@tangle-network/agent-eval 0.183.0 → 0.185.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (248) hide show
  1. package/CHANGELOG.md +28 -0
  2. package/dist/adapters/http.d.ts +2 -2
  3. package/dist/{agent-profile-cell-s__adRnK.d.ts → agent-profile-cell-BHcB-mZX.d.ts} +8 -9
  4. package/dist/{agent-profile-cell-s__adRnK.d.ts.map → agent-profile-cell-BHcB-mZX.d.ts.map} +1 -1
  5. package/dist/{agent-profile-cell-Cv6UA-W_.js → agent-profile-cell-DcFW4vOc.js} +16 -15
  6. package/dist/agent-profile-cell-DcFW4vOc.js.map +1 -0
  7. package/dist/{agent-profile-CivaSsSy.d.ts → agent-profile-uAIQK3Nr.d.ts} +3 -3
  8. package/dist/{agent-profile-CivaSsSy.d.ts.map → agent-profile-uAIQK3Nr.d.ts.map} +1 -1
  9. package/dist/analyst/index.d.ts +11 -336
  10. package/dist/analyst/index.d.ts.map +1 -1
  11. package/dist/analyst/index.js +10 -8
  12. package/dist/analyst/index.js.map +1 -1
  13. package/dist/analyze-runs-CrYhqXop.d.ts +75 -0
  14. package/dist/analyze-runs-CrYhqXop.d.ts.map +1 -0
  15. package/dist/{define-agent-eval-CneiWyuA.js → analyze-runs-sTygaDyD.js} +6 -813
  16. package/dist/analyze-runs-sTygaDyD.js.map +1 -0
  17. package/dist/{benchmark-DQKzykkO.js → benchmark-BeQJyrRC.js} +3 -3
  18. package/dist/{benchmark-DQKzykkO.js.map → benchmark-BeQJyrRC.js.map} +1 -1
  19. package/dist/{benchmark-command-NBrqNwgy.js → benchmark-command-kvq7yFY4.js} +17 -54
  20. package/dist/benchmark-command-kvq7yFY4.js.map +1 -0
  21. package/dist/benchmarks/index.d.ts +4 -4
  22. package/dist/benchmarks/index.js +3 -3
  23. package/dist/builder-eval/index.js +1 -1
  24. package/dist/campaign/index.d.ts +6 -5
  25. package/dist/campaign/index.js +10 -9
  26. package/dist/{campaign-BwcEBC91.js → campaign-eJgtRs-W.js} +16 -15
  27. package/dist/{campaign-BwcEBC91.js.map → campaign-eJgtRs-W.js.map} +1 -1
  28. package/dist/{campaign-evidence-PYPUftRy.js → campaign-evidence-BsrCmBSE.js} +7 -258
  29. package/dist/campaign-evidence-BsrCmBSE.js.map +1 -0
  30. package/dist/{canonical-DPyQ_rpt.js → canonical-BsZVgBhW.js} +2 -4
  31. package/dist/canonical-BsZVgBhW.js.map +1 -0
  32. package/dist/cli.js +2 -2
  33. package/dist/{client-kh2jOjTK.d.ts → client-Bc8LANCG.d.ts} +4 -4
  34. package/dist/{client-kh2jOjTK.d.ts.map → client-Bc8LANCG.d.ts.map} +1 -1
  35. package/dist/contract/index.d.ts +10 -10
  36. package/dist/contract/index.js +15 -284
  37. package/dist/contract/index.js.map +1 -1
  38. package/dist/{default-registry-BwDSWVzg.d.ts → default-registry-CXv9kHMt.d.ts} +6 -6
  39. package/dist/{default-registry-BwDSWVzg.d.ts.map → default-registry-CXv9kHMt.d.ts.map} +1 -1
  40. package/dist/{default-registry-C0-N4PkG.js → default-registry-xiHrldW8.js} +6 -4
  41. package/dist/{default-registry-C0-N4PkG.js.map → default-registry-xiHrldW8.js.map} +1 -1
  42. package/dist/define-agent-eval-D4uumjl3.js +813 -0
  43. package/dist/define-agent-eval-D4uumjl3.js.map +1 -0
  44. package/dist/{define-agent-eval-CwOWWQt_.d.ts → define-agent-eval-DE67ECH1.d.ts} +7 -7
  45. package/dist/{define-agent-eval-CwOWWQt_.d.ts.map → define-agent-eval-DE67ECH1.d.ts.map} +1 -1
