@tangle-network/agent-eval 0.215.1 → 0.216.1

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 (400) hide show
  1. package/CHANGELOG.md +51 -0
  2. package/dist/adapters/http.d.ts +3 -3
  3. package/dist/{agent-profile-DmzL8KaJ.d.ts → agent-profile-VFcQncbW.d.ts} +5 -5
  4. package/dist/{agent-profile-DmzL8KaJ.d.ts.map → agent-profile-VFcQncbW.d.ts.map} +1 -1
  5. package/dist/{agent-profile-cell-BYc0bxsL.js → agent-profile-cell-BK8mlqlq.js} +3 -3
  6. package/dist/{agent-profile-cell-BYc0bxsL.js.map → agent-profile-cell-BK8mlqlq.js.map} +1 -1
  7. package/dist/agreement-irr-BEWLIF3_.js +186 -0
  8. package/dist/agreement-irr-BEWLIF3_.js.map +1 -0
  9. package/dist/analyst/index.d.ts +487 -14
  10. package/dist/analyst/index.d.ts.map +1 -1
  11. package/dist/analyst/index.js +1547 -12
  12. package/dist/analyst/index.js.map +1 -1
  13. package/dist/{analyze-runs-3xpby3p2.js → analyze-runs-C84irm4i.js} +608 -604
  14. package/dist/analyze-runs-C84irm4i.js.map +1 -0
  15. package/dist/{analyze-runs-BhFLOztS.d.ts → analyze-runs-D68fd-su.d.ts} +17 -15
  16. package/dist/analyze-runs-D68fd-su.d.ts.map +1 -0
  17. package/dist/{baseline-Dm8966du.js → baseline-Udi_PLuT.js} +5 -5
  18. package/dist/{baseline-Dm8966du.js.map → baseline-Udi_PLuT.js.map} +1 -1
  19. package/dist/{benchmark-BywPeAUo.js → benchmark-CXmVVHH2.js} +2 -2
  20. package/dist/{benchmark-BywPeAUo.js.map → benchmark-CXmVVHH2.js.map} +1 -1
  21. package/dist/{benchmark-command-CgpGi7vY.js → benchmark-command-By0vh8q7.js} +71 -70
  22. package/dist/benchmark-command-By0vh8q7.js.map +1 -0
  23. package/dist/benchmarks/index.d.ts +5 -5
  24. package/dist/benchmarks/index.js +3 -3
  25. package/dist/builder-eval/index.js +2 -2
  26. package/dist/campaign/index.d.ts +7 -7
  27. package/dist/campaign/index.js +14 -14
  28. package/dist/{campaign-CpnsCgfz.js → campaign-BudZ70zo.js} +245 -247
  29. package/dist/campaign-BudZ70zo.js.map +1 -0
  30. package/dist/{campaign-evidence-OcUciD-N.js → campaign-evidence-Clrn4hBV.js} +8 -8
  31. package/dist/campaign-evidence-Clrn4hBV.js.map +1 -0
  32. package/dist/{canonical-BsZVgBhW.js → canonical-B2RqLofv.js} +2 -2
  33. package/dist/{canonical-BsZVgBhW.js.map → canonical-B2RqLofv.js.map} +1 -1
  34. package/dist/{chat-client-DthPOYGd.js → chat-client-l-MVDlDk.js} +2 -2
  35. package/dist/{chat-client-DthPOYGd.js.map → chat-client-l-MVDlDk.js.map} +1 -1
  36. package/dist/{chat-json-call-BrKsyspg.js → chat-json-call-puMArmtW.js} +2 -2
  37. package/dist/{chat-json-call-BrKsyspg.js.map → chat-json-call-puMArmtW.js.map} +1 -1
  38. package/dist/{claim-BXcZw-jo.js → claim-q4HQhK1G.js} +4 -4
  39. package/dist/claim-q4HQhK1G.js.map +1 -0
  40. package/dist/cli.js +9 -9
  41. package/dist/cli.js.map +1 -1
  42. package/dist/cluster-bootstrap-qcUWYFl-.js +66 -0
  43. package/dist/cluster-bootstrap-qcUWYFl-.js.map +1 -0
  44. package/dist/contract/index.d.ts +11 -11
  45. package/dist/contract/index.js +1023 -1079
  46. package/dist/contract/index.js.map +1 -1
  47. package/dist/{cost-ledger-D2FnJPBF.js → cost-ledger-BP8V4kUQ.js} +835 -816
  48. package/dist/cost-ledger-BP8V4kUQ.js.map +1 -0
  49. package/dist/{cost-ledger-CO7UyEcQ.d.ts → cost-ledger-Jk62OBfF.d.ts} +12 -47
  50. package/dist/cost-ledger-Jk62OBfF.d.ts.map +1 -0
  51. package/dist/{counterfactual-DtMWvn7D.js → counterfactual-Dn3veueY.js} +2 -2
  52. package/dist/{counterfactual-DtMWvn7D.js.map → counterfactual-Dn3veueY.js.map} +1 -1
  53. package/dist/{default-registry-KP4-Tolz.js → default-registry-COhk9g36.js} +744 -759
  54. package/dist/default-registry-COhk9g36.js.map +1 -0
  55. package/dist/{default-registry-B4B7GMvI.d.ts → default-registry-DzHiGruW.d.ts} +6 -6
  56. package/dist/{default-registry-B4B7GMvI.d.ts.map → default-registry-DzHiGruW.d.ts.map} +1 -1
  57. package/dist/{define-agent-eval-BdfnqXeP.js → define-agent-eval-CNTPzIkp.js} +10 -10
  58. package/dist/define-agent-eval-CNTPzIkp.js.map +1 -0
  59. package/dist/{define-agent-eval-DfXvNho5.d.ts → define-agent-eval-WSck0CZJ.d.ts} +8 -8
  60. package/dist/{define-agent-eval-DfXvNho5.d.ts.map → define-agent-eval-WSck0CZJ.d.ts.map} +1 -1
  61. package/dist/{descriptive-CnU8Yq29.js → descriptive-DboxPZrG.js} +2 -2
  62. package/dist/{descriptive-CnU8Yq29.js.map → descriptive-DboxPZrG.js.map} +1 -1
  63. package/dist/diagnosis/index.d.ts +3 -3
  64. package/dist/diagnosis/index.js +4 -4
  65. package/dist/{dspy-rlm-engine-DZfHM7kT.js → dspy-rlm-engine-LRCu61m5.js} +3 -3
  66. package/dist/{dspy-rlm-engine-DZfHM7kT.js.map → dspy-rlm-engine-LRCu61m5.js.map} +1 -1
  67. package/dist/{effect-sizes-DiH8MGOH.js → effect-sizes-BhzEpXMc.js} +2 -2
  68. package/dist/{effect-sizes-DiH8MGOH.js.map → effect-sizes-BhzEpXMc.js.map} +1 -1
  69. package/dist/{engine-C37HS2kr.d.ts → engine-3A_4jOlI.d.ts} +5 -5
  70. package/dist/{engine-C37HS2kr.d.ts.map → engine-3A_4jOlI.d.ts.map} +1 -1
  71. package/dist/{equal-terms-Blx9Zk5I.js → equal-terms-BtdXayCJ.js} +2 -2
  72. package/dist/{equal-terms-Blx9Zk5I.js.map → equal-terms-BtdXayCJ.js.map} +1 -1
  73. package/dist/{errors-Dngq5h35.js → errors-DVZiqS7K.js} +6 -2
  74. package/dist/{errors-Dngq5h35.js.map → errors-DVZiqS7K.js.map} +1 -1
  75. package/dist/{eval-campaign-4byTtqrf.js → eval-campaign-90jeaCAI.js} +6 -6
  76. package/dist/{eval-campaign-4byTtqrf.js.map → eval-campaign-90jeaCAI.js.map} +1 -1
  77. package/dist/evaluation-6Yoi74gM.d.ts +75 -0
  78. package/dist/evaluation-6Yoi74gM.d.ts.map +1 -0
  79. package/dist/evaluation.d.ts +2 -74
  80. package/dist/evaluation.js +1 -1
  81. package/dist/{exact-types-Dlb73wN4.d.ts → exact-types-ClugMJfI.d.ts} +2 -2
  82. package/dist/{exact-types-Dlb73wN4.d.ts.map → exact-types-ClugMJfI.d.ts.map} +1 -1
  83. package/dist/experiment/index.d.ts +404 -3
  84. package/dist/experiment/index.d.ts.map +1 -1
  85. package/dist/experiment/index.js +313 -19
  86. package/dist/experiment/index.js.map +1 -1
  87. package/dist/{external-optimizer-process-GngofQML.js → external-optimizer-process-BjMGUuds.js} +3 -3
  88. package/dist/{external-optimizer-process-GngofQML.js.map → external-optimizer-process-BjMGUuds.js.map} +1 -1
  89. package/dist/{external-optimizer-subprocess-BUxifb8g.js → external-optimizer-subprocess-BqW0hHcI.js} +791 -532
  90. package/dist/external-optimizer-subprocess-BqW0hHcI.js.map +1 -0
  91. package/dist/{feedback-trajectory-BLN_6fEs.d.ts → feedback-trajectory-C4EG0nNH.d.ts} +5 -8
  92. package/dist/feedback-trajectory-C4EG0nNH.d.ts.map +1 -0
  93. package/dist/fuzz.d.ts +1 -1
  94. package/dist/fuzz.js +3 -3
  95. package/dist/hosted/index.js +1 -1
  96. package/dist/{index-CPJJr98o.d.ts → index-BFRO6chs.d.ts} +3 -2
  97. package/dist/index-BFRO6chs.d.ts.map +1 -0
  98. package/dist/{index-Ch9VxCpK.d.ts → index-BO41eGj5.d.ts} +2 -2
  99. package/dist/{index-Ch9VxCpK.d.ts.map → index-BO41eGj5.d.ts.map} +1 -1
  100. package/dist/{index-B3WBm953.d.ts → index-CyszSg4O.d.ts} +165 -162
  101. package/dist/index-CyszSg4O.d.ts.map +1 -0
  102. package/dist/{index-B7m2uzJL.d.ts → index-DPcngtwt.d.ts} +30 -17
  103. package/dist/index-DPcngtwt.d.ts.map +1 -0
  104. package/dist/{index-tYLUvk_w.d.ts → index-DV-jSzxF.d.ts} +9 -9
  105. package/dist/{index-tYLUvk_w.d.ts.map → index-DV-jSzxF.d.ts.map} +1 -1
  106. package/dist/index.d.ts +1969 -1871
  107. package/dist/index.d.ts.map +1 -1
  108. package/dist/index.js +162 -201
  109. package/dist/index.js.map +1 -1
  110. package/dist/{integrity-BvnVARVl.js → integrity-BXOsAb3i.js} +2 -2
  111. package/dist/{integrity-BvnVARVl.js.map → integrity-BXOsAb3i.js.map} +1 -1
  112. package/dist/{integrity-CCAxGVK6.d.ts → integrity-DAMEDFaC.d.ts} +2 -2
  113. package/dist/{integrity-CCAxGVK6.d.ts.map → integrity-DAMEDFaC.d.ts.map} +1 -1
  114. package/dist/{integrity-BV1ZDf42.js → integrity-dLiDe2kD.js} +2 -5
  115. package/dist/integrity-dLiDe2kD.js.map +1 -0
  116. package/dist/{internal-CRHwS9iz.js → internal-Epl-YBZW.js} +3 -3
  117. package/dist/{internal-CRHwS9iz.js.map → internal-Epl-YBZW.js.map} +1 -1
  118. package/dist/interval-CELpye1n.js +219 -0
  119. package/dist/interval-CELpye1n.js.map +1 -0
  120. package/dist/interval-ohuOJurX.d.ts +47 -0
  121. package/dist/interval-ohuOJurX.d.ts.map +1 -0
  122. package/dist/{journal-DvKCvjty.js → journal-DfRz9kps.js} +2 -2
  123. package/dist/{journal-DvKCvjty.js.map → journal-DfRz9kps.js.map} +1 -1
  124. package/dist/{judge-calibration-N44V04Se.js → judge-calibration-BpCy5XQl.js} +3 -3
  125. package/dist/{judge-calibration-N44V04Se.js.map → judge-calibration-BpCy5XQl.js.map} +1 -1
  126. package/dist/{judge-gate-DNcUrRFA.d.ts → judge-gate-CE7CQAgt.d.ts} +3 -3
  127. package/dist/{judge-gate-DNcUrRFA.d.ts.map → judge-gate-CE7CQAgt.d.ts.map} +1 -1
  128. package/dist/{judge-gate-B1qYnkpi.js → judge-gate-iY9s4L44.js} +3 -3
  129. package/dist/judge-gate-iY9s4L44.js.map +1 -0
  130. package/dist/{kind-factory-AaEhUADH.js → kind-factory-BobI5Jy3.js} +5 -5
  131. package/dist/{kind-factory-AaEhUADH.js.map → kind-factory-BobI5Jy3.js.map} +1 -1
  132. package/dist/ledger-core/index.js +3 -3
  133. package/dist/{ledger-core-B2hahzrt.js → ledger-core-Gq7yp8lP.js} +4 -4
  134. package/dist/{ledger-core-B2hahzrt.js.map → ledger-core-Gq7yp8lP.js.map} +1 -1
  135. package/dist/{llm-client-DQveoxSM.js → llm-client-CyBAoYiX.js} +192 -219
  136. package/dist/llm-client-CyBAoYiX.js.map +1 -0
  137. package/dist/{llm-judge-Cd5s5Ei6.js → llm-judge-CG12SQJq.js} +499 -921
  138. package/dist/llm-judge-CG12SQJq.js.map +1 -0
  139. package/dist/matrix/index.d.ts +2 -2
  140. package/dist/matrix/index.js +1 -1
  141. package/dist/{matrix-DXVAJMLD.d.ts → matrix-DSrSxjjy.d.ts} +4 -4
  142. package/dist/{matrix-DXVAJMLD.d.ts.map → matrix-DSrSxjjy.d.ts.map} +1 -1
  143. package/dist/{matrix-_BV9-Lk5.js → matrix-Dm0BQW7U.js} +7 -11
  144. package/dist/matrix-Dm0BQW7U.js.map +1 -0
  145. package/dist/{paired-tests-jbwPdSz5.js → mean-DHoMiudD.js} +11 -5
  146. package/dist/mean-DHoMiudD.js.map +1 -0
  147. package/dist/meta-eval/index.d.ts +3 -3
  148. package/dist/meta-eval/index.js +9 -70
  149. package/dist/meta-eval/index.js.map +1 -1
  150. package/dist/mint-DNU8QPG4.d.ts +81 -0
  151. package/dist/mint-DNU8QPG4.d.ts.map +1 -0
  152. package/dist/{mint-Csor1BLJ.js → mint-QaD5cA0u.js} +3 -3
  153. package/dist/{mint-Csor1BLJ.js.map → mint-QaD5cA0u.js.map} +1 -1
  154. package/dist/{multiplicity-D9cjmStq.js → multiplicity-DQUFcbW4.js} +2 -2
  155. package/dist/{multiplicity-D9cjmStq.js.map → multiplicity-DQUFcbW4.js.map} +1 -1
  156. package/dist/multishot/golden/index.d.ts +1 -1
  157. package/dist/multishot/golden/index.d.ts.map +1 -1
  158. package/dist/multishot/golden/index.js +3 -6
  159. package/dist/multishot/golden/index.js.map +1 -1
  160. package/dist/multishot/index.d.ts +2 -2
  161. package/dist/multishot/index.js +1 -1
  162. package/dist/nonnegative-number-DvdMxg_f.js +9 -0
  163. package/dist/nonnegative-number-DvdMxg_f.js.map +1 -0
  164. package/dist/openapi.json +1 -1
  165. package/dist/{otel-spans-C8YIJWy1.js → otel-spans-B9tUXRHF.js} +3 -3
  166. package/dist/{otel-spans-C8YIJWy1.js.map → otel-spans-B9tUXRHF.js.map} +1 -1
  167. package/dist/{paired-arms-DzKn0TfY.js → paired-arms-CwMGJsrS.js} +5 -5
  168. package/dist/{paired-arms-DzKn0TfY.js.map → paired-arms-CwMGJsrS.js.map} +1 -1
  169. package/dist/{paired-binary-Q36bhn2R.js → paired-binary-CvpFHfKZ.js} +2 -2
  170. package/dist/{paired-binary-Q36bhn2R.js.map → paired-binary-CvpFHfKZ.js.map} +1 -1
  171. package/dist/{paired-promotion-decision-BN6IQTiu.js → paired-promotion-decision-DsAPYGCm.js} +3 -3
  172. package/dist/{paired-promotion-decision-BN6IQTiu.js.map → paired-promotion-decision-DsAPYGCm.js.map} +1 -1
  173. package/dist/{search-kernel-BztlERLv.js → pareto-Cm6ndrWN.js} +883 -861
  174. package/dist/pareto-Cm6ndrWN.js.map +1 -0
  175. package/dist/pipelines/index.d.ts +2 -2
  176. package/dist/pipelines/index.js +3 -3
  177. package/dist/{power-and-mde-fK5Qocwo.js → power-and-mde-Hj2ubb4B.js} +2 -2
  178. package/dist/{power-and-mde-fK5Qocwo.js.map → power-and-mde-Hj2ubb4B.js.map} +1 -1
  179. package/dist/predicate-FdKmElA_.js +38 -0
  180. package/dist/predicate-FdKmElA_.js.map +1 -0
  181. package/dist/{prime-protocol-DcAK9pC8.js → prime-protocol-CHKgA3a6.js} +2 -2
  182. package/dist/{prime-protocol-DcAK9pC8.js.map → prime-protocol-CHKgA3a6.js.map} +1 -1
  183. package/dist/{prime-protocol-BHia9vDz.d.ts → prime-protocol-HiB9HX7E.d.ts} +3 -3
  184. package/dist/{prime-protocol-BHia9vDz.d.ts.map → prime-protocol-HiB9HX7E.d.ts.map} +1 -1
  185. package/dist/{produced-state-BGp4BX8t.js → produced-state-BsDmnwxc.js} +8 -8
  186. package/dist/{produced-state-BGp4BX8t.js.map → produced-state-BsDmnwxc.js.map} +1 -1
  187. package/dist/profile-cell.js +1 -1
  188. package/dist/{release-confidence-BK2QCB5M.d.ts → promotion-gate-BawrdPYr.d.ts} +102 -102
  189. package/dist/promotion-gate-BawrdPYr.d.ts.map +1 -0
  190. package/dist/{promotion-policy-CWW35UWq.js → promotion-policy-KO57gbyV.js} +4 -4
  191. package/dist/{promotion-policy-CWW35UWq.js.map → promotion-policy-KO57gbyV.js.map} +1 -1
  192. package/dist/{proposal-findings-CXblsskH.js → proposal-findings-CY1xKn_a.js} +2 -36
  193. package/dist/proposal-findings-CY1xKn_a.js.map +1 -0
  194. package/dist/{query-B1D1qp2Z.js → query-BUCZCK7F.js} +3 -3
  195. package/dist/{query-B1D1qp2Z.js.map → query-BUCZCK7F.js.map} +1 -1
