@tangle-network/agent-eval 0.95.0 → 0.95.1

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 (961) 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/{trace-analyst/analyst.d.ts → analyst-C8HHvfJp.d.ts} +14 -12
  8. package/dist/{contract/analyze-runs.d.ts → analyze-runs-DtT6F_6T.d.ts} +10 -9
  9. package/dist/authenticity/index.d.ts +16 -15
  10. package/dist/{baseline.d.ts → baseline-Bbid3WoO.d.ts} +44 -8
  11. package/dist/belief-state/index.d.ts +605 -14
  12. package/dist/benchmarks/index.d.ts +5 -23
  13. package/dist/builder-eval/index.d.ts +250 -5
  14. package/dist/calibration-BPmzuVPk.d.ts +101 -0
  15. package/dist/campaign/index.d.ts +1452 -38
  16. package/dist/cli.d.ts +0 -2
  17. package/dist/{contract/intake/code-agent-session.d.ts → code-agent-session-CPHRCb4-.d.ts} +17 -15
  18. package/dist/contract/index.d.ts +688 -106
  19. package/dist/control-Doncu-B_.d.ts +259 -0
  20. package/dist/{control-runtime.d.ts → control-runtime-Acf9CGhw.d.ts} +26 -23
  21. package/dist/control.d.ts +10 -11
  22. package/dist/corpus-D4YW9UoJ.d.ts +560 -0
  23. package/dist/{cost-ledger.d.ts → cost-ledger-DuSqlw5B.d.ts} +11 -10
  24. package/dist/{counterfactual.d.ts → counterfactual-DlOz8PBx.d.ts} +13 -12
  25. package/dist/{dataset.d.ts → dataset-BbGkaN2I.d.ts} +14 -11
  26. package/dist/{analyst/registry.d.ts → default-registry-GyE8X5SP.d.ts} +37 -8
  27. package/dist/diagnose.d.ts +252 -1
  28. package/dist/{trace/emitter.d.ts → emitter-C2rqGH_l.d.ts} +12 -9
  29. package/dist/{errors.d.ts → errors-CzMUYo7b.d.ts} +11 -10
  30. package/dist/failure-cluster-DH9Flgcf.d.ts +76 -0
  31. package/dist/{feedback-trajectory.d.ts → feedback-trajectory-BxY0cKfs.d.ts} +38 -36
  32. package/dist/fuzz.d.ts +547 -1
  33. package/dist/gepa-C1NCIZ9o.d.ts +414 -0
  34. package/dist/governance/index.d.ts +135 -5
  35. package/dist/harness-optimizer-mOl9XX_O.d.ts +106 -0
  36. package/dist/hosted/index.d.ts +239 -10
  37. package/dist/index-_Y4oNOOb.d.ts +159 -0
  38. package/dist/index.d.ts +5660 -277
  39. package/dist/{contract/insight-report.d.ts → insight-report-BnRjTibG.d.ts} +19 -16
  40. package/dist/{trace/integrity.d.ts → integrity-D2t12mMw.d.ts} +14 -11
  41. package/dist/{judge-calibration.d.ts → judge-calibration-0p2QcWNE.d.ts} +17 -16
  42. package/dist/{analyst/kind-factory.d.ts → kind-factory-X3eDYbKn.d.ts} +61 -10
  43. package/dist/knowledge/index.d.ts +103 -3
  44. package/dist/{llm-client.d.ts → llm-client-Bj7g0rqu.d.ts} +23 -21
  45. package/dist/matrix/index.d.ts +30 -12
  46. package/dist/meta-eval/index.d.ts +182 -6
  47. package/dist/{multi-layer-verifier.d.ts → multi-layer-verifier-DUZXrPDA.d.ts} +15 -12
  48. package/dist/multishot/index.d.ts +290 -7
  49. package/dist/{rl/off-policy.d.ts → off-policy-DiwuKKg7.d.ts} +9 -8
  50. package/dist/openapi.json +1 -1
  51. package/dist/{meta-eval/outcome-store.d.ts → outcome-store-rnXLEqSn.d.ts} +8 -7
  52. package/dist/{pareto.d.ts → pareto-E-pembql.d.ts} +10 -9
  53. package/dist/perf/index.d.ts +119 -13
  54. package/dist/pipelines/index.d.ts +173 -8
  55. package/dist/pre-registration-nfUdc9EQ.d.ts +483 -0
  56. package/dist/prm/index.d.ts +104 -5
  57. package/dist/provenance-CncDq9qE.d.ts +426 -0
  58. package/dist/query-B7GGjRox.d.ts +32 -0
  59. package/dist/{trace/raw-provider-sink.d.ts → raw-provider-sink-C46HDghv.d.ts} +14 -13
  60. package/dist/{red-team.d.ts → red-team-BWdoyleI.d.ts} +16 -14
  61. package/dist/{trace/redact.d.ts → redact-B40YG2M_.d.ts} +8 -7
  62. package/dist/release-report-pidWUMZ2.d.ts +233 -0
  63. package/dist/reporting.d.ts +16 -15
  64. package/dist/{eval-campaign.d.ts → researcher-Jr8ME1dZ.d.ts} +156 -34
  65. package/dist/rl.d.ts +1193 -1
  66. package/dist/{prm/rubric.d.ts → rubric-Cc6UHvUb.d.ts} +13 -10
  67. package/dist/{meta-eval/rubric-predictive-validity.d.ts → rubric-predictive-validity-C2hDKM8Z.d.ts} +11 -8
  68. package/dist/run-critic-CmMf05uV.d.ts +56 -0
  69. package/dist/{run-record.d.ts → run-record-CP2ObebC.d.ts} +117 -15
  70. package/dist/runtime-trajectory-BOUUjI0y.d.ts +49 -0
  71. package/dist/{trace/schema.d.ts → schema-m0gsnbt3.d.ts} +30 -29
  72. package/dist/semantic-concept-judge-DSBB2Cfp.d.ts +624 -0
  73. package/dist/{sequential.d.ts → sequential-5iSVfzl2.d.ts} +10 -9
  74. package/dist/{series-convergence.d.ts → series-convergence-D5OWMBg6.d.ts} +5 -4
  75. package/dist/sink-fetch-B1Yg4Til.d.ts +101 -0
  76. package/dist/{statistics.d.ts → statistics-CCJpTGOS.d.ts} +48 -46
  77. package/dist/{trace/store.d.ts → store-BcFXE6LG.d.ts} +12 -21
  78. package/dist/{trace-analyst/types.d.ts → store-C1YxJDEK.d.ts} +74 -18
  79. package/dist/storyboard/index.d.ts +81 -16
  80. package/dist/{summary-report.d.ts → summary-report-CInXwsza.d.ts} +160 -23
  81. package/dist/telemetry/{sink-file.d.ts → file.d.ts} +7 -5
  82. package/dist/telemetry/index.d.ts +35 -17
  83. package/dist/{sandbox-harness.d.ts → test-graded-scenario-DeODGLra.d.ts} +60 -15
  84. package/dist/{locked-jsonl-appender.d.ts → testing-C21CHsq2.d.ts} +4 -3
  85. package/dist/testing.d.ts +1 -5
  86. package/dist/traces.d.ts +976 -4
  87. package/dist/{trajectory.d.ts → trajectory-2TkpSEVh.d.ts} +9 -8
  88. package/dist/{analyst/types.d.ts → types-B5x54y6n.d.ts} +112 -19
  89. package/dist/{matrix/types.d.ts → types-BUxNaJ8c.d.ts} +11 -9
  90. package/dist/{types.d.ts → types-C7DGg5ex.d.ts} +33 -31
  91. package/dist/{campaign/types.d.ts → types-DQRY8ZT-.d.ts} +47 -44
  92. package/dist/{verdict.d.ts → verdict-C9MlYujm.d.ts} +3 -2
  93. package/dist/wire/index.d.ts +570 -13
  94. package/dist/workflow/index.d.ts +496 -22
  95. package/package.json +2 -2
  96. package/dist/action-policy.d.ts +0 -24
  97. package/dist/action-policy.d.ts.map +0 -1
  98. package/dist/action-policy.test.d.ts +0 -2
  99. package/dist/action-policy.test.d.ts.map +0 -1
  100. package/dist/active-learning.d.ts +0 -41
  101. package/dist/active-learning.d.ts.map +0 -1
  102. package/dist/adapters/http.d.ts.map +0 -1
  103. package/dist/adapters/langchain.d.ts.map +0 -1
  104. package/dist/adapters/otel.d.ts.map +0 -1
  105. package/dist/agent-profile-cell.d.ts +0 -101
  106. package/dist/agent-profile-cell.d.ts.map +0 -1
  107. package/dist/agent-profile.d.ts +0 -27
  108. package/dist/agent-profile.d.ts.map +0 -1
  109. package/dist/agent-profile.test.d.ts +0 -2
  110. package/dist/agent-profile.test.d.ts.map +0 -1
  111. package/dist/analyst/adapters.d.ts +0 -62
  112. package/dist/analyst/adapters.d.ts.map +0 -1
  113. package/dist/analyst/analyst.test.d.ts +0 -2
  114. package/dist/analyst/analyst.test.d.ts.map +0 -1
  115. package/dist/analyst/ax-service.d.ts +0 -27
  116. package/dist/analyst/ax-service.d.ts.map +0 -1
  117. package/dist/analyst/behavioral-analyst.d.ts +0 -28
  118. package/dist/analyst/behavioral-analyst.d.ts.map +0 -1
  119. package/dist/analyst/chat-client.d.ts +0 -91
  120. package/dist/analyst/chat-client.d.ts.map +0 -1
  121. package/dist/analyst/default-registry.d.ts +0 -27
  122. package/dist/analyst/default-registry.d.ts.map +0 -1
  123. package/dist/analyst/default-registry.test.d.ts +0 -2
  124. package/dist/analyst/default-registry.test.d.ts.map +0 -1
  125. package/dist/analyst/finding-signature.d.ts +0 -48
  126. package/dist/analyst/finding-signature.d.ts.map +0 -1
  127. package/dist/analyst/finding-subject.d.ts +0 -146
  128. package/dist/analyst/finding-subject.d.ts.map +0 -1
  129. package/dist/analyst/finding-subject.test.d.ts +0 -2
  130. package/dist/analyst/finding-subject.test.d.ts.map +0 -1
  131. package/dist/analyst/findings-store.d.ts +0 -75
  132. package/dist/analyst/findings-store.d.ts.map +0 -1
  133. package/dist/analyst/index.d.ts.map +0 -1
  134. package/dist/analyst/kind-factory.d.ts.map +0 -1
  135. package/dist/analyst/kinds/failure-mode.d.ts +0 -19
  136. package/dist/analyst/kinds/failure-mode.d.ts.map +0 -1
  137. package/dist/analyst/kinds/improvement.d.ts +0 -23
  138. package/dist/analyst/kinds/improvement.d.ts.map +0 -1
  139. package/dist/analyst/kinds/index.d.ts +0 -22
  140. package/dist/analyst/kinds/index.d.ts.map +0 -1
  141. package/dist/analyst/kinds/kinds.test.d.ts +0 -2
  142. package/dist/analyst/kinds/kinds.test.d.ts.map +0 -1
  143. package/dist/analyst/kinds/knowledge-gap.d.ts +0 -28
  144. package/dist/analyst/kinds/knowledge-gap.d.ts.map +0 -1
  145. package/dist/analyst/kinds/knowledge-poisoning.d.ts +0 -22
  146. package/dist/analyst/kinds/knowledge-poisoning.d.ts.map +0 -1
  147. package/dist/analyst/kinds/skill-usage.d.ts +0 -84
  148. package/dist/analyst/kinds/skill-usage.d.ts.map +0 -1
  149. package/dist/analyst/kinds/skill-usage.test.d.ts +0 -2
  150. package/dist/analyst/kinds/skill-usage.test.d.ts.map +0 -1
  151. package/dist/analyst/parse-tolerant.d.ts +0 -26
  152. package/dist/analyst/parse-tolerant.d.ts.map +0 -1
  153. package/dist/analyst/parse-tolerant.test.d.ts +0 -2
  154. package/dist/analyst/parse-tolerant.test.d.ts.map +0 -1
  155. package/dist/analyst/registry.budget.test.d.ts +0 -2
  156. package/dist/analyst/registry.budget.test.d.ts.map +0 -1
  157. package/dist/analyst/registry.d.ts.map +0 -1
  158. package/dist/analyst/steer-firewall.d.ts +0 -35
  159. package/dist/analyst/steer-firewall.d.ts.map +0 -1
  160. package/dist/analyst/steer-firewall.test.d.ts +0 -2
  161. package/dist/analyst/steer-firewall.test.d.ts.map +0 -1
  162. package/dist/analyst/structure-findings.d.ts +0 -37
  163. package/dist/analyst/structure-findings.d.ts.map +0 -1
  164. package/dist/analyst/structure-findings.test.d.ts +0 -2
  165. package/dist/analyst/structure-findings.test.d.ts.map +0 -1
  166. package/dist/analyst/tool-groups.d.ts +0 -34
  167. package/dist/analyst/tool-groups.d.ts.map +0 -1
  168. package/dist/analyst/types.d.ts.map +0 -1
  169. package/dist/anti-slop.d.ts +0 -59
  170. package/dist/anti-slop.d.ts.map +0 -1
  171. package/dist/artifact-validator.d.ts +0 -74
  172. package/dist/artifact-validator.d.ts.map +0 -1
  173. package/dist/attestation.d.ts +0 -63
  174. package/dist/attestation.d.ts.map +0 -1
  175. package/dist/attestation.test.d.ts +0 -2
  176. package/dist/attestation.test.d.ts.map +0 -1
  177. package/dist/authenticity/index.d.ts.map +0 -1
  178. package/dist/authenticity/index.test.d.ts +0 -2
  179. package/dist/authenticity/index.test.d.ts.map +0 -1
  180. package/dist/auto-pr.d.ts +0 -120
  181. package/dist/auto-pr.d.ts.map +0 -1
  182. package/dist/baseline.d.ts.map +0 -1
  183. package/dist/behavior-dsl.d.ts +0 -73
  184. package/dist/behavior-dsl.d.ts.map +0 -1
  185. package/dist/belief-state/calibration.d.ts +0 -10
  186. package/dist/belief-state/calibration.d.ts.map +0 -1
  187. package/dist/belief-state/calibration.test.d.ts +0 -2
  188. package/dist/belief-state/calibration.test.d.ts.map +0 -1
  189. package/dist/belief-state/code-agent-corpus.d.ts +0 -66
  190. package/dist/belief-state/code-agent-corpus.d.ts.map +0 -1
  191. package/dist/belief-state/code-agent-corpus.test.d.ts +0 -2
  192. package/dist/belief-state/code-agent-corpus.test.d.ts.map +0 -1