  46. package/dist/diagnosis/index.d.ts +319 -0
  47. package/dist/diagnosis/index.d.ts.map +1 -0
  48. package/dist/diagnosis/index.js +1328 -0
  49. package/dist/diagnosis/index.js.map +1 -0
  50. package/dist/{dspy-rlm-engine-8pgBSiy7.js → dspy-rlm-engine-BE9lCzzL.js} +3 -3
  51. package/dist/{dspy-rlm-engine-8pgBSiy7.js.map → dspy-rlm-engine-BE9lCzzL.js.map} +1 -1
  52. package/dist/{engine-DS1cysJy.d.ts → engine-cXUK-WHS.d.ts} +5 -77
  53. package/dist/engine-cXUK-WHS.d.ts.map +1 -0
  54. package/dist/equal-terms-Blx9Zk5I.js +25 -0
  55. package/dist/equal-terms-Blx9Zk5I.js.map +1 -0
  56. package/dist/{eval-campaign-CCEkwzoH.js → eval-campaign-XFHEpN86.js} +4 -4
  57. package/dist/{eval-campaign-CCEkwzoH.js.map → eval-campaign-XFHEpN86.js.map} +1 -1
  58. package/dist/evaluation.d.ts +75 -0
  59. package/dist/evaluation.d.ts.map +1 -0
  60. package/dist/evaluation.js +137 -0
  61. package/dist/evaluation.js.map +1 -0
  62. package/dist/{exact-types-BZDe0W2D.d.ts → exact-types-BEHN7tnq.d.ts} +2 -2
  63. package/dist/{exact-types-BZDe0W2D.d.ts.map → exact-types-BEHN7tnq.d.ts.map} +1 -1
  64. package/dist/experiment/index.d.ts +2 -2
  65. package/dist/experiment/index.js +6 -5
  66. package/dist/experiment/index.js.map +1 -1
  67. package/dist/{experiment-tracker-C7PfnF4b.d.ts → experiment-tracker-URRGP9ek.d.ts} +2 -2
  68. package/dist/{experiment-tracker-C7PfnF4b.d.ts.map → experiment-tracker-URRGP9ek.d.ts.map} +1 -1
  69. package/dist/{external-optimizer-process-DIB4YWEe.js → external-optimizer-process-BbhvXTFx.js} +4 -4
  70. package/dist/{external-optimizer-process-DIB4YWEe.js.map → external-optimizer-process-BbhvXTFx.js.map} +1 -1
  71. package/dist/{external-optimizer-subprocess-CjguXN2e.js → external-optimizer-subprocess-DMgmqTB6.js} +3 -3
  72. package/dist/{external-optimizer-subprocess-CjguXN2e.js.map → external-optimizer-subprocess-DMgmqTB6.js.map} +1 -1
  73. package/dist/{feedback-trajectory-CXmtITBo.d.ts → feedback-trajectory-CytTR7u4.d.ts} +3 -3
  74. package/dist/{feedback-trajectory-CXmtITBo.d.ts.map → feedback-trajectory-CytTR7u4.d.ts.map} +1 -1
  75. package/dist/hosted/index.d.ts +2 -2
  76. package/dist/{index-DoykkxW0.d.ts → index-BQWfAFeH.d.ts} +11 -11
  77. package/dist/{index-DoykkxW0.d.ts.map → index-BQWfAFeH.d.ts.map} +1 -1
  78. package/dist/{index-DbG8QbH9.d.ts → index-C8_iU8KL.d.ts} +6 -6
  79. package/dist/{index-DbG8QbH9.d.ts.map → index-C8_iU8KL.d.ts.map} +1 -1
  80. package/dist/{index--_8_Nlpn.d.ts → index-CKZa_rnt.d.ts} +4 -121
  81. package/dist/index-CKZa_rnt.d.ts.map +1 -0
  82. package/dist/{index-Bp_6sj3x.d.ts → index-ww8736Am.d.ts} +20 -315
  83. package/dist/index-ww8736Am.d.ts.map +1 -0
  84. package/dist/index.d.ts +24 -22
  85. package/dist/index.d.ts.map +1 -1
  86. package/dist/index.js +25 -24
  87. package/dist/index.js.map +1 -1
  88. package/dist/{insight-report-D1qa0HWs.d.ts → insight-report-C3Bfl762.d.ts} +3 -3
  89. package/dist/{insight-report-D1qa0HWs.d.ts.map → insight-report-C3Bfl762.d.ts.map} +1 -1
  90. package/dist/jev-protocol.d.ts +74 -0
  91. package/dist/jev-protocol.d.ts.map +1 -0
  92. package/dist/jev-protocol.js +91 -0
