@tangle-network/agent-eval 0.216.1 → 0.217.1

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 (521) hide show
  1. package/CHANGELOG.md +102 -57
  2. package/README.md +2 -2
  3. package/dist/adapters/http.d.ts +2 -2
  4. package/dist/{agent-profile-VFcQncbW.d.ts → agent-profile-DTt8uGP9.d.ts} +4 -17
  5. package/dist/agent-profile-DTt8uGP9.d.ts.map +1 -0
  6. package/dist/{agent-profile-cell-BK8mlqlq.js → agent-profile-cell-B1V1gUZs.js} +4 -53
  7. package/dist/agent-profile-cell-B1V1gUZs.js.map +1 -0
  8. package/dist/{agent-profile-cell-BNp-Js9F.d.ts → agent-profile-cell-BcpFbNd5.d.ts} +2 -22
  9. package/dist/agent-profile-cell-BcpFbNd5.d.ts.map +1 -0
  10. package/dist/{agreement-irr-BEWLIF3_.js → agreement-irr-BhAqjluA.js} +5 -5
  11. package/dist/agreement-irr-BhAqjluA.js.map +1 -0
  12. package/dist/analyst/index.d.ts +36 -169
  13. package/dist/analyst/index.d.ts.map +1 -1
  14. package/dist/analyst/index.js +17 -192
  15. package/dist/analyst/index.js.map +1 -1
  16. package/dist/{analyze-runs-C84irm4i.js → analyze-runs-DqPJaJ4X.js} +225 -21
  17. package/dist/analyze-runs-DqPJaJ4X.js.map +1 -0
  18. package/dist/{analyze-runs-D68fd-su.d.ts → analyze-runs-DyeNcenW.d.ts} +5 -5
  19. package/dist/{analyze-runs-D68fd-su.d.ts.map → analyze-runs-DyeNcenW.d.ts.map} +1 -1
  20. package/dist/authenticity/index.d.ts +0 -23
  21. package/dist/authenticity/index.d.ts.map +1 -1
  22. package/dist/authenticity/index.js +1 -34
  23. package/dist/authenticity/index.js.map +1 -1
  24. package/dist/{benchmark-command-By0vh8q7.js → benchmark-command-xQK8nOI6.js} +554 -89
  25. package/dist/benchmark-command-xQK8nOI6.js.map +1 -0
  26. package/dist/benchmarks/index.d.ts +5 -6
  27. package/dist/benchmarks/index.d.ts.map +1 -1
  28. package/dist/benchmarks/index.js +4 -4
  29. package/dist/benchmarks/index.js.map +1 -1
  30. package/dist/builder-eval/index.d.ts +3 -3
  31. package/dist/builder-eval/index.js +2 -2
  32. package/dist/builder-eval/index.js.map +1 -1
  33. package/dist/campaign/index.d.ts +7 -7
  34. package/dist/campaign/index.js +15 -15
  35. package/dist/{campaign-BudZ70zo.js → campaign-CJGkxAQ6.js} +23 -530
  36. package/dist/campaign-CJGkxAQ6.js.map +1 -0
  37. package/dist/{campaign-evidence-Clrn4hBV.js → campaign-evidence-CQYw8_0H.js} +10 -18
  38. package/dist/{campaign-evidence-Clrn4hBV.js.map → campaign-evidence-CQYw8_0H.js.map} +1 -1
  39. package/dist/{canonical-B2RqLofv.js → canonical-Dq2hOQ8L.js} +2 -2
  40. package/dist/{canonical-B2RqLofv.js.map → canonical-Dq2hOQ8L.js.map} +1 -1
  41. package/dist/cell-spend-DEWR3nj-.js +90 -0
  42. package/dist/cell-spend-DEWR3nj-.js.map +1 -0
  43. package/dist/{chat-client-l-MVDlDk.js → chat-client-ldqxYYTN.js} +2 -2
  44. package/dist/{chat-client-l-MVDlDk.js.map → chat-client-ldqxYYTN.js.map} +1 -1
  45. package/dist/{chat-json-call-puMArmtW.js → chat-json-call-DavzexSp.js} +3 -2
  46. package/dist/chat-json-call-DavzexSp.js.map +1 -0
  47. package/dist/{claim-q4HQhK1G.js → claim-BPVuyHuy.js} +3 -3
  48. package/dist/{claim-q4HQhK1G.js.map → claim-BPVuyHuy.js.map} +1 -1
  49. package/dist/cli.js +113 -12
  50. package/dist/cli.js.map +1 -1
  51. package/dist/{client-BzL9niqi.d.ts → client-hwo3cJjR.d.ts} +3 -16
  52. package/dist/client-hwo3cJjR.d.ts.map +1 -0
  53. package/dist/{cluster-bootstrap-qcUWYFl-.js → cluster-bootstrap-DEwwhl0C.js} +3 -3
  54. package/dist/{cluster-bootstrap-qcUWYFl-.js.map → cluster-bootstrap-DEwwhl0C.js.map} +1 -1
  55. package/dist/{produced-state-BsDmnwxc.js → completion-verifier-B_jw6C9p.js} +9 -64
  56. package/dist/completion-verifier-B_jw6C9p.js.map +1 -0
  57. package/dist/contract/index.d.ts +12 -12
  58. package/dist/contract/index.js +25 -401
  59. package/dist/contract/index.js.map +1 -1
  60. package/dist/{cost-ledger-BP8V4kUQ.js → cost-ledger-D0352Z4d.js} +4 -8
  61. package/dist/cost-ledger-D0352Z4d.js.map +1 -0
  62. package/dist/dataset-DQqhOCPt.d.ts.map +1 -1
  63. package/dist/{default-registry-COhk9g36.js → default-registry-ClYjq3QE.js} +8 -11
  64. package/dist/default-registry-ClYjq3QE.js.map +1 -0
  65. package/dist/{default-registry-DzHiGruW.d.ts → default-registry-DXgLmoDI.d.ts} +7 -50
  66. package/dist/default-registry-DXgLmoDI.d.ts.map +1 -0
  67. package/dist/{define-agent-eval-CNTPzIkp.js → define-agent-eval-BDxNlP3D.js} +10 -10
  68. package/dist/{define-agent-eval-CNTPzIkp.js.map → define-agent-eval-BDxNlP3D.js.map} +1 -1
  69. package/dist/{define-agent-eval-WSck0CZJ.d.ts → define-agent-eval-BIkH_gAR.d.ts} +8 -8
  70. package/dist/define-agent-eval-BIkH_gAR.d.ts.map +1 -0
  71. package/dist/{descriptive-DboxPZrG.js → descriptive-CVNy48rR.js} +2 -2
  72. package/dist/{descriptive-DboxPZrG.js.map → descriptive-CVNy48rR.js.map} +1 -1
  73. package/dist/diagnosis/index.d.ts +4 -14
  74. package/dist/diagnosis/index.d.ts.map +1 -1
  75. package/dist/diagnosis/index.js +67 -28
  76. package/dist/diagnosis/index.js.map +1 -1
  77. package/dist/{dspy-rlm-engine-LRCu61m5.js → dspy-rlm-engine-DhdCr1Dc.js} +5 -10
  78. package/dist/dspy-rlm-engine-DhdCr1Dc.js.map +1 -0
  79. package/dist/{emitter-HcCmFQl_.d.ts → emitter-DABYYTcH.d.ts} +5 -5
  80. package/dist/emitter-DABYYTcH.d.ts.map +1 -0
  81. package/dist/emitter-DbMXUIVP.js.map +1 -1
  82. package/dist/{engine-3A_4jOlI.d.ts → engine-BvZ0R-36.d.ts} +17 -21
  83. package/dist/engine-BvZ0R-36.d.ts.map +1 -0
  84. package/dist/{equal-terms-BtdXayCJ.js → equal-terms-DbPGcVtI.js} +2 -2
  85. package/dist/equal-terms-DbPGcVtI.js.map +1 -0
  86. package/dist/{errors-DVZiqS7K.js → errors-BcOs5MKW.js} +2 -8
  87. package/dist/{errors-DVZiqS7K.js.map → errors-BcOs5MKW.js.map} +1 -1
  88. package/dist/{eval-campaign-90jeaCAI.js → eval-campaign-BLEfBGbA.js} +6 -6
  89. package/dist/{eval-campaign-90jeaCAI.js.map → eval-campaign-BLEfBGbA.js.map} +1 -1
  90. package/dist/{evaluation-6Yoi74gM.d.ts → evaluation-jTjC_QQD.d.ts} +7 -7
  91. package/dist/evaluation-jTjC_QQD.d.ts.map +1 -0
  92. package/dist/evaluation.d.ts +2 -2
  93. package/dist/evaluation.js +1 -1
  94. package/dist/evaluation.js.map +1 -1
  95. package/dist/{exact-types-ClugMJfI.d.ts → exact-types-CflbJ9f3.d.ts} +2 -2
  96. package/dist/{exact-types-ClugMJfI.d.ts.map → exact-types-CflbJ9f3.d.ts.map} +1 -1
  97. package/dist/{exec-ChJ6d5if.js → exec-CpbtOXVt.js} +2 -24
  98. package/dist/{exec-ChJ6d5if.js.map → exec-CpbtOXVt.js.map} +1 -1
  99. package/dist/execution-tracks-CpgFPpS5.js.map +1 -1
  100. package/dist/experiment/index.d.ts +9 -112
  101. package/dist/experiment/index.d.ts.map +1 -1
  102. package/dist/experiment/index.js +149 -241
  103. package/dist/experiment/index.js.map +1 -1
  104. package/dist/{exporters-X4xwla8A.js → exporters-Cl6IeNAu.js} +3 -3
  105. package/dist/exporters-Cl6IeNAu.js.map +1 -0
  106. package/dist/{exporters-ChTnR90l.d.ts → exporters-Dk3oEbK_.d.ts} +2 -4