  196. package/dist/{random-Dn5fPWkt.js → random-ByUSnV9w.js} +2 -2
  197. package/dist/{random-Dn5fPWkt.js.map → random-ByUSnV9w.js.map} +1 -1
  198. package/dist/{raw-provider-sink-rB2DYxr6.js → raw-provider-sink-DoHz4JNG.js} +2 -2
  199. package/dist/{raw-provider-sink-rB2DYxr6.js.map → raw-provider-sink-DoHz4JNG.js.map} +1 -1
  200. package/dist/{redact-Dk1ZdJEm.js → redact-B5xXKqQ3.js} +97 -98
  201. package/dist/redact-B5xXKqQ3.js.map +1 -0
  202. package/dist/{redact-Cveoq6fJ.d.ts → redact-mtCCcT8z.d.ts} +4 -20
  203. package/dist/redact-mtCCcT8z.d.ts.map +1 -0
  204. package/dist/{registry-CMYIaDgS.d.ts → registry-D7_hm2fK.d.ts} +11 -13
  205. package/dist/registry-D7_hm2fK.d.ts.map +1 -0
  206. package/dist/{release-confidence-Dbe9AS_X.js → release-confidence-CpuP65S5.js} +5 -5
  207. package/dist/{release-confidence-Dbe9AS_X.js.map → release-confidence-CpuP65S5.js.map} +1 -1
  208. package/dist/{report-command-CtIIsQKP.js → report-command-SC9nqb4M.js} +128 -177
  209. package/dist/report-command-SC9nqb4M.js.map +1 -0
  210. package/dist/reporting-html-Kvlc8oMt.js +100 -0
  211. package/dist/reporting-html-Kvlc8oMt.js.map +1 -0
  212. package/dist/reporting.d.ts +3 -3
  213. package/dist/reporting.js +6 -6
  214. package/dist/{researcher-f6mZhh58.d.ts → researcher-CEcfku-T.d.ts} +5 -5
  215. package/dist/{researcher-f6mZhh58.d.ts.map → researcher-CEcfku-T.d.ts.map} +1 -1
  216. package/dist/reward-CpCmAM8L.d.ts +78 -0
  217. package/dist/reward-CpCmAM8L.d.ts.map +1 -0
  218. package/dist/{reward-hacking-BFcQgkux.js → reward-hacking-E0d-WVcP.js} +2 -2
  219. package/dist/{reward-hacking-BFcQgkux.js.map → reward-hacking-E0d-WVcP.js.map} +1 -1
  220. package/dist/{reward-hacking-DHA7c1l8.d.ts → reward-hacking-RVdZ6Kkt.d.ts} +2 -2
  221. package/dist/{reward-hacking-DHA7c1l8.d.ts.map → reward-hacking-RVdZ6Kkt.d.ts.map} +1 -1
  222. package/dist/rl.d.ts +8 -8
  223. package/dist/rl.js +14 -14
  224. package/dist/rollout/index.d.ts +3 -2
  225. package/dist/rollout/index.js +2 -2
  226. package/dist/{rollout-D9bKKBnZ.js → rollout-DZV1t2Wg.js} +2 -2
  227. package/dist/{rollout-D9bKKBnZ.js.map → rollout-DZV1t2Wg.js.map} +1 -1
  228. package/dist/round-decimal-COrKglMh.js +10 -0
  229. package/dist/round-decimal-COrKglMh.js.map +1 -0
  230. package/dist/{rubric-predictive-validity-2SgHxoa5.d.ts → rubric-predictive-validity-C03EC7ZE.d.ts} +2 -2
  231. package/dist/{rubric-predictive-validity-2SgHxoa5.d.ts.map → rubric-predictive-validity-C03EC7ZE.d.ts.map} +1 -1
  232. package/dist/{rubric-predictive-validity-BsgkBfT0.js → rubric-predictive-validity-DaFW70y9.js} +3 -3
  233. package/dist/{rubric-predictive-validity-BsgkBfT0.js.map → rubric-predictive-validity-DaFW70y9.js.map} +1 -1
  234. package/dist/{run-record-DlTZ_lqe.js → run-record-Boc5IUTp.js} +4 -4
  235. package/dist/{run-record-DlTZ_lqe.js.map → run-record-Boc5IUTp.js.map} +1 -1
  236. package/dist/{run-record-D1Hw-Jml.d.ts → run-record-C1E24G2A.d.ts} +2 -2
  237. package/dist/{run-record-D1Hw-Jml.d.ts.map → run-record-C1E24G2A.d.ts.map} +1 -1
  238. package/dist/{run-record-ES07wfz9.js → run-record-C5e7FjhJ.js} +3 -5
  239. package/dist/run-record-C5e7FjhJ.js.map +1 -0
  240. package/dist/search/index.d.ts +46 -32
  241. package/dist/search/index.d.ts.map +1 -1
  242. package/dist/search/index.js +3 -3
  243. package/dist/{search-kernel-BqyBZ9Wm.d.ts → search-kernel-DyNEa6IU.d.ts} +14 -10
  244. package/dist/search-kernel-DyNEa6IU.d.ts.map +1 -0
  245. package/dist/{search-ledger-d7V8o9Lk.js → search-ledger-CVUGLDZp.js} +189 -199
  246. package/dist/search-ledger-CVUGLDZp.js.map +1 -0
  247. package/dist/{search-ledger-file-BqOo-bVo.js → search-ledger-file-DDQUgvCg.js} +4 -4
  248. package/dist/{search-ledger-file-BqOo-bVo.js.map → search-ledger-file-DDQUgvCg.js.map} +1 -1
  249. package/dist/search-ledger-types-p2xKfiDJ.d.ts.map +1 -1
  250. package/dist/{search-shipper-er4pK5_Y.js → search-shipper-BOSD46qv.js} +3 -3
  251. package/dist/{search-shipper-er4pK5_Y.js.map → search-shipper-BOSD46qv.js.map} +1 -1
  252. package/dist/{search-summary-CBvz3u_S.js → search-summary-BQCSWf1g.js} +5 -5
  253. package/dist/{search-summary-CBvz3u_S.js.map → search-summary-BQCSWf1g.js.map} +1 -1
  254. package/dist/{semantic-concept-judge-BPqbwIVs.js → semantic-concept-judge-BQF8te0T.js} +12 -14
  255. package/dist/semantic-concept-judge-BQF8te0T.js.map +1 -0
  256. package/dist/{sequential-D-1PPCmM.js → sequential-Bf_pbE50.js} +5 -5
  257. package/dist/{sequential-D-1PPCmM.js.map → sequential-Bf_pbE50.js.map} +1 -1
  258. package/dist/{sequential-eprocess-D1jKoihe.js → sequential-eprocess-DM_qENCl.js} +2 -2
  259. package/dist/{sequential-eprocess-D1jKoihe.js.map → sequential-eprocess-DM_qENCl.js.map} +1 -1
  260. package/dist/{server-DbcC7xMl.js → server-e9oWO_bN.js} +4 -4
  261. package/dist/{server-DbcC7xMl.js.map → server-e9oWO_bN.js.map} +1 -1
  262. package/dist/{single-run-lock-fVu3MZWH.js → single-run-lock-CV9Eyd_q.js} +2 -2
  263. package/dist/{single-run-lock-fVu3MZWH.js.map → single-run-lock-CV9Eyd_q.js.map} +1 -1
  264. package/dist/{skillopt-optimization-method-Dcepr6QL.js → skillopt-optimization-method-BiwA8nfX.js} +427 -12
  265. package/dist/skillopt-optimization-method-BiwA8nfX.js.map +1 -0
  266. package/dist/{spans-D9_ehFuk.d.ts → spans-BHAUXMGN.d.ts} +2 -2
  267. package/dist/{spans-D9_ehFuk.d.ts.map → spans-BHAUXMGN.d.ts.map} +1 -1
  268. package/dist/{statistical-heldout-1Gbmex7v.js → statistical-heldout-42NlCNDe.js} +3 -3
  269. package/dist/{statistical-heldout-1Gbmex7v.js.map → statistical-heldout-42NlCNDe.js.map} +1 -1
  270. package/dist/{statistical-heldout-CJTj2EoS.d.ts → statistical-heldout-CYpGaoDH.d.ts} +3 -3
  271. package/dist/{statistical-heldout-CJTj2EoS.d.ts.map → statistical-heldout-CYpGaoDH.d.ts.map} +1 -1
  272. package/dist/{store-otlp-DWVhUgEj.js → store-otlp-DJ9dtA5x.js} +135 -132
  273. package/dist/store-otlp-DJ9dtA5x.js.map +1 -0
  274. package/dist/{store-tool-spans-BNICrhsb.js → store-tool-spans-CIX5arAG.js} +7 -7
  275. package/dist/{store-tool-spans-BNICrhsb.js.map → store-tool-spans-CIX5arAG.js.map} +1 -1
  276. package/dist/{store-tool-spans-BNzls3OD.d.ts → store-tool-spans-DdCSPa7C.d.ts} +162 -161
  277. package/dist/store-tool-spans-DdCSPa7C.d.ts.map +1 -0
  278. package/dist/storyboard/index.d.ts.map +1 -1
  279. package/dist/storyboard/index.js +20 -3
  280. package/dist/storyboard/index.js.map +1 -1
  281. package/dist/{student-t-BQeKJx2N.js → student-t-DVPW7CdL.js} +2 -2
  282. package/dist/{student-t-BQeKJx2N.js.map → student-t-DVPW7CdL.js.map} +1 -1
  283. package/dist/{summary-report-DtdRucM0.js → summary-report-V_JACR4Y.js} +526 -614
  284. package/dist/summary-report-V_JACR4Y.js.map +1 -0
  285. package/dist/{summary-report-C-CtItCN.d.ts → summary-report-bdFDXq0b.d.ts} +10 -5
  286. package/dist/summary-report-bdFDXq0b.d.ts.map +1 -0
  287. package/dist/supervisor-run/index.d.ts +154 -150
  288. package/dist/supervisor-run/index.d.ts.map +1 -1
  289. package/dist/supervisor-run/index.js +65 -3
  290. package/dist/supervisor-run/index.js.map +1 -1
  291. package/dist/systemone-BomhHr2c.d.ts +135 -0
  292. package/dist/systemone-BomhHr2c.d.ts.map +1 -0
  293. package/dist/systemone-protocol.js +1 -1
  294. package/dist/{systemone-review-C4dZgrgN.d.ts → systemone-review-CUVcQJu6.d.ts} +2 -2
  295. package/dist/{systemone-review-C4dZgrgN.d.ts.map → systemone-review-CUVcQJu6.d.ts.map} +1 -1
  296. package/dist/{systemone-review-DTClwgmE.js → systemone-review-D7sXwsqq.js} +3 -3
  297. package/dist/{systemone-review-DTClwgmE.js.map → systemone-review-D7sXwsqq.js.map} +1 -1
  298. package/dist/systemone.d.ts +4 -134
  299. package/dist/systemone.js +4 -4
  300. package/dist/{task-failure-attributes-DngaLdE4.js → task-failure-attributes-hEmASjix.js} +2 -2
  301. package/dist/{task-failure-attributes-DngaLdE4.js.map → task-failure-attributes-hEmASjix.js.map} +1 -1
  302. package/dist/{terminal-record-BHocwFw1.js → terminal-record-DZyQZ3pz.js} +14 -10
  303. package/dist/terminal-record-DZyQZ3pz.js.map +1 -0
  304. package/dist/{tool-groups-BKRrtqy4.d.ts → tool-groups-C3NEcdCM.d.ts} +3 -3
  305. package/dist/tool-groups-C3NEcdCM.d.ts.map +1 -0
  306. package/dist/{tool-waste-Wi4g6Ekb.d.ts → tool-waste-BKgvd-TP.d.ts} +30 -30
  307. package/dist/tool-waste-BKgvd-TP.d.ts.map +1 -0
  308. package/dist/{tool-waste-BYt5rdt8.js → tool-waste-BlesqTqW.js} +5 -158
  309. package/dist/tool-waste-BlesqTqW.js.map +1 -0
  310. package/dist/{trace-questions-CLx8vt6t.d.ts → trace-questions-CM2CAocU.d.ts} +4 -4
  311. package/dist/{trace-questions-CLx8vt6t.d.ts.map → trace-questions-CM2CAocU.d.ts.map} +1 -1
  312. package/dist/trace-repair/index.d.ts +3 -3
  313. package/dist/trace-repair/index.d.ts.map +1 -1
  314. package/dist/trace-repair/index.js +7 -13
  315. package/dist/trace-repair/index.js.map +1 -1
  316. package/dist/traces.d.ts +76 -58
  317. package/dist/traces.d.ts.map +1 -1
  318. package/dist/traces.js +9 -9
  319. package/dist/trajectory-replay/index.js +3 -3
  320. package/dist/{tree-rXLW3M_4.js → tree-C1XYeXvy.js} +2277 -2339
  321. package/dist/tree-C1XYeXvy.js.map +1 -0
  322. package/dist/{types-DjOMwGY_.d.ts → types-BqkVmHfv.d.ts} +7 -1
  323. package/dist/types-BqkVmHfv.d.ts.map +1 -0
  324. package/dist/{types-DBZ7keha.d.ts → types-C02qtnG8.d.ts} +4 -46
  325. package/dist/types-C02qtnG8.d.ts.map +1 -0
  326. package/dist/{types-DHb5TI36.d.ts → types-C1tnkEgM.d.ts} +347 -6
  327. package/dist/types-C1tnkEgM.d.ts.map +1 -0
  328. package/dist/{types-DinHtRJp.d.ts → types-Cbe2fWBa.d.ts} +3 -3
  329. package/dist/{types-DinHtRJp.d.ts.map → types-Cbe2fWBa.d.ts.map} +1 -1
  330. package/dist/{types-CGEmVYOk.d.ts → types-Cq8mSbq8.d.ts} +78 -72
  331. package/dist/types-Cq8mSbq8.d.ts.map +1 -0
  332. package/dist/{types-CW5b90JH.js → types-IXD2HJzu.js} +16 -3
  333. package/dist/types-IXD2HJzu.js.map +1 -0
  334. package/dist/{verdict-DhPoXApk.js → verdict-Dl0tiT1X.js} +2 -2
  335. package/dist/{verdict-DhPoXApk.js.map → verdict-Dl0tiT1X.js.map} +1 -1
  336. package/dist/{verdict-cache-BQEsVy9n.js → verdict-cache-Bm8EKL2s.js} +2 -2
  337. package/dist/{verdict-cache-BQEsVy9n.js.map → verdict-cache-Bm8EKL2s.js.map} +1 -1
  338. package/dist/wire/index.d.ts +3 -3
  339. package/dist/wire/index.js +1 -1
  340. package/docs/concepts.md +1 -0
  341. package/docs/design/mlbenchmarks-book-review.md +1 -1
  342. package/docs/design/mlbenchmarks-review/probes.mts +23 -31
  343. package/docs/experiment.md +1 -1
  344. package/docs/finding-layers.md +69 -0
  345. package/package.json +1 -1
  346. package/dist/analyze-runs-3xpby3p2.js.map +0 -1
  347. package/dist/analyze-runs-BhFLOztS.d.ts.map +0 -1
  348. package/dist/ast-DJ6Bc7ab.js +0 -558
  349. package/dist/ast-DJ6Bc7ab.js.map +0 -1
  350. package/dist/ast-hI-vjW6J.d.ts +0 -457
  351. package/dist/ast-hI-vjW6J.d.ts.map +0 -1
  352. package/dist/benchmark-command-CgpGi7vY.js.map +0 -1
  353. package/dist/campaign-CpnsCgfz.js.map +0 -1
  354. package/dist/campaign-evidence-OcUciD-N.js.map +0 -1
  355. package/dist/claim-BXcZw-jo.js.map +0 -1
  356. package/dist/cost-ledger-CO7UyEcQ.d.ts.map +0 -1
  357. package/dist/cost-ledger-D2FnJPBF.js.map +0 -1
  358. package/dist/default-registry-KP4-Tolz.js.map +0 -1
  359. package/dist/define-agent-eval-BdfnqXeP.js.map +0 -1
  360. package/dist/evaluation.d.ts.map +0 -1
  361. package/dist/external-optimizer-subprocess-BUxifb8g.js.map +0 -1
  362. package/dist/feedback-trajectory-BLN_6fEs.d.ts.map +0 -1
  363. package/dist/index-B3WBm953.d.ts.map +0 -1
  364. package/dist/index-B7m2uzJL.d.ts.map +0 -1
  365. package/dist/index-CPJJr98o.d.ts.map +0 -1
  366. package/dist/integrity-BV1ZDf42.js.map +0 -1
  367. package/dist/judge-gate-B1qYnkpi.js.map +0 -1
  368. package/dist/llm-client-DQveoxSM.js.map +0 -1
  369. package/dist/llm-judge-Cd5s5Ei6.js.map +0 -1
  370. package/dist/matrix-_BV9-Lk5.js.map +0 -1
  371. package/dist/mint-DCGoBYhB.d.ts +0 -155
  372. package/dist/mint-DCGoBYhB.d.ts.map +0 -1
  373. package/dist/paired-tests-jbwPdSz5.js.map +0 -1
  374. package/dist/proposal-findings-CXblsskH.js.map +0 -1
  375. package/dist/redact-Cveoq6fJ.d.ts.map +0 -1
  376. package/dist/redact-Dk1ZdJEm.js.map +0 -1
  377. package/dist/registry-CMYIaDgS.d.ts.map +0 -1
  378. package/dist/release-confidence-BK2QCB5M.d.ts.map +0 -1
  379. package/dist/report-command-CtIIsQKP.js.map +0 -1
  380. package/dist/run-record-ES07wfz9.js.map +0 -1
  381. package/dist/search-kernel-BqyBZ9Wm.d.ts.map +0 -1
  382. package/dist/search-kernel-BztlERLv.js.map +0 -1
  383. package/dist/search-ledger-d7V8o9Lk.js.map +0 -1
  384. package/dist/semantic-concept-judge-BPqbwIVs.js.map +0 -1
  385. package/dist/skillopt-optimization-method-Dcepr6QL.js.map +0 -1
  386. package/dist/store-otlp-DWVhUgEj.js.map +0 -1
  387. package/dist/store-tool-spans-BNzls3OD.d.ts.map +0 -1
  388. package/dist/summary-report-C-CtItCN.d.ts.map +0 -1
  389. package/dist/summary-report-DtdRucM0.js.map +0 -1
  390. package/dist/systemone.d.ts.map +0 -1
  391. package/dist/terminal-record-BHocwFw1.js.map +0 -1
  392. package/dist/tool-groups-BKRrtqy4.d.ts.map +0 -1
  393. package/dist/tool-waste-BYt5rdt8.js.map +0 -1
  394. package/dist/tool-waste-Wi4g6Ekb.d.ts.map +0 -1
  395. package/dist/tree-rXLW3M_4.js.map +0 -1
  396. package/dist/types-CGEmVYOk.d.ts.map +0 -1
  397. package/dist/types-CW5b90JH.js.map +0 -1
  398. package/dist/types-DBZ7keha.d.ts.map +0 -1