  193. package/dist/belief-state/code-agent-evidence.d.ts +0 -22
  194. package/dist/belief-state/code-agent-evidence.d.ts.map +0 -1
  195. package/dist/belief-state/code-agent-evidence.test.d.ts +0 -2
  196. package/dist/belief-state/code-agent-evidence.test.d.ts.map +0 -1
  197. package/dist/belief-state/extract.d.ts +0 -7
  198. package/dist/belief-state/extract.d.ts.map +0 -1
  199. package/dist/belief-state/extract.test.d.ts +0 -2
  200. package/dist/belief-state/extract.test.d.ts.map +0 -1
  201. package/dist/belief-state/index.d.ts.map +0 -1
  202. package/dist/belief-state/ope.d.ts +0 -17
  203. package/dist/belief-state/ope.d.ts.map +0 -1
  204. package/dist/belief-state/ope.test.d.ts +0 -2
  205. package/dist/belief-state/ope.test.d.ts.map +0 -1
  206. package/dist/belief-state/phase0-measurement.d.ts +0 -55
  207. package/dist/belief-state/phase0-measurement.d.ts.map +0 -1
  208. package/dist/belief-state/report.d.ts +0 -17
  209. package/dist/belief-state/report.d.ts.map +0 -1
  210. package/dist/belief-state/report.test.d.ts +0 -2
  211. package/dist/belief-state/report.test.d.ts.map +0 -1
  212. package/dist/belief-state/research-evidence.d.ts +0 -23
  213. package/dist/belief-state/research-evidence.d.ts.map +0 -1
  214. package/dist/belief-state/research-evidence.test.d.ts +0 -2
  215. package/dist/belief-state/research-evidence.test.d.ts.map +0 -1
  216. package/dist/belief-state/runtime-benchmark-corpus.d.ts +0 -32
  217. package/dist/belief-state/runtime-benchmark-corpus.d.ts.map +0 -1
  218. package/dist/belief-state/runtime-hooks.d.ts +0 -87
  219. package/dist/belief-state/runtime-hooks.d.ts.map +0 -1
  220. package/dist/belief-state/runtime-hooks.test.d.ts +0 -2
  221. package/dist/belief-state/runtime-hooks.test.d.ts.map +0 -1
  222. package/dist/belief-state/selective.d.ts +0 -15
  223. package/dist/belief-state/selective.d.ts.map +0 -1
  224. package/dist/belief-state/selective.test.d.ts +0 -2
  225. package/dist/belief-state/selective.test.d.ts.map +0 -1
  226. package/dist/belief-state/shadow-probe.d.ts +0 -80
  227. package/dist/belief-state/shadow-probe.d.ts.map +0 -1
  228. package/dist/belief-state/shadow-probe.test.d.ts +0 -2
  229. package/dist/belief-state/shadow-probe.test.d.ts.map +0 -1
  230. package/dist/belief-state/types.d.ts +0 -195
  231. package/dist/belief-state/types.d.ts.map +0 -1
  232. package/dist/belief-state/types.test.d.ts +0 -2
  233. package/dist/belief-state/types.test.d.ts.map +0 -1
  234. package/dist/benchmark.d.ts +0 -14
  235. package/dist/benchmark.d.ts.map +0 -1
  236. package/dist/benchmarks/index.d.ts.map +0 -1
  237. package/dist/benchmarks/routing/dataset.d.ts +0 -34
  238. package/dist/benchmarks/routing/dataset.d.ts.map +0 -1
  239. package/dist/benchmarks/routing/index.d.ts +0 -34
  240. package/dist/benchmarks/routing/index.d.ts.map +0 -1
  241. package/dist/benchmarks/types.d.ts +0 -49
  242. package/dist/benchmarks/types.d.ts.map +0 -1
  243. package/dist/bisector.d.ts +0 -81
  244. package/dist/bisector.d.ts.map +0 -1
  245. package/dist/budget-guard.d.ts +0 -31
  246. package/dist/budget-guard.d.ts.map +0 -1
  247. package/dist/builder-eval/builder-session.d.ts +0 -111
  248. package/dist/builder-eval/builder-session.d.ts.map +0 -1
  249. package/dist/builder-eval/correlation.d.ts +0 -32
  250. package/dist/builder-eval/correlation.d.ts.map +0 -1
  251. package/dist/builder-eval/index.d.ts.map +0 -1
  252. package/dist/builder-eval/project-registry.d.ts +0 -51
  253. package/dist/builder-eval/project-registry.d.ts.map +0 -1
  254. package/dist/builder-eval/three-layer-eval.d.ts +0 -55
  255. package/dist/builder-eval/three-layer-eval.d.ts.map +0 -1
  256. package/dist/campaign/analyst-surface.d.ts +0 -108
  257. package/dist/campaign/analyst-surface.d.ts.map +0 -1
  258. package/dist/campaign/analyst-surface.test.d.ts +0 -2
  259. package/dist/campaign/analyst-surface.test.d.ts.map +0 -1
  260. package/dist/campaign/auto-pr.d.ts +0 -46
  261. package/dist/campaign/auto-pr.d.ts.map +0 -1
  262. package/dist/campaign/distillation/agreement-judge.d.ts +0 -69
  263. package/dist/campaign/distillation/agreement-judge.d.ts.map +0 -1
  264. package/dist/campaign/distillation/cli.d.ts +0 -35
  265. package/dist/campaign/distillation/cli.d.ts.map +0 -1
  266. package/dist/campaign/distillation/distillation.test.d.ts +0 -2
  267. package/dist/campaign/distillation/distillation.test.d.ts.map +0 -1
  268. package/dist/campaign/distillation/gold-scenarios.d.ts +0 -54
  269. package/dist/campaign/distillation/gold-scenarios.d.ts.map +0 -1
  270. package/dist/campaign/distillation/run-distillation.d.ts +0 -119
  271. package/dist/campaign/distillation/run-distillation.d.ts.map +0 -1
  272. package/dist/campaign/gates/compose.d.ts +0 -12
  273. package/dist/campaign/gates/compose.d.ts.map +0 -1
  274. package/dist/campaign/gates/default-production-gate.d.ts +0 -58
  275. package/dist/campaign/gates/default-production-gate.d.ts.map +0 -1
  276. package/dist/campaign/gates/heldout-gate.d.ts +0 -12
  277. package/dist/campaign/gates/heldout-gate.d.ts.map +0 -1
  278. package/dist/campaign/gates/promotion-policy.d.ts +0 -125
  279. package/dist/campaign/gates/promotion-policy.d.ts.map +0 -1
  280. package/dist/campaign/gates/promotion-policy.test.d.ts +0 -2
  281. package/dist/campaign/gates/promotion-policy.test.d.ts.map +0 -1
  282. package/dist/campaign/gates/sequential.d.ts +0 -146
  283. package/dist/campaign/gates/sequential.d.ts.map +0 -1
  284. package/dist/campaign/gates/sequential.test.d.ts +0 -2
  285. package/dist/campaign/gates/sequential.test.d.ts.map +0 -1
  286. package/dist/campaign/gates/statistical-heldout.d.ts +0 -99
  287. package/dist/campaign/gates/statistical-heldout.d.ts.map +0 -1
  288. package/dist/campaign/gates/statistical-heldout.test.d.ts +0 -2
  289. package/dist/campaign/gates/statistical-heldout.test.d.ts.map +0 -1
  290. package/dist/campaign/index.d.ts.map +0 -1
  291. package/dist/campaign/labeled-store/fs-adapter.d.ts +0 -59
  292. package/dist/campaign/labeled-store/fs-adapter.d.ts.map +0 -1
  293. package/dist/campaign/presets/compare-proposers.d.ts +0 -146
  294. package/dist/campaign/presets/compare-proposers.d.ts.map +0 -1
  295. package/dist/campaign/presets/playback.d.ts +0 -120
  296. package/dist/campaign/presets/playback.d.ts.map +0 -1
  297. package/dist/campaign/presets/playback.test.d.ts +0 -2
  298. package/dist/campaign/presets/playback.test.d.ts.map +0 -1
  299. package/dist/campaign/presets/run-eval.d.ts +0 -14
  300. package/dist/campaign/presets/run-eval.d.ts.map +0 -1
  301. package/dist/campaign/presets/run-improvement-loop.d.ts +0 -63
  302. package/dist/campaign/presets/run-improvement-loop.d.ts.map +0 -1
  303. package/dist/campaign/presets/run-improvement-loop.test.d.ts +0 -2
  304. package/dist/campaign/presets/run-improvement-loop.test.d.ts.map +0 -1
  305. package/dist/campaign/presets/run-optimization.d.ts +0 -92
  306. package/dist/campaign/presets/run-optimization.d.ts.map +0 -1
  307. package/dist/campaign/presets/run-profile-matrix.d.ts +0 -151
  308. package/dist/campaign/presets/run-profile-matrix.d.ts.map +0 -1
  309. package/dist/campaign/presets/run-skill-opt.d.ts +0 -96
  310. package/dist/campaign/presets/run-skill-opt.d.ts.map +0 -1
  311. package/dist/campaign/proposers/_findings-text.d.ts +0 -22
  312. package/dist/campaign/proposers/_findings-text.d.ts.map +0 -1
  313. package/dist/campaign/proposers/ace.d.ts +0 -33
  314. package/dist/campaign/proposers/ace.d.ts.map +0 -1
  315. package/dist/campaign/proposers/ace.test.d.ts +0 -2
  316. package/dist/campaign/proposers/ace.test.d.ts.map +0 -1
  317. package/dist/campaign/proposers/analysis-edit.d.ts +0 -32
  318. package/dist/campaign/proposers/analysis-edit.d.ts.map +0 -1
  319. package/dist/campaign/proposers/evolutionary.d.ts +0 -20
  320. package/dist/campaign/proposers/evolutionary.d.ts.map +0 -1
  321. package/dist/campaign/proposers/fapo.d.ts +0 -120
  322. package/dist/campaign/proposers/fapo.d.ts.map +0 -1
  323. package/dist/campaign/proposers/gepa.d.ts +0 -86
  324. package/dist/campaign/proposers/gepa.d.ts.map +0 -1
  325. package/dist/campaign/proposers/halo.d.ts +0 -44
  326. package/dist/campaign/proposers/halo.d.ts.map +0 -1
  327. package/dist/campaign/proposers/halo.test.d.ts +0 -2
  328. package/dist/campaign/proposers/halo.test.d.ts.map +0 -1
  329. package/dist/campaign/proposers/memory.d.ts +0 -47
  330. package/dist/campaign/proposers/memory.d.ts.map +0 -1
  331. package/dist/campaign/proposers/memory.test.d.ts +0 -2
  332. package/dist/campaign/proposers/memory.test.d.ts.map +0 -1
  333. package/dist/campaign/proposers/skill-opt.d.ts +0 -88
  334. package/dist/campaign/proposers/skill-opt.d.ts.map +0 -1
  335. package/dist/campaign/proposers/trace-analyst.d.ts +0 -48
  336. package/dist/campaign/proposers/trace-analyst.d.ts.map +0 -1
  337. package/dist/campaign/proposers/trace-analyst.test.d.ts +0 -2
  338. package/dist/campaign/proposers/trace-analyst.test.d.ts.map +0 -1
  339. package/dist/campaign/provenance.d.ts +0 -185
  340. package/dist/campaign/provenance.d.ts.map +0 -1
  341. package/dist/campaign/run-campaign.d.ts +0 -90
  342. package/dist/campaign/run-campaign.d.ts.map +0 -1
  343. package/dist/campaign/score-utils.d.ts +0 -26
  344. package/dist/campaign/score-utils.d.ts.map +0 -1
  345. package/dist/campaign/skill-patch.d.ts +0 -62
  346. package/dist/campaign/skill-patch.d.ts.map +0 -1
  347. package/dist/campaign/storage.d.ts +0 -38
  348. package/dist/campaign/storage.d.ts.map +0 -1
  349. package/dist/campaign/types.d.ts.map +0 -1
  350. package/dist/campaign/worktree/index.d.ts +0 -53
  351. package/dist/campaign/worktree/index.d.ts.map +0 -1
  352. package/dist/canary.d.ts +0 -101
  353. package/dist/canary.d.ts.map +0 -1
  354. package/dist/causal-attribution.d.ts +0 -45
  355. package/dist/causal-attribution.d.ts.map +0 -1
  356. package/dist/ci-gate.d.ts +0 -44
  357. package/dist/ci-gate.d.ts.map +0 -1
  358. package/dist/cli.d.ts.map +0 -1
  359. package/dist/client.d.ts +0 -77
  360. package/dist/client.d.ts.map +0 -1
  361. package/dist/client.test.d.ts +0 -2
  362. package/dist/client.test.d.ts.map +0 -1
  363. package/dist/command-runner.d.ts +0 -74
  364. package/dist/command-runner.d.ts.map +0 -1
  365. package/dist/command-runner.test.d.ts +0 -2
  366. package/dist/command-runner.test.d.ts.map +0 -1
  367. package/dist/completion-verifier.d.ts +0 -147
  368. package/dist/completion-verifier.d.ts.map +0 -1
  369. package/dist/completion-verifier.test.d.ts +0 -9
  370. package/dist/completion-verifier.test.d.ts.map +0 -1
  371. package/dist/concurrency.d.ts +0 -23
  372. package/dist/concurrency.d.ts.map +0 -1
  373. package/dist/contamination-guard.d.ts +0 -81
  374. package/dist/contamination-guard.d.ts.map +0 -1
  375. package/dist/contract/analyze-runs.d.ts.map +0 -1
  376. package/dist/contract/define-agent-eval.d.ts +0 -52
