
  538. package/dist/judge-ensemble.d.ts +0 -66
  539. package/dist/judge-ensemble.d.ts.map +0 -1
  540. package/dist/judge-ensemble.test.d.ts +0 -8
  541. package/dist/judge-ensemble.test.d.ts.map +0 -1
  542. package/dist/judge-families.d.ts +0 -38
  543. package/dist/judge-families.d.ts.map +0 -1
  544. package/dist/judge-panel.d.ts +0 -65
  545. package/dist/judge-panel.d.ts.map +0 -1
  546. package/dist/judge-retry.d.ts +0 -70
  547. package/dist/judge-retry.d.ts.map +0 -1
  548. package/dist/judge-runner.d.ts +0 -36
  549. package/dist/judge-runner.d.ts.map +0 -1
  550. package/dist/judge-runner.test.d.ts +0 -2
  551. package/dist/judge-runner.test.d.ts.map +0 -1
  552. package/dist/judges.d.ts +0 -74
  553. package/dist/judges.d.ts.map +0 -1
  554. package/dist/keyword-coverage-judge.d.ts +0 -89
  555. package/dist/keyword-coverage-judge.d.ts.map +0 -1
  556. package/dist/keyword-coverage-judge.test.d.ts +0 -2
  557. package/dist/keyword-coverage-judge.test.d.ts.map +0 -1
  558. package/dist/knowledge/index.d.ts.map +0 -1
  559. package/dist/knowledge/readiness.d.ts +0 -26
  560. package/dist/knowledge/readiness.d.ts.map +0 -1
  561. package/dist/knowledge/types.d.ts +0 -75
  562. package/dist/knowledge/types.d.ts.map +0 -1
  563. package/dist/live-proof.d.ts +0 -62
  564. package/dist/live-proof.d.ts.map +0 -1
  565. package/dist/llm-client.d.ts.map +0 -1
  566. package/dist/llm-client.test.d.ts +0 -2
  567. package/dist/llm-client.test.d.ts.map +0 -1
  568. package/dist/locked-jsonl-appender.d.ts.map +0 -1
  569. package/dist/matrix/aggregation.d.ts +0 -16
  570. package/dist/matrix/aggregation.d.ts.map +0 -1
  571. package/dist/matrix/index.d.ts.map +0 -1
  572. package/dist/matrix/runner.d.ts +0 -15
  573. package/dist/matrix/runner.d.ts.map +0 -1
  574. package/dist/matrix/types.d.ts.map +0 -1
  575. package/dist/meta-eval/calibration.d.ts +0 -47
  576. package/dist/meta-eval/calibration.d.ts.map +0 -1
  577. package/dist/meta-eval/correlation-study.d.ts +0 -53
  578. package/dist/meta-eval/correlation-study.d.ts.map +0 -1
  579. package/dist/meta-eval/index.d.ts.map +0 -1
  580. package/dist/meta-eval/outcome-store.d.ts.map +0 -1
  581. package/dist/meta-eval/rubric-predictive-validity.d.ts.map +0 -1
  582. package/dist/meta-eval/sentinel.d.ts +0 -169
  583. package/dist/meta-eval/sentinel.d.ts.map +0 -1
  584. package/dist/metrics.d.ts +0 -63
  585. package/dist/metrics.d.ts.map +0 -1
  586. package/dist/model-seats.d.ts +0 -71
  587. package/dist/model-seats.d.ts.map +0 -1
  588. package/dist/model-seats.test.d.ts +0 -2
  589. package/dist/model-seats.test.d.ts.map +0 -1
  590. package/dist/muffled-gate-scanner.d.ts +0 -102
  591. package/dist/muffled-gate-scanner.d.ts.map +0 -1
  592. package/dist/multi-layer-verifier.d.ts.map +0 -1
  593. package/dist/multi-layer-verifier.test.d.ts +0 -2
  594. package/dist/multi-layer-verifier.test.d.ts.map +0 -1
  595. package/dist/multi-toolchain-layer.d.ts +0 -80
  596. package/dist/multi-toolchain-layer.d.ts.map +0 -1
  597. package/dist/multi-toolchain-layer.test.d.ts +0 -2
  598. package/dist/multi-toolchain-layer.test.d.ts.map +0 -1
  599. package/dist/multishot/default-tools.d.ts +0 -34
  600. package/dist/multishot/default-tools.d.ts.map +0 -1
  601. package/dist/multishot/index.d.ts.map +0 -1
  602. package/dist/multishot/judges.d.ts +0 -32
  603. package/dist/multishot/judges.d.ts.map +0 -1
  604. package/dist/multishot/matrix.d.ts +0 -107
  605. package/dist/multishot/matrix.d.ts.map +0 -1
  606. package/dist/multishot/multishot.d.ts +0 -23
  607. package/dist/multishot/multishot.d.ts.map +0 -1
  608. package/dist/multishot/router.d.ts +0 -37
  609. package/dist/multishot/router.d.ts.map +0 -1
  610. package/dist/multishot/types.d.ts +0 -60
  611. package/dist/multishot/types.d.ts.map +0 -1
  612. package/dist/observability.d.ts +0 -71
  613. package/dist/observability.d.ts.map +0 -1
  614. package/dist/oracle.d.ts +0 -55
  615. package/dist/oracle.d.ts.map +0 -1
  616. package/dist/orthogonality.d.ts +0 -35
  617. package/dist/orthogonality.d.ts.map +0 -1
  618. package/dist/otel-pipeline.d.ts +0 -31
  619. package/dist/otel-pipeline.d.ts.map +0 -1
  620. package/dist/paraphrase.d.ts +0 -107
  621. package/dist/paraphrase.d.ts.map +0 -1
  622. package/dist/pareto.d.ts.map +0 -1
  623. package/dist/partition-held-out.d.ts +0 -70
  624. package/dist/partition-held-out.d.ts.map +0 -1
  625. package/dist/partition-held-out.test.d.ts +0 -2
  626. package/dist/partition-held-out.test.d.ts.map +0 -1
  627. package/dist/perf/index.d.ts.map +0 -1
  628. package/dist/perf/integrity.d.ts +0 -30
  629. package/dist/perf/integrity.d.ts.map +0 -1
  630. package/dist/perf/journey.d.ts +0 -45
  631. package/dist/perf/journey.d.ts.map +0 -1
  632. package/dist/perf/ratchet.d.ts +0 -47
  633. package/dist/perf/ratchet.d.ts.map +0 -1
  634. package/dist/pipelines/budget-breach.d.ts +0 -31
  635. package/dist/pipelines/budget-breach.d.ts.map +0 -1
  636. package/dist/pipelines/budget-breach.test.d.ts +0 -2
  637. package/dist/pipelines/budget-breach.test.d.ts.map +0 -1
  638. package/dist/pipelines/failure-cluster.d.ts +0 -38
  639. package/dist/pipelines/failure-cluster.d.ts.map +0 -1
  640. package/dist/pipelines/failure-cluster.test.d.ts +0 -2
  641. package/dist/pipelines/failure-cluster.test.d.ts.map +0 -1
  642. package/dist/pipelines/first-divergence.d.ts +0 -26
  643. package/dist/pipelines/first-divergence.d.ts.map +0 -1
  644. package/dist/pipelines/first-divergence.test.d.ts +0 -2
  645. package/dist/pipelines/first-divergence.test.d.ts.map +0 -1
  646. package/dist/pipelines/index.d.ts.map +0 -1
  647. package/dist/pipelines/judge-agreement.d.ts +0 -26
  648. package/dist/pipelines/judge-agreement.d.ts.map +0 -1
  649. package/dist/pipelines/judge-agreement.test.d.ts +0 -2
  650. package/dist/pipelines/judge-agreement.test.d.ts.map +0 -1
  651. package/dist/pipelines/regression.d.ts +0 -23
  652. package/dist/pipelines/regression.d.ts.map +0 -1
  653. package/dist/pipelines/regression.test.d.ts +0 -2
  654. package/dist/pipelines/regression.test.d.ts.map +0 -1
  655. package/dist/pipelines/stuck-loop.d.ts +0 -32
  656. package/dist/pipelines/stuck-loop.d.ts.map +0 -1
  657. package/dist/pipelines/stuck-loop.test.d.ts +0 -2
  658. package/dist/pipelines/stuck-loop.test.d.ts.map +0 -1
  659. package/dist/pipelines/tool-waste.d.ts +0 -34
  660. package/dist/pipelines/tool-waste.d.ts.map +0 -1
  661. package/dist/pipelines/tool-waste.test.d.ts +0 -2
  662. package/dist/pipelines/tool-waste.test.d.ts.map +0 -1
  663. package/dist/playbook.d.ts +0 -16
  664. package/dist/playbook.d.ts.map +0 -1
  665. package/dist/pr-review-benchmark.d.ts +0 -88
  666. package/dist/pr-review-benchmark.d.ts.map +0 -1
  667. package/dist/pr-review-benchmark.test.d.ts +0 -2
  668. package/dist/pr-review-benchmark.test.d.ts.map +0 -1
  669. package/dist/pre-registration.d.ts +0 -125
  670. package/dist/pre-registration.d.ts.map +0 -1
  671. package/dist/prm/builtin-rubrics.d.ts +0 -33
  672. package/dist/prm/builtin-rubrics.d.ts.map +0 -1
  673. package/dist/prm/index.d.ts.map +0 -1
  674. package/dist/prm/inference.d.ts +0 -29
  675. package/dist/prm/inference.d.ts.map +0 -1
  676. package/dist/prm/inference.test.d.ts +0 -2
  677. package/dist/prm/inference.test.d.ts.map +0 -1
  678. package/dist/prm/rubric.d.ts.map +0 -1
  679. package/dist/prm/training-export.d.ts +0 -38
  680. package/dist/prm/training-export.d.ts.map +0 -1
  681. package/dist/produced-state.d.ts +0 -63
  682. package/dist/produced-state.d.ts.map +0 -1