  93. package/dist/jev-protocol.js.map +1 -0
  94. package/dist/jev.d.ts +131 -0
  95. package/dist/jev.d.ts.map +1 -0
  96. package/dist/jev.js +312 -0
  97. package/dist/jev.js.map +1 -0
  98. package/dist/{journal-Cs9f7385.js → journal-BuRLm6ho.js} +2 -2
  99. package/dist/{journal-Cs9f7385.js.map → journal-BuRLm6ho.js.map} +1 -1
  100. package/dist/{kind-factory-DxTzg32u.js → kind-factory-V79Hw0AU.js} +7 -84
  101. package/dist/kind-factory-V79Hw0AU.js.map +1 -0
  102. package/dist/ledger-core/index.js +2 -2
  103. package/dist/{llm-judge-FfaRgz9X.js → llm-judge-FWWhEb31.js} +11 -9
  104. package/dist/{llm-judge-FfaRgz9X.js.map → llm-judge-FWWhEb31.js.map} +1 -1
  105. package/dist/{matrix-CyhW-vgJ.d.ts → matrix-CPKsRY8Y.d.ts} +2 -2
  106. package/dist/{matrix-CyhW-vgJ.d.ts.map → matrix-CPKsRY8Y.d.ts.map} +1 -1
  107. package/dist/meta-eval/index.d.ts +54 -2
  108. package/dist/meta-eval/index.d.ts.map +1 -1
  109. package/dist/meta-eval/index.js +52 -3
  110. package/dist/meta-eval/index.js.map +1 -1
  111. package/dist/{mint-ySIIkKlV.js → mint-B4zgSAJW.js} +63 -14
  112. package/dist/mint-B4zgSAJW.js.map +1 -0
  113. package/dist/mint-BI14xehJ.d.ts +285 -0
  114. package/dist/mint-BI14xehJ.d.ts.map +1 -0
  115. package/dist/multishot/golden/index.d.ts +1 -1
  116. package/dist/multishot/index.d.ts +2 -2
  117. package/dist/openapi.json +1 -1
  118. package/dist/otel-spans-CQY1pNiN.js +275 -0
  119. package/dist/otel-spans-CQY1pNiN.js.map +1 -0
  120. package/dist/otlp-attributes-BpC_blLF.js +83 -0
  121. package/dist/otlp-attributes-BpC_blLF.js.map +1 -0
  122. package/dist/{run-record-DPa2PeHx.js → paired-promotion-decision-ClJh_9nK.js} +2 -215
  123. package/dist/paired-promotion-decision-ClJh_9nK.js.map +1 -0
  124. package/dist/pipelines/index.js +2 -2
  125. package/dist/prime-bridge-transport-Z3cpKGoV.js +43 -0
  126. package/dist/prime-bridge-transport-Z3cpKGoV.js.map +1 -0
  127. package/dist/prime-protocol-CDBcL9NM.d.ts +331 -0
  128. package/dist/prime-protocol-CDBcL9NM.d.ts.map +1 -0
  129. package/dist/{prime-protocol-6tZTVsWm.js → prime-protocol-DcAK9pC8.js} +2 -23
  130. package/dist/prime-protocol-DcAK9pC8.js.map +1 -0
  131. package/dist/{produced-state-CGoVInrl.js → produced-state-CmB02tr2.js} +23 -19
  132. package/dist/{produced-state-CGoVInrl.js.map → produced-state-CmB02tr2.js.map} +1 -1
  133. package/dist/profile-cell.d.ts +1 -1
  134. package/dist/profile-cell.js +1 -1
  135. package/dist/{promotion-policy-BKFrl7CC.js → promotion-policy-DdrYWHDh.js} +3 -3
  136. package/dist/{promotion-policy-BKFrl7CC.js.map → promotion-policy-DdrYWHDh.js.map} +1 -1
  137. package/dist/{query-D1nLIKt7.js → query-DlV0pdx4.js} +2 -2
  138. package/dist/{query-D1nLIKt7.js.map → query-DlV0pdx4.js.map} +1 -1
  139. package/dist/{registry-BRbB6Y0v.d.ts → registry-D5Ly3GRV.d.ts} +3 -3
  140. package/dist/{registry-BRbB6Y0v.d.ts.map → registry-D5Ly3GRV.d.ts.map} +1 -1
  141. package/dist/{release-confidence-BcqeQTHW.d.ts → release-confidence-DsETiO4f.d.ts} +3 -3
  142. package/dist/{release-confidence-BcqeQTHW.d.ts.map → release-confidence-DsETiO4f.d.ts.map} +1 -1