  377. package/dist/contract/define-agent-eval.d.ts.map +0 -1
  378. package/dist/contract/diff.d.ts +0 -114
  379. package/dist/contract/diff.d.ts.map +0 -1
  380. package/dist/contract/index.d.ts.map +0 -1
  381. package/dist/contract/insight-report.d.ts.map +0 -1
  382. package/dist/contract/insight-types-fwd.d.ts +0 -7
  383. package/dist/contract/insight-types-fwd.d.ts.map +0 -1
  384. package/dist/contract/intake/agent-trace.d.ts +0 -97
  385. package/dist/contract/intake/agent-trace.d.ts.map +0 -1
  386. package/dist/contract/intake/code-agent-session.d.ts.map +0 -1
  387. package/dist/contract/intake/feedback-table.d.ts +0 -87
  388. package/dist/contract/intake/feedback-table.d.ts.map +0 -1
  389. package/dist/contract/intake/index.d.ts +0 -22
  390. package/dist/contract/intake/index.d.ts.map +0 -1
  391. package/dist/contract/intake/otel-spans.d.ts +0 -33
  392. package/dist/contract/intake/otel-spans.d.ts.map +0 -1
  393. package/dist/contract/self-improve.d.ts +0 -284
  394. package/dist/contract/self-improve.d.ts.map +0 -1
  395. package/dist/control-runtime.d.ts.map +0 -1
  396. package/dist/control-runtime.test.d.ts +0 -2
  397. package/dist/control-runtime.test.d.ts.map +0 -1
  398. package/dist/control.d.ts.map +0 -1
  399. package/dist/convergence.d.ts +0 -29
  400. package/dist/convergence.d.ts.map +0 -1
  401. package/dist/cost-ledger.d.ts.map +0 -1
  402. package/dist/cost-ledger.test.d.ts +0 -2
  403. package/dist/cost-ledger.test.d.ts.map +0 -1
  404. package/dist/cost-report.d.ts +0 -43
  405. package/dist/cost-report.d.ts.map +0 -1
  406. package/dist/cost-report.test.d.ts +0 -2
  407. package/dist/cost-report.test.d.ts.map +0 -1
  408. package/dist/cost-tracker.d.ts +0 -76
  409. package/dist/cost-tracker.d.ts.map +0 -1
  410. package/dist/counterfactual.d.ts.map +0 -1
  411. package/dist/cross-trace-diff.d.ts +0 -56
  412. package/dist/cross-trace-diff.d.ts.map +0 -1
  413. package/dist/dataset.d.ts.map +0 -1
  414. package/dist/deploy-gate-layer.d.ts +0 -125
  415. package/dist/deploy-gate-layer.d.ts.map +0 -1
  416. package/dist/deploy-gate-layer.test.d.ts +0 -2
  417. package/dist/deploy-gate-layer.test.d.ts.map +0 -1
  418. package/dist/description-length-gate.d.ts +0 -119
  419. package/dist/description-length-gate.d.ts.map +0 -1
  420. package/dist/detectors/edge.test.d.ts +0 -2
  421. package/dist/detectors/edge.test.d.ts.map +0 -1
  422. package/dist/detectors/index.d.ts +0 -81
  423. package/dist/detectors/index.d.ts.map +0 -1
  424. package/dist/detectors/index.test.d.ts +0 -2
  425. package/dist/detectors/index.test.d.ts.map +0 -1
  426. package/dist/diagnose/causal-sweep.d.ts +0 -100
  427. package/dist/diagnose/causal-sweep.d.ts.map +0 -1
  428. package/dist/diagnose/index.d.ts +0 -36
  429. package/dist/diagnose/index.d.ts.map +0 -1
  430. package/dist/diagnose/remediation.d.ts +0 -68
  431. package/dist/diagnose/remediation.d.ts.map +0 -1
  432. package/dist/diagnose/repair.d.ts +0 -77
  433. package/dist/diagnose/repair.d.ts.map +0 -1
  434. package/dist/discover-personas.d.ts +0 -35
  435. package/dist/discover-personas.d.ts.map +0 -1
  436. package/dist/driver.d.ts +0 -95
  437. package/dist/driver.d.ts.map +0 -1
  438. package/dist/driver.test.d.ts +0 -8
  439. package/dist/driver.test.d.ts.map +0 -1
  440. package/dist/dual-agent-bench.d.ts +0 -81
  441. package/dist/dual-agent-bench.d.ts.map +0 -1
  442. package/dist/error-count-extractor.d.ts +0 -47
  443. package/dist/error-count-extractor.d.ts.map +0 -1
  444. package/dist/error-count-extractor.test.d.ts +0 -2
  445. package/dist/error-count-extractor.test.d.ts.map +0 -1
  446. package/dist/errors.d.ts.map +0 -1
  447. package/dist/eval-campaign.d.ts.map +0 -1
  448. package/dist/eval-campaign.test.d.ts +0 -2
  449. package/dist/eval-campaign.test.d.ts.map +0 -1
  450. package/dist/eval-tools.d.ts +0 -55
  451. package/dist/eval-tools.d.ts.map +0 -1
  452. package/dist/eval-trace-store.d.ts +0 -107
  453. package/dist/eval-trace-store.d.ts.map +0 -1
  454. package/dist/eval-trace-store.test.d.ts +0 -2
  455. package/dist/eval-trace-store.test.d.ts.map +0 -1
  456. package/dist/executor.d.ts +0 -38
  457. package/dist/executor.d.ts.map +0 -1
  458. package/dist/executor.test.d.ts +0 -10
  459. package/dist/executor.test.d.ts.map +0 -1
  460. package/dist/experiment-tracker.d.ts +0 -178
  461. package/dist/experiment-tracker.d.ts.map +0 -1
  462. package/dist/experiment-tracker.test.d.ts +0 -2
  463. package/dist/experiment-tracker.test.d.ts.map +0 -1
  464. package/dist/failure-taxonomy.d.ts +0 -38
  465. package/dist/failure-taxonomy.d.ts.map +0 -1
  466. package/dist/feedback-trajectory.d.ts.map +0 -1
  467. package/dist/feedback-trajectory.test.d.ts +0 -2
  468. package/dist/feedback-trajectory.test.d.ts.map +0 -1
  469. package/dist/flow-layer.d.ts +0 -90
  470. package/dist/flow-layer.d.ts.map +0 -1
  471. package/dist/flow-layer.test.d.ts +0 -2
  472. package/dist/flow-layer.test.d.ts.map +0 -1
  473. package/dist/fuzz/capsule.d.ts +0 -46
  474. package/dist/fuzz/capsule.d.ts.map +0 -1
  475. package/dist/fuzz/cube.d.ts +0 -36
  476. package/dist/fuzz/cube.d.ts.map +0 -1
  477. package/dist/fuzz/explorer-cost.test.d.ts +0 -2
  478. package/dist/fuzz/explorer-cost.test.d.ts.map +0 -1
  479. package/dist/fuzz/explorer.d.ts +0 -64
  480. package/dist/fuzz/explorer.d.ts.map +0 -1
  481. package/dist/fuzz/fuzz-agent.d.ts +0 -16
  482. package/dist/fuzz/fuzz-agent.d.ts.map +0 -1
  483. package/dist/fuzz/fuzz-agent.test.d.ts +0 -2
  484. package/dist/fuzz/fuzz-agent.test.d.ts.map +0 -1
  485. package/dist/fuzz/gates.d.ts +0 -33
  486. package/dist/fuzz/gates.d.ts.map +0 -1
  487. package/dist/fuzz/index.d.ts +0 -26
  488. package/dist/fuzz/index.d.ts.map +0 -1
  489. package/dist/fuzz/policies.d.ts +0 -28
  490. package/dist/fuzz/policies.d.ts.map +0 -1
  491. package/dist/fuzz/tools.d.ts +0 -20
  492. package/dist/fuzz/tools.d.ts.map +0 -1
  493. package/dist/fuzz/types.d.ts +0 -307
  494. package/dist/fuzz/types.d.ts.map +0 -1
  495. package/dist/golden-matcher.d.ts +0 -71
  496. package/dist/golden-matcher.d.ts.map +0 -1
  497. package/dist/governance/eu-ai-act.d.ts +0 -37
  498. package/dist/governance/eu-ai-act.d.ts.map +0 -1
  499. package/dist/governance/index.d.ts.map +0 -1
  500. package/dist/governance/nist-ai-rmf.d.ts +0 -15
  501. package/dist/governance/nist-ai-rmf.d.ts.map +0 -1
  502. package/dist/governance/soc2.d.ts +0 -12
  503. package/dist/governance/soc2.d.ts.map +0 -1
  504. package/dist/governance/types.d.ts +0 -66
  505. package/dist/governance/types.d.ts.map +0 -1
  506. package/dist/harness-optimizer.d.ts +0 -82
  507. package/dist/harness-optimizer.d.ts.map +0 -1
  508. package/dist/held-out-gate.d.ts +0 -135
  509. package/dist/held-out-gate.d.ts.map +0 -1
  510. package/dist/hosted/client.d.ts +0 -73
  511. package/dist/hosted/client.d.ts.map +0 -1
  512. package/dist/hosted/from-env.test.d.ts +0 -8
  513. package/dist/hosted/from-env.test.d.ts.map +0 -1
  514. package/dist/hosted/index.d.ts.map +0 -1
  515. package/dist/hosted/types.d.ts +0 -159
  516. package/dist/hosted/types.d.ts.map +0 -1
  517. package/dist/index.d.ts.map +0 -1
  518. package/dist/integrity/backend-integrity.d.ts +0 -71
  519. package/dist/integrity/backend-integrity.d.ts.map +0 -1
  520. package/dist/integrity/preflight.d.ts +0 -72
  521. package/dist/integrity/preflight.d.ts.map +0 -1
  522. package/dist/integrity/preflight.test.d.ts +0 -2
  523. package/dist/integrity/preflight.test.d.ts.map +0 -1
  524. package/dist/integrity/single-backend.d.ts +0 -67
  525. package/dist/integrity/single-backend.d.ts.map +0 -1
  526. package/dist/intent-match-judge.d.ts +0 -69
  527. package/dist/intent-match-judge.d.ts.map +0 -1
  528. package/dist/intent-match-judge.test.d.ts +0 -2
  529. package/dist/intent-match-judge.test.d.ts.map +0 -1
  530. package/dist/judge-calibration.d.ts.map +0 -1
  531. package/dist/judge-ensemble.d.ts +0 -66
  532. package/dist/judge-ensemble.d.ts.map +0 -1
  533. package/dist/judge-ensemble.test.d.ts +0 -8
  534. package/dist/judge-ensemble.test.d.ts.map +0 -1
  535. package/dist/judge-families.d.ts +0 -38
  536. package/dist/judge-families.d.ts.map +0 -1
  537. package/dist/judge-panel.d.ts +0 -65
  538. package/dist/judge-panel.d.ts.map +0 -1
  539. package/dist/judge-retry.d.ts +0 -70
  540. package/dist/judge-retry.d.ts.map +0 -1
  541. package/dist/judge-runner.d.ts +0 -36
  542. package/dist/judge-runner.d.ts.map +0 -1
  543. package/dist/judge-runner.test.d.ts +0 -2
  544. package/dist/judge-runner.test.d.ts.map +0 -1
  545. package/dist/judges.d.ts +0 -74
  546. package/dist/judges.d.ts.map +0 -1
  547. package/dist/keyword-coverage-judge.d.ts +0 -89
  548. package/dist/keyword-coverage-judge.d.ts.map +0 -1
  549. package/dist/keyword-coverage-judge.test.d.ts +0 -2
  550. package/dist/keyword-coverage-judge.test.d.ts.map +0 -1
  551. package/dist/knowledge/index.d.ts.map +0 -1
  552. package/dist/knowledge/readiness.d.ts +0 -26
  553. package/dist/knowledge/readiness.d.ts.map +0 -1
  554. package/dist/knowledge/types.d.ts +0 -75
  555. package/dist/knowledge/types.d.ts.map +0 -1
  556. package/dist/live-proof.d.ts +0 -62
  557. package/dist/live-proof.d.ts.map +0 -1
  558. package/dist/llm-client.d.ts.map +0 -1
  559. package/dist/llm-client.test.d.ts +0 -2
  560. package/dist/llm-client.test.d.ts.map +0 -1
  561. package/dist/locked-jsonl-appender.d.ts.map +0 -1
  562. package/dist/matrix/aggregation.d.ts +0 -16
  563. package/dist/matrix/aggregation.d.ts.map +0 -1
  564. package/dist/matrix/index.d.ts.map +0 -1
  565. package/dist/matrix/runner.d.ts +0 -15
  566. package/dist/matrix/runner.d.ts.map +0 -1
  567. package/dist/matrix/types.d.ts.map +0 -1
  568. package/dist/meta-eval/calibration.d.ts +0 -47
  569. package/dist/meta-eval/calibration.d.ts.map +0 -1
  570. package/dist/meta-eval/correlation-study.d.ts +0 -53
  571. package/dist/meta-eval/correlation-study.d.ts.map +0 -1
  572. package/dist/meta-eval/index.d.ts.map +0 -1
  573. package/dist/meta-eval/outcome-store.d.ts.map +0 -1
  574. package/dist/meta-eval/rubric-predictive-validity.d.ts.map +0 -1
  575. package/dist/meta-eval/sentinel.d.ts +0 -169
  576. package/dist/meta-eval/sentinel.d.ts.map +0 -1
  577. package/dist/metrics.d.ts +0 -63
  578. package/dist/metrics.d.ts.map +0 -1
  579. package/dist/model-seats.d.ts +0 -71
  580. package/dist/model-seats.d.ts.map +0 -1
  581. package/dist/model-seats.test.d.ts +0 -2
  582. package/dist/model-seats.test.d.ts.map +0 -1
  583. package/dist/muffled-gate-scanner.d.ts +0 -102
  584. package/dist/muffled-gate-scanner.d.ts.map +0 -1
  585. package/dist/multi-layer-verifier.d.ts.map +0 -1
  586. package/dist/multi-layer-verifier.test.d.ts +0 -2
  587. package/dist/multi-layer-verifier.test.d.ts.map +0 -1
  588. package/dist/multi-toolchain-layer.d.ts +0 -80
  589. package/dist/multi-toolchain-layer.d.ts.map +0 -1
  590. package/dist/multi-toolchain-layer.test.d.ts +0 -2