  107. package/dist/{exporters-ChTnR90l.d.ts.map → exporters-Dk3oEbK_.d.ts.map} +1 -1
  108. package/dist/{external-optimizer-process-BjMGUuds.js → external-optimizer-process-KVFGvYX1.js} +4 -3
  109. package/dist/external-optimizer-process-KVFGvYX1.js.map +1 -0
  110. package/dist/{external-optimizer-subprocess-BqW0hHcI.js → external-optimizer-subprocess-XIHBOSKt.js} +9 -11
  111. package/dist/external-optimizer-subprocess-XIHBOSKt.js.map +1 -0
  112. package/dist/{extract-usage-BrnzRJ6H.js → extract-usage-BZLU01kR.js} +2 -23
  113. package/dist/{extract-usage-BrnzRJ6H.js.map → extract-usage-BZLU01kR.js.map} +1 -1
  114. package/dist/{failure-cluster-CbgEDcFJ.d.ts → failure-cluster-DN0t63Z0.d.ts} +4 -43
  115. package/dist/failure-cluster-DN0t63Z0.d.ts.map +1 -0
  116. package/dist/{failure-taxonomy-CjabTATj.js → failure-taxonomy-Bx5oWqaW.js} +2 -19
  117. package/dist/{failure-taxonomy-CjabTATj.js.map → failure-taxonomy-Bx5oWqaW.js.map} +1 -1
  118. package/dist/{feedback-trajectory-C4EG0nNH.d.ts → feedback-trajectory-BWmhRBGu.d.ts} +7 -9
  119. package/dist/feedback-trajectory-BWmhRBGu.d.ts.map +1 -0
  120. package/dist/{semantic-concept-judge-BQF8te0T.js → findings-store-BMVQtEsm.js} +9 -249
  121. package/dist/findings-store-BMVQtEsm.js.map +1 -0
  122. package/dist/fuzz.d.ts +19 -62
  123. package/dist/fuzz.d.ts.map +1 -1
  124. package/dist/fuzz.js +62 -108
  125. package/dist/fuzz.js.map +1 -1
  126. package/dist/{hf-dataset-BR2yhuNn.js → hf-dataset-D9f7-oX7.js} +4 -24
  127. package/dist/hf-dataset-D9f7-oX7.js.map +1 -0
  128. package/dist/hosted/index.d.ts +2 -2
  129. package/dist/hosted/index.js +2 -2
  130. package/dist/{index-DHKEAv3s.d.ts → index-9jJ2fKSX.d.ts} +2 -64
  131. package/dist/index-9jJ2fKSX.d.ts.map +1 -0
  132. package/dist/{index-DV-jSzxF.d.ts → index-BLGqZ81N.d.ts} +10 -15
  133. package/dist/{index-DV-jSzxF.d.ts.map → index-BLGqZ81N.d.ts.map} +1 -1
  134. package/dist/{index-DPcngtwt.d.ts → index-BiT6RtNH.d.ts} +11 -36
  135. package/dist/index-BiT6RtNH.d.ts.map +1 -0
  136. package/dist/{index-BO41eGj5.d.ts → index-CX-GAUpw.d.ts} +3 -7
  137. package/dist/index-CX-GAUpw.d.ts.map +1 -0
  138. package/dist/{index-BFRO6chs.d.ts → index-CiRwLv2R.d.ts} +5 -69
  139. package/dist/index-CiRwLv2R.d.ts.map +1 -0
  140. package/dist/{index-CyszSg4O.d.ts → index-ZuWx52WG.d.ts} +115 -347
  141. package/dist/index-ZuWx52WG.d.ts.map +1 -0
  142. package/dist/index.d.ts +211 -218
  143. package/dist/index.d.ts.map +1 -1
  144. package/dist/index.js +402 -1040
  145. package/dist/index.js.map +1 -1
  146. package/dist/{integrity-dLiDe2kD.js → integrity-DOd9Q86O.js} +7 -9
  147. package/dist/integrity-DOd9Q86O.js.map +1 -0
  148. package/dist/{integrity-BXOsAb3i.js → integrity-W4CADcFc.js} +3 -7
  149. package/dist/{integrity-BXOsAb3i.js.map → integrity-W4CADcFc.js.map} +1 -1
  150. package/dist/{integrity-DAMEDFaC.d.ts → integrity-_qgZLA5p.d.ts} +4 -6
  151. package/dist/{integrity-DAMEDFaC.d.ts.map → integrity-_qgZLA5p.d.ts.map} +1 -1
  152. package/dist/{internal-Epl-YBZW.js → internal-BTK91Bqt.js} +3 -3
  153. package/dist/{internal-Epl-YBZW.js.map → internal-BTK91Bqt.js.map} +1 -1
  154. package/dist/{interval-CELpye1n.js → interval-Dq9gt7jx.js} +3 -3
  155. package/dist/{interval-CELpye1n.js.map → interval-Dq9gt7jx.js.map} +1 -1
  156. package/dist/{journal-DfRz9kps.js → journal-CR4QikYp.js} +110 -110
  157. package/dist/journal-CR4QikYp.js.map +1 -0
  158. package/dist/json-BhdsJaOQ.js +111 -0
  159. package/dist/json-BhdsJaOQ.js.map +1 -0
  160. package/dist/{judge-calibration-BpCy5XQl.js → judge-calibration-BkXMIWA4.js} +3 -3
  161. package/dist/{judge-calibration-BpCy5XQl.js.map → judge-calibration-BkXMIWA4.js.map} +1 -1
  162. package/dist/{judge-gate-iY9s4L44.js → judge-gate-Co5tuh57.js} +4 -27
  163. package/dist/{judge-gate-iY9s4L44.js.map → judge-gate-Co5tuh57.js.map} +1 -1
  164. package/dist/{judge-gate-CE7CQAgt.d.ts → judge-gate-DsL0yFcb.d.ts} +8 -9
  165. package/dist/judge-gate-DsL0yFcb.d.ts.map +1 -0
  166. package/dist/{kind-factory-BobI5Jy3.js → kind-factory-Gy4fXixd.js} +88 -17
  167. package/dist/kind-factory-Gy4fXixd.js.map +1 -0
  168. package/dist/ledger-core/index.d.ts +2 -2
  169. package/dist/ledger-core/index.js +4 -4
  170. package/dist/{ledger-core-Gq7yp8lP.js → ledger-core-DrJhbms9.js} +5 -84
  171. package/dist/ledger-core-DrJhbms9.js.map +1 -0
  172. package/dist/{llm-client-CyBAoYiX.js → llm-client-BshCREn6.js} +6 -108
  173. package/dist/llm-client-BshCREn6.js.map +1 -0
  174. package/dist/{llm-judge-CG12SQJq.js → llm-judge-C8jMTNNG.js} +24 -32
  175. package/dist/llm-judge-C8jMTNNG.js.map +1 -0
  176. package/dist/matrix/index.d.ts +3 -3
  177. package/dist/matrix/index.js +3 -2
  178. package/dist/{matrix-Dm0BQW7U.js → matrix-B8l-Nsxd.js} +3 -88
  179. package/dist/matrix-B8l-Nsxd.js.map +1 -0
  180. package/dist/{matrix-DSrSxjjy.d.ts → matrix-DHe-dwVJ.d.ts} +9 -18
  181. package/dist/matrix-DHe-dwVJ.d.ts.map +1 -0
  182. package/dist/{mean-DHoMiudD.js → mean-BbA_JCzW.js} +39 -5
  183. package/dist/mean-BbA_JCzW.js.map +1 -0
  184. package/dist/meta-eval/index.d.ts +5 -5
  185. package/dist/meta-eval/index.js +10 -10
  186. package/dist/metrics-BKaPBtIG.js.map +1 -1
  187. package/dist/{mint-DNU8QPG4.d.ts → mint-BRZXuwRi.d.ts} +4 -4
  188. package/dist/{mint-DNU8QPG4.d.ts.map → mint-BRZXuwRi.d.ts.map} +1 -1
  189. package/dist/{mint-QaD5cA0u.js → mint-zszyYlpA.js} +4 -4
  190. package/dist/{mint-QaD5cA0u.js.map → mint-zszyYlpA.js.map} +1 -1
  191. package/dist/{multi-layer-verifier-BUaQ4C17.d.ts → multi-layer-verifier-DEgMUk6v.d.ts} +2 -2
  192. package/dist/{multi-layer-verifier-BUaQ4C17.d.ts.map → multi-layer-verifier-DEgMUk6v.d.ts.map} +1 -1
  193. package/dist/{multiplicity-DQUFcbW4.js → multiplicity-Sg0U7h8j.js} +2 -2
  194. package/dist/{multiplicity-DQUFcbW4.js.map → multiplicity-Sg0U7h8j.js.map} +1 -1
  195. package/dist/multishot/golden/index.d.ts +1 -23
  196. package/dist/multishot/golden/index.d.ts.map +1 -1
  197. package/dist/multishot/golden/index.js +3 -3
  198. package/dist/multishot/golden/index.js.map +1 -1
  199. package/dist/multishot/index.d.ts +3 -12
  200. package/dist/multishot/index.d.ts.map +1 -1
  201. package/dist/multishot/index.js +4 -3
  202. package/dist/multishot/index.js.map +1 -1
  203. package/dist/openapi.json +114 -10
  204. package/dist/{otel-spans-B9tUXRHF.js → otel-spans-CdN909k1.js} +3 -3
  205. package/dist/{otel-spans-B9tUXRHF.js.map → otel-spans-CdN909k1.js.map} +1 -1
  206. package/dist/{paired-arms-CwMGJsrS.js → paired-arms-Ct4LTKc1.js} +32 -9
  207. package/dist/paired-arms-Ct4LTKc1.js.map +1 -0
  208. package/dist/{paired-binary-CvpFHfKZ.js → paired-binary-B_nejA44.js} +3 -36
  209. package/dist/paired-binary-B_nejA44.js.map +1 -0
  210. package/dist/{paired-promotion-decision-DPKz8ZnF.d.ts → paired-promotion-decision-CQMJ5xGN.d.ts} +2 -34