  399. package/dist/types-DHb5TI36.d.ts.map +0 -1
  400. package/dist/types-DjOMwGY_.d.ts.map +0 -1
@@ -76,6 +76,34 @@ declare function iqr(xs: number[]): number;
76
76
  */
77
77
  declare function welchsTTest(a: number[], b: number[]): WelchTestResult;
78
78
  //#endregion
79
+ //#region src/tool-use-metrics.d.ts
80
+ interface ToolUseMetrics {
81
+ runId: string;
82
+ totalCalls: number;
83
+ /** Calls whose arguments were captured and can be compared for duplication. */
84
+ callsWithCapturedArgs: number;
85
+ byTool: Record<string, ToolStats>;
86
+ errorRate: number;
87
+ /** Ratio of captured-argument calls already seen with the same tool name and arguments. */
88
+ duplicateRate: number;
89
+ /** Ratio of error calls followed by ≥1 retry on same tool. */
90
+ retryRate: number;
91
+ /** Optional: of the calls agent made, fraction the evaluator marked as "correct selection". */
92
+ selectionAccuracy?: number;
93
+ }
94
+ interface ToolStats {
95
+ calls: number;
96
+ callsWithCapturedArgs: number;
97
+ errors: number;
98
+ avgLatencyMs: number;
99
+ duplicates: number;
100
+ }
101
+ interface ToolUseOptions {
102
+ /** Map of spanId → whether the evaluator judged the tool selection correct. Optional. */
103
+ selectionLabels?: Record<string, boolean>;
104
+ }
105
+ declare function computeToolUseMetrics(store: TraceStore, runId: string, options?: ToolUseOptions): Promise<ToolUseMetrics>;
106
+ //#endregion
79
107
  //#region src/pipelines/budget-breach.d.ts
80
108
  interface BudgetBreachFinding {
81
109
  runId: string;
@@ -117,34 +145,6 @@ interface JudgeAgreementReport {
117
145
  }
118
146
  declare function judgeAgreementView(store: TraceStore): Promise<JudgeAgreementReport>;
119
147
  //#endregion
120
- //#region src/tool-use-metrics.d.ts
121
- interface ToolUseMetrics {
122
- runId: string;
123
- totalCalls: number;
124
- /** Calls whose arguments were captured and can be compared for duplication. */
125
- callsWithCapturedArgs: number;
126
- byTool: Record<string, ToolStats>;
127
- errorRate: number;
128
- /** Ratio of captured-argument calls already seen with the same tool name and arguments. */
129
- duplicateRate: number;
130
- /** Ratio of error calls followed by ≥1 retry on same tool. */
131
- retryRate: number;
132
- /** Optional: of the calls agent made, fraction the evaluator marked as "correct selection". */
133
- selectionAccuracy?: number;
134
- }
135
- interface ToolStats {
136
- calls: number;
137
- callsWithCapturedArgs: number;
138
- errors: number;
139
- avgLatencyMs: number;
140
- duplicates: number;
141
- }
142
- interface ToolUseOptions {
143
- /** Map of spanId → whether the evaluator judged the tool selection correct. Optional. */
144
- selectionLabels?: Record<string, boolean>;
145
- }
146
- declare function computeToolUseMetrics(store: TraceStore, runId: string, options?: ToolUseOptions): Promise<ToolUseMetrics>;
147
- //#endregion
148
148
  //#region src/pipelines/tool-waste.d.ts
149
149
  interface ToolWasteFinding {
150
150
  runId: string;
@@ -164,5 +164,5 @@ interface ToolWasteOptions {
164
164
  }
165
165
  declare function toolWasteView(store: TraceStore, options?: ToolWasteOptions): Promise<ToolWasteReport>;
166
166
  //#endregion
167
- export { WelchTestResult as _, ToolStats as a, welchsTTest as b, computeToolUseMetrics as c, judgeAgreementView as d, BudgetBreachFinding as f, BaselineReport as g, BaselineOptions as h, toolWasteView as i, JudgeAgreementReport as l, budgetBreachView as m, ToolWasteOptions as n, ToolUseMetrics as o, BudgetBreachReport as p, ToolWasteReport as r, ToolUseOptions as s, ToolWasteFinding as t, JudgePair as u, WelchTestStatus as v, iqr as y };
168
- //# sourceMappingURL=tool-waste-Wi4g6Ekb.d.ts.map
167
+ export { WelchTestResult as _, JudgeAgreementReport as a, welchsTTest as b, BudgetBreachFinding as c, ToolStats as d, ToolUseMetrics as f, BaselineReport as g, BaselineOptions as h, toolWasteView as i, BudgetBreachReport as l, computeToolUseMetrics as m, ToolWasteOptions as n, JudgePair as o, ToolUseOptions as p, ToolWasteReport as r, judgeAgreementView as s, ToolWasteFinding as t, budgetBreachView as u, WelchTestStatus as v, iqr as y };
168
+ //# sourceMappingURL=tool-waste-BKgvd-TP.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"tool-waste-BKgvd-TP.d.ts","names":[],"sources":["../src/baseline.ts","../src/tool-use-metrics.ts","../src/pipelines/budget-breach.ts","../src/pipelines/judge-agreement.ts","../src/pipelines/tool-waste.ts"],"mappings":";;;;UA6BiB;EACf;EACA;EACA;EACA;;;EAGA;EACA;EACA;EACA;EACA;;EAEA;EACA;;UAGe;EACf,SAAS;;EAET;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;;KAGU;;;;;;;;;UAUF;EACR;EACA;EACA;;KAGU,mBACP;EACC;EACA;EACA;EACA;EACA;EACA;EACA;MAED;EACC;EACA;EACA;EACA;EACA;EACA;EACA;;;iBA6EU,IAAI;;;;;;;;;iBAcJ,YAAY,aAAa,cAAc;;;UC9KtC;EACf;EACA;;EAEA;EACA,QAAQ,eAAe;EACvB;;EAEA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;;UAGe;;EAEf,kBAAkB;;iBAGE,sBACpB,OAAO,YACP,eACA,UAAS,iBACR,QAAQ;;;UCnCM;EACf;EACA;EACA;EACA,iBAAiB;EACjB;EACA;EACA;EACA;;UAGe;EACf,UAAU;EACV,aAAa;EACb,YAAY;EACZ,WAAW;EACX;EACA;;iBAGoB,iBACpB,OAAO,YACP;EAAW;EAAqB;IAC/B,QAAQ;;;UCnBM;EACf;EACA;EACA;;EAEA;EACA;EACA;;UAGe;EACf,OAAO;EACP;EACA;;iBAGoB,mBAAmB,OAAO,aAAa,QAAQ;;;UCbpD;EACf;EACA;EACA;EACA;;UAGe;EACf,OAAO;EACP;;UAGe;EACf;EACA,eAAe,MAAM,UAAU;IAAS,KAAK,QAAQ,kBAAkB;;;iBAGnD,cACpB,OAAO,YACP,UAAS,mBACR,QAAQ"}
@@ -1,160 +1,7 @@
1
- import { s as ValidationError } from "./errors-Dngq5h35.js";
2
- import { r as pearsonR } from "./descriptive-CnU8Yq29.js";
3
- import { r as continuousAgreement } from "./judge-calibration-N44V04Se.js";
4
- import { a as hasCapturedToolArgs, d as toolSpans, i as groupBy, r as argHash, s as llmSpans } from "./query-B1D1qp2Z.js";
1
+ import { r as pearsonR } from "./descriptive-DboxPZrG.js";
2
+ import { r as interRaterReliability } from "./agreement-irr-BEWLIF3_.js";
3
+ import { a as hasCapturedToolArgs, d as toolSpans, i as groupBy, r as argHash, s as llmSpans } from "./query-BUCZCK7F.js";
5
4
  import { a as classifyFailure, n as DEFAULT_RULES } from "./failure-taxonomy-CjabTATj.js";
6
- //#region src/statistics/agreement-irr.ts
7
- /**
8
- * Inter-rater reliability — Krippendorff's α under the squared-difference
9
- * metric, pooled across dimensions.
10
- *
11
- * Each inner array is one judge's scores. Items are matched by position
12
- * WITHIN a dimension: the k-th score a judge supplies carrying dimension
13
- * `d` is item k of `d`, and the ratings compared against each other are
14
- * the ones different judges gave to the same item. Every judge that scores
15
- * a dimension at all must supply the same number of scores for it —
16
- * ragged input cannot be aligned into items and throws rather than
17
- * comparing mismatched items.
18
- *
19
- * α = 1 − D_observed / D_expected: D_observed averages the squared
20
- * difference over within-item judge pairs, D_expected over every pair of
21
- * ratings irrespective of item. α = 1 is perfect agreement, 0 is chance,
22
- * negative is systematic disagreement.
23
- */
24
- function interRaterReliability(judgeScores) {
25
- if (judgeScores.length < 2) return 1;
26
- const perDimension = /* @__PURE__ */ new Map();
27
- for (let judgeIndex = 0; judgeIndex < judgeScores.length; judgeIndex++) for (const s of judgeScores[judgeIndex]) {
28
- let byJudge = perDimension.get(s.dimension);
29
- if (byJudge === void 0) {
30
- byJudge = Array.from({ length: judgeScores.length }, () => []);
31
- perDimension.set(s.dimension, byJudge);
32
- }
33
- byJudge[judgeIndex].push(s.score);
34
- }
35
- const allValues = [];
36
- const pairDiffs = [];
37
- for (const [dimension, byJudge] of perDimension) {
38
- const scoring = byJudge.filter((scores) => scores.length > 0);
39
- if (scoring.length < 2) continue;
40
- const itemCount = scoring[0].length;
41
- if (scoring.some((scores) => scores.length !== itemCount)) throw new ValidationError(`interRaterReliability: dimension '${dimension}' has judges supplying ${scoring.map((scores) => scores.length).join("/")} scores — items cannot be aligned`);
42
- for (let item = 0; item < itemCount; item++) {
43
- const ratings = scoring.map((scores) => scores[item]);
44
- for (const v of ratings) allValues.push(v);
45
- for (let i = 0; i < ratings.length; i++) for (let j = i + 1; j < ratings.length; j++) pairDiffs.push((ratings[i] - ratings[j]) ** 2);
46
- }
47
- }
48
- if (pairDiffs.length === 0 || allValues.length < 2) return 1;
49
- const observedDisagreement = pairDiffs.reduce((a, b) => a + b, 0) / pairDiffs.length;
50
- let expectedDisagreement = 0;
51
- let expectedCount = 0;
52
- for (let i = 0; i < allValues.length; i++) for (let j = i + 1; j < allValues.length; j++) {
53
- expectedDisagreement += (allValues[i] - allValues[j]) ** 2;
54
- expectedCount++;
55
- }
56
- expectedDisagreement = expectedCount > 0 ? expectedDisagreement / expectedCount : 0;
57
- if (expectedDisagreement === 0) return 1;
58
- return 1 - observedDisagreement / expectedDisagreement;
59
- }
60
- /**
61
- * Corpus-wide inter-rater agreement across N items × M judges × D dimensions.
62
- *
63
- * For each dimension, builds the [n_items][n_judges] matrix of scores
64
- * (keeping only items every judge rated on that dimension), then runs
65
- * `continuousAgreement` to get ICC(2,1), κ_w, Pearson, Spearman, and
66
- * bootstrap CIs. Reports a pooled mean across dimensions as a single
67
- * "is this judge panel reliable on this corpus?" number.
68
- *
69
- * Fail-loud contract:
70
- * - Empty input throws.
71
- * - Fewer than 2 judges or fewer than 2 items per dimension throws.
72
- * - A judge present in some dimensions but with zero scored items on
73
- * another dimension throws (would silently shrink the matrix).
74
- * - Duplicate (itemId, judgeName, dimension) records throw.
75
- */
76
- function corpusInterRaterAgreement(records, opts = {}) {
77
- if (records.length === 0) throw new ValidationError("corpusInterRaterAgreement: no score records supplied");
78
- const judgesSeen = /* @__PURE__ */ new Set();
79
- const dimsSeen = /* @__PURE__ */ new Set();
80
- const grid = /* @__PURE__ */ new Map();
81
- for (const r of records) {
82
- if (!Number.isFinite(r.score)) throw new ValidationError(`corpusInterRaterAgreement: non-finite score for (item=${r.itemId}, judge=${r.judgeName}, dim=${r.dimension})`);
83
- judgesSeen.add(r.judgeName);
84
- dimsSeen.add(r.dimension);
85
- const byJudge = grid.get(r.dimension) ?? /* @__PURE__ */ new Map();
86
- const byItem = byJudge.get(r.judgeName) ?? /* @__PURE__ */ new Map();
87
- if (byItem.has(r.itemId)) throw new ValidationError(`corpusInterRaterAgreement: duplicate record for (item=${r.itemId}, judge=${r.judgeName}, dim=${r.dimension})`);
88
- byItem.set(r.itemId, r.score);
89
- byJudge.set(r.judgeName, byItem);
90
- grid.set(r.dimension, byJudge);
91
- }
92
- const targetDims = opts.dimensions ?? [...dimsSeen].sort();
93
- for (const d of targetDims) if (!dimsSeen.has(d)) throw new ValidationError(`corpusInterRaterAgreement: dimension '${d}' was requested but no records carry it`);
94
- const targetJudges = opts.judges ? [...opts.judges] : [...judgesSeen].sort();
95
- for (const j of targetJudges) if (!judgesSeen.has(j)) throw new ValidationError(`corpusInterRaterAgreement: judge '${j}' was requested but produced no records`);
96
- if (targetJudges.length < 2) throw new ValidationError(`corpusInterRaterAgreement: need ≥2 judges, got ${targetJudges.length}`);
97
- const perDimension = [];
98
- const iccs = [];
99
- const kappas = [];
100
- for (const dim of targetDims) {
101
- const byJudge = grid.get(dim);
102
- const judgeItemCounts = {};
103
- for (const j of targetJudges) judgeItemCounts[j] = byJudge.get(j)?.size ?? 0;
104
- const emptyJudges = targetJudges.filter((j) => judgeItemCounts[j] === 0);
105
- if (emptyJudges.length > 0) throw new ValidationError(`corpusInterRaterAgreement: dimension '${dim}' has no scores from judge(s) ${emptyJudges.join(", ")} (counts: ${JSON.stringify(judgeItemCounts)})`);
106
- let commonItems = null;
107
- for (const j of targetJudges) {
108
- const ids = new Set(byJudge.get(j).keys());
109
- if (commonItems === null) commonItems = ids;
110
- else commonItems = new Set([...commonItems].filter((x) => ids.has(x)));
111
- }
112
- const sortedItems = [...commonItems ?? /* @__PURE__ */ new Set()].sort();
113
- if (sortedItems.length < 2) throw new ValidationError(`corpusInterRaterAgreement: dimension '${dim}' has ${sortedItems.length} item(s) rated by all ${targetJudges.length} judges (need ≥2)`);
114
- const matrix = sortedItems.map((itemId) => targetJudges.map((j) => byJudge.get(j).get(itemId)));
115
- const agreement = continuousAgreement(matrix, opts);
116
- perDimension.push({
117
- ...agreement,
118
- dimension: dim,
119
- itemIds: sortedItems,
120
- judgeIds: [...targetJudges]
121
- });
122
- if (Number.isFinite(agreement.icc)) iccs.push(agreement.icc);
123
- if (Number.isFinite(agreement.weightedKappa)) kappas.push(agreement.weightedKappa);
124
- }
125
- const mean = (xs) => xs.length === 0 ? NaN : xs.reduce((a, b) => a + b, 0) / xs.length;
126
- return {
127
- perDimension,
128
- overallIcc: mean(iccs),
129
- overallWeightedKappa: mean(kappas),
130
- dimensions: targetDims,
131
- judgeIds: targetJudges
132
- };
133
- }
134
- /**
135
- * Convenience adapter for `JudgeScore[]` data keyed externally by item.