  591. package/dist/multi-toolchain-layer.test.d.ts.map +0 -1
  592. package/dist/multishot/default-tools.d.ts +0 -34
  593. package/dist/multishot/default-tools.d.ts.map +0 -1
  594. package/dist/multishot/index.d.ts.map +0 -1
  595. package/dist/multishot/judges.d.ts +0 -32
  596. package/dist/multishot/judges.d.ts.map +0 -1
  597. package/dist/multishot/matrix.d.ts +0 -107
  598. package/dist/multishot/matrix.d.ts.map +0 -1
  599. package/dist/multishot/multishot.d.ts +0 -23
  600. package/dist/multishot/multishot.d.ts.map +0 -1
  601. package/dist/multishot/router.d.ts +0 -37
  602. package/dist/multishot/router.d.ts.map +0 -1
  603. package/dist/multishot/types.d.ts +0 -60
  604. package/dist/multishot/types.d.ts.map +0 -1
  605. package/dist/observability.d.ts +0 -71
  606. package/dist/observability.d.ts.map +0 -1
  607. package/dist/oracle.d.ts +0 -55
  608. package/dist/oracle.d.ts.map +0 -1
  609. package/dist/orthogonality.d.ts +0 -35
  610. package/dist/orthogonality.d.ts.map +0 -1
  611. package/dist/otel-pipeline.d.ts +0 -31
  612. package/dist/otel-pipeline.d.ts.map +0 -1
  613. package/dist/paraphrase.d.ts +0 -107
  614. package/dist/paraphrase.d.ts.map +0 -1
  615. package/dist/pareto.d.ts.map +0 -1
  616. package/dist/partition-held-out.d.ts +0 -70
  617. package/dist/partition-held-out.d.ts.map +0 -1
  618. package/dist/partition-held-out.test.d.ts +0 -2
  619. package/dist/partition-held-out.test.d.ts.map +0 -1
  620. package/dist/perf/index.d.ts.map +0 -1
  621. package/dist/perf/integrity.d.ts +0 -30
  622. package/dist/perf/integrity.d.ts.map +0 -1
  623. package/dist/perf/journey.d.ts +0 -45
  624. package/dist/perf/journey.d.ts.map +0 -1
  625. package/dist/perf/ratchet.d.ts +0 -47
  626. package/dist/perf/ratchet.d.ts.map +0 -1
  627. package/dist/pipelines/budget-breach.d.ts +0 -31
  628. package/dist/pipelines/budget-breach.d.ts.map +0 -1
  629. package/dist/pipelines/budget-breach.test.d.ts +0 -2
  630. package/dist/pipelines/budget-breach.test.d.ts.map +0 -1
  631. package/dist/pipelines/failure-cluster.d.ts +0 -38
  632. package/dist/pipelines/failure-cluster.d.ts.map +0 -1
  633. package/dist/pipelines/failure-cluster.test.d.ts +0 -2
  634. package/dist/pipelines/failure-cluster.test.d.ts.map +0 -1
  635. package/dist/pipelines/first-divergence.d.ts +0 -26
  636. package/dist/pipelines/first-divergence.d.ts.map +0 -1
  637. package/dist/pipelines/first-divergence.test.d.ts +0 -2
  638. package/dist/pipelines/first-divergence.test.d.ts.map +0 -1
  639. package/dist/pipelines/index.d.ts.map +0 -1
  640. package/dist/pipelines/judge-agreement.d.ts +0 -26
  641. package/dist/pipelines/judge-agreement.d.ts.map +0 -1
  642. package/dist/pipelines/judge-agreement.test.d.ts +0 -2
  643. package/dist/pipelines/judge-agreement.test.d.ts.map +0 -1
  644. package/dist/pipelines/regression.d.ts +0 -23
  645. package/dist/pipelines/regression.d.ts.map +0 -1
  646. package/dist/pipelines/regression.test.d.ts +0 -2
  647. package/dist/pipelines/regression.test.d.ts.map +0 -1
  648. package/dist/pipelines/stuck-loop.d.ts +0 -32
  649. package/dist/pipelines/stuck-loop.d.ts.map +0 -1
  650. package/dist/pipelines/stuck-loop.test.d.ts +0 -2
  651. package/dist/pipelines/stuck-loop.test.d.ts.map +0 -1
  652. package/dist/pipelines/tool-waste.d.ts +0 -34
  653. package/dist/pipelines/tool-waste.d.ts.map +0 -1
  654. package/dist/pipelines/tool-waste.test.d.ts +0 -2
  655. package/dist/pipelines/tool-waste.test.d.ts.map +0 -1
  656. package/dist/playbook.d.ts +0 -16
  657. package/dist/playbook.d.ts.map +0 -1
  658. package/dist/pr-review-benchmark.d.ts +0 -88
  659. package/dist/pr-review-benchmark.d.ts.map +0 -1
  660. package/dist/pr-review-benchmark.test.d.ts +0 -2
  661. package/dist/pr-review-benchmark.test.d.ts.map +0 -1
  662. package/dist/pre-registration.d.ts +0 -125
  663. package/dist/pre-registration.d.ts.map +0 -1
  664. package/dist/prm/builtin-rubrics.d.ts +0 -33
  665. package/dist/prm/builtin-rubrics.d.ts.map +0 -1
  666. package/dist/prm/index.d.ts.map +0 -1
  667. package/dist/prm/inference.d.ts +0 -29
  668. package/dist/prm/inference.d.ts.map +0 -1
  669. package/dist/prm/inference.test.d.ts +0 -2
  670. package/dist/prm/inference.test.d.ts.map +0 -1
  671. package/dist/prm/rubric.d.ts.map +0 -1
  672. package/dist/prm/training-export.d.ts +0 -38
  673. package/dist/prm/training-export.d.ts.map +0 -1
  674. package/dist/produced-state.d.ts +0 -63
  675. package/dist/produced-state.d.ts.map +0 -1
  676. package/dist/produced-state.test.d.ts +0 -8
  677. package/dist/produced-state.test.d.ts.map +0 -1
  678. package/dist/profile/baselines.d.ts +0 -37
  679. package/dist/profile/baselines.d.ts.map +0 -1
  680. package/dist/profile/index.d.ts +0 -105
  681. package/dist/profile/index.d.ts.map +0 -1
  682. package/dist/promotion-gate.d.ts +0 -93
  683. package/dist/promotion-gate.d.ts.map +0 -1
  684. package/dist/prompt-registry.d.ts +0 -41
  685. package/dist/prompt-registry.d.ts.map +0 -1
  686. package/dist/propose-review-control.d.ts +0 -49
  687. package/dist/propose-review-control.d.ts.map +0 -1
  688. package/dist/propose-review-control.test.d.ts +0 -2
  689. package/dist/propose-review-control.test.d.ts.map +0 -1
  690. package/dist/propose-review.d.ts +0 -155
  691. package/dist/propose-review.d.ts.map +0 -1
  692. package/dist/red-team.d.ts.map +0 -1
  693. package/dist/reference-replay-steering.d.ts +0 -11
  694. package/dist/reference-replay-steering.d.ts.map +0 -1
  695. package/dist/reference-replay.d.ts +0 -176
  696. package/dist/reference-replay.d.ts.map +0 -1
  697. package/dist/reflective-mutation.d.ts +0 -79
  698. package/dist/reflective-mutation.d.ts.map +0 -1
  699. package/dist/registry.d.ts +0 -31
  700. package/dist/registry.d.ts.map +0 -1
  701. package/dist/release-confidence.d.ts +0 -128
  702. package/dist/release-confidence.d.ts.map +0 -1
  703. package/dist/release-report.d.ts +0 -11
  704. package/dist/release-report.d.ts.map +0 -1
  705. package/dist/replay.d.ts +0 -120
  706. package/dist/replay.d.ts.map +0 -1
  707. package/dist/reporter.d.ts +0 -14
  708. package/dist/reporter.d.ts.map +0 -1
  709. package/dist/reporting.d.ts.map +0 -1
  710. package/dist/researcher.d.ts +0 -140
  711. package/dist/researcher.d.ts.map +0 -1
  712. package/dist/reviewer.d.ts +0 -118
  713. package/dist/reviewer.d.ts.map +0 -1
  714. package/dist/reviewer.test.d.ts +0 -2
  715. package/dist/reviewer.test.d.ts.map +0 -1
  716. package/dist/reward-model-export.d.ts +0 -60
  717. package/dist/reward-model-export.d.ts.map +0 -1
  718. package/dist/rl/active-curriculum.d.ts +0 -110
  719. package/dist/rl/active-curriculum.d.ts.map +0 -1
  720. package/dist/rl/adaptation-eval.d.ts +0 -109
  721. package/dist/rl/adaptation-eval.d.ts.map +0 -1
  722. package/dist/rl/adversarial.d.ts.map +0 -1
  723. package/dist/rl/compute-curves.d.ts +0 -127
  724. package/dist/rl/compute-curves.d.ts.map +0 -1
  725. package/dist/rl/contamination.d.ts +0 -117
  726. package/dist/rl/contamination.d.ts.map +0 -1
  727. package/dist/rl/corpus.d.ts +0 -55
  728. package/dist/rl/corpus.d.ts.map +0 -1
  729. package/dist/rl/corpus.test.d.ts +0 -2
  730. package/dist/rl/corpus.test.d.ts.map +0 -1
  731. package/dist/rl/dataset.d.ts +0 -102
  732. package/dist/rl/dataset.d.ts.map +0 -1
  733. package/dist/rl/dataset.test.d.ts +0 -2
  734. package/dist/rl/dataset.test.d.ts.map +0 -1
  735. package/dist/rl/exporters.d.ts +0 -141
  736. package/dist/rl/exporters.d.ts.map +0 -1
  737. package/dist/rl/index.d.ts +0 -49
  738. package/dist/rl/index.d.ts.map +0 -1
  739. package/dist/rl/off-policy.d.ts.map +0 -1
  740. package/dist/rl/predictive-validity-researcher.d.ts +0 -69
  741. package/dist/rl/predictive-validity-researcher.d.ts.map +0 -1
  742. package/dist/rl/preferences.d.ts +0 -141
  743. package/dist/rl/preferences.d.ts.map +0 -1
  744. package/dist/rl/process-reward.d.ts +0 -122
  745. package/dist/rl/process-reward.d.ts.map +0 -1
  746. package/dist/rl/reward-hacking.d.ts +0 -104
  747. package/dist/rl/reward-hacking.d.ts.map +0 -1
  748. package/dist/rl/rl-campaign.d.ts +0 -85
  749. package/dist/rl/rl-campaign.d.ts.map +0 -1
  750. package/dist/rl/run-record-adapters.d.ts +0 -56
  751. package/dist/rl/run-record-adapters.d.ts.map +0 -1
  752. package/dist/rl/sim-fidelity.d.ts +0 -166
  753. package/dist/rl/sim-fidelity.d.ts.map +0 -1
  754. package/dist/rl/sim-fidelity.test.d.ts +0 -2
  755. package/dist/rl/sim-fidelity.test.d.ts.map +0 -1
  756. package/dist/rl/tournament.d.ts +0 -115
  757. package/dist/rl/tournament.d.ts.map +0 -1
  758. package/dist/rl/verifiable-reward.d.ts +0 -124
  759. package/dist/rl/verifiable-reward.d.ts.map +0 -1
  760. package/dist/run-critic.d.ts +0 -23
  761. package/dist/run-critic.d.ts.map +0 -1
  762. package/dist/run-evidence.d.ts +0 -32
  763. package/dist/run-evidence.d.ts.map +0 -1
  764. package/dist/run-record.d.ts.map +0 -1
  765. package/dist/run-record.test.d.ts +0 -2
  766. package/dist/run-record.test.d.ts.map +0 -1
  767. package/dist/run-score.d.ts +0 -31
  768. package/dist/run-score.d.ts.map +0 -1
  769. package/dist/runtime-trajectory.d.ts +0 -47
  770. package/dist/runtime-trajectory.d.ts.map +0 -1
  771. package/dist/sandbox-harness.d.ts.map +0 -1
  772. package/dist/sandbox-harness.test.d.ts +0 -2
  773. package/dist/sandbox-harness.test.d.ts.map +0 -1
  774. package/dist/sandbox-pool.d.ts +0 -74
  775. package/dist/sandbox-pool.d.ts.map +0 -1
  776. package/dist/sandbox-pool.test.d.ts +0 -2
  777. package/dist/sandbox-pool.test.d.ts.map +0 -1
  778. package/dist/scorecard.d.ts +0 -133
  779. package/dist/scorecard.d.ts.map +0 -1
  780. package/dist/scorecard.test.d.ts +0 -2
  781. package/dist/scorecard.test.d.ts.map +0 -1
  782. package/dist/self-play.d.ts +0 -69
  783. package/dist/self-play.d.ts.map +0 -1
  784. package/dist/semantic-concept-judge.d.ts +0 -135
  785. package/dist/semantic-concept-judge.d.ts.map +0 -1
  786. package/dist/semantic-concept-judge.test.d.ts +0 -2
  787. package/dist/semantic-concept-judge.test.d.ts.map +0 -1
  788. package/dist/sequential.d.ts.map +0 -1
  789. package/dist/series-convergence.d.ts.map +0 -1
  790. package/dist/slo.d.ts +0 -48
  791. package/dist/slo.d.ts.map +0 -1
  792. package/dist/state-continuity.d.ts +0 -47
  793. package/dist/state-continuity.d.ts.map +0 -1
  794. package/dist/statistics.d.ts.map +0 -1
  795. package/dist/statistics.test.d.ts +0 -2
  796. package/dist/statistics.test.d.ts.map +0 -1
  797. package/dist/steering-optimizer.d.ts +0 -58
  798. package/dist/steering-optimizer.d.ts.map +0 -1
  799. package/dist/steering.d.ts +0 -24
  800. package/dist/steering.d.ts.map +0 -1
  801. package/dist/storyboard/code-edit.d.ts +0 -64
  802. package/dist/storyboard/code-edit.d.ts.map +0 -1
  803. package/dist/storyboard/code-edit.test.d.ts +0 -2
  804. package/dist/storyboard/code-edit.test.d.ts.map +0 -1
  805. package/dist/storyboard/index.d.ts.map +0 -1
  806. package/dist/storyboard/index.test.d.ts +0 -2