  211. package/dist/paired-promotion-decision-CQMJ5xGN.d.ts.map +1 -0
  212. package/dist/{paired-promotion-decision-DsAPYGCm.js → paired-promotion-decision-D0SHb-ZL.js} +3 -3
  213. package/dist/{paired-promotion-decision-DsAPYGCm.js.map → paired-promotion-decision-D0SHb-ZL.js.map} +1 -1
  214. package/dist/{pareto-Cm6ndrWN.js → pareto-BzsxidFD.js} +11 -11
  215. package/dist/pareto-BzsxidFD.js.map +1 -0
  216. package/dist/pipelines/index.d.ts +16 -19
  217. package/dist/pipelines/index.d.ts.map +1 -1
  218. package/dist/pipelines/index.js +4 -68
  219. package/dist/pipelines/index.js.map +1 -1
  220. package/dist/{power-and-mde-Hj2ubb4B.js → power-and-mde-N5nU4_Ly.js} +3 -3
  221. package/dist/{power-and-mde-Hj2ubb4B.js.map → power-and-mde-N5nU4_Ly.js.map} +1 -1
  222. package/dist/{prime-protocol-CHKgA3a6.js → prime-protocol-BFgGrEMP.js} +3 -3
  223. package/dist/prime-protocol-BFgGrEMP.js.map +1 -0
  224. package/dist/{prime-protocol-HiB9HX7E.d.ts → prime-protocol-BWG2DQIk.d.ts} +3 -21
  225. package/dist/prime-protocol-BWG2DQIk.d.ts.map +1 -0
  226. package/dist/profile-cell.d.ts +2 -2
  227. package/dist/profile-cell.js +2 -2
  228. package/dist/{promotion-gate-BawrdPYr.d.ts → promotion-gate-DaY95yU6.d.ts} +7 -8
  229. package/dist/{promotion-gate-BawrdPYr.d.ts.map → promotion-gate-DaY95yU6.d.ts.map} +1 -1
  230. package/dist/{promotion-policy-KO57gbyV.js → promotion-policy-6A8osGr_.js} +4 -4
  231. package/dist/{promotion-policy-KO57gbyV.js.map → promotion-policy-6A8osGr_.js.map} +1 -1
  232. package/dist/{proposal-findings-CY1xKn_a.js → proposal-findings-BbWidvUk.js} +3 -3
  233. package/dist/proposal-findings-BbWidvUk.js.map +1 -0
  234. package/dist/{query-BUCZCK7F.js → query-B_RZ36an.js} +5 -5
  235. package/dist/query-B_RZ36an.js.map +1 -0
  236. package/dist/{query-DK1nWcmp.d.ts → query-Cmv_U8vf.d.ts} +4 -24
  237. package/dist/{query-DK1nWcmp.d.ts.map → query-Cmv_U8vf.d.ts.map} +1 -1
  238. package/dist/{random-ByUSnV9w.js → random-DDDcC7YU.js} +2 -2
  239. package/dist/{random-ByUSnV9w.js.map → random-DDDcC7YU.js.map} +1 -1
  240. package/dist/{raw-provider-sink-DoHz4JNG.js → raw-provider-sink-75w4FWBj.js} +2 -2
  241. package/dist/{raw-provider-sink-DoHz4JNG.js.map → raw-provider-sink-75w4FWBj.js.map} +1 -1
  242. package/dist/{redact-mtCCcT8z.d.ts → redact-DXdlFb_U.d.ts} +2 -3
  243. package/dist/redact-DXdlFb_U.d.ts.map +1 -0
  244. package/dist/{redact-B5xXKqQ3.js → redact-mtkTiJVk.js} +2 -7
  245. package/dist/{redact-B5xXKqQ3.js.map → redact-mtkTiJVk.js.map} +1 -1
  246. package/dist/{registry-D7_hm2fK.d.ts → registry-ssNoHzwD.d.ts} +4 -4
  247. package/dist/{registry-D7_hm2fK.d.ts.map → registry-ssNoHzwD.d.ts.map} +1 -1
  248. package/dist/{release-confidence-CpuP65S5.js → release-confidence-CYIFH-fp.js} +6 -11
  249. package/dist/release-confidence-CYIFH-fp.js.map +1 -0
  250. package/dist/{report-command-SC9nqb4M.js → render-C-w2EZ6p.js} +205 -363
  251. package/dist/render-C-w2EZ6p.js.map +1 -0
  252. package/dist/reporting.d.ts +6 -6
  253. package/dist/reporting.js +8 -8
  254. package/dist/{researcher-CEcfku-T.d.ts → researcher-0ABVIdGa.d.ts} +9 -9
  255. package/dist/{researcher-CEcfku-T.d.ts.map → researcher-0ABVIdGa.d.ts.map} +1 -1
  256. package/dist/{reward-CpCmAM8L.d.ts → reward-DrZ71vnf.d.ts} +3 -26
  257. package/dist/reward-DrZ71vnf.d.ts.map +1 -0
  258. package/dist/{reward-hacking-E0d-WVcP.js → reward-hacking-ByuETaIb.js} +3 -3
  259. package/dist/{reward-hacking-E0d-WVcP.js.map → reward-hacking-ByuETaIb.js.map} +1 -1
  260. package/dist/{reward-hacking-RVdZ6Kkt.d.ts → reward-hacking-E52wq86n.d.ts} +5 -18
  261. package/dist/reward-hacking-E52wq86n.d.ts.map +1 -0
  262. package/dist/reward-nw2xZGZG.js.map +1 -1
  263. package/dist/rl.d.ts +117 -295
  264. package/dist/rl.d.ts.map +1 -1
  265. package/dist/rl.js +27 -747
  266. package/dist/rl.js.map +1 -1
  267. package/dist/rollout/index.d.ts +5 -5
  268. package/dist/rollout/index.js +7 -8
  269. package/dist/{rollout-DZV1t2Wg.js → rollout-D6C50mfA.js} +88 -13
  270. package/dist/rollout-D6C50mfA.js.map +1 -0
  271. package/dist/{rubric-predictive-validity-DaFW70y9.js → rubric-predictive-validity-D3_oI72N.js} +3 -3
  272. package/dist/{rubric-predictive-validity-DaFW70y9.js.map → rubric-predictive-validity-D3_oI72N.js.map} +1 -1
  273. package/dist/{rubric-predictive-validity-C03EC7ZE.d.ts → rubric-predictive-validity-DiwT8PbL.d.ts} +2 -2
  274. package/dist/{rubric-predictive-validity-C03EC7ZE.d.ts.map → rubric-predictive-validity-DiwT8PbL.d.ts.map} +1 -1
  275. package/dist/{run-record-C1E24G2A.d.ts → run-record-CaxbLq6Y.d.ts} +4 -10
  276. package/dist/run-record-CaxbLq6Y.d.ts.map +1 -0
  277. package/dist/{run-record-C5e7FjhJ.js → run-record-CysKLB93.js} +3 -3
  278. package/dist/run-record-CysKLB93.js.map +1 -0
  279. package/dist/{run-record-Boc5IUTp.js → run-record-DhEJgoPz.js} +4 -4
  280. package/dist/{run-record-Boc5IUTp.js.map → run-record-DhEJgoPz.js.map} +1 -1
  281. package/dist/sandbox-harness-BlSOu4LX.d.ts.map +1 -1
  282. package/dist/{schema-CSf6qWgZ.js → schema-Bt2vLFm3.js} +2 -14
  283. package/dist/schema-Bt2vLFm3.js.map +1 -0
  284. package/dist/schema-C-ziyJD7.d.ts.map +1 -1
  285. package/dist/{schema-2g7trUwH.d.ts → schema-D6EVKzTB.d.ts} +4 -5
  286. package/dist/schema-D6EVKzTB.d.ts.map +1 -0
  287. package/dist/{schema-BuiujZ-g.js → schema-DxcKTzbi.js} +2 -2
  288. package/dist/schema-DxcKTzbi.js.map +1 -0
  289. package/dist/search/index.d.ts +3 -31
  290. package/dist/search/index.d.ts.map +1 -1
  291. package/dist/search/index.js +4 -4
  292. package/dist/{search-kernel-DyNEa6IU.d.ts → search-kernel-DxqjBxA-.d.ts} +10 -27
  293. package/dist/search-kernel-DxqjBxA-.d.ts.map +1 -0
  294. package/dist/{search-ledger-CVUGLDZp.js → search-ledger-TUvB5gWH.js} +7 -7
  295. package/dist/{search-ledger-CVUGLDZp.js.map → search-ledger-TUvB5gWH.js.map} +1 -1
  296. package/dist/{search-ledger-file-DDQUgvCg.js → search-ledger-file-UKCO2r8a.js} +4 -4
  297. package/dist/search-ledger-file-UKCO2r8a.js.map +1 -0
  298. package/dist/{search-shipper-BOSD46qv.js → search-shipper-Cw2z3-td.js} +4 -4
  299. package/dist/search-shipper-Cw2z3-td.js.map +1 -0
  300. package/dist/{search-summary-BQCSWf1g.js → search-summary-DQIvJsSB.js} +5 -5
  301. package/dist/{search-summary-BQCSWf1g.js.map → search-summary-DQIvJsSB.js.map} +1 -1
  302. package/dist/{sequential-CzK5DarL.js → sequential-BVv1sjsw.js} +2 -32
  303. package/dist/{sequential-CzK5DarL.js.map → sequential-BVv1sjsw.js.map} +1 -1
  304. package/dist/{sequential-Bf_pbE50.js → sequential-CimQrbjv.js} +5 -5
  305. package/dist/sequential-CimQrbjv.js.map +1 -0
  306. package/dist/{sequential-BhsrMupG.d.ts → sequential-KwGspfeR.d.ts} +2 -9
  307. package/dist/{sequential-BhsrMupG.d.ts.map → sequential-KwGspfeR.d.ts.map} +1 -1
  308. package/dist/{sequential-eprocess-DM_qENCl.js → sequential-eprocess-DodI6RoF.js} +2 -2