136
- *
137
- * Use when you have per-item arrays of `JudgeScore[]` (e.g. one
138
- * `ScenarioResult.judgeScores` per scenario) and want corpus-wide
139
- * agreement without manually flattening. `itemId` must be unique per
140
- * row of `itemsScores`.
141
- */
142
- function corpusInterRaterAgreementFromJudgeScores(itemsScores, opts = {}) {
143
- const records = [];
144
- const seen = /* @__PURE__ */ new Set();
145
- for (const { itemId, scores } of itemsScores) {
146
- if (seen.has(itemId)) throw new ValidationError(`corpusInterRaterAgreementFromJudgeScores: duplicate itemId '${itemId}'`);
147
- seen.add(itemId);
148
- for (const s of scores) records.push({
149
- itemId,
150
- judgeName: s.judgeName,
151
- dimension: s.dimension,
152
- score: s.score
153
- });
154
- }
155
- return corpusInterRaterAgreement(records, opts);
156
- }
157
- //#endregion
158
5
  //#region src/pipelines/budget-breach.ts
159
6
  async function budgetBreachView(store, options = {}) {
160
7
  const runs = await store.listRuns({
@@ -516,6 +363,6 @@ function stringify(v) {
516
363
  }
517
364
  }
518
365
  //#endregion
519
- export { budgetBreachView as a, interRaterReliability as c, failureClusterView as i, computeToolUseMetrics as n, corpusInterRaterAgreement as o, judgeAgreementView as r, corpusInterRaterAgreementFromJudgeScores as s, toolWasteView as t };
366
+ export { budgetBreachView as a, failureClusterView as i, computeToolUseMetrics as n, judgeAgreementView as r, toolWasteView as t };
520
367
 
521
- //# sourceMappingURL=tool-waste-BYt5rdt8.js.map
368
+ //# sourceMappingURL=tool-waste-BlesqTqW.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"tool-waste-BlesqTqW.js","names":[],"sources":["../src/pipelines/budget-breach.ts","../src/pipelines/failure-cluster.ts","../src/pipelines/judge-agreement.ts","../src/tool-use-metrics.ts","../src/pipelines/tool-waste.ts"],"sourcesContent":["/**\n * BudgetBreachView — aggregates breach events across the corpus.\n *\n * Answers: which dimensions get hit most often? Which scenarios are\n * underbudgeted? Which variants trigger the most breaches?\n */\n\nimport type { BudgetSpec } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface BudgetBreachFinding {\n runId: string\n scenarioId: string\n variantId?: string\n dimension: keyof BudgetSpec\n limit: number\n consumed: number\n excessRatio: number\n timestamp: number\n}\n\nexport interface BudgetBreachReport {\n findings: BudgetBreachFinding[]\n byDimension: Record<string, number>\n byScenario: Record<string, number>\n byVariant: Record<string, number>\n totalRuns: number\n breachedRunRatio: number\n}\n\nexport async function budgetBreachView(\n store: TraceStore,\n options: { scenarioId?: string; variantId?: string } = {},\n): Promise<BudgetBreachReport> {\n const runs = await store.listRuns({\n scenarioId: options.scenarioId,\n variantId: options.variantId,\n })\n const findings: BudgetBreachFinding[] = []\n const byDimension: Record<string, number> = {}\n const byScenario: Record<string, number> = {}\n const byVariant: Record<string, number> = {}\n\n for (const run of runs) {\n const entries = await store.budget(run.runId)\n for (const e of entries) {\n if (!e.breached) continue\n const excessRatio = e.limit > 0 ? e.consumed / e.limit : Infinity\n findings.push({\n runId: run.runId,\n scenarioId: run.scenarioId,\n variantId: run.variantId,\n dimension: e.dimension,\n limit: e.limit,\n consumed: e.consumed,\n excessRatio,\n timestamp: e.timestamp,\n })\n byDimension[e.dimension] = (byDimension[e.dimension] ?? 0) + 1\n byScenario[run.scenarioId] = (byScenario[run.scenarioId] ?? 0) + 1\n if (run.variantId) byVariant[run.variantId] = (byVariant[run.variantId] ?? 0) + 1\n }\n }\n\n const breachedRuns = new Set(findings.map((f) => f.runId))\n return {\n findings,\n byDimension,\n byScenario,\n byVariant,\n totalRuns: runs.length,\n breachedRunRatio: runs.length > 0 ? breachedRuns.size / runs.length : 0,\n }\n}\n","/**\n * FailureClusterView — groups failed runs by (failureClass, triggerTool,\n * argHash-prefix) so weekly reviews can prioritize the top-N clusters.\n *\n * Each cluster includes: N runs, scenarios affected, representative\n * error message, a proposed mitigation hint (rule → action table).\n */\n\nimport { classifyFailure, DEFAULT_RULES, type FailureRule } from '../failure-taxonomy'\nimport { argHash, hasCapturedToolArgs, toolSpans } from '../trace/query'\nimport type { FailureClass, Span } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface FailureCluster {\n failureClass: FailureClass\n /** Tool name when the trigger was a tool span, else undefined. */\n toolName?: string\n /** First 16 chars of argHash — clusters similar args. */\n argPrefix?: string\n /**\n * Source dimension when the trigger was a judge span (e.g. `'format'`,\n * `'safety'`, `'correctness'`). Lets cross-template aggregators\n * group failures by the dimension that fired without overloading\n * `argPrefix`. Optional — clusters without this field deserialize cleanly.\n */\n dimension?: string\n runCount: number\n scenarioIds: string[]\n exampleError?: string\n exampleRunId: string\n}\n\nexport interface FailureClusterReport {\n clusters: FailureCluster[]\n totalFailures: number\n totalRuns: number\n}\n\nexport async function failureClusterView(\n store: TraceStore,\n options: { rules?: FailureRule[]; minClusterSize?: number } = {},\n): Promise<FailureClusterReport> {\n const rules = options.rules ?? DEFAULT_RULES\n const minSize = options.minClusterSize ?? 1\n const runs = await store.listRuns()\n\n type Key = string\n const clusters = new Map<Key, FailureCluster>()\n let totalFailures = 0\n\n for (const run of runs) {\n if (run.status === 'completed' && run.outcome?.pass !== false) continue\n totalFailures++\n const spans = await store.spans({ runId: run.runId })\n const events = await store.events({ runId: run.runId })\n const cls = classifyFailure({ run, spans, events }, rules)\n\n let toolName: string | undefined\n let argPrefix: string | undefined\n let dimension: string | undefined\n if (cls.triggerSpanId) {\n const trig = spans.find((s) => s.spanId === cls.triggerSpanId)\n if (trig?.kind === 'tool') {\n toolName = trig.toolName\n if (hasCapturedToolArgs(trig)) argPrefix = argHash(trig.args).slice(0, 16)\n } else if (trig?.kind === 'judge') {\n dimension = trig.dimension\n }\n }\n // Fallback: look at the last errored tool span\n if (!toolName) {\n const ts = await toolSpans(store, run.runId)\n const errored = ts.filter((t) => t.status === 'error').pop()\n if (errored) {\n toolName = errored.toolName\n if (hasCapturedToolArgs(errored)) argPrefix = argHash(errored.args).slice(0, 16)\n }\n }\n // Secondary signal: any judge span on the failed run carries a\n // dimension. Useful when the rule classified by judge score but\n // didn't surface the trigger span (or surfaced a non-judge span).\n if (!dimension) {\n const judge = spans.find((s) => s.kind === 'judge' && typeof s.dimension === 'string')\n if (judge?.kind === 'judge') dimension = judge.dimension\n }\n\n const key = `${cls.failureClass}|${toolName ?? ''}|${argPrefix ?? ''}|${dimension ?? ''}`\n let cluster = clusters.get(key)\n if (!cluster) {\n cluster = {\n failureClass: cls.failureClass,\n toolName,\n argPrefix,\n dimension,\n runCount: 0,\n scenarioIds: [],\n exampleRunId: run.runId,\n exampleError: firstErrorMessage(spans) ?? cls.reason,\n }\n clusters.set(key, cluster)\n }\n cluster.runCount++\n if (!cluster.scenarioIds.includes(run.scenarioId)) cluster.scenarioIds.push(run.scenarioId)\n }\n\n const arr = [...clusters.values()]\n .filter((c) => c.runCount >= minSize)\n .sort((a, b) => b.runCount - a.runCount)\n\n return { clusters: arr, totalFailures, totalRuns: runs.length }\n}\n\nfunction firstErrorMessage(spans: Span[]): string | undefined {\n const errored = spans.find((s) => s.status === 'error')\n return errored?.error\n}\n","/**\n * JudgeAgreementView — pairwise agreement between judges across the\n * corpus, grouped by dimension.\n *\n * Output drives two workflows:\n * - Judge robustness audit: \"does Claude agree with GPT at κ ≥ 0.6?\"\n * - Calibration tracking: κ vs golden human labels over time (by\n * providing a `humanGoldenJudgeId`).\n */\n\nimport { interRaterReliability, pearsonR } from '../statistics'\nimport type { JudgeSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface JudgePair {\n judgeA: string\n judgeB: string\n dimension: string\n /** Number of (targetSpanId, dimension) tuples both judges scored. */\n commonItems: number\n pearson: number\n krippendorff: number\n}\n\nexport interface JudgeAgreementReport {\n pairs: JudgePair[]\n dimensions: string[]\n judgeIds: string[]\n}\n\nexport async function judgeAgreementView(store: TraceStore): Promise<JudgeAgreementReport> {\n const all = (await store.spans({ kind: 'judge' })).filter(\n (s): s is JudgeSpan => s.kind === 'judge',\n )\n if (all.length === 0) return { pairs: [], dimensions: [], judgeIds: [] }\n\n const byDimension = new Map<string, JudgeSpan[]>()\n for (const s of all) {\n const arr = byDimension.get(s.dimension) ?? []\n arr.push(s)\n byDimension.set(s.dimension, arr)\n }\n\n const judgeIds = [...new Set(all.map((s) => s.judgeId))].sort()\n const pairs: JudgePair[] = []\n for (const [dim, spans] of byDimension) {\n const byJudge = new Map<string, Map<string, number>>()\n for (const s of spans) {\n const m = byJudge.get(s.judgeId) ?? new Map<string, number>()\n m.set(s.targetSpanId, s.score)\n byJudge.set(s.judgeId, m)\n }\n const judgesHere = [...byJudge.keys()]\n for (let i = 0; i < judgesHere.length; i++) {\n for (let j = i + 1; j < judgesHere.length; j++) {\n const judgeI = judgesHere[i]!\n const judgeJ = judgesHere[j]!\n const a = byJudge.get(judgeI)!\n const b = byJudge.get(judgeJ)!\n const common: Array<[number, number]> = []\n for (const [target, scoreA] of a) {\n const scoreB = b.get(target)\n if (scoreB !== undefined) common.push([scoreA, scoreB])\n }\n if (common.length < 2) continue\n const judgeScores = common.map(\n ([scoreA, scoreB]) =>\n [\n { judgeName: judgeI, dimension: dim, score: scoreA, reasoning: '' },\n { judgeName: judgeJ, dimension: dim, score: scoreB, reasoning: '' },\n ] as const,\n )\n const k = interRaterReliability(\n judgeScores[0]!.map((_, k2) => judgeScores.map((pair) => pair[k2]!)),\n )\n pairs.push({\n judgeA: judgeI,\n judgeB: judgeJ,\n dimension: dim,\n commonItems: common.length,\n pearson: pearsonR(\n common.map((c) => c[0]),\n common.map((c) => c[1]),\n ),\n krippendorff: k,\n })\n }\n }\n }\n\n return {\n pairs: pairs.sort((a, b) => b.commonItems - a.commonItems),\n dimensions: [...byDimension.keys()].sort(),\n judgeIds,\n }\n}\n","/**\n * Tool-use metrics — derived purely from trace data.