  807. package/dist/storyboard/index.test.d.ts.map +0 -1
  808. package/dist/summary-report.d.ts.map +0 -1
  809. package/dist/telemetry/client.d.ts +0 -35
  810. package/dist/telemetry/client.d.ts.map +0 -1
  811. package/dist/telemetry/index.d.ts.map +0 -1
  812. package/dist/telemetry/schema.d.ts +0 -61
  813. package/dist/telemetry/schema.d.ts.map +0 -1
  814. package/dist/telemetry/sink-fetch.d.ts +0 -39
  815. package/dist/telemetry/sink-fetch.d.ts.map +0 -1
  816. package/dist/telemetry/sink-file.d.ts.map +0 -1
  817. package/dist/test-graded-scenario.d.ts +0 -42
  818. package/dist/test-graded-scenario.d.ts.map +0 -1
  819. package/dist/testing.d.ts.map +0 -1
  820. package/dist/tool-use-metrics.d.ts +0 -35
  821. package/dist/tool-use-metrics.d.ts.map +0 -1
  822. package/dist/trace/capture-fetch.d.ts +0 -48
  823. package/dist/trace/capture-fetch.d.ts.map +0 -1
  824. package/dist/trace/capture-fetch.test.d.ts +0 -2
  825. package/dist/trace/capture-fetch.test.d.ts.map +0 -1
  826. package/dist/trace/emitter.d.ts.map +0 -1
  827. package/dist/trace/extract-usage.d.ts +0 -46
  828. package/dist/trace/extract-usage.d.ts.map +0 -1
  829. package/dist/trace/extract-usage.test.d.ts +0 -2
  830. package/dist/trace/extract-usage.test.d.ts.map +0 -1
  831. package/dist/trace/index.d.ts +0 -15
  832. package/dist/trace/index.d.ts.map +0 -1
  833. package/dist/trace/integrity.d.ts.map +0 -1
  834. package/dist/trace/otel-bridge.d.ts +0 -29
  835. package/dist/trace/otel-bridge.d.ts.map +0 -1
  836. package/dist/trace/otel-export.d.ts +0 -52
  837. package/dist/trace/otel-export.d.ts.map +0 -1
  838. package/dist/trace/otel.d.ts +0 -57
  839. package/dist/trace/otel.d.ts.map +0 -1
  840. package/dist/trace/otlp-attributes.d.ts +0 -17
  841. package/dist/trace/otlp-attributes.d.ts.map +0 -1
  842. package/dist/trace/query.d.ts +0 -29
  843. package/dist/trace/query.d.ts.map +0 -1
  844. package/dist/trace/query.test.d.ts +0 -2
  845. package/dist/trace/query.test.d.ts.map +0 -1
  846. package/dist/trace/raw-provider-sink.d.ts.map +0 -1
  847. package/dist/trace/redact.d.ts.map +0 -1
  848. package/dist/trace/schema.d.ts.map +0 -1
  849. package/dist/trace/store-to-otlp.d.ts +0 -72
  850. package/dist/trace/store-to-otlp.d.ts.map +0 -1
  851. package/dist/trace/store-to-otlp.test.d.ts +0 -2
  852. package/dist/trace/store-to-otlp.test.d.ts.map +0 -1
  853. package/dist/trace/store.d.ts.map +0 -1
  854. package/dist/trace/store.test.d.ts +0 -2
  855. package/dist/trace/store.test.d.ts.map +0 -1
  856. package/dist/trace-analyst/analyst.d.ts.map +0 -1
  857. package/dist/trace-analyst/analyst.test.d.ts +0 -2
  858. package/dist/trace-analyst/analyst.test.d.ts.map +0 -1
  859. package/dist/trace-analyst/behavioral-metrics.d.ts +0 -40
  860. package/dist/trace-analyst/behavioral-metrics.d.ts.map +0 -1
  861. package/dist/trace-analyst/behavioral-metrics.test.d.ts +0 -2
  862. package/dist/trace-analyst/behavioral-metrics.test.d.ts.map +0 -1
  863. package/dist/trace-analyst/hook.d.ts +0 -55
  864. package/dist/trace-analyst/hook.d.ts.map +0 -1
  865. package/dist/trace-analyst/index.d.ts +0 -18
  866. package/dist/trace-analyst/index.d.ts.map +0 -1
  867. package/dist/trace-analyst/insights.d.ts +0 -71
  868. package/dist/trace-analyst/insights.d.ts.map +0 -1
  869. package/dist/trace-analyst/insights.test.d.ts +0 -2
  870. package/dist/trace-analyst/insights.test.d.ts.map +0 -1
  871. package/dist/trace-analyst/otlp-flatten.d.ts +0 -42
  872. package/dist/trace-analyst/otlp-flatten.d.ts.map +0 -1
  873. package/dist/trace-analyst/otlp-span.d.ts +0 -85
  874. package/dist/trace-analyst/otlp-span.d.ts.map +0 -1
  875. package/dist/trace-analyst/otlp-span.test.d.ts +0 -2
  876. package/dist/trace-analyst/otlp-span.test.d.ts.map +0 -1
  877. package/dist/trace-analyst/otlp-to-run-records.d.ts +0 -115
  878. package/dist/trace-analyst/otlp-to-run-records.d.ts.map +0 -1
  879. package/dist/trace-analyst/otlp-to-run-records.test.d.ts +0 -2
  880. package/dist/trace-analyst/otlp-to-run-records.test.d.ts.map +0 -1
  881. package/dist/trace-analyst/otlp-to-run-records.timestamps.test.d.ts +0 -2
  882. package/dist/trace-analyst/otlp-to-run-records.timestamps.test.d.ts.map +0 -1
  883. package/dist/trace-analyst/prompts.d.ts +0 -6
  884. package/dist/trace-analyst/prompts.d.ts.map +0 -1
  885. package/dist/trace-analyst/store-otlp.d.ts +0 -126
  886. package/dist/trace-analyst/store-otlp.d.ts.map +0 -1
  887. package/dist/trace-analyst/store-otlp.test.d.ts +0 -8
  888. package/dist/trace-analyst/store-otlp.test.d.ts.map +0 -1
  889. package/dist/trace-analyst/store-otlp.timestamps.test.d.ts +0 -2
  890. package/dist/trace-analyst/store-otlp.timestamps.test.d.ts.map +0 -1
  891. package/dist/trace-analyst/store.d.ts +0 -63
  892. package/dist/trace-analyst/store.d.ts.map +0 -1
  893. package/dist/trace-analyst/tools.d.ts +0 -44
  894. package/dist/trace-analyst/tools.d.ts.map +0 -1
  895. package/dist/trace-analyst/tools.test.d.ts +0 -10
  896. package/dist/trace-analyst/tools.test.d.ts.map +0 -1
  897. package/dist/trace-analyst/types.d.ts.map +0 -1
  898. package/dist/trace-contracts.d.ts +0 -180
  899. package/dist/trace-contracts.d.ts.map +0 -1
  900. package/dist/traced-analyst.d.ts +0 -26
  901. package/dist/traced-analyst.d.ts.map +0 -1
  902. package/dist/traced-judges.d.ts +0 -27
  903. package/dist/traced-judges.d.ts.map +0 -1
  904. package/dist/traces.d.ts.map +0 -1
  905. package/dist/trajectory.d.ts.map +0 -1
  906. package/dist/types.d.ts.map +0 -1
  907. package/dist/ui-finding.d.ts +0 -104
  908. package/dist/ui-finding.d.ts.map +0 -1
  909. package/dist/verdict-cache.d.ts +0 -78
  910. package/dist/verdict-cache.d.ts.map +0 -1
  911. package/dist/verdict-cache.test.d.ts +0 -2
  912. package/dist/verdict-cache.test.d.ts.map +0 -1
  913. package/dist/verdict.d.ts.map +0 -1
  914. package/dist/visual-diff.d.ts +0 -32
  915. package/dist/visual-diff.d.ts.map +0 -1
  916. package/dist/wire/handlers.d.ts +0 -54
  917. package/dist/wire/handlers.d.ts.map +0 -1
  918. package/dist/wire/index.d.ts.map +0 -1
  919. package/dist/wire/openapi.d.ts +0 -3
  920. package/dist/wire/openapi.d.ts.map +0 -1
  921. package/dist/wire/rpc.d.ts +0 -21
  922. package/dist/wire/rpc.d.ts.map +0 -1
  923. package/dist/wire/rubrics.d.ts +0 -34
  924. package/dist/wire/rubrics.d.ts.map +0 -1
  925. package/dist/wire/schemas.d.ts +0 -410
  926. package/dist/wire/schemas.d.ts.map +0 -1
  927. package/dist/wire/server.d.ts +0 -60
  928. package/dist/wire/server.d.ts.map +0 -1
  929. package/dist/workflow/event-schema.d.ts +0 -5
  930. package/dist/workflow/event-schema.d.ts.map +0 -1
  931. package/dist/workflow/feedback-pack.d.ts +0 -99
  932. package/dist/workflow/feedback-pack.d.ts.map +0 -1
  933. package/dist/workflow/index.d.ts.map +0 -1
  934. package/dist/workflow/intelligence-export.d.ts +0 -62
  935. package/dist/workflow/intelligence-export.d.ts.map +0 -1
  936. package/dist/workflow/partner-report.d.ts +0 -49
  937. package/dist/workflow/partner-report.d.ts.map +0 -1
  938. package/dist/workflow/phase-graph.d.ts +0 -43
  939. package/dist/workflow/phase-graph.d.ts.map +0 -1
  940. package/dist/workflow/promotion-gate.d.ts +0 -61
  941. package/dist/workflow/promotion-gate.d.ts.map +0 -1
  942. package/dist/workflow/run-record.d.ts +0 -12
  943. package/dist/workflow/run-record.d.ts.map +0 -1
  944. package/dist/workflow/runtime-adapter.d.ts +0 -20
  945. package/dist/workflow/runtime-adapter.d.ts.map +0 -1
  946. package/dist/workflow/sanitize.d.ts +0 -21
  947. package/dist/workflow/sanitize.d.ts.map +0 -1
  948. package/dist/workflow/schema.d.ts +0 -5
  949. package/dist/workflow/schema.d.ts.map +0 -1
  950. package/dist/workflow/summary.d.ts +0 -43
  951. package/dist/workflow/summary.d.ts.map +0 -1
  952. package/dist/workflow/trace-event-fields.d.ts +0 -6
  953. package/dist/workflow/trace-event-fields.d.ts.map +0 -1
  954. package/dist/workflow/trajectory.d.ts +0 -15
  955. package/dist/workflow/trajectory.d.ts.map +0 -1
  956. package/dist/workflow/types.d.ts +0 -68
  957. package/dist/workflow/types.d.ts.map +0 -1
  958. package/dist/workspace-inspector.d.ts +0 -67
  959. package/dist/workspace-inspector.d.ts.map +0 -1
  960. package/dist/wrangler-deploy-runner.test.d.ts +0 -2
  961. package/dist/wrangler-deploy-runner.test.d.ts.map +0 -1
@@ -1,125 +0,0 @@
1
- /**
2
- * Deploy gate layer — would the agent's build actually publish?
3
- *
4
- * The product Blueprint Agent fronts promises "go from idea to live URL."
5
- * Pre-Gen-48 the eval stopped at install/typecheck/build/serve — every
6
- * one of which can pass while `vite build` (or `next build`, etc) fails
7
- * on a production-only constraint (env-var requirement, dynamic import
8
- * not statically resolvable, missing public asset).
9
- *
10
- * Deploy gate runs the production build via the supplied {@link DeployRunner}
11
- * and asserts:
12
- * - command exited 0
13
- * - artifact dir contains an entry point (index.html for static SPAs,
14
- * equivalent per framework family)
15
- *
16
- * Ships with a canonical `vite` runner. Additional runners
17
- * (wrangler-deploy --dry-run, next-build, etc.) plug in as factories.
18
- */
19
- import type { Layer } from './multi-layer-verifier';
20
- export type DeployFamily = 'frontend-static' | 'nextjs' | 'remix' | 'fullstack-ts';
21
- export interface DeployRunResult {
22
- ok: boolean;
23
- /** Stdout/stderr tail surfaced as evidence. Bounded in caller. */
24
- output?: string;
25
- /** Wall-clock duration of the build command. */
26
- durationMs?: number;
27
- /** Path to artifact directory the runner expects (dist/, .next/, build/, etc). */
28
- artifactDir?: string;
29
- /** True iff artifactDir contains the family's expected entry point. */
30
- artifactValid?: boolean;
31
- }
32
- export interface DeployRunner {
33
- /** Run the production build. The runner owns command + cwd. */
34
- run(): Promise<DeployRunResult>;
35
- }
36
- export interface DeployGateLayerInput {
37
- /** Build the runner per call. */
38
- runner: () => DeployRunner | Promise<DeployRunner>;
39
- /** Family hint — for logging, surfaced in diagnostics. */
40
- family?: DeployFamily;
41
- /** Layer name. Default `deploy`. */
42
- name?: string;
43
- /** Layer dependencies — default `['build']`. */
44
- dependsOn?: string[];
45
- /** Weight in blendedScore. Default 1. */
46
- weight?: number;
47
- /** Cap (ms). Default 120s — prod builds are slower than dev. */
48
- capMs?: number;
49
- /** When true, treat artifactValid=false as a fail (default true). */
50
- requireArtifact?: boolean;
51
- }
52
- /**
53
- * Build a deploy gate layer that runs the production build and verifies
54
- * the artifact. Pass: ok && artifactValid. Score: 1.0 (pass) or 0 (fail).
55
- *
56
- * For families where artifact-validation isn't applicable (e.g. a
57
- * server-rendered build that prints a manifest), set `requireArtifact:
58
- * false` and rely on the runner's own ok signal.