  309. package/dist/{sequential-eprocess-DM_qENCl.js.map → sequential-eprocess-DodI6RoF.js.map} +1 -1
  310. package/dist/{server-e9oWO_bN.js → server-DoVbqqtJ.js} +101 -101
  311. package/dist/server-DoVbqqtJ.js.map +1 -0
  312. package/dist/{single-run-lock-CV9Eyd_q.js → single-run-lock-U58yf8N6.js} +2 -2
  313. package/dist/{single-run-lock-CV9Eyd_q.js.map → single-run-lock-U58yf8N6.js.map} +1 -1
  314. package/dist/{skillopt-optimization-method-BiwA8nfX.js → skillopt-optimization-method-kyL94-ri.js} +17 -95
  315. package/dist/skillopt-optimization-method-kyL94-ri.js.map +1 -0
  316. package/dist/{spans-BHAUXMGN.d.ts → spans-CVngGij3.d.ts} +2 -2
  317. package/dist/{spans-BHAUXMGN.d.ts.map → spans-CVngGij3.d.ts.map} +1 -1
  318. package/dist/{statistical-heldout-42NlCNDe.js → statistical-heldout-Bksv92Th.js} +3 -3
  319. package/dist/statistical-heldout-Bksv92Th.js.map +1 -0
  320. package/dist/{statistical-heldout-CYpGaoDH.d.ts → statistical-heldout-C9AbgREN.d.ts} +8 -67
  321. package/dist/statistical-heldout-C9AbgREN.d.ts.map +1 -0
  322. package/dist/{steps-Bg5-KVx1.d.ts → steps-DlepJPwD.d.ts} +2 -56
  323. package/dist/steps-DlepJPwD.d.ts.map +1 -0
  324. package/dist/{store-C4avFjIi.d.ts → store-B3nWRy3B.d.ts} +2 -2
  325. package/dist/{store-C4avFjIi.d.ts.map → store-B3nWRy3B.d.ts.map} +1 -1
  326. package/dist/{store-otlp-DJ9dtA5x.js → store-otlp-BteDAbHK.js} +3 -3
  327. package/dist/{store-otlp-DJ9dtA5x.js.map → store-otlp-BteDAbHK.js.map} +1 -1
  328. package/dist/{store-tool-spans-CIX5arAG.js → store-tool-spans-D5l_OfTJ.js} +8 -8
  329. package/dist/store-tool-spans-D5l_OfTJ.js.map +1 -0
  330. package/dist/{store-tool-spans-DdCSPa7C.d.ts → store-tool-spans-DZS0HWVf.d.ts} +9 -18
  331. package/dist/store-tool-spans-DZS0HWVf.d.ts.map +1 -0
  332. package/dist/storyboard/index.d.ts +4 -11
  333. package/dist/storyboard/index.d.ts.map +1 -1
  334. package/dist/storyboard/index.js +1 -16
  335. package/dist/storyboard/index.js.map +1 -1
  336. package/dist/{summary-report-V_JACR4Y.js → summary-report-BsXJoYX5.js} +88 -10
  337. package/dist/summary-report-BsXJoYX5.js.map +1 -0
  338. package/dist/{summary-report-bdFDXq0b.d.ts → summary-report-DM7tlml1.d.ts} +10 -19
  339. package/dist/summary-report-DM7tlml1.d.ts.map +1 -0
  340. package/dist/supervisor-run/index.d.ts +60 -28
  341. package/dist/supervisor-run/index.d.ts.map +1 -1
  342. package/dist/supervisor-run/index.js +4 -458
  343. package/dist/supervisor-run/index.js.map +1 -1
  344. package/dist/{systemone-BomhHr2c.d.ts → systemone-Bp_Ac0p1.d.ts} +5 -5
  345. package/dist/{systemone-BomhHr2c.d.ts.map → systemone-Bp_Ac0p1.d.ts.map} +1 -1
  346. package/dist/systemone-protocol.js +1 -1
  347. package/dist/{systemone-review-CUVcQJu6.d.ts → systemone-review-CwE6JGIp.d.ts} +2 -2
  348. package/dist/{systemone-review-CUVcQJu6.d.ts.map → systemone-review-CwE6JGIp.d.ts.map} +1 -1
  349. package/dist/{systemone-review-D7sXwsqq.js → systemone-review-DzdOJ8I6.js} +3 -3
  350. package/dist/{systemone-review-D7sXwsqq.js.map → systemone-review-DzdOJ8I6.js.map} +1 -1
  351. package/dist/systemone.d.ts +5 -5
  352. package/dist/systemone.js +5 -5
  353. package/dist/systemone.js.map +1 -1
  354. package/dist/{task-failure-attributes-hEmASjix.js → task-failure-attributes-BMsblaDh.js} +3 -3
  355. package/dist/task-failure-attributes-BMsblaDh.js.map +1 -0
  356. package/dist/{terminal-record-DZyQZ3pz.js → terminal-record-78g6E4vh.js} +5 -5
  357. package/dist/terminal-record-78g6E4vh.js.map +1 -0
  358. package/dist/{tool-groups-C3NEcdCM.d.ts → tool-groups-DyNz2JZ_.d.ts} +3 -3
  359. package/dist/tool-groups-DyNz2JZ_.d.ts.map +1 -0
  360. package/dist/{tool-waste-BKgvd-TP.d.ts → tool-waste-CkjzWmeY.d.ts} +25 -28
  361. package/dist/tool-waste-CkjzWmeY.d.ts.map +1 -0
  362. package/dist/{tool-waste-BlesqTqW.js → tool-waste-DLsjS_p-.js} +5 -5
  363. package/dist/{tool-waste-BlesqTqW.js.map → tool-waste-DLsjS_p-.js.map} +1 -1
  364. package/dist/{trace-questions-CM2CAocU.d.ts → trace-questions-xxwgsuMi.d.ts} +4 -9
  365. package/dist/{trace-questions-CM2CAocU.d.ts.map → trace-questions-xxwgsuMi.d.ts.map} +1 -1
  366. package/dist/trace-repair/index.d.ts +14 -211
  367. package/dist/trace-repair/index.d.ts.map +1 -1
  368. package/dist/trace-repair/index.js +865 -981
  369. package/dist/trace-repair/index.js.map +1 -1
  370. package/dist/traces.d.ts +18 -46
  371. package/dist/traces.d.ts.map +1 -1
  372. package/dist/traces.js +13 -13
  373. package/dist/traces.js.map +1 -1
  374. package/dist/trajectory-replay/index.d.ts +3 -117
  375. package/dist/trajectory-replay/index.d.ts.map +1 -1
  376. package/dist/trajectory-replay/index.js +98 -56
  377. package/dist/trajectory-replay/index.js.map +1 -1
  378. package/dist/{tree-C1XYeXvy.js → tree-BGRglkKd.js} +13 -10
  379. package/dist/tree-BGRglkKd.js.map +1 -0
  380. package/dist/{types-B3jzCp0p.js → types-BXgASBCL.js} +2 -2
  381. package/dist/{types-B3jzCp0p.js.map → types-BXgASBCL.js.map} +1 -1
  382. package/dist/{types-C1tnkEgM.d.ts → types-CW8ewATZ.d.ts} +4 -113
  383. package/dist/types-CW8ewATZ.d.ts.map +1 -0
  384. package/dist/{types-BqkVmHfv.d.ts → types-CmXY0PQf.d.ts} +2 -4
  385. package/dist/types-CmXY0PQf.d.ts.map +1 -0
  386. package/dist/{types-C02qtnG8.d.ts → types-D-mgRzlh.d.ts} +7 -28
  387. package/dist/types-D-mgRzlh.d.ts.map +1 -0
  388. package/dist/{types-Cq8mSbq8.d.ts → types-D8GhJoHk.d.ts} +7 -9
  389. package/dist/{types-Cq8mSbq8.d.ts.map → types-D8GhJoHk.d.ts.map} +1 -1
  390. package/dist/{types-IXD2HJzu.js → types-DqFMNTgK.js} +2 -4
  391. package/dist/types-DqFMNTgK.js.map +1 -0
  392. package/dist/{types-Cbe2fWBa.d.ts → types-DtMbFw6U.d.ts} +3 -3
  393. package/dist/{types-Cbe2fWBa.d.ts.map → types-DtMbFw6U.d.ts.map} +1 -1
  394. package/dist/{verdict-Dl0tiT1X.js → verdict-BGQmxQ8A.js} +3 -3
  395. package/dist/verdict-BGQmxQ8A.js.map +1 -0
  396. package/dist/{verdict-E4eRNf7-.d.ts → verdict-C24NELtN.d.ts} +2 -15
  397. package/dist/{verdict-E4eRNf7-.d.ts.map → verdict-C24NELtN.d.ts.map} +1 -1
  398. package/dist/{verdict-cache-Bm8EKL2s.js → verdict-cache-CL_7tePD.js} +2 -2
  399. package/dist/verdict-cache-CL_7tePD.js.map +1 -0
  400. package/dist/wire/index.d.ts +31 -410
  401. package/dist/wire/index.d.ts.map +1 -1
  402. package/dist/wire/index.js +2 -2
  403. package/docs/changes/jev-evidence-review.md +1 -1
  404. package/docs/changes/jev-product-adoption.md +1 -1
  405. package/docs/changes/jev-recorded-evidence.md +1 -1
  406. package/docs/design/statistics-decisions.md +17 -17
  407. package/docs/design.md +1 -1
  408. package/docs/evaluation-integrity.md +1 -1
  409. package/docs/proofs/campaign-attempt-retention-20261001/README.md +2 -2
  410. package/docs/public-api.md +138 -529
  411. package/docs/rollout.md +1 -1
  412. package/docs/trace-repair-admission.md +1 -1
  413. package/docs/wire-protocol.md +1 -1