  683. package/dist/produced-state.test.d.ts +0 -8
  684. package/dist/produced-state.test.d.ts.map +0 -1
  685. package/dist/profile/baselines.d.ts +0 -37
  686. package/dist/profile/baselines.d.ts.map +0 -1
  687. package/dist/profile/index.d.ts +0 -105
  688. package/dist/profile/index.d.ts.map +0 -1
  689. package/dist/promotion-gate.d.ts +0 -93
  690. package/dist/promotion-gate.d.ts.map +0 -1
  691. package/dist/prompt-registry.d.ts +0 -41
  692. package/dist/prompt-registry.d.ts.map +0 -1
  693. package/dist/propose-review-control.d.ts +0 -49
  694. package/dist/propose-review-control.d.ts.map +0 -1
  695. package/dist/propose-review-control.test.d.ts +0 -2
  696. package/dist/propose-review-control.test.d.ts.map +0 -1
  697. package/dist/propose-review.d.ts +0 -155
  698. package/dist/propose-review.d.ts.map +0 -1
  699. package/dist/red-team.d.ts.map +0 -1
  700. package/dist/reference-replay-steering.d.ts +0 -11
  701. package/dist/reference-replay-steering.d.ts.map +0 -1
  702. package/dist/reference-replay.d.ts +0 -176
  703. package/dist/reference-replay.d.ts.map +0 -1
  704. package/dist/reflective-mutation.d.ts +0 -79
  705. package/dist/reflective-mutation.d.ts.map +0 -1
  706. package/dist/registry.d.ts +0 -31
  707. package/dist/registry.d.ts.map +0 -1
  708. package/dist/release-confidence.d.ts +0 -128
  709. package/dist/release-confidence.d.ts.map +0 -1
  710. package/dist/release-report.d.ts +0 -11
  711. package/dist/release-report.d.ts.map +0 -1
  712. package/dist/replay.d.ts +0 -120
  713. package/dist/replay.d.ts.map +0 -1
  714. package/dist/reporter.d.ts +0 -14
  715. package/dist/reporter.d.ts.map +0 -1
  716. package/dist/reporting.d.ts.map +0 -1
  717. package/dist/researcher.d.ts +0 -140
  718. package/dist/researcher.d.ts.map +0 -1
  719. package/dist/reviewer.d.ts +0 -118
  720. package/dist/reviewer.d.ts.map +0 -1
  721. package/dist/reviewer.test.d.ts +0 -2
  722. package/dist/reviewer.test.d.ts.map +0 -1
  723. package/dist/reward-model-export.d.ts +0 -60
  724. package/dist/reward-model-export.d.ts.map +0 -1
  725. package/dist/rl/active-curriculum.d.ts +0 -110
  726. package/dist/rl/active-curriculum.d.ts.map +0 -1
  727. package/dist/rl/adaptation-eval.d.ts +0 -109
  728. package/dist/rl/adaptation-eval.d.ts.map +0 -1
  729. package/dist/rl/adversarial.d.ts.map +0 -1
  730. package/dist/rl/compute-curves.d.ts +0 -127
  731. package/dist/rl/compute-curves.d.ts.map +0 -1
  732. package/dist/rl/contamination.d.ts +0 -117
  733. package/dist/rl/contamination.d.ts.map +0 -1
  734. package/dist/rl/corpus.d.ts +0 -55
  735. package/dist/rl/corpus.d.ts.map +0 -1
  736. package/dist/rl/corpus.test.d.ts +0 -2
  737. package/dist/rl/corpus.test.d.ts.map +0 -1
  738. package/dist/rl/dataset.d.ts +0 -102
  739. package/dist/rl/dataset.d.ts.map +0 -1
  740. package/dist/rl/dataset.test.d.ts +0 -2
  741. package/dist/rl/dataset.test.d.ts.map +0 -1
  742. package/dist/rl/exporters.d.ts +0 -141
  743. package/dist/rl/exporters.d.ts.map +0 -1
  744. package/dist/rl/index.d.ts +0 -49
  745. package/dist/rl/index.d.ts.map +0 -1
  746. package/dist/rl/off-policy.d.ts.map +0 -1
  747. package/dist/rl/predictive-validity-researcher.d.ts +0 -69
  748. package/dist/rl/predictive-validity-researcher.d.ts.map +0 -1
  749. package/dist/rl/preferences.d.ts +0 -141
  750. package/dist/rl/preferences.d.ts.map +0 -1
  751. package/dist/rl/process-reward.d.ts +0 -122
  752. package/dist/rl/process-reward.d.ts.map +0 -1
  753. package/dist/rl/reward-hacking.d.ts +0 -104
  754. package/dist/rl/reward-hacking.d.ts.map +0 -1
  755. package/dist/rl/rl-campaign.d.ts +0 -85
  756. package/dist/rl/rl-campaign.d.ts.map +0 -1
  757. package/dist/rl/run-record-adapters.d.ts +0 -56
  758. package/dist/rl/run-record-adapters.d.ts.map +0 -1
  759. package/dist/rl/sim-fidelity.d.ts +0 -166
  760. package/dist/rl/sim-fidelity.d.ts.map +0 -1
  761. package/dist/rl/sim-fidelity.test.d.ts +0 -2
  762. package/dist/rl/sim-fidelity.test.d.ts.map +0 -1
  763. package/dist/rl/tournament.d.ts +0 -115
  764. package/dist/rl/tournament.d.ts.map +0 -1
  765. package/dist/rl/verifiable-reward.d.ts +0 -124
  766. package/dist/rl/verifiable-reward.d.ts.map +0 -1
  767. package/dist/run-critic.d.ts +0 -23
  768. package/dist/run-critic.d.ts.map +0 -1
  769. package/dist/run-evidence.d.ts +0 -32
  770. package/dist/run-evidence.d.ts.map +0 -1
  771. package/dist/run-record.d.ts.map +0 -1
  772. package/dist/run-record.test.d.ts +0 -2
  773. package/dist/run-record.test.d.ts.map +0 -1
  774. package/dist/run-score.d.ts +0 -31
  775. package/dist/run-score.d.ts.map +0 -1
  776. package/dist/runtime-trajectory.d.ts +0 -47
  777. package/dist/runtime-trajectory.d.ts.map +0 -1
  778. package/dist/sandbox-harness.d.ts.map +0 -1
  779. package/dist/sandbox-harness.test.d.ts +0 -2
  780. package/dist/sandbox-harness.test.d.ts.map +0 -1
  781. package/dist/sandbox-pool.d.ts +0 -74
  782. package/dist/sandbox-pool.d.ts.map +0 -1
  783. package/dist/sandbox-pool.test.d.ts +0 -2
  784. package/dist/sandbox-pool.test.d.ts.map +0 -1
  785. package/dist/scorecard.d.ts +0 -133
  786. package/dist/scorecard.d.ts.map +0 -1
  787. package/dist/scorecard.test.d.ts +0 -2
  788. package/dist/scorecard.test.d.ts.map +0 -1
  789. package/dist/self-play.d.ts +0 -69
  790. package/dist/self-play.d.ts.map +0 -1
  791. package/dist/semantic-concept-judge.d.ts +0 -135
  792. package/dist/semantic-concept-judge.d.ts.map +0 -1
  793. package/dist/semantic-concept-judge.test.d.ts +0 -2
  794. package/dist/semantic-concept-judge.test.d.ts.map +0 -1
  795. package/dist/sequential.d.ts.map +0 -1
  796. package/dist/series-convergence.d.ts.map +0 -1
  797. package/dist/slo.d.ts +0 -48
  798. package/dist/slo.d.ts.map +0 -1
  799. package/dist/state-continuity.d.ts +0 -47
  800. package/dist/state-continuity.d.ts.map +0 -1
  801. package/dist/statistics.d.ts.map +0 -1
  802. package/dist/statistics.test.d.ts +0 -2
  803. package/dist/statistics.test.d.ts.map +0 -1
  804. package/dist/steering-optimizer.d.ts +0 -58
  805. package/dist/steering-optimizer.d.ts.map +0 -1
  806. package/dist/steering.d.ts +0 -24
  807. package/dist/steering.d.ts.map +0 -1
  808. package/dist/storyboard/code-edit.d.ts +0 -64
  809. package/dist/storyboard/code-edit.d.ts.map +0 -1
  810. package/dist/storyboard/code-edit.test.d.ts +0 -2
  811. package/dist/storyboard/code-edit.test.d.ts.map +0 -1
  812. package/dist/storyboard/index.d.ts.map +0 -1
  813. package/dist/storyboard/index.test.d.ts +0 -2
  814. package/dist/storyboard/index.test.d.ts.map +0 -1
  815. package/dist/summary-report.d.ts.map +0 -1
  816. package/dist/telemetry/client.d.ts +0 -35
  817. package/dist/telemetry/client.d.ts.map +0 -1
  818. package/dist/telemetry/index.d.ts.map +0 -1
  819. package/dist/telemetry/schema.d.ts +0 -61
  820. package/dist/telemetry/schema.d.ts.map +0 -1
  821. package/dist/telemetry/sink-fetch.d.ts +0 -39
  822. package/dist/telemetry/sink-fetch.d.ts.map +0 -1
  823. package/dist/telemetry/sink-file.d.ts.map +0 -1
  824. package/dist/test-graded-scenario.d.ts +0 -42
  825. package/dist/test-graded-scenario.d.ts.map +0 -1
  826. package/dist/testing.d.ts.map +0 -1
  827. package/dist/tool-use-metrics.d.ts +0 -35
  828. package/dist/tool-use-metrics.d.ts.map +0 -1
  829. package/dist/trace/capture-fetch.d.ts +0 -48
  830. package/dist/trace/capture-fetch.d.ts.map +0 -1
  831. package/dist/trace/capture-fetch.test.d.ts +0 -2
  832. package/dist/trace/capture-fetch.test.d.ts.map +0 -1
  833. package/dist/trace/emitter.d.ts.map +0 -1
  834. package/dist/trace/extract-usage.d.ts +0 -46