  143. package/dist/{release-confidence-DMg8n18l.js → release-confidence-RanRUXRW.js} +2 -2
  144. package/dist/{release-confidence-DMg8n18l.js.map → release-confidence-RanRUXRW.js.map} +1 -1
  145. package/dist/reporting.d.ts +3 -3
  146. package/dist/reporting.js +2 -2
  147. package/dist/{researcher-64T49THL.d.ts → researcher-CJkH-6Ty.d.ts} +4 -4
  148. package/dist/{researcher-64T49THL.d.ts.map → researcher-CJkH-6Ty.d.ts.map} +1 -1
  149. package/dist/{reward-hacking-uzO_ihep.d.ts → reward-hacking-BIQ2BWLG.d.ts} +2 -2
  150. package/dist/{reward-hacking-uzO_ihep.d.ts.map → reward-hacking-BIQ2BWLG.d.ts.map} +1 -1
  151. package/dist/rl.d.ts +10 -6
  152. package/dist/rl.d.ts.map +1 -1
  153. package/dist/rl.js +22 -7
  154. package/dist/rl.js.map +1 -1
  155. package/dist/rollout/index.d.ts +2 -2
  156. package/dist/rollout/index.js +2 -2
  157. package/dist/{rollout-D_54FZVu.js → rollout-C0Ifk3vK.js} +2 -2
  158. package/dist/{rollout-D_54FZVu.js.map → rollout-C0Ifk3vK.js.map} +1 -1
  159. package/dist/{rubric-predictive-validity-Bmj2_cll.d.ts → rubric-predictive-validity-CTU8y8Cv.d.ts} +2 -2
  160. package/dist/{rubric-predictive-validity-Bmj2_cll.d.ts.map → rubric-predictive-validity-CTU8y8Cv.d.ts.map} +1 -1
  161. package/dist/run-record-CNNzWGpB.js +216 -0
  162. package/dist/run-record-CNNzWGpB.js.map +1 -0
  163. package/dist/{run-record-BiTWauyO.d.ts → run-record-D9wS35fK.d.ts} +2 -2
  164. package/dist/{run-record-BiTWauyO.d.ts.map → run-record-D9wS35fK.d.ts.map} +1 -1
  165. package/dist/{run-record-DualPTn2.js → run-record-tg_5H2IK.js} +2 -2
  166. package/dist/{run-record-DualPTn2.js.map → run-record-tg_5H2IK.js.map} +1 -1
  167. package/dist/search-ledger-types-D7KDVg-o.d.ts +302 -0
  168. package/dist/search-ledger-types-D7KDVg-o.d.ts.map +1 -0
  169. package/dist/{semantic-concept-judge-_3Fp6Q_3.js → semantic-concept-judge-Cnt0TWp0.js} +4 -4
  170. package/dist/{semantic-concept-judge-_3Fp6Q_3.js.map → semantic-concept-judge-Cnt0TWp0.js.map} +1 -1
  171. package/dist/{sequential-D_xHrBZe.js → sequential-PgLEqcTJ.js} +3 -3
  172. package/dist/{sequential-D_xHrBZe.js.map → sequential-PgLEqcTJ.js.map} +1 -1
  173. package/dist/{server-BZQR1f0C.js → server-DmpcFiVK.js} +2 -2
  174. package/dist/{server-BZQR1f0C.js.map → server-DmpcFiVK.js.map} +1 -1
  175. package/dist/{skillopt-optimization-method-DY1aQ-aL.js → skillopt-optimization-method-FSk8b5FQ.js} +8 -7
  176. package/dist/{skillopt-optimization-method-DY1aQ-aL.js.map → skillopt-optimization-method-FSk8b5FQ.js.map} +1 -1
  177. package/dist/{statistical-heldout-CpVd6FmY.d.ts → statistical-heldout-Bu4oNOJh.d.ts} +3 -3
  178. package/dist/{statistical-heldout-CpVd6FmY.d.ts.map → statistical-heldout-Bu4oNOJh.d.ts.map} +1 -1
  179. package/dist/statistical-heldout-CeWT_W7y.js +256 -0
  180. package/dist/statistical-heldout-CeWT_W7y.js.map +1 -0
  181. package/dist/{store-otlp-DAaEbMLf.js → store-otlp-2gYCy-v0.js} +2 -2
  182. package/dist/{store-otlp-DAaEbMLf.js.map → store-otlp-2gYCy-v0.js.map} +1 -1
  183. package/dist/{store-tool-spans-BZ8MPKFe.js → store-tool-spans-B1e6Auc2.js} +69 -6