  414. package/package.json +1 -1
  415. package/dist/active-curriculum-B_Lc9kfO.js +0 -61
  416. package/dist/active-curriculum-B_Lc9kfO.js.map +0 -1
  417. package/dist/agent-profile-VFcQncbW.d.ts.map +0 -1
  418. package/dist/agent-profile-cell-BK8mlqlq.js.map +0 -1
  419. package/dist/agent-profile-cell-BNp-Js9F.d.ts.map +0 -1
  420. package/dist/agreement-irr-BEWLIF3_.js.map +0 -1
  421. package/dist/analyze-runs-C84irm4i.js.map +0 -1
  422. package/dist/baseline-Udi_PLuT.js +0 -259
  423. package/dist/baseline-Udi_PLuT.js.map +0 -1
  424. package/dist/benchmark-CXmVVHH2.js +0 -562
  425. package/dist/benchmark-CXmVVHH2.js.map +0 -1
  426. package/dist/benchmark-command-By0vh8q7.js.map +0 -1
  427. package/dist/campaign-BudZ70zo.js.map +0 -1
  428. package/dist/chat-json-call-puMArmtW.js.map +0 -1
  429. package/dist/claude-jsonl-dZir12m6.js +0 -86
  430. package/dist/claude-jsonl-dZir12m6.js.map +0 -1
  431. package/dist/client-BzL9niqi.d.ts.map +0 -1
  432. package/dist/cost-ledger-BP8V4kUQ.js.map +0 -1
  433. package/dist/counterfactual-CPaWN99-.d.ts +0 -60
  434. package/dist/counterfactual-CPaWN99-.d.ts.map +0 -1
  435. package/dist/counterfactual-Dn3veueY.js +0 -98
  436. package/dist/counterfactual-Dn3veueY.js.map +0 -1
  437. package/dist/default-registry-COhk9g36.js.map +0 -1
  438. package/dist/default-registry-DzHiGruW.d.ts.map +0 -1
  439. package/dist/define-agent-eval-WSck0CZJ.d.ts.map +0 -1
  440. package/dist/dspy-rlm-engine-LRCu61m5.js.map +0 -1
  441. package/dist/effect-sizes-BhzEpXMc.js +0 -82
  442. package/dist/effect-sizes-BhzEpXMc.js.map +0 -1
  443. package/dist/emitter-HcCmFQl_.d.ts.map +0 -1
  444. package/dist/engine-3A_4jOlI.d.ts.map +0 -1
  445. package/dist/equal-terms-BtdXayCJ.js.map +0 -1
  446. package/dist/evaluation-6Yoi74gM.d.ts.map +0 -1
  447. package/dist/exporters-X4xwla8A.js.map +0 -1
  448. package/dist/external-optimizer-process-BjMGUuds.js.map +0 -1
  449. package/dist/external-optimizer-subprocess-BqW0hHcI.js.map +0 -1
  450. package/dist/failure-cluster-CbgEDcFJ.d.ts.map +0 -1
  451. package/dist/feedback-trajectory-C4EG0nNH.d.ts.map +0 -1
  452. package/dist/hf-dataset-BR2yhuNn.js.map +0 -1
  453. package/dist/index-BFRO6chs.d.ts.map +0 -1
  454. package/dist/index-BO41eGj5.d.ts.map +0 -1
  455. package/dist/index-CyszSg4O.d.ts.map +0 -1
  456. package/dist/index-DHKEAv3s.d.ts.map +0 -1
  457. package/dist/index-DPcngtwt.d.ts.map +0 -1
  458. package/dist/integrity-dLiDe2kD.js.map +0 -1
  459. package/dist/journal-DfRz9kps.js.map +0 -1
  460. package/dist/judge-gate-CE7CQAgt.d.ts.map +0 -1
  461. package/dist/kind-factory-BobI5Jy3.js.map +0 -1
  462. package/dist/ledger-core-Gq7yp8lP.js.map +0 -1
  463. package/dist/llm-client-CyBAoYiX.js.map +0 -1
  464. package/dist/llm-judge-CG12SQJq.js.map +0 -1
  465. package/dist/matrix-DSrSxjjy.d.ts.map +0 -1
  466. package/dist/matrix-Dm0BQW7U.js.map +0 -1
  467. package/dist/mean-DHoMiudD.js.map +0 -1
  468. package/dist/normal-DsJZtrJc.js +0 -25
  469. package/dist/normal-DsJZtrJc.js.map +0 -1
  470. package/dist/paired-arms-CwMGJsrS.js.map +0 -1
  471. package/dist/paired-binary-CvpFHfKZ.js.map +0 -1
  472. package/dist/paired-promotion-decision-DPKz8ZnF.d.ts.map +0 -1
  473. package/dist/pareto-Cm6ndrWN.js.map +0 -1
  474. package/dist/prime-protocol-CHKgA3a6.js.map +0 -1
  475. package/dist/prime-protocol-HiB9HX7E.d.ts.map +0 -1
  476. package/dist/produced-state-BsDmnwxc.js.map +0 -1
  477. package/dist/proposal-findings-CY1xKn_a.js.map +0 -1
  478. package/dist/query-BUCZCK7F.js.map +0 -1
  479. package/dist/redact-mtCCcT8z.d.ts.map +0 -1
  480. package/dist/release-confidence-CpuP65S5.js.map +0 -1
  481. package/dist/report-command-SC9nqb4M.js.map +0 -1
  482. package/dist/reward-CpCmAM8L.d.ts.map +0 -1
  483. package/dist/reward-hacking-RVdZ6Kkt.d.ts.map +0 -1
  484. package/dist/rollout-DZV1t2Wg.js.map +0 -1
  485. package/dist/run-record-C1E24G2A.d.ts.map +0 -1
  486. package/dist/run-record-C5e7FjhJ.js.map +0 -1
  487. package/dist/schema-2g7trUwH.d.ts.map +0 -1
  488. package/dist/schema-BuiujZ-g.js.map +0 -1
  489. package/dist/schema-CSf6qWgZ.js.map +0 -1
  490. package/dist/search-kernel-DyNEa6IU.d.ts.map +0 -1
  491. package/dist/search-ledger-file-DDQUgvCg.js.map +0 -1
  492. package/dist/search-shipper-BOSD46qv.js.map +0 -1
  493. package/dist/semantic-concept-judge-BQF8te0T.js.map +0 -1
  494. package/dist/sequential-Bf_pbE50.js.map +0 -1
  495. package/dist/server-e9oWO_bN.js.map +0 -1
  496. package/dist/skillopt-optimization-method-BiwA8nfX.js.map +0 -1
  497. package/dist/statistical-heldout-42NlCNDe.js.map +0 -1
  498. package/dist/statistical-heldout-CYpGaoDH.d.ts.map +0 -1
  499. package/dist/steps-Bg5-KVx1.d.ts.map +0 -1
  500. package/dist/store-tool-spans-CIX5arAG.js.map +0 -1
  501. package/dist/store-tool-spans-DdCSPa7C.d.ts.map +0 -1
  502. package/dist/student-t-DVPW7CdL.js +0 -39
  503. package/dist/student-t-DVPW7CdL.js.map +0 -1
  504. package/dist/summary-report-V_JACR4Y.js.map +0 -1
  505. package/dist/summary-report-bdFDXq0b.d.ts.map +0 -1
  506. package/dist/task-failure-attributes-hEmASjix.js.map +0 -1
  507. package/dist/terminal-record-DZyQZ3pz.js.map +0 -1
  508. package/dist/tool-groups-C3NEcdCM.d.ts.map +0 -1
  509. package/dist/tool-waste-BKgvd-TP.d.ts.map +0 -1
  510. package/dist/trajectory-Y1fmdlw7.d.ts +0 -24
  511. package/dist/trajectory-Y1fmdlw7.d.ts.map +0 -1
  512. package/dist/tree-C1XYeXvy.js.map +0 -1
  513. package/dist/types-BqkVmHfv.d.ts.map +0 -1
  514. package/dist/types-C02qtnG8.d.ts.map +0 -1
  515. package/dist/types-C1tnkEgM.d.ts.map +0 -1
  516. package/dist/types-IXD2HJzu.js.map +0 -1
  517. package/dist/usage-receipt-Cx3xNeyB.js +0 -76
  518. package/dist/usage-receipt-Cx3xNeyB.js.map +0 -1
  519. package/dist/verdict-Dl0tiT1X.js.map +0 -1
  520. package/dist/verdict-cache-Bm8EKL2s.js.map +0 -1
  521. package/docs/design/self-improvement-evidence-audit.md +0 -263
@@ -1,104 +1,22 @@
1
- import { n as CaptureIntegrityError, s as ValidationError } from "../errors-DVZiqS7K.js";
2
- import { a as hashCanonical, t as LEDGER_HASH_PATTERN } from "../canonical-B2RqLofv.js";
3
- import { f as evaluateHypothesis, g as verifyManifest, h as signManifest, m as manifestContentDigest, p as hashJson } from "../agent-profile-cell-BK8mlqlq.js";
4
- import { t as mulberry32 } from "../random-ByUSnV9w.js";
5
- import { n as bonferroni, r as holm, t as benjaminiHochberg } from "../multiplicity-DQUFcbW4.js";
6
- import { a as pairedRiskDifferenceExact, c as wilson, i as pairedRiskDifference, n as mcnemar, o as pairedRiskDifferenceScore } from "../paired-binary-CvpFHfKZ.js";
7
- import { n as studentTQuantile, t as studentTCdf } from "../student-t-DVPW7CdL.js";
8
- import { i as pairedBootstrap, n as BOOTSTRAP_GATE_MIN_N } from "../mean-DHoMiudD.js";
9