  835. package/dist/trace/extract-usage.d.ts.map +0 -1
  836. package/dist/trace/extract-usage.test.d.ts +0 -2
  837. package/dist/trace/extract-usage.test.d.ts.map +0 -1
  838. package/dist/trace/index.d.ts +0 -15
  839. package/dist/trace/index.d.ts.map +0 -1
  840. package/dist/trace/integrity.d.ts.map +0 -1
  841. package/dist/trace/otel-bridge.d.ts +0 -29
  842. package/dist/trace/otel-bridge.d.ts.map +0 -1
  843. package/dist/trace/otel-export.d.ts +0 -52
  844. package/dist/trace/otel-export.d.ts.map +0 -1
  845. package/dist/trace/otel.d.ts +0 -57
  846. package/dist/trace/otel.d.ts.map +0 -1
  847. package/dist/trace/otlp-attributes.d.ts +0 -17
  848. package/dist/trace/otlp-attributes.d.ts.map +0 -1
  849. package/dist/trace/query.d.ts +0 -29
  850. package/dist/trace/query.d.ts.map +0 -1
  851. package/dist/trace/query.test.d.ts +0 -2
  852. package/dist/trace/query.test.d.ts.map +0 -1
  853. package/dist/trace/raw-provider-sink.d.ts.map +0 -1
  854. package/dist/trace/redact.d.ts.map +0 -1
  855. package/dist/trace/schema.d.ts.map +0 -1
  856. package/dist/trace/store-to-otlp.d.ts +0 -72
  857. package/dist/trace/store-to-otlp.d.ts.map +0 -1
  858. package/dist/trace/store-to-otlp.test.d.ts +0 -2
  859. package/dist/trace/store-to-otlp.test.d.ts.map +0 -1
  860. package/dist/trace/store.d.ts.map +0 -1
  861. package/dist/trace/store.test.d.ts +0 -2
  862. package/dist/trace/store.test.d.ts.map +0 -1
  863. package/dist/trace-analyst/analyst.d.ts.map +0 -1
  864. package/dist/trace-analyst/analyst.test.d.ts +0 -2
  865. package/dist/trace-analyst/analyst.test.d.ts.map +0 -1
  866. package/dist/trace-analyst/behavioral-metrics.d.ts +0 -40
  867. package/dist/trace-analyst/behavioral-metrics.d.ts.map +0 -1
  868. package/dist/trace-analyst/behavioral-metrics.test.d.ts +0 -2
  869. package/dist/trace-analyst/behavioral-metrics.test.d.ts.map +0 -1
  870. package/dist/trace-analyst/hook.d.ts +0 -55
  871. package/dist/trace-analyst/hook.d.ts.map +0 -1
  872. package/dist/trace-analyst/index.d.ts +0 -18
  873. package/dist/trace-analyst/index.d.ts.map +0 -1
  874. package/dist/trace-analyst/insights.d.ts +0 -71
  875. package/dist/trace-analyst/insights.d.ts.map +0 -1
  876. package/dist/trace-analyst/insights.test.d.ts +0 -2
  877. package/dist/trace-analyst/insights.test.d.ts.map +0 -1
  878. package/dist/trace-analyst/otlp-flatten.d.ts +0 -42
  879. package/dist/trace-analyst/otlp-flatten.d.ts.map +0 -1
  880. package/dist/trace-analyst/otlp-span.d.ts +0 -85
  881. package/dist/trace-analyst/otlp-span.d.ts.map +0 -1
  882. package/dist/trace-analyst/otlp-span.test.d.ts +0 -2
  883. package/dist/trace-analyst/otlp-span.test.d.ts.map +0 -1
  884. package/dist/trace-analyst/otlp-to-run-records.d.ts +0 -115
  885. package/dist/trace-analyst/otlp-to-run-records.d.ts.map +0 -1
  886. package/dist/trace-analyst/otlp-to-run-records.test.d.ts +0 -2
  887. package/dist/trace-analyst/otlp-to-run-records.test.d.ts.map +0 -1
  888. package/dist/trace-analyst/otlp-to-run-records.timestamps.test.d.ts +0 -2
  889. package/dist/trace-analyst/otlp-to-run-records.timestamps.test.d.ts.map +0 -1
  890. package/dist/trace-analyst/prompts.d.ts +0 -6
  891. package/dist/trace-analyst/prompts.d.ts.map +0 -1
  892. package/dist/trace-analyst/store-otlp.d.ts +0 -126
  893. package/dist/trace-analyst/store-otlp.d.ts.map +0 -1
  894. package/dist/trace-analyst/store-otlp.test.d.ts +0 -8
  895. package/dist/trace-analyst/store-otlp.test.d.ts.map +0 -1
  896. package/dist/trace-analyst/store-otlp.timestamps.test.d.ts +0 -2
  897. package/dist/trace-analyst/store-otlp.timestamps.test.d.ts.map +0 -1
  898. package/dist/trace-analyst/store.d.ts +0 -63
  899. package/dist/trace-analyst/store.d.ts.map +0 -1
  900. package/dist/trace-analyst/tools.d.ts +0 -44
  901. package/dist/trace-analyst/tools.d.ts.map +0 -1
  902. package/dist/trace-analyst/tools.test.d.ts +0 -10
  903. package/dist/trace-analyst/tools.test.d.ts.map +0 -1
  904. package/dist/trace-analyst/types.d.ts.map +0 -1
  905. package/dist/trace-contracts.d.ts +0 -180
  906. package/dist/trace-contracts.d.ts.map +0 -1
  907. package/dist/traced-analyst.d.ts +0 -26
  908. package/dist/traced-analyst.d.ts.map +0 -1
  909. package/dist/traced-judges.d.ts +0 -27
  910. package/dist/traced-judges.d.ts.map +0 -1
  911. package/dist/traces.d.ts.map +0 -1
  912. package/dist/trajectory.d.ts.map +0 -1
  913. package/dist/types.d.ts.map +0 -1
  914. package/dist/ui-finding.d.ts +0 -104
  915. package/dist/ui-finding.d.ts.map +0 -1
  916. package/dist/verdict-cache.d.ts +0 -78
  917. package/dist/verdict-cache.d.ts.map +0 -1
  918. package/dist/verdict-cache.test.d.ts +0 -2
  919. package/dist/verdict-cache.test.d.ts.map +0 -1
  920. package/dist/verdict.d.ts.map +0 -1
  921. package/dist/visual-diff.d.ts +0 -32
  922. package/dist/visual-diff.d.ts.map +0 -1
  923. package/dist/wire/handlers.d.ts +0 -54
  924. package/dist/wire/handlers.d.ts.map +0 -1
  925. package/dist/wire/index.d.ts.map +0 -1
  926. package/dist/wire/openapi.d.ts +0 -3
  927. package/dist/wire/openapi.d.ts.map +0 -1
  928. package/dist/wire/rpc.d.ts +0 -21
  929. package/dist/wire/rpc.d.ts.map +0 -1
  930. package/dist/wire/rubrics.d.ts +0 -34
  931. package/dist/wire/rubrics.d.ts.map +0 -1
  932. package/dist/wire/schemas.d.ts +0 -410
  933. package/dist/wire/schemas.d.ts.map +0 -1
  934. package/dist/wire/server.d.ts +0 -60
  935. package/dist/wire/server.d.ts.map +0 -1
  936. package/dist/workflow/event-schema.d.ts +0 -5
  937. package/dist/workflow/event-schema.d.ts.map +0 -1
  938. package/dist/workflow/feedback-pack.d.ts +0 -99
  939. package/dist/workflow/feedback-pack.d.ts.map +0 -1
  940. package/dist/workflow/index.d.ts.map +0 -1
  941. package/dist/workflow/intelligence-export.d.ts +0 -62
  942. package/dist/workflow/intelligence-export.d.ts.map +0 -1
  943. package/dist/workflow/partner-report.d.ts +0 -49
  944. package/dist/workflow/partner-report.d.ts.map +0 -1
  945. package/dist/workflow/phase-graph.d.ts +0 -43
  946. package/dist/workflow/phase-graph.d.ts.map +0 -1
  947. package/dist/workflow/promotion-gate.d.ts +0 -61
  948. package/dist/workflow/promotion-gate.d.ts.map +0 -1
  949. package/dist/workflow/run-record.d.ts +0 -12
  950. package/dist/workflow/run-record.d.ts.map +0 -1
  951. package/dist/workflow/runtime-adapter.d.ts +0 -20
  952. package/dist/workflow/runtime-adapter.d.ts.map +0 -1
  953. package/dist/workflow/sanitize.d.ts +0 -21
  954. package/dist/workflow/sanitize.d.ts.map +0 -1
  955. package/dist/workflow/schema.d.ts +0 -5
  956. package/dist/workflow/schema.d.ts.map +0 -1
  957. package/dist/workflow/summary.d.ts +0 -43
  958. package/dist/workflow/summary.d.ts.map +0 -1
  959. package/dist/workflow/trace-event-fields.d.ts +0 -6
  960. package/dist/workflow/trace-event-fields.d.ts.map +0 -1
  961. package/dist/workflow/trajectory.d.ts +0 -15
  962. package/dist/workflow/trajectory.d.ts.map +0 -1
  963. package/dist/workflow/types.d.ts +0 -68
  964. package/dist/workflow/types.d.ts.map +0 -1
  965. package/dist/workspace-inspector.d.ts +0 -67
  966. package/dist/workspace-inspector.d.ts.map +0 -1
  967. package/dist/wrangler-deploy-runner.test.d.ts +0 -2
  968. package/dist/wrangler-deploy-runner.test.d.ts.map +0 -1
@@ -1,72 +0,0 @@
1
- /**
2
- * Backend preflight: verify the models a campaign is about to spend tokens
3
- * against are actually served by the router BEFORE the run starts. The PRE-hoc
4
- * complement to `assertRealBackend` (which inspects RunRecords AFTER the run to
5
- * catch a stub/unconfigured backend).