59
- */
60
- export declare function deployGateLayer<Env = unknown>(input: DeployGateLayerInput): Layer<Env>;
61
- export interface ViteDeployRunnerInput {
62
- /** Workdir to build. The runner cd's here. */
63
- workdir: string;
64
- /**
65
- * Function to run a shell command in `workdir`. Same shape as
66
- * agent-eval's CommandRunner.run for compositional reuse.
67
- */
68
- exec: (cmd: string, opts?: {
69
- cwd?: string;
70
- timeoutMs?: number;
71
- }) => Promise<{
72
- stdout: string;
73
- stderr: string;
74
- exitCode: number;
75
- }>;
76
- /**
77
- * Function to test whether a path exists in the workdir. Inject
78
- * `(p) => existsSync(join(workdir, p))` for host runs.
79
- */
80
- exists: (relativePath: string) => boolean | Promise<boolean>;
81
- /** Build command. Default `npm run build`. */
82
- buildCommand?: string;
83
- /** Artifact directory to validate. Default `dist`. */
84
- artifactDir?: string;
85
- /** Entry-point file under artifactDir. Default `index.html`. */
86
- artifactEntry?: string;
87
- /** Per-build cap (ms). Default 90s. */
88
- timeoutMs?: number;
89
- }
90
- /**
91
- * Canonical runner for `frontend-static` family — runs the build script,
92
- * validates `<artifactDir>/<artifactEntry>` exists. Use as the `runner:`
93
- * factory for {@link deployGateLayer}.
94
- */
95
- export declare function viteDeployRunner(input: ViteDeployRunnerInput): DeployRunner;
96
- export interface WranglerDeployRunnerInput {
97
- workdir: string;
98
- exec: (cmd: string, opts?: {
99
- cwd?: string;
100
- timeoutMs?: number;
101
- }) => Promise<{
102
- stdout: string;
103
- stderr: string;
104
- exitCode: number;
105
- }>;
106
- exists: (relativePath: string) => boolean | Promise<boolean>;
107
- /** Build command. Default `npm run build`. */
108
- buildCommand?: string;
109
- /** Wrangler dry-run command. Default `npx wrangler deploy --dry-run --outdir dist`. */
110
- dryRunCommand?: string;
111
- /** Per-step cap (ms). Default 120s. */
112
- timeoutMs?: number;
113
- }
114
- /**
115
- * Canonical runner for the `fullstack-ts` family on Cloudflare Workers
116
- * (Remix / React Router v7 / Hono on Workers). Detects wrangler.toml or
117
- * wrangler.jsonc in the workdir, builds, then `wrangler deploy --dry-run`
118
- * to catch missing bindings, syntax errors in wrangler config, and
119
- * import-time crashes that don't surface in `tsc`.
120
- *
121
- * No wrangler config = skip with "no wrangler" evidence (not a failure
122
- * — the gate caller decides whether to require deploy validation).
123
- */
124
- export declare function wranglerDeployRunner(input: WranglerDeployRunnerInput): DeployRunner;
125
- //# sourceMappingURL=deploy-gate-layer.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"deploy-gate-layer.d.ts","sourceRoot":"","sources":["../src/deploy-gate-layer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAEH,OAAO,KAAK,EAAE,KAAK,EAAe,MAAM,wBAAwB,CAAA;AAIhE,MAAM,MAAM,YAAY,GAAG,iBAAiB,GAAG,QAAQ,GAAG,OAAO,GAAG,cAAc,CAAA;AAElF,MAAM,WAAW,eAAe;IAC9B,EAAE,EAAE,OAAO,CAAA;IACX,kEAAkE;IAClE,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,gDAAgD;IAChD,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,kFAAkF;IAClF,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,uEAAuE;IACvE,aAAa,CAAC,EAAE,OAAO,CAAA;CACxB;AAED,MAAM,WAAW,YAAY;IAC3B,+DAA+D;IAC/D,GAAG,IAAI,OAAO,CAAC,eAAe,CAAC,CAAA;CAChC;AAED,MAAM,WAAW,oBAAoB;IACnC,iCAAiC;IACjC,MAAM,EAAE,MAAM,YAAY,GAAG,OAAO,CAAC,YAAY,CAAC,CAAA;IAClD,0DAA0D;IAC1D,MAAM,CAAC,EAAE,YAAY,CAAA;IACrB,oCAAoC;IACpC,IAAI,CAAC,EAAE,MAAM,CAAA;IACb,gDAAgD;IAChD,SAAS,CAAC,EAAE,MAAM,EAAE,CAAA;IACpB,yCAAyC;IACzC,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,gEAAgE;IAChE,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,qEAAqE;IACrE,eAAe,CAAC,EAAE,OAAO,CAAA;CAC1B;AAID;;;;;;;GAOG;AACH,wBAAgB,eAAe,CAAC,GAAG,GAAG,OAAO,EAAE,KAAK,EAAE,oBAAoB,GAAG,KAAK,CAAC,GAAG,CAAC,CAwFtF;AAID,MAAM,WAAW,qBAAqB;IACpC,8CAA8C;IAC9C,OAAO,EAAE,MAAM,CAAA;IACf;;;OAGG;IACH,IAAI,EAAE,CACJ,GAAG,EAAE,MAAM,EACX,IAAI,CAAC,EAAE;QAAE,GAAG,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAA;KAAE,KACxC,OAAO,CAAC;QAAE,MAAM,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;IAClE;;;OAGG;IACH,MAAM,EAAE,CAAC,YAAY,EAAE,MAAM,KAAK,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,CAAA;IAC5D,8CAA8C;IAC9C,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,sDAAsD;IACtD,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,gEAAgE;IAChE,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,uCAAuC;IACvC,SAAS,CAAC,EAAE,MAAM,CAAA;CACnB;AAED;;;;GAIG;AACH,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,qBAAqB,GAAG,YAAY,CA2B3E;AAID,MAAM,WAAW,yBAAyB;IACxC,OAAO,EAAE,MAAM,CAAA;IACf,IAAI,EAAE,CACJ,GAAG,EAAE,MAAM,EACX,IAAI,CAAC,EAAE;QAAE,GAAG,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAA;KAAE,KACxC,OAAO,CAAC;QAAE,MAAM,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;IAClE,MAAM,EAAE,CAAC,YAAY,EAAE,MAAM,KAAK,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,CAAA;IAC5D,8CAA8C;IAC9C,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,uFAAuF;IACvF,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,uCAAuC;IACvC,SAAS,CAAC,EAAE,MAAM,CAAA;CACnB;AAED;;;;;;;;;GASG;AACH,wBAAgB,oBAAoB,CAAC,KAAK,EAAE,yBAAyB,GAAG,YAAY,CAsDnF"}
@@ -1,2 +0,0 @@
1
- export {};
2
- //# sourceMappingURL=deploy-gate-layer.test.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"deploy-gate-layer.test.d.ts","sourceRoot":"","sources":["../src/deploy-gate-layer.test.ts"],"names":[],"mappings":""}
@@ -1,119 +0,0 @@
1
- import type { RunRecord } from './run-record';
2
- /**
3
- * DescriptionLengthGate — a Minimum-Description-Length promotion gate, the
4
- * Builder/Breaker acceptance rule from Wang & Buehler, "Self-Revising Discovery
5
- * Systems for Science" (arXiv:2606.01444, MIT LAMM), eq. 5:
6
- *
7
- * L(M, D) = L_model(M) + L_data(D | M)
8
- * accept M' over M iff L(M', D∪E) < L(M, D∪E)
9
- *
10
- * Both candidate and baseline are scored on the SAME enlarged evidence set
11
- * (every accumulated task — NOT just the held-out split), and the candidate is
12
- * accepted only if it lowers the TOTAL bit cost. This is the gate's whole
13
- * point and what distinguishes it from a monotone held-out delta:
14
- *
15
- * - L_model(M) — the candidate's own description length: the compressed size
16
- * of its model text (a prompt, skill, profile, or symbolic model). A bigger
17
- * model pays more bits.
18
- * - L_data(D|M) — the residual: bits of "surprise" that the model did not
19
- * simply succeed, −Σ_i log2(s_i) over the model's per-task score s_i.
20
- * Perfect scores cost 0 bits; failure costs a lot (capped, not infinite).
21
- *
22
- * A candidate that merely memorizes new counterexamples grows L_model faster
23
- * than it shrinks L_data and LOSES — a principled, complexity-penalized
24
- * alternative to HeldOutGate's held-out paired delta. Use this gate
25
- * when the model text whose size you want to penalize is available; use
26
- * HeldOutGate when promotion should turn on held-out generalization with an
27
- * overfit-gap check instead.
28
- *
29
- * Scale / calibration: a gzip'd prose model is hundreds–thousands of bits;
30
- * a single task contributes at most −log2(scoreFloor) data bits (≈10). So with
31
- * little evidence the model term dominates and the gate is conservative about
32
- * model GROWTH — it promotes a larger model only once accumulated evidence
33
- * genuinely pays for the added bits (exactly the paper's regime, where D∪E
34
- * grows). `lambda` is the lever: λ<1 discounts model bits (more permissive),
35
- * λ>1 is stricter. A shrinking-or-equal model that does no worse always wins.
36
- *
37
- * Stateless: construct once with the description-length budget, call
38
- * `evaluate` per (candidate, baseline) pair.
39
- */
40
- export type DescriptionLengthRejectionCode = 'few_tasks' | 'no_total_gain' | 'model_bloat';
41
- export interface DescriptionLengthConfig {
42
- /** Stable label of the baseline. Required — paper-grade evaluation never
43
- * compares two unlabelled candidates. */
44
- baselineKey: string;
45
- /** Weight on model bits relative to data bits (the description-length
46
- * budget λ). 1 = bits are bits. >1 = more complexity-averse. Default 1. */
47
- lambda?: number;
48
- /** The candidate must beat the baseline by at least this many bits to
49
- * promote — a robustness margin against measurement noise. Default 0
50
- * (strict `<`, as the paper). */
51
- marginBits?: number;
52
- /** Per-task score floor for the residual code: −log2(max(s, floor)). Caps a
53
- * total-failure task's surprise instead of letting it diverge. Default
54
- * 2^-10 (a failed task costs 10 bits, not ∞). */
55
- scoreFloor?: number;
56
- /** Minimum number of shared (candidate, baseline) tasks before the gate will
57
- * consider promoting. Default 3. */
58
- minTasks?: number;
59
- }
60
- export interface DescriptionLengthEvidence {
61
- /** Shared tasks scored on both sides (the enlarged evidence D∪E). */
62
- tasks: number;
63
- /** Compressed-model bits — L_model. */
64
- modelBits: {
65
- candidate: number;
66
- baseline: number;
67
- };
68
- /** Residual surprise bits — L_data(D|M). */
69
- dataBits: {
70
- candidate: number;
71
- baseline: number;
72
- };
73
- /** λ·L_model + L_data — the quantity the gate minimizes. */
74
- totalBits: {
75
- candidate: number;
76
- baseline: number;
77
- };
78
- /** candidate − baseline total. Negative = candidate compresses better. */
79
- deltaBits: number;
80
- /** Per-component deltas, for audit: did the win come from a smaller model,
81
- * better outcomes, or both? */
82
- modelBitsDelta: number;
83
- dataBitsDelta: number;
84
- }
85
- export interface DescriptionLengthDecision {
86
- promote: boolean;
87
- candidateId: string;
88
- baselineId: string;
89
- evidence: DescriptionLengthEvidence;
90
- reason: string;
91
- rejectionCode: DescriptionLengthRejectionCode | null;
92
- }
93
- export interface DescriptionLengthCandidate {
94
- /** The model text whose size is L_model (a prompt, skill, profile, or
95
- * symbolic model; concatenated if several files). */
96
- content: string;
97
- /** Runs whose per-task scores form L_data. */
98
- runs: RunRecord[];
99
- }
100
- /** Compressed-model bits — the model's description length L_model. gzip is a
101
- * deterministic, dependency-free stand-in for Kolmogorov complexity; it
102
- * rewards genuine compactness and penalizes boilerplate padding. */
103
- export declare function modelDescriptionBits(content: string): number;
104
- /** Residual surprise L_data(D|M) = −Σ_i log2(max(s_i, floor)) over the given
105
- * per-task scores. Lower = the model more reliably succeeds. */
106
- export declare function dataDescriptionBits(scoreByTask: Map<string, number>, keys: Iterable<string>, scoreFloor: number): number;
107
- export declare class DescriptionLengthGate {
108
- private readonly baselineKey;
109
- private readonly lambda;
110
- private readonly marginBits;
111
- private readonly scoreFloor;
112
- private readonly minTasks;
113
- constructor(config: DescriptionLengthConfig);
114
- /** Decide whether `candidate` should replace `baseline`. Both are scored on
115
- * the shared task set (the enlarged evidence); the candidate promotes only
116
- * if λ·L_model + L_data is strictly lower by at least `marginBits`. */
117
- evaluate(candidate: DescriptionLengthCandidate, baseline: DescriptionLengthCandidate): DescriptionLengthDecision;
118
- }
119
- //# sourceMappingURL=description-length-gate.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"description-length-gate.d.ts","sourceRoot":"","sources":["../src/description-length-gate.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,cAAc,CAAA;AAE7C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqCG;AAEH,MAAM,MAAM,8BAA8B,GAAG,WAAW,GAAG,eAAe,GAAG,aAAa,CAAA;AAE1F,MAAM,WAAW,uBAAuB;IACtC;8CAC0C;IAC1C,WAAW,EAAE,MAAM,CAAA;IACnB;gFAC4E;IAC5E,MAAM,CAAC,EAAE,MAAM,CAAA;IACf;;sCAEkC;IAClC,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB;;sDAEkD;IAClD,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB;yCACqC;IACrC,QAAQ,CAAC,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,yBAAyB;IACxC,qEAAqE;IACrE,KAAK,EAAE,MAAM,CAAA;IACb,uCAAuC;IACvC,SAAS,EAAE;QAAE,SAAS,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,CAAA;IAClD,4CAA4C;IAC5C,QAAQ,EAAE;QAAE,SAAS,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,CAAA;IACjD,4DAA4D;IAC5D,SAAS,EAAE;QAAE,SAAS,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,CAAA;IAClD,0EAA0E;IAC1E,SAAS,EAAE,MAAM,CAAA;IACjB;oCACgC;IAChC,cAAc,EAAE,MAAM,CAAA;IACtB,aAAa,EAAE,MAAM,CAAA;CACtB;AAED,MAAM,WAAW,yBAAyB;IACxC,OAAO,EAAE,OAAO,CAAA;IAChB,WAAW,EAAE,MAAM,CAAA;IACnB,UAAU,EAAE,MAAM,CAAA;IAClB,QAAQ,EAAE,yBAAyB,CAAA;IACnC,MAAM,EAAE,MAAM,CAAA;IACd,aAAa,EAAE,8BAA8B,GAAG,IAAI,CAAA;CACrD;AAED,MAAM,WAAW,0BAA0B;IACzC;0DACsD;IACtD,OAAO,EAAE,MAAM,CAAA;IACf,8CAA8C;IAC9C,IAAI,EAAE,SAAS,EAAE,CAAA;CAClB;AA8BD;;qEAEqE;AACrE,wBAAgB,oBAAoB,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAE5D;AAED;iEACiE;AACjE,wBAAgB,mBAAmB,CACjC,WAAW,EAAE,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,IAAI,EAAE,QAAQ,CAAC,MAAM,CAAC,EACtB,UAAU,EAAE,MAAM,GACjB,MAAM,CAQR;AAED,qBAAa,qBAAqB;IAChC,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAQ;IACpC,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAQ;IAC/B,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAQ;IACnC,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAQ;IACnC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAQ;gBAErB,MAAM,EAAE,uBAAuB;IAY3C;;4EAEwE;IACxE,QAAQ,CACN,SAAS,EAAE,0BAA0B,EACrC,QAAQ,EAAE,0BAA0B,GACnC,yBAAyB;CA+D7B"}
@@ -1,2 +0,0 @@
1
- export {};
2
- //# sourceMappingURL=edge.test.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"edge.test.d.ts","sourceRoot":"","sources":["../../src/detectors/edge.test.ts"],"names":[],"mappings":""}
@@ -1,81 +0,0 @@
1
- /**
2
- * Streaming detectors — the canonical, online failure-mode kernel.