\n *\n * No scoring assumptions: consumers supply optional ground-truth tool\n * selections per turn + optional \"information used downstream\" signals.\n * Without those, we still compute descriptive metrics (error rate,\n * retry rate, duplicate-call rate) that are useful on their own.\n */\n\nimport { argHash, groupBy, hasCapturedToolArgs, toolSpans } from './trace/query'\nimport type { Span } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\nexport interface ToolUseMetrics {\n runId: string\n totalCalls: number\n /** Calls whose arguments were captured and can be compared for duplication. */\n callsWithCapturedArgs: number\n byTool: Record<string, ToolStats>\n errorRate: number\n /** Ratio of captured-argument calls already seen with the same tool name and arguments. */\n duplicateRate: number\n /** Ratio of error calls followed by ≥1 retry on same tool. */\n retryRate: number\n /** Optional: of the calls agent made, fraction the evaluator marked as \"correct selection\". */\n selectionAccuracy?: number\n}\n\nexport interface ToolStats {\n calls: number\n callsWithCapturedArgs: number\n errors: number\n avgLatencyMs: number\n duplicates: number\n}\n\nexport interface ToolUseOptions {\n /** Map of spanId → whether the evaluator judged the tool selection correct. Optional. */\n selectionLabels?: Record<string, boolean>\n}\n\nexport async function computeToolUseMetrics(\n store: TraceStore,\n runId: string,\n options: ToolUseOptions = {},\n): Promise<ToolUseMetrics> {\n const tools = await toolSpans(store, runId)\n if (tools.length === 0) {\n return {\n runId,\n totalCalls: 0,\n callsWithCapturedArgs: 0,\n byTool: {},\n errorRate: 0,\n duplicateRate: 0,\n retryRate: 0,\n }\n }\n\n const byTool: Record<string, ToolStats> = {}\n let totalErrors = 0\n let totalDuplicates = 0\n let callsWithCapturedArgs = 0\n const sortedTools = [...tools].sort((a, b) => a.startedAt - b.startedAt)\n const seenSignatures = new Set<string>()\n\n // duplicate detection + per-tool aggregation\n for (const t of sortedTools) {\n byTool[t.toolName] ??= {\n calls: 0,\n callsWithCapturedArgs: 0,\n errors: 0,\n avgLatencyMs: 0,\n duplicates: 0,\n }\n const stat = byTool[t.toolName]!\n stat.calls += 1\n if (t.status === 'error') {\n stat.errors += 1\n totalErrors += 1\n }\n if (typeof t.latencyMs === 'number') stat.avgLatencyMs += t.latencyMs\n if (hasCapturedToolArgs(t)) {\n callsWithCapturedArgs += 1\n stat.callsWithCapturedArgs += 1\n const sig = `${t.toolName}|${argHash(t.args)}`\n if (seenSignatures.has(sig)) {\n stat.duplicates += 1\n totalDuplicates += 1\n }\n seenSignatures.add(sig)\n }\n }\n\n for (const stat of Object.values(byTool)) {\n stat.avgLatencyMs = stat.calls > 0 ? stat.avgLatencyMs / stat.calls : 0\n }\n\n // retry detection: per-tool chronological adjacency where error → next same-tool call\n let retryOpportunities = 0\n let retriesFollowed = 0\n for (const [, arr] of groupBy(sortedTools, (t) => t.toolName)) {\n for (let i = 0; i < arr.length; i++) {\n if (arr[i]!.status !== 'error') continue\n retryOpportunities += 1\n if (arr[i + 1]) retriesFollowed += 1\n }\n }\n const retryRate = retryOpportunities > 0 ? retriesFollowed / retryOpportunities : 0\n\n let selectionAccuracy: number | undefined\n if (options.selectionLabels) {\n const labeled = sortedTools.filter((t) => t.spanId in options.selectionLabels!)\n if (labeled.length > 0) {\n selectionAccuracy =\n labeled.filter((t) => options.selectionLabels![t.spanId]).length / labeled.length\n }\n }\n\n return {\n runId,\n totalCalls: sortedTools.length,\n callsWithCapturedArgs,\n byTool,\n errorRate: totalErrors / sortedTools.length,\n duplicateRate: callsWithCapturedArgs > 0 ? totalDuplicates / callsWithCapturedArgs : 0,\n retryRate,\n selectionAccuracy,\n }\n}\n\nexport type { Span }\n","/**\n * ToolWasteView — fraction of tool calls whose results weren't used\n * downstream. Without a \"used\" signal we fall back to structural\n * proxies: error calls, duplicate calls, and tool calls followed by\n * zero subsequent LLM spans are all considered waste.\n *\n * Consumers can pass a `usageOracle` that inspects a tool span and\n * returns true iff the tool's result appears in a later LLM message,\n * artifact, or state mutation — that's the canonical definition; the\n * default heuristic is a reasonable fallback.\n */\n\nimport { computeToolUseMetrics } from '../tool-use-metrics'\nimport { llmSpans, toolSpans } from '../trace/query'\nimport type { LlmSpan, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface ToolWasteFinding {\n runId: string\n wastedCalls: number\n totalCalls: number\n wasteRate: number\n}\n\nexport interface ToolWasteReport {\n byRun: ToolWasteFinding[]\n overallWasteRate: number\n}\n\nexport interface ToolWasteOptions {\n runId?: string\n usageOracle?: (tool: ToolSpan, later: { llm: Awaited<ReturnType<typeof llmSpans>> }) => boolean\n}\n\nexport async function toolWasteView(\n store: TraceStore,\n options: ToolWasteOptions = {},\n): Promise<ToolWasteReport> {\n const runs = options.runId ? [options.runId] : (await store.listRuns()).map((r) => r.runId)\n\n const byRun: ToolWasteFinding[] = []\n let totalCalls = 0\n let totalWasted = 0\n for (const runId of runs) {\n const tools = await toolSpans(store, runId)\n if (tools.length === 0) {\n byRun.push({ runId, wastedCalls: 0, totalCalls: 0, wasteRate: 0 })\n continue\n }\n const llms = await llmSpans(store, runId)\n // Sort LLM spans once by start time, then build a suffix index of the\n // concatenated message text. `suffixText[i]` is the haystack of every\n // string message content in spans[i..]. Per tool we binary-search the\n // first span started strictly after the tool, then test that one suffix\n // — turning the per-tool O(llms × messages × content) scan into a single\n // O(log llms) lookup over precomputed text.\n const sortedLlm = [...llms].sort((a, b) => a.startedAt - b.startedAt)\n const startTimes = sortedLlm.map((l) => l.startedAt)\n const suffixText = buildSuffixText(sortedLlm)\n let wasted = 0\n for (const t of tools) {\n if (t.status === 'error') {\n wasted++\n continue\n }\n // First LLM span started strictly after this tool (upper-bound search).\n const cutoff = upperBound(startTimes, t.startedAt)\n if (options.usageOracle) {\n if (!options.usageOracle(t, { llm: sortedLlm.slice(cutoff) })) wasted++\n } else {\n // Default heuristic: a tool whose result is NOT mentioned in any\n // later LLM input message is likely wasted. An empty/null result has\n // no payload to propagate downstream — there is nothing to find in a\n // later message, so it is not evidence of waste; skip it.\n const resultStr = stringify(t.result)\n if (resultStr === '') continue\n const haystack = suffixText[cutoff] ?? ''\n const used = haystack.includes(resultStr.slice(0, 120))\n if (!used) wasted++\n }\n }\n const wasteRate = wasted / tools.length\n byRun.push({ runId, wastedCalls: wasted, totalCalls: tools.length, wasteRate })\n totalCalls += tools.length\n totalWasted += wasted\n }\n return { byRun, overallWasteRate: totalCalls > 0 ? totalWasted / totalCalls : 0 }\n}\n\n/**\n * Build per-position suffix haystacks: result[i] is the concatenation of every\n * string message content in spans[i..end]. Built back-to-front so each entry\n * reuses the next one — O(total message text) rather than O(spans²).\n */\nfunction buildSuffixText(spans: LlmSpan[]): string[] {\n const result = new Array<string>(spans.length + 1)\n result[spans.length] = ''\n for (let i = spans.length - 1; i >= 0; i--) {\n const own = spans[i]!.messages.map((m) =>\n typeof m.content === 'string' ? m.content : '',\n ).join('\\n')\n result[i] = `${own}\\n${result[i + 1]}`\n }\n return result\n}\n\n/** Index of the first element strictly greater than `target` in a sorted array. */\nfunction upperBound(sorted: number[], target: number): number {\n let lo = 0\n let hi = sorted.length\n while (lo < hi) {\n const mid = (lo + hi) >>> 1\n if (sorted[mid]! <= target) lo = mid + 1\n else hi = mid\n }\n return lo\n}\n\nfunction stringify(v: unknown): string {\n if (v === null || v === undefined) return ''\n if (typeof v === 'string') return v\n try {\n return JSON.stringify(v)\n } catch {\n return String(v)\n }\n}\n\n// Re-export for convenience in consumers that want both descriptive and usage metrics.\nexport { computeToolUseMetrics }\n"],"mappings":";;;;;AA8BA,eAAsB,iBACpB,OACA,UAAuD,CAAC,GAC3B;CAC7B,MAAM,OAAO,MAAM,MAAM,SAAS;EAChC,YAAY,QAAQ;EACpB,WAAW,QAAQ;CACrB,CAAC;CACD,MAAM,WAAkC,CAAC;CACzC,MAAM,cAAsC,CAAC;CAC7C,MAAM,aAAqC,CAAC;CAC5C,MAAM,YAAoC,CAAC;CAE3C,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,UAAU,MAAM,MAAM,OAAO,IAAI,KAAK;EAC5C,KAAK,MAAM,KAAK,SAAS;GACvB,IAAI,CAAC,EAAE,UAAU;GACjB,MAAM,cAAc,EAAE,QAAQ,IAAI,EAAE,WAAW,EAAE,QAAQ;GACzD,SAAS,KAAK;IACZ,OAAO,IAAI;IACX,YAAY,IAAI;IAChB,WAAW,IAAI;IACf,WAAW,EAAE;IACb,OAAO,EAAE;IACT,UAAU,EAAE;IACZ;IACA,WAAW,EAAE;GACf,CAAC;GACD,YAAY,EAAE,cAAc,YAAY,EAAE,cAAc,KAAK;GAC7D,WAAW,IAAI,eAAe,WAAW,IAAI,eAAe,KAAK;GACjE,IAAI,IAAI,WAAW,UAAU,IAAI,cAAc,UAAU,IAAI,cAAc,KAAK;EAClF;CACF;CAEA,MAAM,eAAe,IAAI,IAAI,SAAS,KAAK,MAAM,EAAE,KAAK,CAAC;CACzD,OAAO;EACL;EACA;EACA;EACA;EACA,WAAW,KAAK;EAChB,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;CACxE;AACF;;;;;;;;;;ACnCA,eAAsB,mBACpB,OACA,UAA8D,CAAC,GAChC;CAC/B,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,UAAU,QAAQ,kBAAkB;CAC1C,MAAM,OAAO,MAAM,MAAM,SAAS;CAGlC,MAAM,2BAAW,IAAI,IAAyB;CAC9C,IAAI,gBAAgB;CAEpB,KAAK,MAAM,OAAO,MAAM;EACtB,IAAI,IAAI,WAAW,eAAe,IAAI,SAAS,SAAS,OAAO;EAC/D;EACA,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;EACpD,MAAM,SAAS,MAAM,MAAM,OAAO,EAAE,OAAO,IAAI,MAAM,CAAC;EACtD,MAAM,MAAM,gBAAgB;GAAE;GAAK;GAAO;EAAO,GAAG,KAAK;EAEzD,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI,IAAI,eAAe;GACrB,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,WAAW,IAAI,aAAa;GAC7D,IAAI,MAAM,SAAS,QAAQ;IACzB,WAAW,KAAK;IAChB,IAAI,oBAAoB,IAAI,GAAG,YAAY,QAAQ,KAAK,IAAI,CAAC,CAAC,MAAM,GAAG,EAAE;GAC3E,OAAO,IAAI,MAAM,SAAS,SACxB,YAAY,KAAK;EAErB;EAEA,IAAI,CAAC,UAAU;GAEb,MAAM,WAAU,MADC,UAAU,OAAO,IAAI,KAAK,EAAA,CACxB,QAAQ,MAAM,EAAE,WAAW,OAAO,CAAC,CAAC,IAAI;GAC3D,IAAI,SAAS;IACX,WAAW,QAAQ;IACnB,IAAI,oBAAoB,OAAO,GAAG,YAAY,QAAQ,QAAQ,IAAI,CAAC,CAAC,MAAM,