- import { a as requiredSampleSize, i as requiredPairedSampleSize, n as mcnemarRequiredN, r as pairedMde, t as mcnemarPower } from "../power-and-mde-Hj2ubb4B.js";
10
- import { n as pairArms, r as pairRunRecords, t as comparePairedArms } from "../paired-arms-CwMGJsrS.js";
11
- import { t as eProcess } from "../sequential-eprocess-DM_qENCl.js";
12
- import { a as heldoutSignificance, o as pairHoldout } from "../statistical-heldout-42NlCNDe.js";
13
- import { D as FinalEvidenceError, E as FinalEvidenceConflictError, O as openFinalEvidenceLedger, T as powerPreflight, a as verifyEvidenceReceipt, i as isIndependentEvidence, n as EVIDENCE_RECEIPT_VERSION, r as createEvidenceReceipt, t as createCampaignEvidenceReceipt } from "../campaign-evidence-Clrn4hBV.js";
1
+ import { n as CaptureIntegrityError, s as ValidationError } from "../errors-BcOs5MKW.js";
2
+ import { a as hashCanonical, t as LEDGER_HASH_PATTERN } from "../canonical-Dq2hOQ8L.js";
3
+ import { d as signManifest, f as verifyManifest, l as hashJson, u as manifestContentDigest } from "../agent-profile-cell-B1V1gUZs.js";
4
+ import { t as mulberry32 } from "../random-DDDcC7YU.js";
5
+ import { n as pairArms, r as pairRunRecords, t as comparePairedArms } from "../paired-arms-Ct4LTKc1.js";
6
+ import { n as bonferroni, r as holm, t as benjaminiHochberg } from "../multiplicity-Sg0U7h8j.js";
7
+ import { a as pairedRiskDifferenceScore, i as pairedRiskDifferenceExact, r as pairedRiskDifference, s as wilson, t as mcnemar } from "../paired-binary-B_nejA44.js";
8
+ import { c as studentTCdf, i as pairedBootstrap, l as studentTQuantile, n as BOOTSTRAP_GATE_MIN_N } from "../mean-BbA_JCzW.js";
9
+ import { a as requiredSampleSize, i as requiredPairedSampleSize, n as mcnemarRequiredN, r as pairedMde, t as mcnemarPower } from "../power-and-mde-N5nU4_Ly.js";
10
+ import { t as eProcess } from "../sequential-eprocess-DodI6RoF.js";
11
+ import { a as heldoutSignificance, o as pairHoldout } from "../statistical-heldout-Bksv92Th.js";
12
+ import { S as openFinalEvidenceLedger, b as FinalEvidenceConflictError, i as verifyEvidenceReceipt, n as createEvidenceReceipt, r as isIndependentEvidence, t as createCampaignEvidenceReceipt, x as FinalEvidenceError, y as powerPreflight } from "../campaign-evidence-CQYw8_0H.js";
14
13
  import { n as evaluatePredicate, r as readField, t as compareValues } from "../predicate-FdKmElA_.js";
15
- import { r as summarizeEvaluationUnits, t as defineEvaluationClaim } from "../claim-q4HQhK1G.js";
16
- import { n as pairedEvalueSequence, r as sequentialCrossingHorizon } from "../sequential-CzK5DarL.js";
17
- import { n as intervalSpecProblems, r as computeEstimand, t as computeInterval } from "../interval-CELpye1n.js";
18
- import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-KO57gbyV.js";
19
- import { n as sequentialPairedGate, t as sequentialDecide } from "../sequential-Bf_pbE50.js";
14
+ import { r as summarizeEvaluationUnits, t as defineEvaluationClaim } from "../claim-BPVuyHuy.js";
15
+ import { n as pairedEvalueSequence } from "../sequential-BVv1sjsw.js";
16
+ import { n as intervalSpecProblems, r as computeEstimand, t as computeInterval } from "../interval-Dq9gt7jx.js";
17
+ import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-6A8osGr_.js";
18
+ import { n as sequentialPairedGate, t as sequentialDecide } from "../sequential-CimQrbjv.js";
20
19
  import { z } from "zod";
21
- //#region src/experiment/ast/budget.ts
22
- function evaluateHaltRule(halt, gates) {
23
- const seen = new Map(gates.map((g) => [g.id, g]));
24
- const missing = halt.when.gates.filter((id) => !seen.has(id));
25
- if (missing.length > 0) throw new ValidationError(`evaluateHaltRule: halt references gates that were not evaluated: [${missing.join(", ")}]`);
26
- const failed = halt.when.gates.filter((id) => !seen.get(id).passed);
27
- return failed.length > 0 ? {
28
- fired: true,
29
- action: halt.action,
30
- failedGates: failed
31
- } : {
32
- fired: false,
33
- action: null,
34
- failedGates: []
35
- };
36
- }
37
- /**
38
- * Execute the uniform-pass schedule against measured pass costs. Pass 1 always
39
- * runs; each later pass runs only when the cumulative spend plus the last
40
- * measured pass cost stays at or under the registered ceiling. The registered
41
- * ledger is the pre-spend the ceiling counts.
42
- */
43
- function runUniformPassBudget(rule, measuredPassCosts) {
44
- let cumulative = rule.ledger.reduce((sum, entry) => sum + entry.usd, 0);
45
- const decisions = [];
46
- let uniformN = 0;
47
- for (let pass = 1; pass <= rule.maxPasses; pass++) {
48
- if (pass === 1) {
49
- if (measuredPassCosts[0] === void 0) break;
50
- cumulative += measuredPassCosts[0];
51
- uniformN = 1;
52
- continue;
53
- }
54
- const projected = measuredPassCosts[pass - 2];
55
- if (projected === void 0) break;
56
- const go = cumulative + projected <= rule.ceilingUsd;
57
- decisions.push({
58
- pass,
59
- cumulativeBefore: cumulative,
60
- projected,
61
- go
62
- });
63
- if (!go || measuredPassCosts[pass - 1] === void 0) break;
64
- cumulative += measuredPassCosts[pass - 1];
65
- uniformN = pass;
66
- }
67
- return {
68
- decisions,
69
- uniformN
70
- };
71
- }
72
- /**
73
- * Walk the registered n-ladder and pick the first affordable step. When no
74
- * step fits the ceiling, the rule refuses and reports the projection instead
75
- * of shrinking the row set — "never subset rows" is the registered invariant.
76
- */
77
- function projectNLadderBudget(rule, measured) {
78
- const projections = rule.steps.map((n) => {
79
- const projectedUsd = measured.unitCostUsd * measured.rows * n;
80
- return {
81
- n,
82
- projectedUsd,
83
- affordable: projectedUsd <= rule.ceilingUsd
84
- };
85
- });
86
- const first = projections.find((p) => p.affordable);
87
- if (first) return {
88
- chosenN: first.n,
89
- projections,
90
- refusal: null
91
- };
92
- return {
93
- chosenN: null,
94
- projections,
95
- refusal: {
96
- onExhaust: rule.onExhaust,
97
- reason: `no ladder step fits the ${rule.ceilingUsd} USD ceiling at ${measured.rows} rows x ${measured.unitCostUsd} USD per unit`
98
- }
99
- };
100
- }
101
- //#endregion
102
20
  //#region src/experiment/ast/decision.ts
103
21
  /** A decision rule's branches did not cover the evidence. */
104
22
  var DecisionTableNotTotalError = class extends ValidationError {};
@@ -137,67 +55,6 @@ function executeDecisionRule(rule, evidence) {
137
55
  throw new DecisionTableNotTotalError("executeDecisionRule: decision table is not total — no branch matched the evidence");
138
56
  }
139
57
  //#endregion
140
- //#region src/experiment/ast/reissue.ts
141
- /**
142
- * Classify one rollout event under the registered reissue policy. A carrier
143
- * event within the issue budget is reissued; a model outcome always stands;
144
- * a carrier event past `maxIssues` is exhausted and reported, never retried.