6
- *
7
- * Two checks, increasing in cost:
8
- * - membership (free): GET `{baseUrl}/models` once; a model is `listed` when
9
- * its id is in the served set.
10
- * - probe (spends a tiny number of tokens): POST `{baseUrl}/chat/completions`
11
- * per model with a 1-message, 5-token request; `served` is whether the
12
- * router returns 2xx, with the HTTP `status` and the body's `error.message`
13
- * captured in `detail`.
14
- *
15
- * A default model the router cannot serve is a config bug. Gate a campaign on
16
- * `assertModelsServed` and it surfaces every dead id with its status + detail
17
- * instead of silently producing a stub run.
18
- */
19
- import { AgentEvalError } from '../errors';
20
- export interface ModelPreflight {
21
- /** The model id as supplied by the caller. */
22
- model: string;
23
- /** Membership in the `{baseUrl}/models` served set. */
24
- listed: boolean;
25
- /** 2xx on a 1-token chat probe. `null` when `probe` was not requested. */
26
- served: boolean | null;
27
- /** HTTP status of the probe. `null` when not probed. */
28
- status: number | null;
29
- /** Probe body's `error.message` when present, else `null`. */
30
- detail: string | null;
31
- }
32
- export interface PreflightModelsOptions {
33
- /** Router base URL, e.g. `https://router.tangle.tools/v1`. Trailing slash tolerated. */
34
- baseUrl: string;
35
- /** Bearer token sent as `Authorization: Bearer <apiKey>`. */
36
- apiKey: string;
37
- /** Model ids to check. */
38
- models: string[];
39
- /** When true, additionally spend a 1-token chat probe per model. Default false. */
40
- probe?: boolean;
41
- /** Injectable fetch for tests; defaults to the global. */
42
- fetchImpl?: typeof fetch;
43
- }
44
- export interface PreflightOutcome {
45
- succeeded: boolean;
46
- value: ModelPreflight[] | null;
47
- error: string | null;
48
- }
49
- /**
50
- * Check that `models` are reachable on the router. Returns a typed outcome —
51
- * a network failure yields `{ succeeded: false, error }`, never a throw and
52
- * never a partial result silently reported as success. No retries, no
53
- * fallbacks.
54
- *
55
- * The membership check (one GET) always runs. When `probe` is true, each model
56
- * additionally gets a 1-token chat probe so a model that is listed but
57
- * unconfigured (a 401 `model_not_found` from the router) is caught.
58
- */
59
- export declare function preflightModels(opts: PreflightModelsOptions): Promise<PreflightOutcome>;
60
- export declare class ModelsUnreachableError extends AgentEvalError {
61
- readonly results: ReadonlyArray<ModelPreflight>;
62
- constructor(message: string, results: ReadonlyArray<ModelPreflight>);
63
- }
64
- /**
65
- * Throw `ModelsUnreachableError` naming EVERY model that is unlisted or (when
66
- * probed) failed its probe — with status + detail per model. A model is dead
67
- * if it is unlisted, or if `served === false`. Callers gate a campaign on this
68
- * before spending tokens. When the network call itself fails the underlying
69
- * outcome error is rethrown — there is no partial silent pass.
70
- */
71
- export declare function assertModelsServed(opts: PreflightModelsOptions): Promise<ModelPreflight[]>;
72
- //# sourceMappingURL=preflight.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"preflight.d.ts","sourceRoot":"","sources":["../../src/integrity/preflight.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAEH,OAAO,EAAE,cAAc,EAAe,MAAM,WAAW,CAAA;AAEvD,MAAM,WAAW,cAAc;IAC7B,8CAA8C;IAC9C,KAAK,EAAE,MAAM,CAAA;IACb,uDAAuD;IACvD,MAAM,EAAE,OAAO,CAAA;IACf,0EAA0E;IAC1E,MAAM,EAAE,OAAO,GAAG,IAAI,CAAA;IACtB,wDAAwD;IACxD,MAAM,EAAE,MAAM,GAAG,IAAI,CAAA;IACrB,8DAA8D;IAC9D,MAAM,EAAE,MAAM,GAAG,IAAI,CAAA;CACtB;AAED,MAAM,WAAW,sBAAsB;IACrC,wFAAwF;IACxF,OAAO,EAAE,MAAM,CAAA;IACf,6DAA6D;IAC7D,MAAM,EAAE,MAAM,CAAA;IACd,0BAA0B;IAC1B,MAAM,EAAE,MAAM,EAAE,CAAA;IAChB,mFAAmF;IACnF,KAAK,CAAC,EAAE,OAAO,CAAA;IACf,0DAA0D;IAC1D,SAAS,CAAC,EAAE,OAAO,KAAK,CAAA;CACzB;AAED,MAAM,WAAW,gBAAgB;IAC/B,SAAS,EAAE,OAAO,CAAA;IAClB,KAAK,EAAE,cAAc,EAAE,GAAG,IAAI,CAAA;IAC9B,KAAK,EAAE,MAAM,GAAG,IAAI,CAAA;CACrB;AAwBD;;;;;;;;;GASG;AACH,wBAAsB,eAAe,CAAC,IAAI,EAAE,sBAAsB,GAAG,OAAO,CAAC,gBAAgB,CAAC,CA4D7F;AAED,qBAAa,sBAAuB,SAAQ,cAAc;aAGtC,OAAO,EAAE,aAAa,CAAC,cAAc,CAAC;gBADtD,OAAO,EAAE,MAAM,EACC,OAAO,EAAE,aAAa,CAAC,cAAc,CAAC;CAKzD;AAYD;;;;;;GAMG;AACH,wBAAsB,kBAAkB,CAAC,IAAI,EAAE,sBAAsB,GAAG,OAAO,CAAC,cAAc,EAAE,CAAC,CAiBhG"}
@@ -1,2 +0,0 @@
1
- export {};
2
- //# sourceMappingURL=preflight.test.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"preflight.test.d.ts","sourceRoot":"","sources":["../../src/integrity/preflight.test.ts"],"names":[],"mappings":""}
@@ -1,67 +0,0 @@
1
- /**
2
- * Single-backend guard: assert the agent and the rubric judge run through the
3
- * SAME backend config, so the judge can't silently re-route through a
4
- * different (often paid) backend than the agent.
5
- *
6
- * The bug class: `--backend cli-bridge` rewires the agent, but the judge still
7
- * reads `process.env.TANGLE_API_KEY` → router. Cost is billed against the
8
- * router, the eval reports the cli-bridge model, and the data is unusable.
9
- * Four consumers hand-roll this comparison (legal at `canonical.ts:702-795`);
10
- * this is the one substrate copy.
11
- *
12
- * Complements `assertRealBackend` (records → stub vs real) and
13
- * `assertCrossFamily` (judge-ensemble family diversity): this one compares two
14
- * backend *configs* before the run.
15
- */
16
- import { AgentEvalError } from '../errors';
17
- /**
18
- * Minimal backend-config shape the assertion reads. Consumers may pass richer
19
- * types — only these five fields are inspected.
20
- */
21
- export interface BackendDescriptor {
22
- /** Backend route — e.g. `'tcloud' | 'cli-bridge' | 'sandbox' | 'direct-provider'`;
23
- * free-form for consumer extensibility. */
24
- kind: string;
25
- /** Resolved base URL. Compared lexically (trailing slash stripped). */
26
- baseUrl: string;
27
- /** Model id (with snapshot suffix). Compared lexically. */
28
- model: string;
29
- /** Optional provider override. Compared when both set; flagged when only
30
- * one side sets it. */
31
- provider?: string;
32
- /** Bearer token. Values are NEVER compared (security) — only that EITHER
33
- * both are set OR both are empty. Mismatched presence is a divergence. */
34
- apiKey?: string;
35
- }
36
- export interface AssertSingleBackendOptions {
37
- /** When true, ANY field divergence fails. When false (default), only
38
- * `kind` / `baseUrl` / `provider` / `apiKeyPresence` divergence throws —
39
- * a different judge `model` on the same route is allowed (the legal
40
- * pattern: a cheaper judge model). */
41
- strict?: boolean;
42
- agentLabel?: string;
43
- judgeLabel?: string;
44
- }
45
- export type SingleBackendField = 'kind' | 'baseUrl' | 'model' | 'provider' | 'apiKeyPresence';
46
- export interface SingleBackendDivergence {
47
- field: SingleBackendField;
48
- agent: string | undefined;
49
- judge: string | undefined;
50
- }
51
- export interface SingleBackendReport {
52
- /** True when agent + judge agree per the configured strictness. */
53
- ok: boolean;
54
- /** Every divergence detected (includes `model` even when non-blocking). */
55
- divergences: ReadonlyArray<SingleBackendDivergence>;
56
- }
57
- export declare class SingleBackendError extends AgentEvalError {
58
- readonly report: SingleBackendReport;
59
- constructor(message: string, report: SingleBackendReport);
60
- }
61
- /**
62
- * Throw `SingleBackendError` when the agent and judge backends diverge in a
63
- * way that would re-route the judge through a different backend than the
64
- * agent. Returns the report so callers can log it in either case.