  184. package/dist/store-tool-spans-B1e6Auc2.js.map +1 -0
  185. package/dist/{store-tool-spans-Dt-YdAuE.d.ts → store-tool-spans-BrdfPaVh.d.ts} +54 -6
  186. package/dist/store-tool-spans-BrdfPaVh.d.ts.map +1 -0
  187. package/dist/{summary-report-IGAOtgPd.js → summary-report-BMvou2oZ.js} +2 -2
  188. package/dist/{summary-report-IGAOtgPd.js.map → summary-report-BMvou2oZ.js.map} +1 -1
  189. package/dist/{summary-report-D1h4dlrK.d.ts → summary-report-BzFS5UhZ.d.ts} +2 -2
  190. package/dist/{summary-report-D1h4dlrK.d.ts.map → summary-report-BzFS5UhZ.d.ts.map} +1 -1
  191. package/dist/{task-failure-attributes-B_JjrOVo.js → task-failure-attributes-B34TiMBe.js} +2 -2
  192. package/dist/{task-failure-attributes-B_JjrOVo.js.map → task-failure-attributes-B34TiMBe.js.map} +1 -1
  193. package/dist/{tool-groups-B2bSNaJB.d.ts → tool-groups-eOOgYGaM.d.ts} +3 -3
  194. package/dist/tool-groups-eOOgYGaM.d.ts.map +1 -0
  195. package/dist/{tool-waste-C7MU9u1e.js → tool-waste-nbHPAxei.js} +2 -2
  196. package/dist/{tool-waste-C7MU9u1e.js.map → tool-waste-nbHPAxei.js.map} +1 -1
  197. package/dist/trace-repair/index.d.ts +2 -2
  198. package/dist/trace-repair/index.js +5 -4
  199. package/dist/trace-repair/index.js.map +1 -1
  200. package/dist/traces.d.ts +5 -5
  201. package/dist/traces.js +8 -7
  202. package/dist/traces.js.map +1 -1
  203. package/dist/trajectory-replay/index.js +1 -1
  204. package/dist/{types-CS0qk_Yp.d.ts → types-BTRcIqFr.d.ts} +3 -3
  205. package/dist/{types-CS0qk_Yp.d.ts.map → types-BTRcIqFr.d.ts.map} +1 -1
  206. package/dist/{types-D7gEdPoQ.d.ts → types-BjB2wZM9.d.ts} +2 -2
  207. package/dist/{types-D7gEdPoQ.d.ts.map → types-BjB2wZM9.d.ts.map} +1 -1
  208. package/dist/types-C8VBbiw6.js +91 -0
  209. package/dist/types-C8VBbiw6.js.map +1 -0
  210. package/dist/{types-BvZoPTGa.d.ts → types-c3ZRTbw9.d.ts} +3 -3
  211. package/dist/{types-BvZoPTGa.d.ts.map → types-c3ZRTbw9.d.ts.map} +1 -1
  212. package/dist/{types-DQ0e2E7y.js → usage-receipt-Cx3xNeyB.js} +2 -89
  213. package/dist/usage-receipt-Cx3xNeyB.js.map +1 -0
  214. package/dist/{verdict-BQ3pCFf8.js → verdict-DhPoXApk.js} +2 -2
  215. package/dist/{verdict-BQ3pCFf8.js.map → verdict-DhPoXApk.js.map} +1 -1
  216. package/dist/{verdict-cache-B3eCVQtY.js → verdict-cache-BQEsVy9n.js} +2 -2
  217. package/dist/{verdict-cache-B3eCVQtY.js.map → verdict-cache-BQEsVy9n.js.map} +1 -1
  218. package/dist/wire/index.d.ts +1 -1
  219. package/dist/wire/index.js +1 -1
  220. package/docs/changes/evaluation-definition-identity.md +27 -0
  221. package/docs/changes/jev-evidence-review.md +18 -0
  222. package/docs/changes/jev-product-adoption.md +18 -0
  223. package/docs/changes/probability-policy-audit.md +81 -0
  224. package/docs/diagnosis.md +42 -0
  225. package/docs/jev-behavior-review.md +174 -0
  226. package/docs/jev-integrations.md +50 -0
  227. package/docs/jev.md +140 -0
  228. package/docs/produced-state.md +13 -0
  229. package/docs/profile-json-identity.md +33 -0
  230. package/package.json +22 -1
  231. package/dist/agent-profile-cell-Cv6UA-W_.js.map +0 -1
  232. package/dist/benchmark-command-NBrqNwgy.js.map +0 -1