145
- */
146
- function classifyReissue(policy, event, issuesSoFar) {
147
- if (!policy.carrierEvents.includes(event)) return "stands";
148
- return issuesSoFar < policy.maxIssues ? "reissue" : "exhausted";
149
- }
150
- //#endregion
151
- //#region src/experiment/ast/selection.ts
152
- /**
153
- * Execute a selection rule.
154
- *
155
- * Round-robin walks groups in lexicographic order and takes ids in
156
- * within-group order until `take` ids are chosen. Filter-of keeps the ids of
157
- * `bases[rule.base]` whose record satisfies the predicate, in the registered
158
- * order. Ids absent from `records` are evaluated on their id alone (fields
159
- * derived from the id via `idFields`), so a sealed base outlives its source
160
- * records.
161
- */
162
- function runSelectionRule(rule, records, options) {
163
- if (rule.kind === "round-robin") {
164
- const allowed = new Set(rule.reads);
165
- for (const field of [rule.groupBy, rule.withinOrder.field]) if (!allowed.has(field)) throw new ValidationError(`runSelectionRule: round-robin reads '${field}' but its closed read set is [${rule.reads.join(", ")}]`);
166
- const byGroup = /* @__PURE__ */ new Map();
167
- for (const record of records) {
168
- const group = String(readField(record, rule.groupBy));
169
- const id = String(readField(record, rule.withinOrder.field));
170
- const bucket = byGroup.get(group);
171
- if (bucket) bucket.push(id);
172
- else byGroup.set(group, [id]);
173
- }
174
- for (const ids of byGroup.values()) {
175
- ids.sort();
176
- if (rule.withinOrder.dir === "desc") ids.reverse();
177
- }
178
- const groups = [...byGroup.keys()].sort();
179
- const chosen = [];
180
- let cursor = 0;
181
- while (chosen.length < rule.take && groups.some((g) => byGroup.get(g).length > 0)) {
182
- const group = groups[cursor % groups.length];
183
- const ids = byGroup.get(group);
184
- if (ids.length > 0) chosen.push(ids.shift());
185
- cursor += 1;
186
- }
187
- return chosen;
188
- }
189
- const base = options.bases?.[rule.base];
190
- if (!base) throw new ValidationError(`runSelectionRule: filter-of base '${rule.base}' was not provided`);
191
- const index = new Map(records.map((r) => [String(readField(r, options.idField)), r]));
192
- const sorted = [...base.filter((id) => {
193
- const record = index.get(id) ?? options.idFields?.(id);
194
- if (!record) throw new ValidationError(`runSelectionRule: base id '${id}' has no record and no idFields derivation`);
195
- return evaluatePredicate(rule.keep, record);
196
- })].sort();
197
- if (rule.order.dir === "desc") sorted.reverse();
198
- return sorted;
199
- }
200
- //#endregion
201
58
  //#region src/experiment/ast/validity.ts
202
59
  /**
203
60
  * Replicate-flip counting over graded states. A state whose replicates split
@@ -355,6 +212,137 @@ function assertMatchedBudgets(rule, arms) {
355
212
  return verdict;
356
213
  }
357
214
  //#endregion
215
+ //#region src/experiment/ast/budget.ts
216
+ function evaluateHaltRule(halt, gates) {
217
+ const seen = new Map(gates.map((g) => [g.id, g]));
218
+ const missing = halt.when.gates.filter((id) => !seen.has(id));
219
+ if (missing.length > 0) throw new ValidationError(`evaluateHaltRule: halt references gates that were not evaluated: [${missing.join(", ")}]`);
220
+ const failed = halt.when.gates.filter((id) => !seen.get(id).passed);
221
+ return failed.length > 0 ? {
222
+ fired: true,
223
+ action: halt.action,
224
+ failedGates: failed
225
+ } : {
226
+ fired: false,
227
+ action: null,
228
+ failedGates: []
229
+ };
230
+ }
231
+ /**
232
+ * Execute the uniform-pass schedule against measured pass costs. Pass 1 always
233
+ * runs; each later pass runs only when the cumulative spend plus the last
234
+ * measured pass cost stays at or under the registered ceiling. The registered
235
+ * ledger is the pre-spend the ceiling counts.
236
+ */
237
+ function runUniformPassBudget(rule, measuredPassCosts) {
238
+ let cumulative = rule.ledger.reduce((sum, entry) => sum + entry.usd, 0);
239
+ const decisions = [];
240
+ let uniformN = 0;
241
+ for (let pass = 1; pass <= rule.maxPasses; pass++) {
242
+ if (pass === 1) {
243
+ if (measuredPassCosts[0] === void 0) break;
244
+ cumulative += measuredPassCosts[0];
245
+ uniformN = 1;
246
+ continue;
247
+ }
248
+ const projected = measuredPassCosts[pass - 2];
249
+ if (projected === void 0) break;
250
+ const go = cumulative + projected <= rule.ceilingUsd;
251
+ decisions.push({
252
+ pass,
253
+ cumulativeBefore: cumulative,
254
+ projected,
255
+ go
256
+ });
257
+ if (!go || measuredPassCosts[pass - 1] === void 0) break;
258
+ cumulative += measuredPassCosts[pass - 1];
259
+ uniformN = pass;
260
+ }
261
+ return {
262
+ decisions,
263
+ uniformN
264
+ };
265
+ }
266
+ /**
267
+ * Walk the registered n-ladder and pick the first affordable step. When no
268
+ * step fits the ceiling, the rule refuses and reports the projection instead
269
+ * of shrinking the row set — "never subset rows" is the registered invariant.
270
+ */
271
+ function projectNLadderBudget(rule, measured) {
272
+ const projections = rule.steps.map((n) => {
273
+ const projectedUsd = measured.unitCostUsd * measured.rows * n;
274
+ return {
275
+ n,
276
+ projectedUsd,
277
+ affordable: projectedUsd <= rule.ceilingUsd
278
+ };
279
+ });
280
+ const first = projections.find((p) => p.affordable);
281
+ if (first) return {
282
+ chosenN: first.n,
283
+ projections,
284
+ refusal: null
285
+ };
286
+ return {
287
+ chosenN: null,
288
+ projections,
289
+ refusal: {
290
+ onExhaust: rule.onExhaust,
291
+ reason: `no ladder step fits the ${rule.ceilingUsd} USD ceiling at ${measured.rows} rows x ${measured.unitCostUsd} USD per unit`
292
+ }
293
+ };
294
+ }
295
+ //#endregion
296
+ //#region src/experiment/ast/selection.ts
297
+ /**
298
+ * Execute a selection rule.
299
+ *
300
+ * Round-robin walks groups in lexicographic order and takes ids in
301
+ * within-group order until `take` ids are chosen. Filter-of keeps the ids of
302
+ * `bases[rule.base]` whose record satisfies the predicate, in the registered
303
+ * order. Ids absent from `records` are evaluated on their id alone (fields
304
+ * derived from the id via `idFields`), so a sealed base outlives its source
305
+ * records.
306
+ */
307
+ function runSelectionRule(rule, records, options) {
308
+ if (rule.kind === "round-robin") {
309
+ const allowed = new Set(rule.reads);
310
+ for (const field of [rule.groupBy, rule.withinOrder.field]) if (!allowed.has(field)) throw new ValidationError(`runSelectionRule: round-robin reads '${field}' but its closed read set is [${rule.reads.join(", ")}]`);
311
+ const byGroup = /* @__PURE__ */ new Map();
312
+ for (const record of records) {
313
+ const group = String(readField(record, rule.groupBy));
314
+ const id = String(readField(record, rule.withinOrder.field));
315
+ const bucket = byGroup.get(group);
316
+ if (bucket) bucket.push(id);
317
+ else byGroup.set(group, [id]);
318
+ }
319
+ for (const ids of byGroup.values()) {
320
+ ids.sort();
321
+ if (rule.withinOrder.dir === "desc") ids.reverse();
322
+ }
323
+ const groups = [...byGroup.keys()].sort();
324
+ const chosen = [];
325
+ let cursor = 0;
326
+ while (chosen.length < rule.take && groups.some((g) => byGroup.get(g).length > 0)) {
327
+ const group = groups[cursor % groups.length];
328
+ const ids = byGroup.get(group);
329
+ if (ids.length > 0) chosen.push(ids.shift());
330
+ cursor += 1;
331
+ }
332
+ return chosen;
333
+ }
334
+ const base = options.bases?.[rule.base];
335
+ if (!base) throw new ValidationError(`runSelectionRule: filter-of base '${rule.base}' was not provided`);
336
+ const index = new Map(records.map((r) => [String(readField(r, options.idField)), r]));
337
+ const sorted = [...base.filter((id) => {
338
+ const record = index.get(id) ?? options.idFields?.(id);
339
+ if (!record) throw new ValidationError(`runSelectionRule: base id '${id}' has no record and no idFields derivation`);
340
+ return evaluatePredicate(rule.keep, record);
341
+ })].sort();
342
+ if (rule.order.dir === "desc") sorted.reverse();
343
+ return sorted;
344
+ }
345
+ //#endregion
358
346
  //#region src/experiment/funnel.ts
359
347
  /**
360
348
  * The denominator chain as a first-class object.