65
- */
66
- export declare function assertSingleBackend(agent: BackendDescriptor, judge: BackendDescriptor, opts?: AssertSingleBackendOptions): SingleBackendReport;
67
- //# sourceMappingURL=single-backend.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"single-backend.d.ts","sourceRoot":"","sources":["../../src/integrity/single-backend.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,OAAO,EAAE,cAAc,EAAE,MAAM,WAAW,CAAA;AAE1C;;;GAGG;AACH,MAAM,WAAW,iBAAiB;IAChC;gDAC4C;IAC5C,IAAI,EAAE,MAAM,CAAA;IACZ,uEAAuE;IACvE,OAAO,EAAE,MAAM,CAAA;IACf,2DAA2D;IAC3D,KAAK,EAAE,MAAM,CAAA;IACb;4BACwB;IACxB,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB;+EAC2E;IAC3E,MAAM,CAAC,EAAE,MAAM,CAAA;CAChB;AAED,MAAM,WAAW,0BAA0B;IACzC;;;2CAGuC;IACvC,MAAM,CAAC,EAAE,OAAO,CAAA;IAChB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,UAAU,CAAC,EAAE,MAAM,CAAA;CACpB;AAED,MAAM,MAAM,kBAAkB,GAAG,MAAM,GAAG,SAAS,GAAG,OAAO,GAAG,UAAU,GAAG,gBAAgB,CAAA;AAE7F,MAAM,WAAW,uBAAuB;IACtC,KAAK,EAAE,kBAAkB,CAAA;IACzB,KAAK,EAAE,MAAM,GAAG,SAAS,CAAA;IACzB,KAAK,EAAE,MAAM,GAAG,SAAS,CAAA;CAC1B;AAED,MAAM,WAAW,mBAAmB;IAClC,mEAAmE;IACnE,EAAE,EAAE,OAAO,CAAA;IACX,2EAA2E;IAC3E,WAAW,EAAE,aAAa,CAAC,uBAAuB,CAAC,CAAA;CACpD;AAED,qBAAa,kBAAmB,SAAQ,cAAc;aAGlC,MAAM,EAAE,mBAAmB;gBAD3C,OAAO,EAAE,MAAM,EACC,MAAM,EAAE,mBAAmB;CAK9C;AAMD;;;;GAIG;AACH,wBAAgB,mBAAmB,CACjC,KAAK,EAAE,iBAAiB,EACxB,KAAK,EAAE,iBAAiB,EACxB,IAAI,GAAE,0BAA+B,GACpC,mBAAmB,CA6CrB"}
@@ -1,69 +0,0 @@
1
- /**
2
- * Intent-match judge — "did the agent build the right APP, ignoring
3
- * whether every feature is wired up?"
4
- *
5
- * Distinct from {@link runSemanticConceptJudge} which scores per-concept
6
- * presence. The semantic judge can return 0/4 concepts present even
7
- * when the agent built a thoughtful, polished, on-brief app that just
8
- * lacks one or two features. The semantic judge can also return 4/4
9
- * present even when the agent shipped the wrong project (keyword-rich
10
- * stub).
11
- *
12
- * Intent-match asks ONE question:
13
- * "Looking at the agent's work as a whole — independent of feature
14
- * coverage — is this an honest attempt at the user's request?"
15
- *
16
- * Returns a 0–1 score and a 1-sentence evidence string. Use as a sanity
17
- * check on `completenessScore`-style metrics: if intent-match is high
18
- * and concept count is low, the agent built the right thing but is
19
- * missing features (ship and iterate). If intent-match is low, the
20
- * agent built the wrong thing (reject regardless of concept count).
21
- *
22
- * Soft-fails on LLM/JSON error (`available: false`) so callers can
23
- * treat failure as "judge skipped."
24
- */
25
- import { type LlmClientOptions } from './llm-client';
26
- export declare const INTENT_MATCH_JUDGE_VERSION = "intent-match-judge-v1-2026-04-24";
27
- export interface IntentMatchInput {
28
- /** The full natural-language prompt the agent was handed. */
29
- userRequest: string;
30
- /** Top-level source files from the agent's workdir. */
31
- sourceFiles: Array<{
32
- path: string;
33
- content: string;
34
- }>;
35
- /** Rendered HTML the preview returned, when available. */
36
- servedHtml?: string;
37
- /** Optional metadata to inject (id, vertical, difficulty). */
38
- artifactLabel?: string;
39
- artifactDescription?: string;
40
- }
41
- export interface IntentMatchResult {
42
- kind: 'intent-match';
43
- version: string;
44
- /** 0..1 — 1 = unmistakably the right app, 0 = unrelated to the brief. */
45
- score: number;
46
- /** One-sentence rationale citing concrete evidence (file or HTML). */
47
- evidence: string;
48
- durationMs: number;
49
- costUsd: number | null;
50
- available: boolean;
51
- error?: string;
52
- }
53
- export interface IntentMatchOptions {
54
- model?: string;
55
- timeoutMs?: number;
56
- maxSourceChars?: number;
57
- maxPerFileChars?: number;
58
- maxHtmlChars?: number;
59
- llm?: LlmClientOptions;
60
- }
61
- /**
62
- * Run the intent-match judge. Soft-fails to available=false on error.
63
- */
64
- export declare function runIntentMatchJudge(input: IntentMatchInput, options?: IntentMatchOptions): Promise<IntentMatchResult>;
65
- /**
66
- * Factory: pin LLM options once, return a closure.
67
- */
68
- export declare function createIntentMatchJudge(options?: IntentMatchOptions): (input: IntentMatchInput) => Promise<IntentMatchResult>;
69
- //# sourceMappingURL=intent-match-judge.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"intent-match-judge.d.ts","sourceRoot":"","sources":["../src/intent-match-judge.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AAEH,OAAO,EAAe,KAAK,gBAAgB,EAAE,MAAM,cAAc,CAAA;AAEjE,eAAO,MAAM,0BAA0B,qCAAqC,CAAA;AAE5E,MAAM,WAAW,gBAAgB;IAC/B,6DAA6D;IAC7D,WAAW,EAAE,MAAM,CAAA;IACnB,uDAAuD;IACvD,WAAW,EAAE,KAAK,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;IACrD,0DAA0D;IAC1D,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,8DAA8D;IAC9D,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,mBAAmB,CAAC,EAAE,MAAM,CAAA;CAC7B;AAED,MAAM,WAAW,iBAAiB;IAChC,IAAI,EAAE,cAAc,CAAA;IACpB,OAAO,EAAE,MAAM,CAAA;IACf,yEAAyE;IACzE,KAAK,EAAE,MAAM,CAAA;IACb,sEAAsE;IACtE,QAAQ,EAAE,MAAM,CAAA;IAChB,UAAU,EAAE,MAAM,CAAA;IAClB,OAAO,EAAE,MAAM,GAAG,IAAI,CAAA;IACtB,SAAS,EAAE,OAAO,CAAA;IAClB,KAAK,CAAC,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,kBAAkB;IACjC,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,cAAc,CAAC,EAAE,MAAM,CAAA;IACvB,eAAe,CAAC,EAAE,MAAM,CAAA;IACxB,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,GAAG,CAAC,EAAE,gBAAgB,CAAA;CACvB;AAmED;;GAEG;AACH,wBAAsB,mBAAmB,CACvC,KAAK,EAAE,gBAAgB,EACvB,OAAO,GAAE,kBAAuB,GAC/B,OAAO,CAAC,iBAAiB,CAAC,CAiE5B;AAED;;GAEG;AACH,wBAAgB,sBAAsB,CACpC,OAAO,GAAE,kBAAuB,GAC/B,CAAC,KAAK,EAAE,gBAAgB,KAAK,OAAO,CAAC,iBAAiB,CAAC,CAEzD"}
@@ -1,2 +0,0 @@
1
- export {};
2
- //# sourceMappingURL=intent-match-judge.test.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"intent-match-judge.test.d.ts","sourceRoot":"","sources":["../src/intent-match-judge.test.ts"],"names":[],"mappings":""}
@@ -1 +0,0 @@
1
- {"version":3,"file":"judge-calibration.d.ts","sourceRoot":"","sources":["../src/judge-calibration.