GAAG,EAAE;GACjF;EACF;EAIA,IAAI,CAAC,WAAW;GACd,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,SAAS,WAAW,OAAO,EAAE,cAAc,QAAQ;GACrF,IAAI,OAAO,SAAS,SAAS,YAAY,MAAM;EACjD;EAEA,MAAM,MAAM,GAAG,IAAI,aAAa,GAAG,YAAY,GAAG,GAAG,aAAa,GAAG,GAAG,aAAa;EACrF,IAAI,UAAU,SAAS,IAAI,GAAG;EAC9B,IAAI,CAAC,SAAS;GACZ,UAAU;IACR,cAAc,IAAI;IAClB;IACA;IACA;IACA,UAAU;IACV,aAAa,CAAC;IACd,cAAc,IAAI;IAClB,cAAc,kBAAkB,KAAK,KAAK,IAAI;GAChD;GACA,SAAS,IAAI,KAAK,OAAO;EAC3B;EACA,QAAQ;EACR,IAAI,CAAC,QAAQ,YAAY,SAAS,IAAI,UAAU,GAAG,QAAQ,YAAY,KAAK,IAAI,UAAU;CAC5F;CAMA,OAAO;EAAE,UAJG,CAAC,GAAG,SAAS,OAAO,CAAC,CAAC,CAC/B,QAAQ,MAAM,EAAE,YAAY,OAAO,CAAC,CACpC,MAAM,GAAG,MAAM,EAAE,WAAW,EAAE,QAEZ;EAAG;EAAe,WAAW,KAAK;CAAO;AAChE;AAEA,SAAS,kBAAkB,OAAmC;CAE5D,OADgB,MAAM,MAAM,MAAM,EAAE,WAAW,OAClC,CAAC,EAAE;AAClB;;;;;;;;;;;;ACrFA,eAAsB,mBAAmB,OAAkD;CACzF,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,MAAM,QAAQ,CAAC,EAAA,CAAG,QAChD,MAAsB,EAAE,SAAS,OACpC;CACA,IAAI,IAAI,WAAW,GAAG,OAAO;EAAE,OAAO,CAAC;EAAG,YAAY,CAAC;EAAG,UAAU,CAAC;CAAE;CAEvE,MAAM,8BAAc,IAAI,IAAyB;CACjD,KAAK,MAAM,KAAK,KAAK;EACnB,MAAM,MAAM,YAAY,IAAI,EAAE,SAAS,KAAK,CAAC;EAC7C,IAAI,KAAK,CAAC;EACV,YAAY,IAAI,EAAE,WAAW,GAAG;CAClC;CAEA,MAAM,WAAW,CAAC,GAAG,IAAI,IAAI,IAAI,KAAK,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK;CAC9D,MAAM,QAAqB,CAAC;CAC5B,KAAK,MAAM,CAAC,KAAK,UAAU,aAAa;EACtC,MAAM,0BAAU,IAAI,IAAiC;EACrD,KAAK,MAAM,KAAK,OAAO;GACrB,MAAM,IAAI,QAAQ,IAAI,EAAE,OAAO,qBAAK,IAAI,IAAoB;GAC5D,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK;GAC7B,QAAQ,IAAI,EAAE,SAAS,CAAC;EAC1B;EACA,MAAM,aAAa,CAAC,GAAG,QAAQ,KAAK,CAAC;EACrC,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KACrC,KAAK,IAAI,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;GAC9C,MAAM,SAAS,WAAW;GAC1B,MAAM,SAAS,WAAW;GAC1B,MAAM,IAAI,QAAQ,IAAI,MAAM;GAC5B,MAAM,IAAI,QAAQ,IAAI,MAAM;GAC5B,MAAM,SAAkC,CAAC;GACzC,KAAK,MAAM,CAAC,QAAQ,WAAW,GAAG;IAChC,MAAM,SAAS,EAAE,IAAI,MAAM;IAC3B,IAAI,WAAW,KAAA,GAAW,OAAO,KAAK,CAAC,QAAQ,MAAM,CAAC;GACxD;GACA,IAAI,OAAO,SAAS,GAAG;GACvB,MAAM,cAAc,OAAO,KACxB,CAAC,QAAQ,YACR,CACE;IAAE,WAAW;IAAQ,WAAW;IAAK,OAAO;IAAQ,WAAW;GAAG,GAClE;IAAE,WAAW;IAAQ,WAAW;IAAK,OAAO;IAAQ,WAAW;GAAG,CACpE,CACJ;GACA,MAAM,IAAI,sBACR,YAAY,EAAE,CAAE,KAAK,GAAG,OAAO,YAAY,KAAK,SAAS,KAAK,GAAI,CAAC,CACrE;GACA,MAAM,KAAK;IACT,QAAQ;IACR,QAAQ;IACR,WAAW;IACX,aAAa,OAAO;IACpB,SAAS,SACP,OAAO,KAAK,MAAM,EAAE,EAAE,GACtB,OAAO,KAAK,MAAM,EAAE,EAAE,CACxB;IACA,cAAc;GAChB,CAAC;EACH;CAEJ;CAEA,OAAO;EACL,OAAO,MAAM,MAAM,GAAG,MAAM,EAAE,cAAc,EAAE,WAAW;EACzD,YAAY,CAAC,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK;EACzC;CACF;AACF;;;;;;;;;;;ACtDA,eAAsB,sBACpB,OACA,OACA,UAA0B,CAAC,GACF;CACzB,MAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;CAC1C,IAAI,MAAM,WAAW,GACnB,OAAO;EACL;EACA,YAAY;EACZ,uBAAuB;EACvB,QAAQ,CAAC;EACT,WAAW;EACX,eAAe;EACf,WAAW;CACb;CAGF,MAAM,SAAoC,CAAC;CAC3C,IAAI,cAAc;CAClB,IAAI,kBAAkB;CACtB,IAAI,wBAAwB;CAC5B,MAAM,cAAc,CAAC,GAAG,KAAK,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;CACvE,MAAM,iCAAiB,IAAI,IAAY;CAGvC,KAAK,MAAM,KAAK,aAAa;EAC3B,OAAO,EAAE,cAAc;GACrB,OAAO;GACP,uBAAuB;GACvB,QAAQ;GACR,cAAc;GACd,YAAY;EACd;EACA,MAAM,OAAO,OAAO,EAAE;EACtB,KAAK,SAAS;EACd,IAAI,EAAE,WAAW,SAAS;GACxB,KAAK,UAAU;GACf,eAAe;EACjB;EACA,IAAI,OAAO,EAAE,cAAc,UAAU,KAAK,gBAAgB,EAAE;EAC5D,IAAI,oBAAoB,CAAC,GAAG;GAC1B,yBAAyB;GACzB,KAAK,yBAAyB;GAC9B,MAAM,MAAM,GAAG,EAAE,SAAS,GAAG,QAAQ,EAAE,IAAI;GAC3C,IAAI,eAAe,IAAI,GAAG,GAAG;IAC3B,KAAK,cAAc;IACnB,mBAAmB;GACrB;GACA,eAAe,IAAI,GAAG;EACxB;CACF;CAEA,KAAK,MAAM,QAAQ,OAAO,OAAO,MAAM,GACrC,KAAK,eAAe,KAAK,QAAQ,IAAI,KAAK,eAAe,KAAK,QAAQ;CAIxE,IAAI,qBAAqB;CACzB,IAAI,kBAAkB;CACtB,KAAK,MAAM,GAAG,QAAQ,QAAQ,cAAc,MAAM,EAAE,QAAQ,GAC1D,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;EACnC,IAAI,IAAI,EAAE,CAAE,WAAW,SAAS;EAChC,sBAAsB;EACtB,IAAI,IAAI,IAAI,IAAI,mBAAmB;CACrC;CAEF,MAAM,YAAY,qBAAqB,IAAI,kBAAkB,qBAAqB;CAElF,IAAI;CACJ,IAAI,QAAQ,iBAAiB;EAC3B,MAAM,UAAU,YAAY,QAAQ,MAAM,EAAE,UAAU,QAAQ,eAAgB;EAC9E,IAAI,QAAQ,SAAS,GACnB,oBACE,QAAQ,QAAQ,MAAM,QAAQ,gBAAiB,EAAE,OAAO,CAAC,CAAC,SAAS,QAAQ;CAEjF;CAEA,OAAO;EACL;EACA,YAAY,YAAY;EACxB;EACA;EACA,WAAW,cAAc,YAAY;EACrC,eAAe,wBAAwB,IAAI,kBAAkB,wBAAwB;EACrF;EACA;CACF;AACF;;;;;;;;;;;;;;AC/FA,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;CAC1B,MAAM,OAAO,QAAQ,QAAQ,CAAC,QAAQ,KAAK,KAAK,MAAM,MAAM,SAAS,EAAA,CAAG,KAAK,MAAM,EAAE,KAAK;CAE1F,MAAM,QAA4B,CAAC;CACnC,IAAI,aAAa;CACjB,IAAI,cAAc;CAClB,KAAK,MAAM,SAAS,MAAM;EACxB,MAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;EAC1C,IAAI,MAAM,WAAW,GAAG;GACtB,MAAM,KAAK;IAAE;IAAO,aAAa;IAAG,YAAY;IAAG,WAAW;GAAE,CAAC;GACjE;EACF;EAQA,MAAM,YAAY,CAAC,GAAG,MAPH,SAAS,OAAO,KAAK,CAOd,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;EACpE,MAAM,aAAa,UAAU,KAAK,MAAM,EAAE,SAAS;EACnD,MAAM,aAAa,gBAAgB,SAAS;EAC5C,IAAI,SAAS;EACb,KAAK,MAAM,KAAK,OAAO;GACrB,IAAI,EAAE,WAAW,SAAS;IACxB;IACA;GACF;GAEA,MAAM,SAAS,WAAW,YAAY,EAAE,SAAS;GACjD,IAAI,QAAQ,aACN;QAAA,CAAC,QAAQ,YAAY,GAAG,EAAE,KAAK,UAAU,MAAM,MAAM,EAAE,CAAC,GAAG;GAAA,OAC1D;IAKL,MAAM,YAAY,UAAU,EAAE,MAAM;IACpC,IAAI,cAAc,IAAI;IAGtB,IAAI,EAFa,WAAW,WAAW,GAAA,CACjB,SAAS,UAAU,MAAM,GAAG,GAAG,CAC7C,GAAG;GACb;EACF;EACA,MAAM,YAAY,SAAS,MAAM;EACjC,MAAM,KAAK;GAAE;GAAO,aAAa;GAAQ,YAAY,MAAM;GAAQ;EAAU,CAAC;EAC9E,cAAc,MAAM;EACpB,eAAe;CACjB;CACA,OAAO;EAAE;EAAO,kBAAkB,aAAa,IAAI,cAAc,aAAa;CAAE;AAClF;;;;;;AAOA,SAAS,gBAAgB,OAA4B;CACnD,MAAM,SAAS,IAAI,MAAc,MAAM,SAAS,CAAC;CACjD,OAAO,MAAM,UAAU;CACvB,KAAK,IAAI,IAAI,MAAM,SAAS,GAAG,KAAK,GAAG,KAAK;EAC1C,MAAM,MAAM,MAAM,EAAE,CAAE,SAAS,KAAK,MAClC,OAAO,EAAE,YAAY,WAAW,EAAE,UAAU,EAC9C,CAAC,CAAC,KAAK,IAAI;EACX,OAAO,KAAK,GAAG,IAAI,IAAI,OAAO,IAAI;CACpC;CACA,OAAO;AACT;;AAGA,SAAS,WAAW,QAAkB,QAAwB;CAC5D,IAAI,KAAK;CACT,IAAI,KAAK,OAAO;CAChB,OAAO,KAAK,IAAI;EACd,MAAM,MAAO,KAAK,OAAQ;EAC1B,IAAI,OAAO,QAAS,QAAQ,KAAK,MAAM;OAClC,KAAK;CACZ;CACA,OAAO;AACT;AAEA,SAAS,UAAU,GAAoB;CACrC,IAAI,MAAM,QAAQ,MAAM,KAAA,GAAW,OAAO;CAC1C,IAAI,OAAO,MAAM,UAAU,OAAO;CAClC,IAAI;EACF,OAAO,KAAK,UAAU,CAAC;CACzB,QAAQ;EACN,OAAO,OAAO,CAAC;CACjB;AACF"}
@@ -1,6 +1,6 @@
1
- import { c as CostLedgerHandle } from "./cost-ledger-CO7UyEcQ.js";
2
- import { O as TraceAnalysisStore } from "./types-DHb5TI36.js";
3
- import { a as TraceAnalystLimits, l as RawAnalystFinding, n as TraceAnalysisEngine } from "./engine-C37HS2kr.js";
1
+ import { a as CostLedgerHandle } from "./cost-ledger-Jk62OBfF.js";
2
+ import { X as TraceAnalysisStore } from "./types-C1tnkEgM.js";
3
+ import { a as TraceAnalystLimits, l as RawAnalystFinding, n as TraceAnalysisEngine } from "./engine-3A_4jOlI.js";
4
4
  //#region src/analyst/trace-questions.d.ts
5
5
  /** One question to ask of the store. */
6
6
  interface TraceQuestionSpec {
@@ -65,4 +65,4 @@ declare function askTraceQuestions(opts: AskTraceQuestionsOptions): Promise<Trac
65
65
  declare function citedSpansOnly(store: TraceAnalysisStore, allowed: ReadonlyMap<string, ReadonlySet<string>>): TraceAnalysisStore;
66
66
  //#endregion
67
67
  export { askTraceQuestions as a, VerifiedTraceFinding as i, TraceQuestionOutcome as n, citedSpansOnly as o, TraceQuestionSpec as r, AskTraceQuestionsOptions as t };
68
- //# sourceMappingURL=trace-questions-CLx8vt6t.d.ts.map
68
+ //# sourceMappingURL=trace-questions-CM2CAocU.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"trace-questions-CLx8vt6t.d.ts","names":[],"sources":["../src/analyst/trace-questions.ts"],"mappings":";;;;;UA6BiB;;WAEN;WACA;;WAEA;;UAGM;WACN,oBAAoB;WACpB,OAAO;WACP,QAAQ;;WAER,YAAY;;WAEZ;;WAEA;;WAEA;;WAEA;;WAEA;WACA,SAAS,QAAQ;WACjB,SAAS;WACT,OAAO,iBAAiB,SAAS;;;UAI3B;WACN,SAAS;;WAET;;WAEA;;WAEA;;UAGM;WACN;WACA;WACA;WACA;;WAEA;WACA,mBAAmB;;WAEnB;WACA;;;;;;;iBAyBW,kBACpB,MAAM,2BACL,QAAQ;;;;;iBA2JK,eACd,OAAO,oBACP,SAAS,oBAAoB,uBAC5B"}
1
+ {"version":3,"file":"trace-questions-CM2CAocU.d.ts","names":[],"sources":["../src/analyst/trace-questions.ts"],"mappings":";;;;;UA6BiB;;WAEN;WACA;;WAEA;;UAGM;WACN,oBAAoB;WACpB,OAAO;WACP,QAAQ;;WAER,YAAY;;WAEZ;;WAEA;;WAEA;;WAEA;;WAEA;WACA,SAAS,QAAQ;WACjB,SAAS;WACT,OAAO,iBAAiB,SAAS;;;UAI3B;WACN,SAAS;;WAET;;WAEA;;WAEA;;UAGM;WACN;WACA;WACA;WACA;;WAEA;WACA,mBAAmB;;WAEnB;WACA;;;;;;;iBAyBW,kBACpB,MAAM,2BACL,QAAQ;;;;;iBA2JK,eACd,OAAO,oBACP,SAAS,oBAAoB,uBAC5B"}
@@ -1,8 +1,8 @@
1
+ import { a as CostLedgerHandle, l as CostProvenance, p as CustomTokenPricing } from "../cost-ledger-Jk62OBfF.js";
1
2
  import { r as CaptureIntegrityError } from "../errors-DEE6u6ot.js";
2
- import { b as CustomTokenPricing, c as CostLedgerHandle, p as CostProvenance } from "../cost-ledger-CO7UyEcQ.js";
3
- import { d as RunTokenUsage } from "../run-record-D1Hw-Jml.js";
3
+ import { d as RunTokenUsage } from "../run-record-C1E24G2A.js";
4
4
  import { t as DefaultVerdict } from "../verdict-E4eRNf7-.js";
5
- import { a as TraceAnalystLimits, n as TraceAnalysisEngine } from "../engine-C37HS2kr.js";
5
+ import { a as TraceAnalystLimits, n as TraceAnalysisEngine } from "../engine-3A_4jOlI.js";
6
6
  import { t as PrimeBridgeTransport } from "../prime-bridge-transport-6feEglLf.js";
7
7
  import { a as RecordedTrajectoryStep, h as isRecordedTimeout } from "../steps-Bg5-KVx1.js";
8
8
  //#region src/trace-repair/mini-swe-scaffold.d.ts
@@ -1 +1 @@
1