  233. package/dist/campaign-evidence-PYPUftRy.js.map +0 -1
  234. package/dist/canonical-DPyQ_rpt.js.map +0 -1
  235. package/dist/define-agent-eval-CneiWyuA.js.map +0 -1
  236. package/dist/engine-DS1cysJy.d.ts.map +0 -1
  237. package/dist/exporters-D98GPi1n.d.ts +0 -161
  238. package/dist/exporters-D98GPi1n.d.ts.map +0 -1
  239. package/dist/index--_8_Nlpn.d.ts.map +0 -1
  240. package/dist/index-Bp_6sj3x.d.ts.map +0 -1
  241. package/dist/kind-factory-DxTzg32u.js.map +0 -1
  242. package/dist/mint-ySIIkKlV.js.map +0 -1
  243. package/dist/prime-protocol-6tZTVsWm.js.map +0 -1
  244. package/dist/run-record-DPa2PeHx.js.map +0 -1
  245. package/dist/store-tool-spans-BZ8MPKFe.js.map +0 -1
  246. package/dist/store-tool-spans-Dt-YdAuE.d.ts.map +0 -1
  247. package/dist/tool-groups-B2bSNaJB.d.ts.map +0 -1
  248. package/dist/types-DQ0e2E7y.js.map +0 -1
@@ -1,6 +1,6 @@
1
- import { a as RunRecord } from "./run-record-BiTWauyO.js";
1
+ import { a as RunRecord } from "./run-record-D9wS35fK.js";
2
2
  import { a as PairedPromotionDecision, d as PairedBootstrapResult, i as PairedMcNemarEvidence, r as PairedDecisionStatistic, t as PairedDecisionMethod, u as PairedBootstrapOptions } from "./paired-promotion-decision-DPsMQm-0.js";
3
- import { R as Scenario, a as CampaignResult, b as GenerationRecord, h as GateContext, j as MutableSurface, p as Gate, v as GateResult, w as JudgeScore } from "./types-CS0qk_Yp.js";
3
+ import { R as Scenario, a as CampaignResult, b as GenerationRecord, h as GateContext, j as MutableSurface, p as Gate, v as GateResult, w as JudgeScore } from "./types-BTRcIqFr.js";
4
4
  import { r as LedgerHash } from "./canonical-CFpojCN5.js";
5
5
  import { z } from "zod";
6
6
  //#region src/statistics/paired-binary.d.ts
@@ -1340,4 +1340,4 @@ declare function dimensionRegressions(candidate: Map<string, Record<string, Judg
1340
1340
  }): DimensionRegression[];
1341
1341
  //#endregion
1342
1342
  export { paretoSignificanceGate as $, wilson as $t, PairArmsOptions as A, FinalEvidenceReservation as At, PowerPreflight as B, eProcess as Bt, hashJson as C, verifyAttestation as Ct, ComparePairedArmsOptions as D, FinalEvidenceMeasurement as Dt, verifyManifest as E, FinalEvidenceLedger as Et, PairedCorrectness as F, summarizeEvaluationUnits as Ft, BuildEvidenceVectorOptions as G, ScoreRiskDifferenceResult as Gt, powerPreflight as H, McNemarResult as Ht, PairedMetricDelta as I, EProcess as It, ParetoSignificanceGateOptions as J, pairedBinaryScale as Jt, EvidenceVector as K, isBinaryOutcomeVector as Kt, comparePairedArms as L, EProcessOptions as Lt, PairRunRecordsResult as M, EvaluationClaim as Mt, PairedArmRow as N, EvaluationUnitSummary as Nt, MatchedPair as O, FinalEvidenceOutcome as Ot, PairedArmsComparison as P, defineEvaluationClaim as Pt, paretoPolicy as Q, passAtK as Qt, pairArms as R, EProcessState as Rt, evaluateHypothesis as S, attest as St, signManifest as T, FinalEvidenceError as Tt, AxisEvidence as U, ProportionInterval as Ut, PowerPreflightOptions as V, ExactRiskDifferenceResult as Vt, AxisVerdict as W, RiskDifferenceResult as Wt, PromotionPolicy