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH,MAAM,WAAW,UAAU;IACzB,MAAM,EAAE,MAAM,CAAA;IACd,UAAU,EAAE,MAAM,CAAA;IAClB,mFAAmF;IACnF,KAAK,CAAC,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,cAAc;IAC7B,MAAM,EAAE,MAAM,CAAA;IACd,KAAK,EAAE,MAAM,CAAA;IACb,uEAAuE;IACvE,gBAAgB,CAAC,EAAE,OAAO,GAAG,QAAQ,CAAA;CACtC;AAED,MAAM,WAAW,iBAAiB;IAChC,CAAC,EAAE,MAAM,CAAA;IACT,OAAO,EAAE,MAAM,CAAA;IACf,oEAAoE;IACpE,KAAK,EAAE,MAAM,CAAA;IACb,oCAAoC;IACpC,GAAG,EAAE,MAAM,CAAA;IACX,yDAAyD;IACzD,UAAU,EAAE,KAAK,CAAC;QAAE,MAAM,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;CACnF;AAED,wBAAgB,cAAc,CAC5B,MAAM,EAAE,UAAU,EAAE,EACpB,SAAS,EAAE,cAAc,EAAE,GAC1B,iBAAiB,CAwBnB;AAED,MAAM,WAAW,oBAAoB;IACnC;;;OAGG;IACH,QAAQ,EAAE,MAAM,CAAA;IAChB,CAAC,EAAE,MAAM,CAAA;CACV;AAED;;;GAGG;AACH,wBAAgB,cAAc,CAAC,MAAM,EAAE,cAAc,EAAE,GAAG,oBAAoB,CAc7E;AAED,MAAM,WAAW,mBAAmB;IAClC,6FAA6F;IAC7F,OAAO,EAAE,MAAM,CAAA;IACf,CAAC,EAAE,MAAM,CAAA;CACV;AAED,wBAAgB,aAAa,CAC3B,OAAO,EAAE,KAAK,CAAC;IAAE,SAAS,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,CAAC,GACnD,mBAAmB,CAUrB;AAED,MAAM,WAAW,oBAAoB;IACnC,oEAAoE;IACpE,YAAY,EAAE,MAAM,CAAA;IACpB,eAAe,EAAE,MAAM,CAAA;IACvB,SAAS,EAAE,MAAM,CAAA;IACjB,CAAC,EAAE,MAAM,CAAA;CACV;AAED;;;;GAIG;AACH,wBAAgB,cAAc,CAC5B,OAAO,EAAE,KAAK,CAAC;IAAE,KAAK,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,OAAO,CAAA;CAAE,CAAC,GACnD,oBAAoB,CAatB;AAmFD,MAAM,WAAW,mBAAmB;IAClC,wEAAwE;IACxE,aAAa,EAAE,MAAM,CAAA;IACrB,0EAA0E;IAC1E,GAAG,EAAE,MAAM,CAAA;IACX,6EAA6E;IAC7E,OAAO,EAAE,MAAM,CAAA;IACf,oEAAoE;IACpE,QAAQ,EAAE,MAAM,CAAA;IAChB,+CAA+C;IAC/C,EAAE,EAAE;QACF,GAAG,EAAE,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;QACrB,aAAa,EAAE,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;KAChC,CAAA;IACD,uDAAuD;IACvD,CAAC,EAAE,MAAM,CAAA;IACT,wBAAwB;IACxB,MAAM,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,0BAA0B;IACzC,kFAAkF;IAClF,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,+CAA+C;IAC/C,OAAO,CAAC,EAAE,QAAQ,GAAG,WAAW,CAAA;IAChC,8DAA8D;IAC9D,IAAI,CAAC,EAAE,MAAM,CAAA;IACb,wDAAwD;IACxD,OAAO,CAAC,EAAE,MAAM,CAAA;CACjB;AAED;;;;;;GAMG;AACH,wBAAgB,mBAAmB,CACjC,MAAM,EAAE,MAAM,EAAE,EAAE,EAClB,IAAI,GAAE,0BAA+B,GACpC,mBAAmB,CAmErB;AAED,MAAM,WAAW,2BAA4B,SAAQ,iBAAiB;IACpE,uDAAuD;IACvD,uBAAuB,EAAE,MAAM,CAAA;IAC/B,0DAA0D;IAC1D,GAAG,EAAE,MAAM,CAAA;IACX,QAAQ,EAAE,MAAM,CAAA;IAChB,EAAE,EAAE;QACF,GAAG,EAAE,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;QACrB,aAAa,EAAE,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;KAChC,CAAA;CACF;AAED;;;;GAIG;AACH,wBAAgB,wBAAwB,CACtC,MAAM,EAAE,UAAU,EAAE,EACpB,SAAS,EAAE,cAAc,EAAE,EAC3B,IAAI,GAAE,0BAA+B,GACpC,2BAA2B,CAoB7B"}
@@ -1,66 +0,0 @@
1
- /**
2
- * Judge-ensemble reducer — folds N independent judge verdicts on the same
3
- * artifact into one aggregate score.
4
- *
5
- * The pattern every multi-model judge re-implements: run K uncorrelated judges
6
- * (different model families remove within-family bias), then reduce their
7
- * per-dimension verdicts to a single composite + an inter-rater disagreement
8
- * signal. This is the pure reduction — no LLM, no I/O — so a lift it produces
9
- * is attributable to the scores, not to where they came from.
10
- *
11
- * Fail-loud: a judge that errored or returned malformed output is recorded in
12
- * `failedJudges` with `perDimension: null`, never folded into a zero. If EVERY
13
- * judge failed the reducer throws — a silent zero here would corrupt the gate's
14
- * number. A failed judge still burned tokens, so its `costUsd` is still summed.
15
- */
16
- /** One judge's verdict. `perDimension: null` ⇒ that judge call failed (threw or
17
- * returned malformed output) — recorded, never folded into a zero. */
18
- export interface JudgeVerdict<D extends string = string> {
19
- /** Judge identity (typically the model id) — the `perJudge` / `failedJudges` key. */
20
- model: string;
21
- /** Per-dimension scores, or `null` when the judge failed. */
22
- perDimension: Record<D, number> | null;
23
- /** Optional one-line rationale; the first non-empty one becomes the aggregate's. */
24
- rationale?: string;
25
- /** Optional reported cost — summed across ALL verdicts (failed included). */
26
- costUsd?: number;
27
- /** Optional per-dimension reasoning/evidence. Carried through to
28
- * `EnsembleAggregate.verdicts` verbatim — never folded into the math. */
29
- detail?: Partial<Record<D, {
30
- reasoning?: string;
31
- evidence?: string;
32
- }>>;
33
- }
34
- /** The aggregated ensemble result. */
35
- export interface EnsembleAggregate<D extends string = string> {
36
- /** Mean over SURVIVING judges, per dimension (absent-everywhere ⇒ 0). */
37
- perDimension: Record<D, number>;
38
- /** Weighted mean of `perDimension` (uniform unless `weights` given). */
39
- composite: number;
40
- /** Each surviving judge's clamped per-dimension scores, for trace drill-down. */
41
- perJudge: Record<string, Record<D, number>>;
42
- /** Max over dimensions of (max − min) across survivors — the inter-rater signal. */
43
- maxDisagreement: number;
44
- /** Models whose verdict was null (failed). */
45
- failedJudges: string[];
46
- /** Sum of `costUsd` over ALL verdicts, failed included. */
47
- costUsd: number;
48
- /** First non-empty survivor rationale, or `'llm-judge'`. */
49
- rationale: string;
50
- /** The input verdicts, verbatim — drill-down to raw scores, `detail`
51
- * reasoning/evidence, and per-verdict cost without re-running judges. */
52
- verdicts: JudgeVerdict<D>[];
53
- }
54
- /**
55
- * Reduce per-judge verdicts to one aggregate. Generic over the rubric: pass the
56
- * stable-ordered `dimensionKeys` the judges scored.
57
- *
58
- * - Per-dimension score = mean over surviving judges (out-of-range clamped to [0,1]).
59
- * - `composite` = `weightedComposite` of `perDimension`. No `weights` ⇒ uniform
60
- * over every dimension. A partial `weights` map selects AND weights exactly the
61
- * named dimensions (others excluded from the composite) — the substrate's
62
- * sum-normalized weighting, not re-implemented here.
63
- * - Throws if `verdicts` or `dimensionKeys` is empty, or if every judge failed.
64
- */
65
- export declare function aggregateJudgeVerdicts<D extends string>(verdicts: readonly JudgeVerdict<D>[], dimensionKeys: readonly D[], weights?: Partial<Record<D, number>>): EnsembleAggregate<D>;
66
- //# sourceMappingURL=judge-ensemble.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"judge-ensemble.d.ts","sourceRoot":"","sources":["../src/judge-ensemble.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAKH;uEACuE;AACvE,MAAM,WAAW,YAAY,CAAC,CAAC,SAAS,MAAM,GAAG,MAAM;IACrD,qFAAqF;IACrF,KAAK,EAAE,MAAM,CAAA;IACb,6DAA6D;IAC7D,YAAY,EAAE,MAAM,CAAC,CAAC,EAAE,MAAM,CAAC,GAAG,IAAI,CAAA;IACtC,oFAAoF;IACpF,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,6EAA6E;IAC7E,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB;8EAC0E;IAC1E,MAAM,CAAC,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC,EAAE;QAAE,SAAS,CAAC,EAAE,MAAM,CAAC;QAAC,QAAQ,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC,CAAA;CACvE;AAED,sCAAsC;AACtC,MAAM,WAAW,iBAAiB,CAAC,CAAC,SAAS,MAAM,GAAG,MAAM;IAC1D,yEAAyE;IACzE,YAAY,EAAE,MAAM,CAAC,CAAC,EAAE,MAAM,CAAC,CAAA;IAC/B,wEAAwE;IACxE,SAAS,EAAE,MAAM,CAAA;IACjB,iFAAiF;IACjF,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC,CAAA;IAC3C,oFAAoF;IACpF,eAAe,EAAE,MAAM,CAAA;IACvB,8CAA8C;IAC9C,YAAY,EAAE,MAAM,EAAE,CAAA;IACtB,2DAA2D;IAC3D,OAAO,EAAE,MAAM,CAAA;IACf,4DAA4D;IAC5D,SAAS,EAAE,MAAM,CAAA;IACjB;8EAC0E;IAC1E,QAAQ,EAAE,YAAY,CAAC,CAAC,CAAC,EAAE,CAAA;CAC5B;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,sBAAsB,CAAC,CAAC,SAAS,MAAM,EACrD,QAAQ,EAAE,SAAS,YAAY,CAAC,CAAC,CAAC,EAAE,EACpC,aAAa,EAAE,SAAS,CAAC,EAAE,EAC3B,OAAO,CAAC,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC,GACnC,iBAAiB,CAAC,CAAC,CAAC,CA6EtB"}
@@ -1,8 +0,0 @@
1
- /**
2
- * The ensemble reducer's job is to fold N judge verdicts into one trustworthy
3
- * number. These tests pin the invariants a silent zero would break: a failed
4
- * judge never counts as zero, all-failed throws, cost includes failures, and
5
- * the composite/disagreement signals match the per-dimension survivors.