- {"version":3,"file":"index.d.ts","names":["RepairArm"],"sources":["../../src/trace-repair/mini-swe-scaffold.ts","../../src/trace-repair/action-budget.ts","../../src/trace-repair/continuation-records.ts","../../src/trace-repair/control-policy.ts","../../src/trace-repair/admission-records.ts","../../src/trace-repair/continuation-policy.ts","../../src/trace-repair/oracle-determinism.ts","../../src/trace-repair/admission.ts","../../src/trace-repair/admission-contract.ts","../../src/trace-repair/admission-report.ts","../../src/trace-repair/analyst-response.ts","../../src/trace-repair/blinding.ts","../../src/trace-repair/analyst-arm.ts","../../src/trace-repair/arm-completion.ts","../../src/trace-repair/arm-dspy.ts","../../src/trace-repair/degenerate-strategies.ts","../../src/trace-repair/ports.ts","../../src/trace-repair/funnel.ts","../../src/trace-repair/grade.ts","../../src/trace-repair/delta-repair.ts","../../src/trace-repair/docker-environment.ts","../../src/trace-repair/repair-prompt.ts","../../src/trace-repair/test-oracle.ts"],"mappings":";;;;;;;;;qBAyBa;;qBAGA;qBAKA;UAiBI;;EAEf;;;;;EAKA;;;wBAIc,sBAAsB,OAAO;KAiFjC;EACN;EAAgB;;EAChB;EAAsB;;;;;;;wBASZ,YAAY,2BAA2B;UAMtC;EACf;EACA;;EAEA;;;;;;;wBAQc,kBAAkB,QAAQ;;wBAsB1B,yBAAyB,iBAAiB;;wBAS1C,6BAA6B;;;;;;wBAyB7B,aAAa,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;KCxMzB;UAEK;;WAEN;;WAEA;;WAEA;;;qBAIE,8BAA8B;;;;;;;;;;wBA4C3B,gBAAgB;EAAmB;EAAsB;;;;wBA8QzD,sBAAsB,iBAAiB;;;KAM3C;UAQK;EACf;EACA;EACA;;EAEA,SAAS;;EAET,UAAU;;KAGA;WACG;WAA2B,aAAa;;WAExC;WACA,WAAW;WACX;WACA,aAAa;;;;;;;;;;;;;;;wBAqBZ,wBACd,gBACA,cAAc,mBACd,SAAQ,qBACP;;;;;;wBA8Da,6BAA6B;;;;KCtcjC;;UAGK;EACf;EACA;;;KAIU;UAOK;EACf;EACA;;EAEA;EACA;EACA;;UAGe;;EAEf;EACA;EACA;;EAEA,OAAO;;EAEP;EACA;EACA;;UAGe;;EAEf;EACA;;EAEA;;;;;;EAMA;;EAEA,WAAW;EACX,OAAO;;EAEP;;UAGe;;;;;EAKf;;EAEA;;EAEA;EACA;EACA;EACA;EACA;EACA;;UAGe;;EAEf;EACA;;UAGe;EACf;EACA,KAAK;;EAEL;;EAEA;;EAEA;;EAEA;EACA;EACA;EACA,aAAa;EACb,OAAO;EACP,YAAY;;EAEZ;EACA,OAAO;EACP,gBAAgB;EAChB;EACA;EACA;;EAEA;;;UAIe;EACf;EACA;;;UAIe;EACf;EACA;EACA,OAAO;EACP;;;;;;;;;wBAUc,gBAAgB,mBAAmB,wBAAwB;;wBA+B3D,qBAAqB,SAAS,sBAAsB;;;;;;;;;wBAiBpD,cAAc,SAAS;;;;qBCvK1B,iCAAiC;;;;;;;;;;;KAYlC;qBAEC,kCAAkC;UAE9B;;WAEN;;WAEA;;WAEA;;WAEA;;WAEA;;;;;;;;;;UAWM;WACN;WACA;;WAEA;;UAGM,sBAAsB,oBAAoB;;;WAGhD;;;;;;WAMA;;;;;;;wBAQK,iBAAiB;wBAIjB,oBAAoB,OAAO,qBAAqB;;;;;;;;;wBAiDhD,wBACd,QAAQ,mBACR,WAAW;;;;qBC9HA,kCAAkC;;qBAGlC,mCAAmC;;;;;;;;UAW/B;;EAEf;;EAEA;;EAEA;;;;;EAKA;;;;;;EAMA;;;;;;;;;wBAkBc,yBAAyB,eAAe;;;;;;;;;;;;KAyB5C;qBAEC,2BAA2B;;wBAOxB,UAAU,iCAAiC;;;;;;;;KAe/C;qBAiBC,oCAAoC;;qBA6BpC,6BAA6B,SAAS,OAAO;KAsB9C;;UAGK;;EAEf;EACA;;KAGU;EACN;EAAkB,SAAS;;EAE3B;EACA;EACA;EACA;;EAEA;EAAsB;EAAiB;EAAkB;;EACzD;EAA0B;EAAiB;;EAE3C;EACA,KAAK;EACL;EACA;EACA,YAAY;;;UAID;EACf,KAAK;EACL;EACA;EACA;EACA,YAAY;EACZ;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA,SAAS;;;;;EAKT;EACA,kBAAkB;;;EAGlB;EACA;;EAEA,YAAY;;EAEZ;;EAEA,QAAQ;EACR,UAAU;;UAKK;EACf,QAAQ;EACR;EACA;EACA;;UAGe;;EAEf,eAAe;EACf;EACA,QAAQ;EACR;;UAGe;EACf;EACA,SAAS;EACT,WAAW;EACX,cAAc,OAAO;;EAErB;;EAEA,sBAAsB;;;;;;;;;wBAUR,sBACd,mBAAmB,uBACnB,sBAAsB,qBACrB;;wBAsDa,sBAAsB,UAAU;;;;qBC7SnC,yCAAyC;;qBAGzC,kCAAkC;UAE9B;;WAEN;;WAEA;;WAEA;;WAEA;WACA;WACA;;WAEA;;WAEA;;WAEA;WACA;;UAqBM,sCACP,QAAQ,KAAK;;EAErB;;EAEA;;wBAGc,+BACd,OAAO,gCACN;;;;;;wBA+Ba,yBAAyB,QAAQ;;;;;wBAejC,iBAAiB,oBAAoB,eAAe;UAYnD;EACf;EACA,UAAU;EACV;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA,OAAO;;EAEP;EACA;;KAGU,qBACV,SAAS,6BACN,QAAQ;UAEI;EACf;EACA;;EAEA;;EAEA;;UAGe;;EAEf;EACA,YAAY,QAAQ;EACpB,KAAK,iBAAiB;IAAW;MAA2B,QAAQ;EACpE,WAAW;;UAGI;EACf;EACA,KAAK;EACL;;UAGe;EACf;;;;;;EAMA,OAAO,SAAS,iCAAiC,QAAQ;;UAG1C;EACf,QAAQ;EACR,KAAK;;EAEL;;;;;;EAMA,iBAAiB;;EAEjB;;;;;;;EAOA;EACA,OAAO;EACP,cAAc;;EAEd;;;;;;;;;wBAUoB,gBACpB,SAAS,yBACR,QAAQ;;;;;;wBAuOK,WAAW,gBAAgB,2BAA2B;;;;;wBA8CtD,UAAU,gBAAgB,2BAA2B;;;;;;wBAerD,kBAAkB,mBAAmB;;;;qBClfxC,oCAAoC;;;KAIrC;;KAGA;;qBAGC;UAEI;;WAEN;WACA;;UAGM;;WAEN;;;;;;WAMA;WACA;WACA;;;WAGA,qBAAqB;;UAGf;WACN,OAAO;WACP,MAAM;;;;;WAKN,qBAAqB;;UAGf;WACN;;;;WAIA;;;;;;;WAOA;WACA,iBAAiB;WACjB;;;UAIM;;WAEN;WACA;WACA;WACA;;UAGM;WACN,OAAO;WACP;;WAEA;WACA;;WAEA;;WAEA;WACA,kBAAkB;;;WAGlB;;WAEA;;WAEA;;UAGM;WACN;WACA;WACA;;WAEA;WACA;;WAEA;WACA,kBAAkB;WAClB;;WAEA;;;;;;;;;wBAaK,kBAAkB,UAAU,4BAA4B;;;;;;;;KA+H5D,qBAAqB,oBAAoB;;;;;;;;;UAUpC;WACN;WACA,uBAAuB;;wBAGlB,wBAAwB,oBAAoB;wBAc5C,mBACd,mBAAmB,6BAClB;;wBAca,0BAA0B,SAAS;;;;;;;KC3PvC,iBAAiB;EACvB;EAAiB,OAAO;;EACxB;EAAkB;;UAEP;;EAEf;EACA;EACA;;UAGe;;EAEf;;EAEA,4BAA4B;;UAGb;;EAEf;;EAEA,OAAO,KAAK,eAAe,QAAQ,iBAAiB;;UAGrC;EACf;;EAEA;;UAGe;EACf;;EAEA,MAAM,KAAK,eAAe,QAAQ,iBAAiB;;UAGpC;EACf,KAAK;EACL,KAAK;;EAEL;;EAEA,WAAW;;UAGI;;EAEf,OAAO;;EAEP,SAAS;;UAGM;EACf;;;;;EAKA,IAAI,SAAS,0BAA0B,QAAQ,iBAAiB;;UAKjD;;EAEf;;EAEA;;;;;;EAMA,uBAAuB;;EAEvB;;;;;EAKA,mBAAmB;;EAEnB;;UAGe;WACN;WACA;WACA,sBAAsB;WACtB;WACA,kBAAkB;WAClB;;qBAGE,2BAA2B;wBASxB,uBAAuB,QAAO,uBAA4B;;;;;;;wBA6C1D,kBACd,oBACA,eACA,sBACA;UAiBe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;;;EAGA;EACA,kBAAkB;;;EAGlB,gBAAgB,SAAS;;UAGV;EACf,QAAQ;EACR,YAAY;;EAEZ,eAAe;;EAEf,QAAQ,SAAS,OAAO;EACxB,OAAO;;EAEP,aAAa;;EAEb;EACA;;UAGe;EACf,eAAe;;EAEf,QAAQ;EACR,UAAU;EACV,QAAQ;EACR,UAAU;;;;;;;EAOV,aAAa;EACb,SAAS;;EAET;;;;;;;;;wBAUoB,aAAa,SAAS,sBAAsB,QAAQ;;wBAoV1D,eACd,QAAQ,iBACR,SAAS;wBAKK,cAAc,QAAQ;UAIrB;EACf,QAAQ;;EAER,iBAAiB;;EAEjB;;EAEA;;;;;;;;;;;wBAYc,wBAAwB,OAAO;;;;UC9mB9B;;;;WAIN;;WAEA;;WAEA,kBAAkB;;qBAGhB,8BAA8B;UAM1B;;WAEN;;WAEA;;UAGM;WACN;;WAEA;;;WAGA;;UAGM;WACN;;;WAGA;;WAEA;;WAEA;;WAEA;WACA,gBAAgB;;;;;;WAMhB,mBAAmB;;WAEnB,eAAe;WACf,gBAAgB;;WAEhB;;WAEA;WACA,cAAc;WACd,aAAa;;;KAIZ;cAWE;;;;;;;;;;UAWG;YACL;WACD;WACA;WACA;WACA;WACA;WACA,gBAAgB;WAChB,UAAU;WACV;WACA;;;WAGA,eAAe;WACf,kBAAkB;;;WAGlB;;;WAGA;WACA;;;;UAKM;WACN,eAAe;WACf,kBAAkB;WAClB;WACA;WACA;WACA;;KAGC;WAEG;WACA,WAAW;WACX,KAAK;;WAGL;WACA;WACA,WAAW;WACX,WAAW;WACX;;;;;;;;;;;;;;wBAeC,SACd,UAAU,mBACV,WAAU,oBACT;;;UC7Kc;EACf;EACA;EACA;;EAEA;EACA,QAAQ;EACR,YAAY;EACZ,OAAO;;EAEP,UAAU,OAAO;EACjB;IAAe;IAAc;;;EAE7B,eAAe;;;wBAID,kBAAkB,QAAQ,kBAAkB;UAoB3C;;EAEf;;;wBAIc,sBACd,UAAU,mBACV,UAAS;;;;qBChDE;UAEI;;WAEN,MAAM;;WAEN;;UAGM;WACN;;;WAGA;;;WAGA;WACA,cAAc;;UAGR;WACN;;KAGC,kBAAkB,gBAAgB;;KAGlC;KAQA;WACG;WAA0B,OAAO;;WACjC;WAA2B,SAAS;WAA+B;;;;;;;;;wBASlE,qBAAqB,gBAAgB;;wBAoFrC,cAAc;EAC5B;EACA;EACA,cAAc;IACZ;;;UCpIa;WACN;WACA;WACA;;UAGM;WACN;;WAEA;WACA,gBAAgB;;;WAGhB;;WAEA;;UAGM;;;WAGN;;wBAGK,gBACd,KAAK,aACL,UAAS,yBACR;;;;;qBAyBU;;;;;;;;;;;KCtBD;;;;;;;;;;;;;;;;;;KAmBA;WACG;WAAuB;;WAEvB;;WAEA;;WAEA;;WAEA;;UAGE;WACN;;;WAGA;WACA,eAAe;;WAEf,QAAQ;;WAER;WACA,sBAAsB;;;;;WAKtB;;;;;UAMM;WACN,QAAQ;WACR,SAAS;;;;UAKH;WACN;WACA;WACA;WACA;;UAGM;WACN;;WAEA;;;UAIM;WACN;WACA;;;;;;;;;;KAWC;WAEG;WACA;WACA;WACA,cAAc;WACd;WACA;WACA,uBAAuB;WACvB,QAAQ;WACR,OAAO;;WAGP;WACA;WACA;WACA,uBAAuB;WACvB,QAAQ;WACR,OAAO;;WAGP;WACA;WACA;WACA,uBAAuB;WACvB,QAAQ;WACR,OAAO;;UAGL;WACN,aAAa;EACtB,IAAI,SAAS,mBAAmB,QAAQ;;;;;;;;;;KAW9B;WACG;WAA2B,aAAa;;WAExC;WACA,WAAW;WACX;WACA,aAAa;;UAGX;WACN;WACA;;WAEA;WACA,OAAO;;WAEP,QAAQ;WACR;;UAGM;WACN,KAAK;WACL,KAAK;;WAEL;WACA,SAAS;;WAET;;;;;;;;wBASW,aAAa,SAAS,sBAAsB,QAAQ;;;;;;;;wBAuC1D,kBAAkB,QAAQ,kBAAkB;;UAc3C;WACN;;WAEA,2BAA2B;;WAE3B,6BAA6B;WAC7B,eAAe;;;;WAIf;;UAGM;WACN;;;WAGA,4BAA4B;WAC5B,sBAAsB;;;WAGtB;WACA,QAAQ;WACR;;;WAGA;;;;;;;;;;;;;;;wBAgBK,qBACd,eAAe,aACf;WAAoB,SAAS;IAC5B;;;UCpQc;WACN;;WAEA;;WAEA;WACA;;WAEA,WAAW;;WAEX;;WAEA,SAAS;WACT,SAAS;;WAET,sBAAsB;WACtB,eAAe;;wBAGV,0BAA0B,SAAS,6BAA6B;;;;;;;;;;qBChBnE;;qBAGA;;;;;;;;;wBAaG,uBACd,SAAQ;UAqBO;;WAEN,QAAQ;WACR,QAAQ;WACR,YAAY;WACZ;WACA;WACA,WAAW;WACX,SAAS;WACT;WACA,OAAO,iBAAiB,SAAS;;wBAG5B,oBAAoB,SAAS,uBAAuB;UAuJnD;;;;;WAKN;;;wBA6JK,uBAAuB,QAAQ;;;;;;;;;;;;;;;;;KCrZnC;UAEK;WACN;;WAEA;;WAEA;WACA,YAAY;;WAEZ;;qBAGE;WAEL;WAEF;WAEA;WACU;WACA;;WAGR;WAEF;WAEA;WACU;WACA;;WAGR;WAEF;WAEA;WACU;WACA;;WAGR;WAEF;WAEA;WACU;WACA;;WAGR;WAEF;WAEA;WACU;WACA;;WAGR;WAEF;WAEA;WACU;WACA;;WAGR;WAEF;WAEA;WACU;WACA;;WAGR;WAEF;WAEA;WACU;WACA;;KAIJ,+BAA+B;;;;;;;;KCjF/BA;UAQK;EACf;EACA;EACA;;EAEA;;UAGe;;WAEN;EACT,KAAK,iBAAiB,oBAAoB,QAAQ;EAClD,SAAS;;UAGM;EACf;;;EAGA;EACA,KAAKA;;EAEL;;UAGe;;EAEf,KAAK,SAAS,uBAAuB,QAAQ;;UAG9B;EACf;EACA,KAAKA;EACL;;UAGe;;EAEf;EACA;EACA;;EAEA;EACA;;;;;;;;;;UAWe;EACf,MAAM,SAAS,eAAe,SAAS,oBAAoB,QAAQ;;UAGpD;EACf;EACA,KAAKA;EACL;;EAEA,SAAS;;EAET,gBAAgB;;;EAGhB;;;;;;;EAOA,UAAU;EACV;;UAGe;EACf;EACA;EACA;EACA;;UAGe;;EAEf;;;EAGA;;EAEA;;EAEA;EACA;;;KAIU,4BACV,SAAS,8BACN,QAAQ;;;;KCtGD;WACG;WAA0B,QAAQ;WAA+B;;WAEjE;WACA,QAAQ;WACR;WACA,aAAa;;WAGb;WACA;WACA;;UAGE;;WAEN;;WAEA;WACA;;;;;;;;;KAUC;;;WAIG;WACA;;WAGA;WACA;WACA;WACA;WACA;WACA;WACA,QAAQ;;UAGN;WACN;WACA;WACA;WACA;WACA;WACA,QAAQ;;UAGF;WACN;WACA;WACA;;WAEA;;;;;;;;;;KAWC;WAEG;WACA;WACA;WACA,cAAc;WACd,OAAO;;WAGP;WACA;WACA;WACA;;UAGE;;WAEN;;WAEA;;WAEA;;;WAGA;WACA,0BAA0B;;;;;;;;KASzB;WACG;WAA8B,WAAW;;WACzC;;WAEA;WACA;WACA,cAAc;;WAGd;WACA;WACA,cAAc;WACd,WAAW;;WAGX;WACA;WACA,cAAc;WACd,WAAW;WACX,WAAW;WACX,QAAQ;;;;;;;;;;UAWN;WACN;WACA;;WAEA;;qBAaE;;wBASG,aAAa,OAAO,cAAc;UAoBjC;WACN;WACA;WACA;WACA;WACA;WACA;WACA;;WAEA;;WAEA;;wBAGK,YAAY,iBAAiB,gBAAgB;;;UCnK5C;WACN,KAAK;WACL,UAAU;WACV,UAAU;WACV,QAAQ;WACR,cAAc;;;WAGd;WACA,SAAS;;WAET;;;;;;;;WAQA;WACA,cAAc;;;;;;;;;;UAWR,wBAAwB;WAC9B;WACA,OAAO;WACP,QAAQ;;;WAGR;WACA;WACA;;;WAGA,kBAAkB;WAClB;WACA;;WAEA;WACA;;wBAMW,eAAe,SAAS,qBAAqB,QAAQ;;;UCvF1D;;WAEN;WACA;WACA;;UAGM;WACN;WACA;WACA;WACA;WACA;WACA;WACA;;;WAGA;;KAGC;UAUK;WACN,IAAI;WACJ;;WAEA;;UAGM;WACN;WACA,QAAQ;;WAER;WACA;WACA,aAAa;;;;WAIb,cAAc;WACd;WACA,qBAAqB;WACrB,kBAAkB;;wBAGb,YACd,qBAAqB,mBACrB,UAAS,qBACR;;;wBAwKa,wBAAwB,QAAQ;;;UCtO/B;EACf;;EAEA;;UAGe;;EAEf;EACA;;EAEA;;;KAIU,iBAAiB,SAAS,mBAAmB,QAAQ;;;;;;qBAOpD,mBAAmB;UAuDf;;EAEf;;EAEA;;EAEA,MAAM;;EAEN;;EAEA;EACA,YAAY;;EAEZ;;;;;;;;wBAqBc,cAAc;EAC5B;EACA;EACA;EACA;EACA;;wBAmBc,oCACd,SAAS,uCACR;;;qBCxIU;;;;;wBAOG,iBACd,SAAQ;;qBAkBG;;;qBAgBA;;wBASG,uBAAuB,QAAQ;wBAc/B,uBAAuB,QAAQ;;;wBAM/B,qBACd,QAAQ,yBACR,SAAQ;;;;;;;;;wBAkBM,qBACd,SAAQ;;;;;;;;;wBAsBM,sBACd,QAAQ,oBACR;;;;qBC5GW,+BAA+B;;qBAG/B,wBAAwB;UAEpB;;WAEN;WACA;;WAEA;;UAGM;;WAEN,gBAAgB;;WAEhB;;;WAGA;WACA;WACA;;;;;;wBAUK,gBAAgB,gBAAgB;wBAWhC,mBAAmB,SAAS,4BAA4B"}
1
+ {"version":3,"file":"index.d.ts","names":["RepairArm"],"sources":["../../src/trace-repair/mini-swe-scaffold.ts","../../src/trace-repair/action-budget.ts","../../src/trace-repair/continuation-records.ts","../../src/trace-repair/control-policy.ts","../../src/trace-repair/admission-records.ts","../../src/trace-repair/continuation-policy.ts","../../src/trace-repair/oracle-determinism.ts","../../src/trace-repair/admission.ts","../../src/trace-repair/admission-contract.ts","../../src/trace-repair/admission-report.ts","../../src/trace-repair/analyst-response.ts","../../src/trace-repair/blinding.ts","../../src/trace-repair/analyst-arm.ts","../../src/trace-repair/arm-completion.ts","../../src/trace-repair/arm-dspy.ts","../../src/trace-repair/degenerate-strategies.ts","../../src/trace-repair/ports.ts","../../src/trace-repair/funnel.ts","../../src/trace-repair/grade.ts","../../src/trace-repair/delta-repair.ts","../../src/trace-repair/docker-environment.ts","../../src/trace-repair/repair-prompt.ts","../../src/trace-repair/test-oracle.ts"],"mappings":";;;;;;;;;qBAyBa;;qBAGA;qBAKA;UAiBI;;EAEf;;;;;EAKA;;;wBAIc,sBAAsB,OAAO;KAiFjC;EACN;EAAgB;;EAChB;EAAsB;;;;;;;wBASZ,YAAY,2BAA2B;UAMtC;EACf;EACA;;EAEA;;;;;;;wBAQc,kBAAkB,QAAQ;;wBAsB1B,yBAAyB,iBAAiB;;wBAS1C,6BAA6B;;;;;;wBAyB7B,aAAa,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;KCxMzB;UAEK;;WAEN;;WAEA;;WAEA;;;qBAIE,8BAA8B;;;;;;;;;;wBA4C3B,gBAAgB;EAAmB;EAAsB;;;;wBA8QzD,sBAAsB,iBAAiB;;;KAM3C;UAQK;EACf;EACA;EACA;;EAEA,SAAS;;EAET,UAAU;;KAGA;WACG;WAA2B,aAAa;;WAExC;WACA,WAAW;WACX;WACA,aAAa;;;;;;;;;;;;;;;wBAqBZ,wBACd,gBACA,cAAc,mBACd,SAAQ,qBACP;;;;;;wBA8Da,6BAA6B;;;;KCtcjC;;UAGK;EACf;EACA;;;KAIU;UAOK;EACf;EACA;;EAEA;EACA;EACA;;UAGe;;EAEf;EACA;EACA;;EAEA,OAAO;;EAEP;EACA;EACA;;UAGe;;EAEf;EACA;;EAEA;;;;;;EAMA;;EAEA,WAAW;EACX,OAAO;;EAEP;;UAGe;;;;;EAKf;;EAEA;;EAEA;EACA;EACA;EACA;EACA;EACA;;UAGe;;EAEf;EACA;;UAGe;EACf;EACA,KAAK;;EAEL;;EAEA;;EAEA;;EAEA;EACA;EACA;EACA,aAAa;EACb,OAAO;EACP,YAAY;;EAEZ;EACA,OAAO;EACP,gBAAgB;EAChB;EACA;EACA;;EAEA;;;UAIe;EACf;EACA;;;UAIe;EACf;EACA;EACA,OAAO;EACP;;;;;;;;;wBAUc,gBAAgB,mBAAmB,wBAAwB;;