@@ -890,86 +878,6 @@ const evidenceRegistryRecordSchema = z.strictObject({
890
878
  supersedes: z.array(z.string().regex(idPattern)).optional(),
891
879
  notes: z.string().min(1).optional()
892
880
  });
893
- /** Parse one record; throws with the zod issue list on any violation. */
894
- function parseEvidenceRegistryRecord(raw) {
895
- return evidenceRegistryRecordSchema.parse(raw);
896
- }
897
- var EvidenceRegistryError = class extends Error {};
898
- /**
899
- * Validate a set of records as one registry: every record parses, ids are
900
- * unique, and every `supersedes` target exists. Returns records sorted in
901
- * the registry's canonical order (trust ladder, then date desc, then id).
902
- */
903
- function validateEvidenceRegistry(raws) {
904
- const records = raws.map((raw) => parseEvidenceRegistryRecord(raw));
905
- const seen = /* @__PURE__ */ new Set();
906
- for (const record of records) {
907
- if (seen.has(record.id)) throw new EvidenceRegistryError(`duplicate evidence record id: ${record.id}`);
908
- seen.add(record.id);
909
- }
910
- for (const record of records) for (const target of record.supersedes ?? []) if (!seen.has(target)) throw new EvidenceRegistryError(`record ${record.id} supersedes unknown record: ${target}`);
911
- const stateRank = new Map(EVIDENCE_STATES.map((state, index) => [state, index]));
912
- return [...records].sort((a, b) => {
913
- const byState = (stateRank.get(a.evidenceState) ?? 0) - (stateRank.get(b.evidenceState) ?? 0);
914
- if (byState !== 0) return byState;
915
- if (a.date !== b.date) return a.date < b.date ? 1 : -1;
916
- return a.id < b.id ? -1 : 1;
917
- });
918
- }
919
- function mdEscape(text) {
920
- return text.replace(/\|/g, "\\|").replace(/\n/g, " ");
921
- }
922
- function formatCost(costUsd) {
923
- if (costUsd === null) return "not captured";
924
- return `$${costUsd.toFixed(2)}`;
925
- }
926
- function formatN(n) {
927
- const base = `${n.value} ${n.unit}`;
928
- return n.detail ? `${base} (${n.detail})` : base;
929
- }
930
- /**
931
- * Render the registry's human index from records. Deterministic: identical
932
- * records always produce identical output, so a byte diff against the
933
- * committed index is the freshness check. No clocks, no environment.
934
- */
935
- function renderEvidenceIndex(raws) {
936
- const records = validateEvidenceRegistry(raws);
937
- const lines = [
938
- "# Evidence index",
939
- "",
940
- "<!-- GENERATED by scripts/render-evidence-index.ts from evidence/records/*.json.",
941
- " Edit the records, run `pnpm run evidence:render`, and commit both.",
942
- " `pnpm run evidence:check` (inside verify:package) fails when this file drifts. -->",
943
- "",
944
- `${records.length} records. States: ${EVIDENCE_STATES.join(" > ")}.`,
945
- "",
946
- "| state | date | id | claim | result |",
947
- "| --- | --- | --- | --- | --- |"
948
- ];
949
- for (const r of records) lines.push(`| ${r.evidenceState} | ${r.date} | [\`${r.id}\`](#${r.id}) | ${mdEscape(r.claim)} | ${mdEscape(r.result)} |`);
950
- for (const r of records) {
951
- lines.push("", `## ${r.id}`, "");
952
- lines.push(`**${r.evidenceState}** · ${r.date} · ${r.domain} · source repo \`${r.sourceRepo}\``);
953
- lines.push("", `**Claim.** ${r.claim}`, "", `**Result.** ${r.result}`, "");
954
- lines.push(`- **Instrument**: ${r.instrument}`);
955
- lines.push(`- **Command**: ${r.command === null ? "not preserved" : `\`${r.command}\``}`);
956
- lines.push(`- **Arms**: ${r.arms.length === 0 ? "single-arm" : r.arms.join(" vs ")}`);
957
- lines.push(`- **n**: ${formatN(r.n)}`);
958
- lines.push(`- **Cost**: ${formatCost(r.costUsd)}`);
959
- if (r.experimentDigest) lines.push(`- **Experiment seal**: \`${r.experimentDigest}\``);
960
- lines.push(`- **Artifacts**:`);
961
- for (const artifact of r.artifacts) lines.push(` - ${artifact}`);
962
- if (r.confounds.length === 0) lines.push("- **Confounds**: none recorded");
963
- else {
964
- lines.push("- **Confounds**:");
965
- for (const confound of r.confounds) lines.push(` - ${confound}`);
966
- }
967
- if (r.supersedes && r.supersedes.length > 0) lines.push(`- **Supersedes**: ${r.supersedes.map((s) => `\`${s}\``).join(", ")}`);
968
- if (r.notes) lines.push(`- **Notes**: ${r.notes}`);
969
- }
970
- lines.push("");
971
- return lines.join("\n");
972
- }
973
881
  //#endregion
974
882
  //#region src/experiment/power.ts
975
883
  /**
@@ -1531,6 +1439,6 @@ function decideRandomizedCanary(sealed, cohort, rows, verifyCohort) {
1531
1439
  return result;
1532
1440
  }
1533
1441
  //#endregion
1534
- export { BOOTSTRAP_GATE_MIN_N, DesignRefusalError, EVIDENCE_RECEIPT_VERSION, EVIDENCE_STATES, FinalEvidenceConflictError, FinalEvidenceError, FunnelIntegrityError, MatchedBudgetError, SealIntegrityError, amendExperiment, assertDesignAdequate, assertFunnelReconciles, assertMatchedBudgets, benjaminiHochberg, bonferroni, buildEvidenceVector, buildFunnel, classifyReissue, clusteredPower, comparePairedArms, composeFunnels, computeEstimand, computeInterval, createCampaignEvidenceReceipt, createEvidenceReceipt, decideRandomizedCanary, defineEvaluationClaim, defineExperiment, eProcess, evaluateCondition, evaluateHaltRule, evaluateHypothesis, evaluateIdentityGate, evaluateOracleDeterminismGate, evaluatePopulationReproducibilityGate, evaluatePowerFloorGate, evaluatePredicate, evaluateProvenanceGate, evidenceRegistryRecordSchema, executeAdmissionRule, executeDecisionRule, hashJson, heldoutSignificance, holm, isIndependentEvidence, joinProspectiveCanaryReceipts, manifestContentDigest, mcnemar, mcnemarPower, mcnemarRequiredN, mulberry32, openFinalEvidenceLedger, openSealedExperiment, pairArms, pairHoldout, pairRunRecords, pairedBootstrap, pairedEvalueSequence, pairedMde, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, paretoPolicy, paretoSignificanceGate, parseEvidenceRegistryRecord, powerPreflight, projectNLadderBudget, randomizedCanaryObservationDigest, randomizedCanaryRosterDigest, readField, renderEvidenceIndex, renderFunnelTable, requiredPairedSampleSize, requiredSampleSize, runSelectionRule, runUniformPassBudget, sealExperiment, sealRandomizedCanaryRule, sequentialCrossingHorizon, sequentialDecide, sequentialPairedGate, signManifest, summarizeEvaluationUnits, validateEvidenceRegistry, verifyEvidenceReceipt, verifyManifest, verifyMatchedBudgets, verifySealedExperiment, wilson };
1442
+ export { BOOTSTRAP_GATE_MIN_N, EVIDENCE_STATES, FinalEvidenceConflictError, FinalEvidenceError, amendExperiment, assertDesignAdequate, assertMatchedBudgets, benjaminiHochberg, bonferroni, buildEvidenceVector, buildFunnel, clusteredPower, comparePairedArms, composeFunnels, computeInterval, createCampaignEvidenceReceipt, createEvidenceReceipt, decideRandomizedCanary, defineEvaluationClaim, defineExperiment, eProcess, evaluateCondition, evaluatePowerFloorGate, evidenceRegistryRecordSchema, executeAdmissionRule, executeDecisionRule, hashJson, heldoutSignificance, holm, isIndependentEvidence, joinProspectiveCanaryReceipts, manifestContentDigest, mcnemar, mcnemarPower, mcnemarRequiredN, mulberry32, openFinalEvidenceLedger, openSealedExperiment, pairArms, pairHoldout, pairRunRecords, pairedBootstrap, pairedEvalueSequence, pairedMde, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, paretoPolicy, paretoSignificanceGate, powerPreflight, randomizedCanaryObservationDigest, randomizedCanaryRosterDigest, readField, renderFunnelTable, requiredPairedSampleSize, requiredSampleSize, sealExperiment, sealRandomizedCanaryRule, sequentialDecide, sequentialPairedGate, signManifest, summarizeEvaluationUnits, verifyEvidenceReceipt, verifyManifest, verifyMatchedBudgets, verifySealedExperiment, wilson };
1535
1443
 
1536
1444
  //# sourceMappingURL=index.js.map