3
- *
4
- * A `StreamingDetector` is a pure incremental reducer: fold one normalized step, get a signal the
5
- * moment a threshold trips (else null). The SAME kernel runs in two places, so the logic lives once:
6
- * - ONLINE, over a live agent pipe (a worker's tool-call stream) → raise a finding mid-run.
7
- * - INSIDE the control loop (`control-runtime`) → its `stopOn*` policies fold these detectors.
8
- *
9
- * Detectors are stateful but self-contained; `streak` is exposed for telemetry and `reset` clears it.
10
- * Fingerprinting is the CALLER's job — the control loop hashes state/action with `stableFingerprint`,
11
- * a tool-call pipe hashes args with `argHash` — so a detector stays agnostic to what it's watching.
12
- */
13
- import type { FailureClass } from '../trace/schema';
14
- export type DetectorSeverity = 'info' | 'warn' | 'critical';
15
- /** A normalized step a detector folds over. Fields are optional; each detector reads only what it
16
- * needs (repeated-action → `actionFingerprint`; no-progress → `stateFingerprint` + `score`;
17
- * error-streak → `status`). */
18
- export interface DetectorEvent {
19
- /** Fingerprint of the action/tool-call this step took (caller pre-hashes). */
20
- readonly actionFingerprint?: string;
21
- /** Fingerprint of observable state AFTER this step. */
22
- readonly stateFingerprint?: string;
23
- /** Score after this step (the score-flat half of no-progress). */
24
- readonly score?: number;
25
- /** Whether this step errored. */
26
- readonly status?: 'ok' | 'error';
27
- /** Free-form label carried into the signal evidence (e.g. the tool name). */
28
- readonly label?: string;
29
- }
30
- export interface DetectorSignal {
31
- readonly detector: string;
32
- readonly severity: DetectorSeverity;
33
- readonly failureClass?: FailureClass;
34
- readonly reason: string;
35
- /** Consecutive matching steps at the moment the signal fired. */
36
- readonly streak: number;
37
- readonly evidence?: Record<string, unknown>;
38
- }
39
- export interface StreamingDetector {
40
- readonly id: string;
41
- /** Current streak (consecutive matching steps) — for telemetry/introspection between observes. */
42
- readonly streak: number;
43
- /** Fold one event; return a signal when the threshold trips, else null. */
44
- observe(event: DetectorEvent): DetectorSignal | null;
45
- /** Clear all state. */
46
- reset(): void;
47
- }
48
- export interface RepeatedActionOptions {
49
- /** Signal once the same action fingerprint repeats this many CONSECUTIVE steps (default 3).
50
- * `<= 0` disables the signal (the streak is still tracked, for telemetry). */
51
- readonly maxRepeated?: number;
52
- readonly severity?: DetectorSeverity;
53
- /** Failure class to stamp on the signal (default `tool_recovery_failure`). */
54
- readonly failureClass?: FailureClass;
55
- }
56
- /** Same action fingerprint N consecutive steps = a stuck loop (the #1 long-horizon failure mode). */
57
- export declare function repeatedActionDetector(opts?: RepeatedActionOptions): StreamingDetector;
58
- export interface NoProgressOptions {
59
- /** Signal once state is unchanged AND score is flat for this many CONSECUTIVE steps (default 3).
60
- * `<= 0` disables the signal. */
61
- readonly maxNoProgress?: number;
62
- /** Minimum |score change| that counts as progress (default 0.001). */
63
- readonly minScoreDelta?: number;
64
- readonly severity?: DetectorSeverity;
65
- readonly failureClass?: FailureClass;
66
- }
67
- /** State + score unchanged across N steps = spinning wheels. Compares each step to the previous one,
68
- * so prime with the initial state (observe it once before the first real step) to detect on step 1. */
69
- export declare function noProgressDetector(opts?: NoProgressOptions): StreamingDetector;
70
- export interface ErrorStreakOptions {
71
- /** Signal once this many CONSECUTIVE steps errored (default 3). `<= 0` disables. */
72
- readonly maxErrors?: number;
73
- readonly severity?: DetectorSeverity;
74
- readonly failureClass?: FailureClass;
75
- }
76
- /** N consecutive tool errors = the worker is hammering a broken approach. */
77
- export declare function errorStreakDetector(opts?: ErrorStreakOptions): StreamingDetector;
78
- /** Fold one event through many detectors at once; returns every signal that fired this step. The
79
- * natural shape for an online pipe watching with a whole panel of detectors. */
80
- export declare function observeAll(detectors: ReadonlyArray<StreamingDetector>, event: DetectorEvent): DetectorSignal[];
81
- //# sourceMappingURL=index.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/detectors/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,iBAAiB,CAAA;AAEnD,MAAM,MAAM,gBAAgB,GAAG,MAAM,GAAG,MAAM,GAAG,UAAU,CAAA;AAE3D;;gCAEgC;AAChC,MAAM,WAAW,aAAa;IAC5B,8EAA8E;IAC9E,QAAQ,CAAC,iBAAiB,CAAC,EAAE,MAAM,CAAA;IACnC,uDAAuD;IACvD,QAAQ,CAAC,gBAAgB,CAAC,EAAE,MAAM,CAAA;IAClC,kEAAkE;IAClE,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAA;IACvB,iCAAiC;IACjC,QAAQ,CAAC,MAAM,CAAC,EAAE,IAAI,GAAG,OAAO,CAAA;IAChC,6EAA6E;IAC7E,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAA;CACxB;AAED,MAAM,WAAW,cAAc;IAC7B,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAA;IACzB,QAAQ,CAAC,QAAQ,EAAE,gBAAgB,CAAA;IACnC,QAAQ,CAAC,YAAY,CAAC,EAAE,YAAY,CAAA;IACpC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAA;IACvB,iEAAiE;IACjE,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAA;IACvB,QAAQ,CAAC,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CAC5C;AAED,MAAM,WAAW,iBAAiB;IAChC,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAA;IACnB,kGAAkG;IAClG,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAA;IACvB,2EAA2E;IAC3E,OAAO,CAAC,KAAK,EAAE,aAAa,GAAG,cAAc,GAAG,IAAI,CAAA;IACpD,uBAAuB;IACvB,KAAK,IAAI,IAAI,CAAA;CACd;AAED,MAAM,WAAW,qBAAqB;IACpC;mFAC+E;IAC/E,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAA;IAC7B,QAAQ,CAAC,QAAQ,CAAC,EAAE,gBAAgB,CAAA;IACpC,8EAA8E;IAC9E,QAAQ,CAAC,YAAY,CAAC,EAAE,YAAY,CAAA;CACrC;AAED,qGAAqG;AACrG,wBAAgB,sBAAsB,CAAC,IAAI,GAAE,qBAA0B,GAAG,iBAAiB,CA+B1F;AAED,MAAM,WAAW,iBAAiB;IAChC;sCACkC;IAClC,QAAQ,CAAC,aAAa,CAAC,EAAE,MAAM,CAAA;IAC/B,sEAAsE;IACtE,QAAQ,CAAC,aAAa,CAAC,EAAE,MAAM,CAAA;IAC/B,QAAQ,CAAC,QAAQ,CAAC,EAAE,gBAAgB,CAAA;IACpC,QAAQ,CAAC,YAAY,CAAC,EAAE,YAAY,CAAA;CACrC;AAED;wGACwG;AACxG,wBAAgB,kBAAkB,CAAC,IAAI,GAAE,iBAAsB,GAAG,iBAAiB,CAmClF;AAED,MAAM,WAAW,kBAAkB;IACjC,oFAAoF;IACpF,QAAQ,CAAC,SAAS,CAAC,EAAE,MAAM,CAAA;IAC3B,QAAQ,CAAC,QAAQ,CAAC,EAAE,gBAAgB,CAAA;IACpC,QAAQ,CAAC,YAAY,CAAC,EAAE,YAAY,CAAA;CACrC;AAED,6EAA6E;AAC7E,wBAAgB,mBAAmB,CAAC,IAAI,GAAE,kBAAuB,GAAG,iBAAiB,CA2BpF;AAED;iFACiF;AACjF,wBAAgB,UAAU,CACxB,SAAS,EAAE,aAAa,CAAC,iBAAiB,CAAC,EAC3C,KAAK,EAAE,aAAa,GACnB,cAAc,EAAE,CAOlB"}
@@ -1,2 +0,0 @@
1
- export {};
2
- //# sourceMappingURL=index.test.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"index.test.d.ts","sourceRoot":"","sources":["../../src/detectors/index.test.ts"],"names":[],"mappings":""}
@@ -1,100 +0,0 @@
1
- /**
2
- * Causal sweep — WHY did this run fail?
3
- *
4
- * Orchestrates the dormant counterfactual primitives into a responsibility
5
- * report: for each candidate step, run `reps` counterfactual replays per
6
- * mutation (via `runCounterfactual` — the consumer's `CounterfactualRunner`
7
- * is the execution seam) and reduce the per-rep score deltas into a mean
8
- * effect + bootstrap confidence interval (via `confidenceInterval`).
9
- *
10
- * Why `reps` is REQUIRED: a single intervention delta is one stochastic
11
- * draw — LLM re-execution from a prefix is sampled, so one replay cannot
12
- * distinguish "this step caused the failure" from sampling noise. The
13
- * signal is the distribution of deltas across reps; the CI over that
14
- * distribution is what lets a caller say "this step's effect excludes
15
- * zero" instead of eyeballing a point estimate.
16
- *
17
- * Budget discipline: the sweep never silently drops cells. When the
18
- * remaining budget cannot fund a full `reps`-sized cell, the sweep halts
19
- * and every step not fully probed is named in `uncovered`.
20
- */
21
- import { attributeCounterfactuals, type CounterfactualMutation, type CounterfactualRunner } from '../counterfactual';
22
- import type { Span } from '../trace/schema';
23
- import type { TraceStore } from '../trace/store';
24
- import { type TrajectoryStep } from '../trajectory';
25
- /** Stable reference to a trajectory step — carried through reports,
26
- * findings, and corpus records so evidence stays addressable. */
27
- export interface StepRef {
28
- index: number;
29
- spanId: string;
30
- kind: Span['kind'];
31
- name: string;
32
- }
33
- export declare function stepRefOf(step: TrajectoryStep): StepRef;
34
- export interface CausalSweepOptions {
35
- store: TraceStore;
36
- /** The failed run to diagnose. Its `outcome.score` is the baseline every
37
- * counterfactual delta is measured against. */
38
- runId: string;
39
- /** Execution seam — identical contract to `runCounterfactual`: re-runs the
40
- * agent from the mutation point and MUST `endRun` with a numeric score. */
41
- runner: CounterfactualRunner;
42
- /** Trajectory indices to probe. Default: every llm + tool span — the kinds
43
- * the existing `CounterfactualMutation` set targets. */
44
- candidateSteps?: number[];
45
- /**
46
- * Mutations to probe a given step with. Returned mutations MUST target
47
- * `step.index`. Default probes are the payload-free existing kinds:
48
- * - tool span → `swap-tool-result` with `newResult: null` (knockout:
49
- * how much did the run depend on this tool's information?)
50
- * - llm span → `truncate-after` (re-roll: how much did the realized
51
- * turn deviate from the policy's typical continuation?)
52
- * `swap-model` / `inject-system-message` need consumer payloads, so they
53
- * are opt-in via this callback.