6
- */
7
- export {};
8
- //# sourceMappingURL=judge-ensemble.test.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"judge-ensemble.test.d.ts","sourceRoot":"","sources":["../src/judge-ensemble.test.ts"],"names":[],"mappings":"AAAA;;;;;GAKG"}
@@ -1,38 +0,0 @@
1
- /**
2
- * Judge model-family classification + cross-family enforcement.
3
- *
4
- * A judge ensemble built entirely from one provider family shares that
5
- * family's blind spots and self-preference — its "agreement" is correlated
6
- * bias, not independent signal. `assertCrossFamily` makes the consumer prove
7
- * the ensemble spans ≥2 families; `judgeFamily` is the single regex map that
8
- * replaces the per-consumer copies (tax/legal/creative/gtm each ship one).
9
- */
10
- /** Provider family a model belongs to. `unknown` when no rule matches. */
11
- export type JudgeFamily = 'anthropic' | 'openai' | 'google' | 'meta' | 'mistral' | 'deepseek' | 'xai' | 'qwen' | 'cohere' | 'amazon' | 'moonshot' | 'zhipu' | 'unknown';
12
- /**
13
- * Classify a model id into its provider family. Strips a `@snapshot` suffix
14
- * and prefers an explicit `provider/...` prefix; otherwise matches the model
15
- * name. Returns `unknown` when nothing matches (callers decide whether that's
16
- * acceptable — `assertCrossFamily` counts it as its own family).
17
- */
18
- export declare function judgeFamily(modelId: string): JudgeFamily;
19
- export interface AssertCrossFamilyOptions {
20
- /** Minimum number of distinct families the ensemble must span. Default 2. */
21
- minFamilies?: number;
22
- /** When false (default), `unknown`-family models do NOT count toward the
23
- * family total — an ensemble of all-unclassifiable models is not provably
24
- * cross-family. Set true to count `unknown` as one shared family. */
25
- allowUnknown?: boolean;
26
- }
27
- export declare class CrossFamilyError extends Error {
28
- readonly families: JudgeFamily[];
29
- readonly models: string[];
30
- constructor(message: string, families: JudgeFamily[], models: string[]);
31
- }
32
- /**
33
- * Throw unless the judge models span at least `minFamilies` distinct provider
34
- * families. Pass the model ids backing your judge ensemble. Fail-loud by
35
- * design — a correlated single-family ensemble silently inflates agreement.
36
- */
37
- export declare function assertCrossFamily(models: string[], opts?: AssertCrossFamilyOptions): JudgeFamily[];
38
- //# sourceMappingURL=judge-families.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"judge-families.d.ts","sourceRoot":"","sources":["../src/judge-families.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,0EAA0E;AAC1E,MAAM,MAAM,WAAW,GACnB,WAAW,GACX,QAAQ,GACR,QAAQ,GACR,MAAM,GACN,SAAS,GACT,UAAU,GACV,KAAK,GACL,MAAM,GACN,QAAQ,GACR,QAAQ,GACR,UAAU,GACV,OAAO,GACP,SAAS,CAAA;AA+Cb;;;;;GAKG;AACH,wBAAgB,WAAW,CAAC,OAAO,EAAE,MAAM,GAAG,WAAW,CAYxD;AAED,MAAM,WAAW,wBAAwB;IACvC,6EAA6E;IAC7E,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB;;0EAEsE;IACtE,YAAY,CAAC,EAAE,OAAO,CAAA;CACvB;AAED,qBAAa,gBAAiB,SAAQ,KAAK;aAGvB,QAAQ,EAAE,WAAW,EAAE;aACvB,MAAM,EAAE,MAAM,EAAE;gBAFhC,OAAO,EAAE,MAAM,EACC,QAAQ,EAAE,WAAW,EAAE,EACvB,MAAM,EAAE,MAAM,EAAE;CAKnC;AAED;;;;GAIG;AACH,wBAAgB,iBAAiB,CAC/B,MAAM,EAAE,MAAM,EAAE,EAChB,IAAI,GAAE,wBAA6B,GAClC,WAAW,EAAE,CAmBf"}
@@ -1,65 +0,0 @@
1
- /**
2
- * Multi-model judge panel — `ensembleJudge` builds a campaign `JudgeConfig`
3
- * that fans one artifact out to K judge models and reduces their verdicts
4
- * through `aggregateJudgeVerdicts` (src/judge-ensemble.ts).
5
- *
6
- * The panel is the fail-loud composition of the substrate's existing judge
7
- * primitives:
8
- * - `assertCrossFamily` (construction-time) — a single-family panel is
9
- * correlated bias, not independent signal.
10
- * - `withJudgeRetry` (per model, opt-in) — transient-fault retry with a
11
- * typed outcome; a judge that exhausts retries is recorded as failed,
12
- * never folded into a zero.
13
- * - `aggregateJudgeVerdicts` — the pure reducer; throws when EVERY judge
14
- * failed so a silent zero can't reach the gate.
15
- *
16
- * The returned `JudgeScore` is on the campaign [0,1] scale and carries the
17
- * ensemble extras (`maxDisagreement`, `failedJudges`, `perJudge`) declared
18
- * on the canonical `JudgeScore` in src/campaign/types.ts.
19
- */
20
- import type { JudgeConfig } from './campaign/types';
21
- import { type JudgeVerdict } from './judge-ensemble';
22
- import { type JudgeRetryPolicy } from './judge-retry';
23
- export interface EnsembleJudgeOptions<D extends string> {
24
- /** Judge name — becomes the returned `JudgeConfig.name`. */
25
- name: string;
26
- /** Rubric dimensions every model scores. Keys of the verdict's `perDimension`. */
27
- dimensions: D[];
28
- /** Judge model ids — one `scoreWith` call per entry. List a model twice to
29
- * sample it twice (votes are suffix-keyed `model#2` so none overwrite). */
30
- models: string[];
31
- /**
32
- * Score the artifact with one model. Throw (or reject) on failure — the
33
- * panel records that model as a failed judge; it is never folded into a
34
- * zero. Verdict scores are clamped to [0,1] by the reducer.
35
- */
36
- scoreWith: (model: string, input: {
37
- artifact: unknown;
38
- scenario?: unknown;
39
- }) => Promise<JudgeVerdict<D>>;
40
- /**
41
- * Per-model retry policy, applied via `withJudgeRetry`. The panel's
42
- * `models` list drives the fan-out, so `retry.models` (the fallback
43
- * rotation) is overridden to each panel model in turn.
44
- */
45
- retry?: JudgeRetryPolicy;
46
- /** Enforce `assertCrossFamily` over `models` at construction. Default true.
47
- * Opt out only for deliberate single-family panels (e.g. self-consistency
48
- * sampling of one model). */
49
- crossFamily?: boolean;
50
- /** Composite weights forwarded to `aggregateJudgeVerdicts`: a partial map
51
- * selects AND weights exactly the named dimensions. Omit for uniform. */
52
- weights?: Partial<Record<D, number>>;
53
- }
54
- /**
55
- * Build a campaign-shaped `JudgeConfig` whose `score()` runs every panel
56
- * model in parallel and reduces the surviving verdicts to one canonical
57
- * `JudgeScore` in [0,1].
58
- *
59
- * Failure semantics: a model whose `scoreWith` throws (or exhausts `retry`)
60
- * lands in `failedJudges` and is excluded from the means. When EVERY model
61
- * fails, `aggregateJudgeVerdicts` throws — the campaign engine records a
62
- * failed cell instead of averaging a fabricated zero.
63
- */
64
- export declare function ensembleJudge<D extends string>(opts: EnsembleJudgeOptions<D>): JudgeConfig<unknown>;
65
- //# sourceMappingURL=judge-panel.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"judge-panel.d.ts","sourceRoot":"","sources":["../src/judge-panel.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAEH,OAAO,KAAK,EAAE,WAAW,EAAc,MAAM,kBAAkB,CAAA;AAC/D,OAAO,EAA0B,KAAK,YAAY,EAAE,MAAM,kBAAkB,CAAA;AAE5E,OAAO,EAAE,KAAK,gBAAgB,EAAkB,MAAM,eAAe,CAAA;AAErE,MAAM,WAAW,oBAAoB,CAAC,CAAC,SAAS,MAAM;IACpD,4DAA4D;IAC5D,IAAI,EAAE,MAAM,CAAA;IACZ,kFAAkF;IAClF,UAAU,EAAE,CAAC,EAAE,CAAA;IACf;gFAC4E;IAC5E,MAAM,EAAE,MAAM,EAAE,CAAA;IAChB;;;;OAIG;IACH,SAAS,EAAE,CACT,KAAK,EAAE,MAAM,EACb,KAAK,EAAE;QAAE,QAAQ,EAAE,OAAO,CAAC;QAAC,QAAQ,CAAC,EAAE,OAAO,CAAA;KAAE,KAC7C,OAAO,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,CAAA;IAC7B;;;;OAIG;IACH,KAAK,CAAC,EAAE,gBAAgB,CAAA;IACxB;;kCAE8B;IAC9B,WAAW,CAAC,EAAE,OAAO,CAAA;IACrB;8EAC0E;IAC1E,OAAO,CAAC,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC,CAAA;CACrC;AAED;;;;;;;;;GASG;AACH,wBAAgB,aAAa,CAAC,CAAC,SAAS,MAAM,EAC5C,IAAI,EAAE,oBAAoB,CAAC,CAAC,CAAC,GAC5B,WAAW,CAAC,OAAO,CAAC,CA2DtB"}
@@ -1,70 +0,0 @@
1
- /**
2
- * Wrap a single judge LLM call with retry, optional fallback-model
3
- * rotation, exponential backoff, and a typed `JudgeRetryOutcome`. Callers
4
- * MUST inspect `succeeded` before using `value`; on failure the library
5
- * returns `value: null` rather than substituting a default, so a judge
6
- * abort cannot silently corrupt a downstream composite.