as X, pairedRiskDifferenceExact as Xt, PromotionObjective as Y, pairedRiskDifference as Yt, buildEvidenceVector as Z, pairedRiskDifferenceScore as Zt, sequentialPairedGate as _, verifyEvidenceReceipt as _t, detectScale as a, createCampaignEvidenceReceipt as at, SignedManifest as b, AttestationVerification as bt, pairHoldout as c, EVIDENCE_RECEIPT_VERSION as ct, SequentialDecision as d, EvidenceBinding as dt, Objective as et, SequentialObservation as f, EvidenceReceipt as ft, sequentialDecide as g, isIndependentEvidence as gt, SequentialStreamState as h, createEvidenceReceipt as ht, PairedHoldout as i, CampaignEvidenceContext as it, PairArmsResult as j, openFinalEvidenceLedger as jt, MatchedRunRecordPair as k, FinalEvidenceRecord as kt, SequentialDecideFn as l, EvidenceAuthority as lt, SequentialPairedGateOptions as m, INDEPENDENT_EVIDENCE_AUTHORITY_KINDS as mt, HeldoutSignificance as n, dominates as nt, dimensionRegressions as o, CreateEvidenceReceiptInput as ot, SequentialPairedGate as p, EvidenceReceiptVerification as pt, ObjectiveSource as q, mcnemar as qt, HeldoutSignificanceOptions as r, paretoFrontier as rt, heldoutSignificance as s, EVIDENCE_AUTHORITY_KINDS as st, DimensionRegression as t, ParetoResult as tt, SequentialDecideOptions as u, EvidenceAuthorityKind as ut, HypothesisManifest as v, ATTESTATION_ALGORITHM as vt, manifestContentDigest as w, FinalEvidenceConflictError as wt, SignedManifestAlgo as x, AttestedReport as xt, HypothesisResult as y, AttestationProvenance as yt, pairRunRecords as z, EProcessStep as zt };
1343
- //# sourceMappingURL=statistical-heldout-CpVd6FmY.d.ts.map
1343
+ //# sourceMappingURL=statistical-heldout-Bu4oNOJh.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"statistical-heldout-CpVd6FmY.d.ts","names":[],"sources":["../src/statistics/paired-binary.ts","../src/statistics/sequential-eprocess.ts","../src/experiment/claim.ts","../src/experiment/final-evidence.ts","../src/attestation.ts","../src/experiment/evidence-receipt.ts","../src/experiment/campaign-evidence.ts","../src/pareto.ts","../src/campaign/gates/promotion-policy.ts","../src/campaign/gates/power-preflight.ts","../src/paired-arms.ts","../src/pre-registration.ts","../src/campaign/gates/sequential.ts","../src/campaign/gates/statistical-heldout.ts"],"mappings":";;;;;;;UAgBiB;;EAEf;;EAEA;;EAEA;;;;;;;;;iBAUc,OAAO,mBAAmB,WAAW,sBAAoB;;;;;;;;;;;;;;;;;;;;;;;;;;iBA2CzD,sBAAsB,QAAQ;;UAU7B;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;iBAec,QACd,SAAS,6BACT,WAAW,8BACV;;UAmBc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;iBAkBc,qBACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;;UAkCc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA+Bc,0BACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;UAyEc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA8Dc,0BACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;;;;;;;;;;;;;;;;;;iBAiFa,kBACd,QAAQ,mBACR,OAAO;;;;;;;;;;;iBA0BO,QAAQ,WAAW,WAAW;;;UCrgB7B;;;EAGf;;;;EAIA;;;;EAIA;;;;;;;;EAQA,SAAS;;UAGM;;EAEf;;EAEA;;EAEA;;;;;;;;;;UAWe,sBAAsB;EACrC;EACA;EACA;;EAEA;;EAEA;;EAEA;;;EAGA;;UAGe;;;EAGf,OAAO,YAAY;EACnB,SAAS;