54
- */
55
- mutationsPerStep?: (step: TrajectoryStep) => CounterfactualMutation[];
56
- /** Replays per (step, mutation) cell. Minimum 2 — see module doc. */
57
- reps: number;
58
- /** Hard cap on total counterfactual replays across the whole sweep. */
59
- budget: number;
60
- /** Seed for the bootstrap CI resampler. Deterministic default so two
61
- * sweeps over the same deltas report identical intervals. */
62
- ciSeed?: number;
63
- /** Bootstrap CI confidence level. Default 0.95. */
64
- ciConfidence?: number;
65
- }
66
- export interface StepResponsibility {
67
- stepRef: StepRef;
68
- mutationKind: CounterfactualMutation['kind'];
69
- /** Mean of per-rep score deltas (counterfactual − original). */
70
- meanEffect: number;
71
- /** Bootstrap CI over the per-rep deltas. */
72
- ci: {
73
- mean: number;
74
- lower: number;
75
- upper: number;
76
- };
77
- /** `ci.lower > 0 || ci.upper < 0` — the effect is distinguishable from noise. */
78
- ciExcludesZero: boolean;
79
- reps: number;
80
- /** Raw per-rep deltas — downstream evidence, never re-derived. */
81
- deltas: number[];
82
- /** Replay run ids (layer='meta', parentRunId=original) for audit. */
83
- counterfactualRunIds: string[];
84
- }
85
- export interface CausalResponsibilityReport {
86
- runId: string;
87
- originalScore: number;
88
- /** Ranked by |meanEffect| descending — the blame ordering. */
89
- steps: StepResponsibility[];
90
- /** Kind-level aggregate from the existing `attributeCounterfactuals`. */
91
- byMutationKind: ReturnType<typeof attributeCounterfactuals>;
92
- replaysUsed: number;
93
- budget: number;
94
- /** Steps planned but not fully probed before the budget ran out.
95
- * Named, never silent: an absent step is "no effect found"; an
96
- * uncovered step is "not measured". */
97
- uncovered: StepRef[];
98
- }
99
- export declare function causalSweep(opts: CausalSweepOptions): Promise<CausalResponsibilityReport>;
100
- //# sourceMappingURL=causal-sweep.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"causal-sweep.d.ts","sourceRoot":"","sources":["../../src/diagnose/causal-sweep.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH,OAAO,EACL,wBAAwB,EACxB,KAAK,sBAAsB,EAE3B,KAAK,oBAAoB,EAE1B,MAAM,mBAAmB,CAAA;AAG1B,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,iBAAiB,CAAA;AAC3C,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,gBAAgB,CAAA;AAChD,OAAO,EAAoC,KAAK,cAAc,EAAE,MAAM,eAAe,CAAA;AAErF;kEACkE;AAClE,MAAM,WAAW,OAAO;IACtB,KAAK,EAAE,MAAM,CAAA;IACb,MAAM,EAAE,MAAM,CAAA;IACd,IAAI,EAAE,IAAI,CAAC,MAAM,CAAC,CAAA;IAClB,IAAI,EAAE,MAAM,CAAA;CACb;AAED,wBAAgB,SAAS,CAAC,IAAI,EAAE,cAAc,GAAG,OAAO,CAOvD;AAED,MAAM,WAAW,kBAAkB;IACjC,KAAK,EAAE,UAAU,CAAA;IACjB;oDACgD;IAChD,KAAK,EAAE,MAAM,CAAA;IACb;gFAC4E;IAC5E,MAAM,EAAE,oBAAoB,CAAA;IAC5B;6DACyD;IACzD,cAAc,CAAC,EAAE,MAAM,EAAE,CAAA;IACzB;;;;;;;;;OASG;IACH,gBAAgB,CAAC,EAAE,CAAC,IAAI,EAAE,cAAc,KAAK,sBAAsB,EAAE,CAAA;IACrE,qEAAqE;IACrE,IAAI,EAAE,MAAM,CAAA;IACZ,uEAAuE;IACvE,MAAM,EAAE,MAAM,CAAA;IACd;kEAC8D;IAC9D,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,mDAAmD;IACnD,YAAY,CAAC,EAAE,MAAM,CAAA;CACtB;AAED,MAAM,WAAW,kBAAkB;IACjC,OAAO,EAAE,OAAO,CAAA;IAChB,YAAY,EAAE,sBAAsB,CAAC,MAAM,CAAC,CAAA;IAC5C,gEAAgE;IAChE,UAAU,EAAE,MAAM,CAAA;IAClB,4CAA4C;IAC5C,EAAE,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,CAAA;IAClD,iFAAiF;IACjF,cAAc,EAAE,OAAO,CAAA;IACvB,IAAI,EAAE,MAAM,CAAA;IACZ,kEAAkE;IAClE,MAAM,EAAE,MAAM,EAAE,CAAA;IAChB,qEAAqE;IACrE,oBAAoB,EAAE,MAAM,EAAE,CAAA;CAC/B;AAED,MAAM,WAAW,0BAA0B;IACzC,KAAK,EAAE,MAAM,CAAA;IACb,aAAa,EAAE,MAAM,CAAA;IACrB,8DAA8D;IAC9D,KAAK,EAAE,kBAAkB,EAAE,CAAA;IAC3B,yEAAyE;IACzE,cAAc,EAAE,UAAU,CAAC,OAAO,wBAAwB,CAAC,CAAA;IAC3D,WAAW,EAAE,MAAM,CAAA;IACnB,MAAM,EAAE,MAAM,CAAA;IACd;;4CAEwC;IACxC,SAAS,EAAE,OAAO,EAAE,CAAA;CACrB;AAcD,wBAAsB,WAAW,CAAC,IAAI,EAAE,kBAAkB,GAAG,OAAO,CAAC,0BAA0B,CAAC,CAmG/F"}
@@ -1,36 +0,0 @@
1
- /**
2
- * Diagnose chain — WHY a run failed, WHAT should have happened, HOW to
3
- * make it happen.
4
- *
5
- * The full remediation pipeline this subpath closes:
6
- *
7
- * fuzz finds → sweep blames → repair prescribes (validated) →
8
- * findings / corpus / invariant remediate → gates verify
9
- *
10
- * Three stages, all orchestration over existing primitives — nothing here
11
- * re-implements replay, mutation, or attribution:
12
- *
13
- * 1. `causalSweep` — WHY. Runs `reps` counterfactual replays per
14
- * (step, mutation) cell through `runCounterfactual` (the consumer's
15
- * `CounterfactualRunner` is the execution seam) and reduces the
16
- * per-rep deltas into a responsibility ranking with bootstrap CIs
17
- * (`confidenceInterval`). Budget-bounded; unprobed steps are named
18
- * in `uncovered`, never dropped.
19
- * 2. `prescribeRepair` — WHAT SHOULD HAVE HAPPENED. Consumer-supplied
20
- * `proposeFix` (LLM-backed in live use) proposes candidate mutations
21
- * for the blamed steps; each candidate is machine-verified by
22
- * replaying WITH it. Only candidates whose every validation rep
23
- * crosses `flipThreshold` become repairs; the rest are rejected
24
- * with a typed reason.
25
- * 3. Remediation adapters — HOW. `toAnalystFindings` feeds the analyst
26
- * registry, `toCorpusRecord` pins the failure as a permanent corpus
27
- * scenario, `suggestInvariant` emits the trace-contracts hint shape.
28
- */
29
- export type { CounterfactualContext, CounterfactualMutation, CounterfactualResult, CounterfactualRunner, } from '../counterfactual';
30
- export type { CausalResponsibilityReport, CausalSweepOptions, StepRef, StepResponsibility, } from './causal-sweep';
31
- export { causalSweep, stepRefOf } from './causal-sweep';
32
- export type { InvariantHint } from './remediation';
33
- export { DIAGNOSE_ANALYST_ID, describeMutation, severityFromEffect, suggestInvariant, toAnalystFindings, toCorpusRecord, } from './remediation';
34
- export type { PrescribeRepairOptions, RejectedRepair, RepairContext, RepairReport, ValidatedRepair, } from './repair';
35
- export { prescribeRepair } from './repair';
36
- //# sourceMappingURL=index.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/diagnose/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AAIH,YAAY,EACV,qBAAqB,EACrB,sBAAsB,EACtB,oBAAoB,EACpB,oBAAoB,GACrB,MAAM,mBAAmB,CAAA;AAC1B,YAAY,EACV,0BAA0B,EAC1B,kBAAkB,EAClB,OAAO,EACP,kBAAkB,GACnB,MAAM,gBAAgB,CAAA;AACvB,OAAO,EAAE,WAAW,EAAE,SAAS,EAAE,MAAM,gBAAgB,CAAA;AACvD,YAAY,EAAE,aAAa,EAAE,MAAM,eAAe,CAAA;AAClD,OAAO,EACL,mBAAmB,EACnB,gBAAgB,EAChB,kBAAkB,EAClB,gBAAgB,EAChB,iBAAiB,EACjB,cAAc,GACf,MAAM,eAAe,CAAA;AACtB,YAAY,EACV,sBAAsB,EACtB,cAAc,EACd,aAAa,EACb,YAAY,EACZ,eAAe,GAChB,MAAM,UAAU,CAAA;AACjB,OAAO,EAAE,eAAe,EAAE,MAAM,UAAU,CAAA"}
@@ -1,68 +0,0 @@
1
- /**
2
- * Remediation adapters — HOW DO WE MAKE IT HAPPEN?
3
- *
4
- * The diagnose chain ends by feeding existing improvement machinery,
5
- * not by building new machinery:
6
- *
7
- * - `toAnalystFindings` → the analyst contract (`makeFinding`), so
8
- * responsibility evidence flows into the same registry / steering /
9
- * diff pipeline every other analyst feeds.
10
- * - `toCorpusRecord` → the RL corpus (`CorpusRecord`), pinning the
11
- * diagnosed failure + validated repair as a permanent scenario.
12
- * - `suggestInvariant` → a plain-data hint in the shape the
13
- * trace-contracts machinery consumes (`never` / `without` clauses).
14
- */
15
- import type { AnalystFinding, AnalystSeverity } from '../analyst/types';
16
- import type { CounterfactualMutation } from '../counterfactual';
17
- import type { CorpusRecord } from '../rl/corpus';
18
- import type { RunRecord } from '../run-record';
19
- import type { CausalResponsibilityReport, StepResponsibility } from './causal-sweep';
20
- import type { RepairReport, ValidatedRepair } from './repair';
21
- export declare const DIAGNOSE_ANALYST_ID = "diagnose-causal-sweep";
22
- /** Severity from causal effect size. Effects whose CI includes zero are
23
- * 'info' regardless of magnitude — an indistinguishable-from-noise effect
24
- * must not steer remediation priority. */
25
- export declare function severityFromEffect(responsibility: StepResponsibility): AnalystSeverity;
26
- /** Deterministic human-readable rendering of a mutation — used in
27
- * recommended actions, corpus completions, and invariant hints. */
28
- export declare function describeMutation(mutation: CounterfactualMutation): string;
29
- /**
30
- * Lift a responsibility report (and optionally its validated repairs) into
31
- * `AnalystFinding`s via the real `makeFinding` factory. One finding per
32
- * probed step; a validated repair for that step upgrades the finding with
33
- * a `recommended_action` + the replay-validation evidence.
34
- *
35
- * Findings are OBSERVED causal probes (replay deltas), not judge verdicts,
36
- * so `derived_from_judge` stays unset and they may steer.
37
- */
38
- export declare function toAnalystFindings(report: CausalResponsibilityReport, repairs?: RepairReport): AnalystFinding[];
39
- /**
40
- * Pin the diagnosed failure as a permanent corpus scenario. Takes the
41
- * original run's `RunRecord` projection plus a validated repair and emits
42
- * a fresh `CorpusRecord` (new runId, so corpus dedup keeps both the raw
43
- * failure and the diagnosed entry).
44
- *
45
- * `completion` defaults to the validated mutation's rendering — "what
46
- * should have happened" in machine-derived form. Supply `prompt` (and
47
- * optionally a richer `completion`) when the trajectory text is available
48
- * so the record is harvestable by `buildDatasetFromCorpus`.
49
- */
50
- export declare function toCorpusRecord(run: RunRecord, repair: ValidatedRepair, opts?: {
51
- prompt?: string;
52
- completion?: string;
53
- }): CorpusRecord;
54
- /** Plain-data invariant hint. The trace-contracts machinery consumes this
55
- * shape: `never` is a pattern that must not appear in a passing trace;
56
- * `without` is a guard whose absence makes the failure reachable. */
57
- export interface InvariantHint {
58
- description: string;
59
- never?: string;
60
- without?: string;
61
- }
62
- /**
63
- * Derive an invariant hint from a validated repair. Deterministic per
64
- * mutation kind — the hint names the contract a trace must satisfy so
65
- * the diagnosed failure cannot silently recur.
66
- */
67
- export declare function suggestInvariant(repair: ValidatedRepair): InvariantHint;
68
- //# sourceMappingURL=remediation.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"remediation.d.ts","sourceRoot":"","sources":["../../src/diagnose/remediation.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,KAAK,EAAE,cAAc,EAAE,eAAe,EAAe,MAAM,kBAAkB,CAAA;AAEpF,OAAO,KAAK,EAAE,sBAAsB,EAAE,MAAM,mBAAmB,CAAA;AAE/D,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,cAAc,CAAA;AAChD,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,eAAe,CAAA;AAE9C,OAAO,KAAK,EAAE,0BAA0B,EAAE,kBAAkB,EAAE,MAAM,gBAAgB,CAAA;AACpF,OAAO,KAAK,EAAE,YAAY,EAAE,eAAe,EAAE,MAAM,UAAU,CAAA;AAE7D,eAAO,MAAM,mBAAmB,0BAA0B,CAAA;AAE1D;;2CAE2C;AAC3C,wBAAgB,kBAAkB,CAAC,cAAc,EAAE,kBAAkB,GAAG,eAAe,CAOtF;AAED;oEACoE;AACpE,wBAAgB,gBAAgB,CAAC,QAAQ,EAAE,sBAAsB,GAAG,MAAM,CAazE;AAED;;;;;;;;GAQG;AACH,wBAAgB,iBAAiB,CAC/B,MAAM,EAAE,0BAA0B,EAClC,OAAO,CAAC,EAAE,YAAY,GACrB,cAAc,EAAE,CA+ClB;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,cAAc,CAC5B,GAAG,EAAE,SAAS,EACd,MAAM,EAAE,eAAe,EACvB,IAAI,GAAE;IAAE,MAAM,CAAC,EAAE,MAAM,CAAC;IAAC,UAAU,CAAC,EAAE,MAAM,CAAA;CAAO,GAClD,YAAY,CAqBd;AAED;;sEAEsE;AACtE,MAAM,WAAW,aAAa;IAC5B,WAAW,EAAE,MAAM,CAAA;IACnB,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,OAAO,CAAC,EAAE,MAAM,CAAA;CACjB;AAED;;;;GAIG;AACH,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,eAAe,GAAG,aAAa,CAoCvE"}