7
- *
8
- * Reporting contract: callers ship `TrialResult.judgeSucceeded = succeeded`
9
- * and `TrialResult.judgeAttempts = attempts` so `aggregateTrialsByMode`
10
- * with `mode: 'exclude-failed'` drops the trial.
11
- */
12
- /** Retry policy for judge LLM calls. */
13
- export interface JudgeRetryPolicy {
14
- /** Max attempts per model. Default 3 (one initial + two retries). */
15
- maxAttempts?: number;
16
- /** Per-attempt timeout in ms. Default 300_000. */
17
- timeoutMs?: number;
18
- /**
19
- * Models to try, in order. The first model is the primary; subsequent
20
- * models are fallbacks invoked only when ALL retries on the previous
21
- * model have been exhausted. Example: `['claude-code/sonnet', 'kimi-code/k2p6']`
22
- * runs claude-code up to maxAttempts times, then falls back to kimi.
23
- * If omitted, the caller's judge function controls model selection and
24
- * the retries apply to that single model.
25
- */
26
- models?: readonly string[];
27
- /** Exponential backoff function, default `attempt → min(500 * 2^attempt, 16_000)`. */
28
- backoffMs?: (attempt: number) => number;
29
- /**
30
- * Predicate deciding whether an error should trigger a retry. Defaults to
31
- * `isTransientLlmError` — the package-wide classifier shared with
32
- * `callLlm` — which retries aborts/timeouts, network faults, HTTP/2
33
- * transport faults, and any `LlmCallError` with status in {429,502,503,504}.
34
- * JSON-parse and schema-rejection errors are NOT retriable (the model
35
- * needs prompt adjustment, not another shot).
36
- */
37
- isRetryable?: (err: unknown) => boolean;
38
- }
39
- /** Outcome of a wrapped judge invocation. */
40
- export interface JudgeRetryOutcome<T> {
41
- /** The judge's returned value when `succeeded === true`. */
42
- value: T | null;
43
- /** True iff one of the attempts completed without throwing. */
44
- succeeded: boolean;
45
- /** Total attempts made across all models. */
46
- attempts: number;
47
- /** Which model the successful attempt used (when succeeded). */
48
- modelUsed?: string;
49
- /** Last error captured when `succeeded === false`. */
50
- error?: Error;
51
- /** Per-attempt error log for forensics. */
52
- attemptErrors: Array<{
53
- attempt: number;
54
- model: string;
55
- error: string;
56
- }>;
57
- }
58
- /**
59
- * Wrap a judge call with retry + fallback-model + typed outcome semantics.
60
- *
61
- * The `judgeFn` signature is `(model: string, signal: AbortSignal) => Promise<T>`.
62
- * The signal will be aborted at `timeoutMs`. Callers should pass the signal
63
- * to their underlying fetch/SDK call so the abort actually fires.
64
- *
65
- * Returns a typed outcome — callers MUST inspect `succeeded` before using
66
- * `value`. The library refuses to default to a silent zero score because a
67
- * synthetic zero is indistinguishable from a real low score downstream.
68
- */
69
- export declare function withJudgeRetry<T>(judgeFn: (model: string, signal: AbortSignal) => Promise<T>, policy?: JudgeRetryPolicy): Promise<JudgeRetryOutcome<T>>;
70
- //# sourceMappingURL=judge-retry.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"judge-retry.d.ts","sourceRoot":"","sources":["../src/judge-retry.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAIH,wCAAwC;AACxC,MAAM,WAAW,gBAAgB;IAC/B,qEAAqE;IACrE,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,kDAAkD;IAClD,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB;;;;;;;OAOG;IACH,MAAM,CAAC,EAAE,SAAS,MAAM,EAAE,CAAA;IAC1B,sFAAsF;IACtF,SAAS,CAAC,EAAE,CAAC,OAAO,EAAE,MAAM,KAAK,MAAM,CAAA;IACvC;;;;;;;OAOG;IACH,WAAW,CAAC,EAAE,CAAC,GAAG,EAAE,OAAO,KAAK,OAAO,CAAA;CACxC;AAED,6CAA6C;AAC7C,MAAM,WAAW,iBAAiB,CAAC,CAAC;IAClC,4DAA4D;IAC5D,KAAK,EAAE,CAAC,GAAG,IAAI,CAAA;IACf,+DAA+D;IAC/D,SAAS,EAAE,OAAO,CAAA;IAClB,6CAA6C;IAC7C,QAAQ,EAAE,MAAM,CAAA;IAChB,gEAAgE;IAChE,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,sDAAsD;IACtD,KAAK,CAAC,EAAE,KAAK,CAAA;IACb,2CAA2C;IAC3C,aAAa,EAAE,KAAK,CAAC;QAAE,OAAO,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;CACxE;AASD;;;;;;;;;;GAUG;AACH,wBAAsB,cAAc,CAAC,CAAC,EACpC,OAAO,EAAE,CAAC,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,WAAW,KAAK,OAAO,CAAC,CAAC,CAAC,EAC3D,MAAM,GAAE,gBAAqB,GAC5B,OAAO,CAAC,iBAAiB,CAAC,CAAC,CAAC,CAAC,CAiE/B"}
@@ -1,36 +0,0 @@
1
- import { type HarnessConfig, type SandboxDriver, type SandboxHarnessResult } from './sandbox-harness';
2
- export type SandboxJudgeKind = 'compiler' | 'test' | 'linter' | 'security';
3
- export interface SandboxJudgeSpec {
4
- id: string;
5
- kind: SandboxJudgeKind;
6
- config: HarnessConfig;
7
- }
8
- export interface SandboxJudgeResult {
9
- id: string;
10
- kind: SandboxJudgeKind;
11
- passed: boolean;
12
- score: number;
13
- summary: string;
14
- detail: SandboxHarnessResult;
15
- }
16
- export interface JudgeFleetOptions {
17
- driver?: SandboxDriver;
18
- parallel?: boolean;
19
- /**
20
- * Max concurrent judge subprocesses. Each spec spawns its own subprocess
21
- * via the driver, so unbounded fan-out exhausts file descriptors / PIDs.
22
- * Defaults to the host CPU count. Ignored when `parallel === false`.
23
- */
24
- concurrency?: number;
25
- }
26
- export declare class JudgeRunner {
27
- private readonly driver;
28
- constructor(driver?: SandboxDriver);
29
- run(spec: SandboxJudgeSpec): Promise<SandboxJudgeResult>;
30
- }
31
- export declare function runJudgeFleet(specs: SandboxJudgeSpec[], options?: JudgeFleetOptions): Promise<SandboxJudgeResult[]>;
32
- export declare function compilerJudge(id: string, config: HarnessConfig): SandboxJudgeSpec;
33
- export declare function testJudge(id: string, config: HarnessConfig): SandboxJudgeSpec;
34
- export declare function linterJudge(id: string, config: HarnessConfig): SandboxJudgeSpec;
35
- export declare function securityJudge(id: string, config: HarnessConfig): SandboxJudgeSpec;
36
- //# sourceMappingURL=judge-runner.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"judge-runner.d.ts","sourceRoot":"","sources":["../src/judge-runner.ts"],"names":[],"mappings":"AACA,OAAO,EACL,KAAK,aAAa,EAClB,KAAK,aAAa,EAElB,KAAK,oBAAoB,EAE1B,MAAM,mBAAmB,CAAA;AAI1B,MAAM,MAAM,gBAAgB,GAAG,UAAU,GAAG,MAAM,GAAG,QAAQ,GAAG,UAAU,CAAA;AAE1E,MAAM,WAAW,gBAAgB;IAC/B,EAAE,EAAE,MAAM,CAAA;IACV,IAAI,EAAE,gBAAgB,CAAA;IACtB,MAAM,EAAE,aAAa,CAAA;CACtB;AAED,MAAM,WAAW,kBAAkB;IACjC,EAAE,EAAE,MAAM,CAAA;IACV,IAAI,EAAE,gBAAgB,CAAA;IACtB,MAAM,EAAE,OAAO,CAAA;IACf,KAAK,EAAE,MAAM,CAAA;IACb,OAAO,EAAE,MAAM,CAAA;IACf,MAAM,EAAE,oBAAoB,CAAA;CAC7B;AAED,MAAM,WAAW,iBAAiB;IAChC,MAAM,CAAC,EAAE,aAAa,CAAA;IACtB,QAAQ,CAAC,EAAE,OAAO,CAAA;IAClB;;;;OAIG;IACH,WAAW,CAAC,EAAE,MAAM,CAAA;CACrB;AAED,qBAAa,WAAW;IACtB,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAe;gBAE1B,MAAM,GAAE,aAA6C;IAI3D,GAAG,CAAC,IAAI,EAAE,gBAAgB,GAAG,OAAO,CAAC,kBAAkB,CAAC;CAoB/D;AAED,wBAAsB,aAAa,CACjC,KAAK,EAAE,gBAAgB,EAAE,EACzB,OAAO,GAAE,iBAAsB,GAC9B,OAAO,CAAC,kBAAkB,EAAE,CAAC,CAmB/B;AAED,wBAAgB,aAAa,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,aAAa,GAAG,gBAAgB,CAEjF;AAED,wBAAgB,SAAS,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,aAAa,GAAG,gBAAgB,CAE7E;AAED,wBAAgB,WAAW,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,aAAa,GAAG,gBAAgB,CAE/E;AAED,wBAAgB,aAAa,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,aAAa,GAAG,gBAAgB,CAEjF"}
@@ -1,2 +0,0 @@
1
- export {};
2
- //# sourceMappingURL=judge-runner.test.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"judge-runner.test.d.ts","sourceRoot":"","sources":["../src/judge-runner.test.ts"],"names":[],"mappings":""}