wBA+B3D,qBAAqB,SAAS,sBAAsB;;;;;;;;;wBAiBpD,cAAc,SAAS;;;;qBCvK1B,iCAAiC;;;;;;;;;;;KAYlC;qBAEC,kCAAkC;UAE9B;;WAEN;;WAEA;;WAEA;;WAEA;;WAEA;;;;;;;;;;UAWM;WACN;WACA;;WAEA;;UAGM,sBAAsB,oBAAoB;;;WAGhD;;;;;;WAMA;;;;;;;wBAQK,iBAAiB;wBAIjB,oBAAoB,OAAO,qBAAqB;;;;;;;;;wBAiDhD,wBACd,QAAQ,mBACR,WAAW;;;;qBC9HA,kCAAkC;;qBAGlC,mCAAmC;;;;;;;;UAW/B;;EAEf;;EAEA;;EAEA;;;;;EAKA;;;;;;EAMA;;;;;;;;;wBAkBc,yBAAyB,eAAe;;;;;;;;;;;;KAyB5C;qBAEC,2BAA2B;;wBAOxB,UAAU,iCAAiC;;;;;;;;KAe/C;qBAiBC,oCAAoC;;qBA6BpC,6BAA6B,SAAS,OAAO;KAsB9C;;UAGK;;EAEf;EACA;;KAGU;EACN;EAAkB,SAAS;;EAE3B;EACA;EACA;EACA;;EAEA;EAAsB;EAAiB;EAAkB;;EACzD;EAA0B;EAAiB;;EAE3C;EACA,KAAK;EACL;EACA;EACA,YAAY;;;UAID;EACf,KAAK;EACL;EACA;EACA;EACA,YAAY;EACZ;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA,SAAS;;;;;EAKT;EACA,kBAAkB;;;EAGlB;EACA;;EAEA,YAAY;;EAEZ;;EAEA,QAAQ;EACR,UAAU;;UAKK;EACf,QAAQ;EACR;EACA;EACA;;UAGe;;EAEf,eAAe;EACf;EACA,QAAQ;EACR;;UAGe;EACf;EACA,SAAS;EACT,WAAW;EACX,cAAc,OAAO;;EAErB;;EAEA,sBAAsB;;;;;;;;;wBAUR,sBACd,mBAAmB,uBACnB,sBAAsB,qBACrB;;wBAsDa,sBAAsB,UAAU;;;;qBC7SnC,yCAAyC;;qBAGzC,kCAAkC;UAE9B;;WAEN;;WAEA;;WAEA;;WAEA;WACA;WACA;;WAEA;;WAEA;;WAEA;WACA;;UAqBM,sCACP,QAAQ,KAAK;;EAErB;;EAEA;;wBAGc,+BACd,OAAO,gCACN;;;;;;wBA+Ba,yBAAyB,QAAQ;;;;;wBAejC,iBAAiB,oBAAoB,eAAe;UAYnD;EACf;EACA,UAAU;EACV;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA,OAAO;;EAEP;EACA;;KAGU,qBACV,SAAS,6BACN,QAAQ;UAEI;EACf;EACA;;EAEA;;EAEA;;UAGe;;EAEf;EACA,YAAY,QAAQ;EACpB,KAAK,iBAAiB;IAAW;MAA2B,QAAQ;EACpE,WAAW;;UAGI;EACf;EACA,KAAK;EACL;;UAGe;EACf;;;;;;EAMA,OAAO,SAAS,iCAAiC,QAAQ;;UAG1C;EACf,QAAQ;EACR,KAAK;;EAEL;;;;;;EAMA,iBAAiB;;EAEjB;;;;;;;EAOA;EACA,OAAO;EACP,cAAc;;EAEd;;;;;;;;;wBAUoB,gBACpB,SAAS,yBACR,QAAQ;;;;;;wBAmOK,WAAW,gBAAgB,2BAA2B;;;;;wBA8CtD,UAAU,gBAAgB,2BAA2B;;;;;;wBAerD,kBAAkB,mBAAmB;;;;qBC9exC,oCAAoC;;;KAIrC;;KAGA;;qBAGC;UAEI;;WAEN;WACA;;UAGM;;WAEN;;;;;;WAMA;WACA;WACA;;;WAGA,qBAAqB;;UAGf;WACN,OAAO;WACP,MAAM;;;;;WAKN,qBAAqB;;UAGf;WACN;;;;WAIA;;;;;;;WAOA;WACA,iBAAiB;WACjB;;;UAIM;;WAEN;WACA;WACA;WACA;;UAGM;WACN,OAAO;WACP;;WAEA;WACA;;WAEA;;WAEA;WACA,kBAAkB;;;WAGlB;;WAEA;;WAEA;;UAGM;WACN;WACA;WACA;;WAEA;WACA;;WAEA;WACA,kBAAkB;WAClB;;WAEA;;;;;;;;;wBAaK,kBAAkB,UAAU,4BAA4B;;;;;;;;KA+H5D,qBAAqB,oBAAoB;;;;;;;;;UAUpC;WACN;WACA,uBAAuB;;wBAGlB,wBAAwB,oBAAoB;wBAc5C,mBACd,mBAAmB,6BAClB;;wBAca,0BAA0B,SAAS;;;;;;;KC3PvC,iBAAiB;EACvB;EAAiB,OAAO;;EACxB;EAAkB;;UAEP;;EAEf;EACA;EACA;;UAGe;;EAEf;;EAEA,4BAA4B;;UAGb;;EAEf;;EAEA,OAAO,KAAK,eAAe,QAAQ,iBAAiB;;UAGrC;EACf;;EAEA;;UAGe;EACf;;EAEA,MAAM,KAAK,eAAe,QAAQ,iBAAiB;;UAGpC;EACf,KAAK;EACL,KAAK;;EAEL;;EAEA,WAAW;;UAGI;;EAEf,OAAO;;EAEP,SAAS;;UAGM;EACf;;;;;EAKA,IAAI,SAAS,0BAA0B,QAAQ,iBAAiB;;UAKjD;;EAEf;;EAEA;;;;;;EAMA,uBAAuB;;EAEvB;;;;;EAKA,mBAAmB;;EAEnB;;UAGe;WACN;WACA;WACA,sBAAsB;WACtB;WACA,kBAAkB;WAClB;;qBAGE,2BAA2B;wBASxB,uBAAuB,QAAO,uBAA4B;;;;;;;wBA6C1D,kBACd,oBACA,eACA,sBACA;UAiBe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;;;EAGA;EACA,kBAAkB;;;EAGlB,gBAAgB,SAAS;;UAGV;EACf,QAAQ;EACR,YAAY;;EAEZ,eAAe;;EAEf,QAAQ,SAAS,OAAO;EACxB,OAAO;;EAEP,aAAa;;EAEb;EACA;;UAGe;EACf,eAAe;;EAEf,QAAQ;EACR,UAAU;EACV,QAAQ;EACR,UAAU;;;;;;;EAOV,aAAa;EACb,SAAS;;EAET;;;;;;;;;wBAUoB,aAAa,SAAS,sBAAsB,QAAQ;;wBAoV1D,eACd,QAAQ,iBACR,SAAS;wBAKK,cAAc,QAAQ;UAIrB;EACf,QAAQ;;EAER,iBAAiB;;EAEjB;;EAEA;;;;;;;;;;;wBAYc,wBAAwB,OAAO;;;;UC9mB9B;;;;WAIN;;WAEA;;WAEA,kBAAkB;;qBAGhB,8BAA8B;UAM1B;;WAEN;;WAEA;;UAGM;WACN;;WAEA;;;WAGA;;UAGM;WACN;;;WAGA;;WAEA;;WAEA;;WAEA;WACA,gBAAgB;;;;;;WAMhB,mBAAmB;;WAEnB,eAAe;WACf,gBAAgB;;WAEhB;;WAEA;WACA,cAAc;WACd,aAAa;;;KAIZ;cAWE;;;;;;;;;;UAWG;YACL;WACD;WACA;WACA;WACA;WACA;WACA,gBAAgB;WAChB,UAAU;WACV;WACA;;;WAGA,eAAe;WACf,kBAAkB;;;WAGlB;;;WAGA;WACA;;;;UAKM;WACN,eAAe;WACf,kBAAkB;WAClB;WACA;WACA;WACA;;KAGC;WAEG;WACA,WAAW;WACX,KAAK;;WAGL;WACA;WACA,WAAW;WACX,WAAW;WACX;;;;;;;;;;;;;;wBAeC,SACd,UAAU,mBACV,WAAU,oBACT;;;UC7Kc;EACf;EACA;EACA;;EAEA;EACA,QAAQ;EACR,YAAY;EACZ,OAAO;;EAEP,UAAU,OAAO;EACjB;IAAe;IAAc;;;EAE7B,eAAe;;;wBAID,kBAAkB,QAAQ,kBAAkB;UAoB3C;;EAEf;;;wBAIc,sBACd,UAAU,mBACV,UAAS;;;;qBChDE;UAEI;;WAEN,MAAM;;WAEN;;UAGM;WACN;;;WAGA;;;WAGA;WACA,cAAc;;UAGR;WACN;;KAGC,kBAAkB,gBAAgB;;KAGlC;KAQA;WACG;WAA0B,OAAO;;WACjC;WAA2B,SAAS;WAA+B;;;;;;;;;wBASlE,qBAAqB,gBAAgB;;wBAoFrC,cAAc;EAC5B;EACA;EACA,cAAc;IACZ;;;UCpIa;WACN;WACA;WACA;;UAGM;WACN;;WAEA;WACA,gBAAgB;;;WAGhB;;WAEA;;UAGM;;;WAGN;;wBAGK,gBACd,KAAK,aACL,UAAS,yBACR;;;;;qBAyBU;;;;;;;;;;;KCtBD;;;;;;;;;;;;;;;;;;KAmBA;WACG;WAAuB;;WAEvB;;WAEA;;WAEA;;WAEA;;UAGE;WACN;;;WAGA;WACA,eAAe;;WAEf,QAAQ;;WAER;WACA,sBAAsB;;;;;WAKtB;;;;;UAMM;WACN,QAAQ;WACR,SAAS;;;;UAKH;WACN;WACA;WACA;WACA;;UAGM;WACN;;WAEA;;;UAIM;WACN;WACA;;;;;;;;;;KAWC;WAEG;WACA;WACA;WACA,cAAc;WACd;WACA;WACA,uBAAuB;WACvB,QAAQ;WACR,OAAO;;WAGP;WACA;WACA;WACA,uBAAuB;WACvB,QAAQ;WACR,OAAO;;WAGP;WACA;WACA;WACA,uBAAuB;WACvB,QAAQ;WACR,OAAO;;UAGL;WACN,aAAa;EACtB,IAAI,SAAS,mBAAmB,QAAQ;;;;;;;;;;KAW9B;WACG;WAA2B,aAAa;;WAExC;WACA,WAAW;WACX;WACA,aAAa;;UAGX;WACN;WACA;;WAEA;WACA,OAAO;;WAEP,QAAQ;WACR;;UAGM;WACN,KAAK;WACL,KAAK;;WAEL;WACA,SAAS;;WAET;;;;;;;;wBASW,aAAa,SAAS,sBAAsB,QAAQ;;;;;;;;wBAuC1D,kBAAkB,QAAQ,kBAAkB;;UAc3C;WACN;;WAEA,2BAA2B;;WAE3B,6BAA6B;WAC7B,eAAe;;;;WAIf;;UAGM;WACN;;;WAGA,4BAA4B;WAC5B,sBAAsB;;;WAGtB;WACA,QAAQ;WACR;;;WAGA;;;;;;;;;;;;;;;wBAgBK,qBACd,eAAe,aACf;WAAoB,SAAS;IAC5B;;;UCpQc;WACN;;WAEA;;WAEA;WACA;;WAEA,WAAW;;WAEX;;WAEA,SAAS;WACT,SAAS;;WAET,sBAAsB;WACtB,eAAe;;wBAGV,0BAA0B,SAAS,6BAA6B;;;;;;;;;;qBChBnE;;qBAGA;;;;;;;;;wBAaG,uBACd,SAAQ;UAqBO;;WAEN,QAAQ;WACR,QAAQ;WACR,YAAY;WACZ;WACA;WACA,WAAW;WACX,SAAS;WACT;WACA,OAAO,iBAAiB,SAAS;;wBAG5B,oBAAoB,SAAS,uBAAuB;UAuJnD;;;;;WAKN;;;wBA6JK,uBAAuB,QAAQ;;;;;;;;;;;;;;;;;KCrZnC;UAEK;WACN;;WAEA;;WAEA;WACA,YAAY;;WAEZ;;qBAGE;WAEL;WAEF;WAEA;WACU;WACA;;WAGR;WAEF;WAEA;WACU;WACA;;WAGR;WAEF;WAEA;WACU;WACA;;WAGR;WAEF;WAEA;WACU;WACA;;WAGR;WAEF;WAEA;WACU;WACA;;WAGR;WAEF;WAEA;WACU;WACA;;WAGR;WAEF;WAEA;WACU;WACA;;WAGR;WAEF;WAEA;WACU;WACA;;KAIJ,+BAA+B;;;;;;;;KCjF/BA;UAQK;EACf;EACA;EACA;;EAEA;;UAGe;;WAEN;EACT,KAAK,iBAAiB,oBAAoB,QAAQ;EAClD,SAAS;;UAGM;EACf;;;EAGA;EACA,KAAKA;;EAEL;;UAGe;;EAEf,KAAK,SAAS,uBAAuB,QAAQ;;UAG9B;EACf;EACA,KAAKA;EACL;;UAGe;;EAEf;EACA;EACA;;EAEA;EACA;;;;;;;;;;UAWe;EACf,MAAM,SAAS,eAAe,SAAS,oBAAoB,QAAQ;;UAGpD;EACf;EACA,KAAKA;EACL;;EAEA,SAAS;;EAET,gBAAgB;;;EAGhB;;;;;;;EAOA,UAAU;EACV;;UAGe;EACf;EACA;EACA;EACA;;UAGe;;EAEf;;;EAGA;;EAEA;;EAEA;EACA;;;KAIU,4BACV,SAAS,8BACN,QAAQ;;;;KCtGD;WACG;WAA0B,QAAQ;WAA+B;;WAEjE;WACA,QAAQ;WACR;WACA,aAAa;;WAGb;WACA;WACA;;UAGE;;WAEN;;WAEA;WACA;;;;;;;;;KAUC;;;WAIG;WACA;;WAGA;WACA;WACA;WACA;WACA;WACA;WACA,QAAQ;;UAGN;WACN;WACA;WACA;WACA;WACA;WACA,QAAQ;;UAGF;WACN;WACA;WACA;;WAEA;;;;;;;;;;KAWC;WAEG;WACA;WACA;WACA,cAAc;WACd,OAAO;;WAGP;WACA;WACA;WACA;;UAGE;;WAEN;;WAEA;;WAEA;;;WAGA;WACA,0BAA0B;;;;;;;;KASzB;WACG;WAA8B,WAAW;;WACzC;;WAEA;WACA;WACA,cAAc;;WAGd;WACA;WACA,cAAc;WACd,WAAW;;WAGX;WACA;WACA,cAAc;WACd,WAAW;WACX,WAAW;WACX,QAAQ;;;;;;;;;;UAWN;WACN;WACA;;WAEA;;qBAaE;;wBASG,aAAa,OAAO,cAAc;UAoBjC;WACN;WACA;WACA;WACA;WACA;WACA;WACA;;WAEA;;WAEA;;wBAGK,YAAY,iBAAiB,gBAAgB;;;UCnK5C;WACN,KAAK;WACL,UAAU;WACV,UAAU;WACV,QAAQ;WACR,cAAc;;;WAGd;WACA,SAAS;;WAET;;;;;;;;WAQA;WACA,cAAc;;;;;;;;;;UAWR,wBAAwB;WAC9B;WACA,OAAO;WACP,QAAQ;;;WAGR;WACA;WACA;;;WAGA,kBAAkB;WAClB;WACA;;WAEA;WACA;;wBAMW,eAAe,SAAS,qBAAqB,QAAQ;;;UCtF1D;;WAEN;WACA;WACA;;UAGM;WACN;WACA;WACA;WACA;WACA;WACA;WACA;;;WAGA;;KAGC;UAUK;WACN,IAAI;WACJ;;WAEA;;UAGM;WACN;WACA,QAAQ;;WAER;WACA;WACA,aAAa;;;;WAIb,cAAc;WACd;WACA,qBAAqB;WACrB,kBAAkB;;wBAGb,YACd,qBAAqB,mBACrB,UAAS,qBACR;;;wBAoKa,wBAAwB,QAAQ;;;UCnO/B;EACf;;EAEA;;UAGe;;EAEf;EACA;;EAEA;;;KAIU,iBAAiB,SAAS,mBAAmB,QAAQ;;;;;;qBAOpD,mBAAmB;UAuDf;;EAEf;;EAEA;;EAEA,MAAM;;EAEN;;EAEA;EACA,YAAY;;EAEZ;;;;;;;;wBAqBc,cAAc;EAC5B;EACA;EACA;EACA;EACA;;wBAmBc,oCACd,SAAS,uCACR;;;qBCxIU;;;;;wBAOG,iBACd,SAAQ;;qBAkBG;;;qBAgBA;;wBASG,uBAAuB,QAAQ;wBAc/B,uBAAuB,QAAQ;;;wBAM/B,qBACd,QAAQ,yBACR,SAAQ;;;;;;;;;wBAkBM,qBACd,SAAQ;;;;;;;;;wBAsBM,sBACd,QAAQ,oBACR;;;;qBC5GW,+BAA+B;;qBAG/B,wBAAwB;UAEpB;;WAEN;WACA;;WAEA;;UAGM;;WAEN,gBAAgB;;WAEhB;;;WAGA;WACA;WACA;;;;;;wBAUK,gBAAgB,gBAAgB;wBAWhC,mBAAmB,SAAS,4BAA4B"}
@@ -1,11 +1,11 @@
1
- import { n as CaptureIntegrityError, s as ValidationError } from "../errors-Dngq5h35.js";
2
- import { i as compareCodeUnits } from "../canonical-BsZVgBhW.js";
3
- import { r as pairedBootstrap } from "../paired-tests-jbwPdSz5.js";
4
- import { n as contentHash } from "../verdict-cache-BQEsVy9n.js";
5
- import { t as certificationEvidenceDigest } from "../verdict-DhPoXApk.js";
1
+ import { l as errorMessage, n as CaptureIntegrityError, s as ValidationError } from "../errors-DVZiqS7K.js";
2
+ import { i as compareCodeUnits } from "../canonical-B2RqLofv.js";
3
+ import { i as pairedBootstrap, t as mean } from "../mean-DHoMiudD.js";
4
+ import { n as contentHash } from "../verdict-cache-Bm8EKL2s.js";
5
+ import { t as certificationEvidenceDigest } from "../verdict-Dl0tiT1X.js";
6
6
  import { t as packageVersion } from "../package-version-D7lQHt_-.js";
7
- import { t as assertEqualDeclarativeTerms } from "../equal-terms-Blx9Zk5I.js";
8
- import { d as runPrimeExchange, n as buildPrimePrompt, t as analystUsageReceiptFromPrimeUsage } from "../prime-protocol-DcAK9pC8.js";
7
+ import { t as assertEqualDeclarativeTerms } from "../equal-terms-BtdXayCJ.js";
8
+ import { d as runPrimeExchange, n as buildPrimePrompt, t as analystUsageReceiptFromPrimeUsage } from "../prime-protocol-CHKgA3a6.js";
9
9
  import { c as deriveFailureSignature, d as isRecordedTimeout, h as parseRecordedReturncode, i as TIMEOUT_OBSERVATION_MARKER, n as FORMAT_ERROR_OBSERVATION_PREFIX, r as SUBMIT_ACTION_SIGNATURE, t as wrapActionForExec } from "../exec-ChJ6d5if.js";
10
10
  import { createHash } from "node:crypto";
11
11
  import { constants } from "node:os";
@@ -983,9 +983,6 @@ async function runOneRollout(input) {
983
983
  await environment.dispose();
984
984
  }
985
985
  }
986
- function errorMessage(error) {
987
- return error instanceof Error ? error.message : String(error);
988
- }
989
986
  /**
990
987
  * Sum only what the provider reported. A call with no usage raises
991
988
  * `callsWithUsage` short of `calls` and clears `captured`, so a partially
@@ -3060,9 +3057,6 @@ function measuredPrefixDivergences(grade) {
3060
3057
  if (grade.outcome === "not-reproduced" && grade.reproduction.basis !== "no-recorded-observation") return grade.reproduction.prefix.divergences;
3061
3058
  return 0;
3062
3059
  }
3063
- function mean(values) {
3064
- return values.reduce((sum, value) => sum + value, 0) / values.length;
3065
- }
3066
3060
  /** Markdown report: provenance, the funnel, every per-row column, the
3067
3061
  * distribution, and the threats. */
3068
3062
  function renderDeltaRepairReport(report) {