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA8BK,SAAS,OAAM,kBAAuB;;;cClFhD,aAAW,EAAA;;;;;;;;;;;;;;;;;GASN,EAAA,KAAA;;KAGC,kBAAkB,EAAE,aAAa;;iBAG7B,sBAAsB,OAAO,kBAAkB;UAO9C;EACf;EACA;EACA,OAAO;IAAQ;IAAY;;;;iBAIb,yBACd,OAAO,iBACP,0BACC;;;cCjCG,mBAAiB,EAAA;;;;;;GAQZ,EAAA,KAAA;;KAGC,2BAA2B,EAAE,aAAa;cAEhD,mBAAiB,EAAA;;;GAKZ,EAAA,KAAA;KAEC,2BAA2B,EAAE,aAAa;UAgCrC;EACf,aAAa;EACb,iBAAiB;EACjB;IAAY,aAAa;IAA0B,WAAW;;;KAGpD,qBAAqB;EAC3B;EAAiB,OAAO;;EACxB;EAAkB;IAAS;IAA8C;;;UAE9D;EACf,QACE,OAAO,2BACN,QAAQ;IAAuB,QAAQ;IAAqB;;EAC/D,OACE,mBACA,aAAa,2BACZ,QAAQ;IAAuB,QAAQ;IAAqB;;EAC/D,QAAQ,QAAQ,qBAAqB;;;cAI1B,2BAA2B;WAE3B;EADX,YACW,8CACT;;;cAQS,mCAAmC;EAC9C,YAAY;;;iBA6HE,wBAAwB;EAAW;IAAiB;;;;;;;;;;;;;;;;;;;;;;;cCrMvD;UAEI;;EAEf,eAAe;;EAEf;;;EAGA;;EAEA;;EAEA;;;EAGA;;UAGe;;EAEf;EACA,YAAY;EACZ,kBAAkB;;EAElB;;UAGe;EACf;;EAEA;;;;;;;;iBAiBc,OAAO,iBAAiB,YAAY,wBAAwB;;;;;;;;;;iBAoB5D,kBACd,iBACA,UAAU,iBACT;;;cC1EU;;cAGA;KAQD,gCAAgC;cAE/B;UAOI;WACN,MAAM;;WAEN;;UAGM;WACN,sBAAsB;;WAEtB;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;WACA,WAAW;;WAEX;;WAEA;;UAGM;WACN,SAAS;WACT,aAAa;;UAGP;WACN;WACA;;UAGM,mCAAmC,KAAK;;;;;;;iBAQzC,sBACd,OAAO,4BACP,YAAY,wBACX;;;;;iBA4Ba,sBAAsB,SAAS,kBAAkB;;;;;;iBAkCjD,sBAAsB,SAAS;;;;KCjJnC,0BAA0B,KACpC;EAOI,YAAY,KAAK;;;iBAGP,8BAA8B,UAAU,UAAU,GAAG;EACnE,UAAU,eAAe,GAAG;EAC5B,SAAS;EACT,SAAS;IAAA;;;;;;;;;;;;;;;;;;KCPC;UAEK,UAAU;;EAEzB;EACA,WAAW;EACX,QAAQ,WAAW;;UAGJ,aAAa;EAC5B,UAAU;EACV,WAAW;;EAEX,cAAc;IAAQ,WAAW;IAAG,WAAW;;;;iBAIjC,UAAU,GAAG,GAAG,GAAG,GAAG,GAAG,YAAY,UAAU;;;;;;iBAmB/C,eAAe,GAAG,YAAY,KAAK,YAAY,UAAU,OAAO,aAAa;;;;;KCxBjF;EAAoB;;EAAwB;EAAmB;;UAE1D;;EAEf;EACA,QAAQ;;;EAGR,WAAW;;;;EAIX;;;;EAIA;;;;;EAKA;;;;KAKU;UAEK;EACf;EACA,QAAQ;EACR,WAAW;;;;;;;;EAQX,WAAW;;;;;EAKX;;;EAGA;IAAM;IAAa;;;EAEnB,mBAAmB;;EAEnB,SAAS;;;EAGT;;EAEA;EACA;EACA,gBAAgB;EAChB;EACA;EACA,SAAS;;UAGM;;EAEf,MAAM;;;EAGN;;;EAGA;IAAQ;IAAmB;;;;;;KAMjB,mBAAmB,IAAI,mBAAmB;UAErC;;;;EAIf;;EAEA;;EAEA;;EAEA;;;EAGA;;;;;;;;iBASc,oBAAoB,WAAW,kBAAkB,UAC/D,KAAK,YAAY,WAAW,YAC5B,YAAY,sBACZ,OAAM,6BACL;;;;;;cAkIU,cAAc;UA+EV,sCAAsC;;EAErD,YAAY;;;EAGZ,SAAS;;EAET;;;;;;;iBAQc,uBAAuB,qBAAqB,kBAAkB,WAAW,UACvF,SAAS,gCACR,KAAK,WAAW;;;;;;;;;;;;;;;;;UCzVF;;EAEf;;;EAGA;;EAEA;;EAEA;;;;EAIA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA;;;EAGA;EACA;EACA;;EAEA;;EAEA;;;iBAYc,eAAe,MAAM,wBAAwB;;;;;UCzB5C;;;EAGf;;;;;EAKA;;EAEA;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;;EAEA;;;UAIe;EACf;;;;EAIA;EACA,UAAU;EACV,WAAW;;UAGI;;EAEf,OAAO;;;EAGP,kBAAkB;;EAElB,mBAAmB;;;;;;;;;;;;;;;;;;;;iBAqBL,SAAS,eAAe,gBAAgB,MAAM,kBAAkB;;UA8G/D;;EAEf;;EAEA;;EAEA,SAAS;;EAET,gBAAgB;;;UAID;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,aAAa;;EAEb;IAAY;IAAW;;;UAGR,iCAAiC;;;;EAIhD;;EAEA,YAAY;;UAGG;EACf;EACA;EACA;;;EAGA,aAAa;EACb,cAAc;;;;;;;;;;;;;;;iBAgBA,kBACd,eAAe,gBACf,MAAM,2BACL;UAmEc;EACf;EACA;EACA,UAAU;EACV,WAAW;;UAGI;EACf,OAAO;EACP,kBAAkB;EAClB,mBAAmB;;;;;;;;;iBAeL,eACd,uBAAuB,aACvB,wBAAwB,cACvB;;;;;;;;;;;;;;;;UCrWc;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;;;KAIU;UAEK,uBAAuB;;EAEtC;;EAEA,MAAM;;UAGS;EACf,UAAU;EACV;EACA;EACA;;;EAGA;;EAEA,kBAAkB;EAGlB;;;;;;;;;;;;;;;;;iBAkBoB,SAAS,GAAG,KAAK,IAAI;;;;;iBAQ3B,sBAAsB,UAAU;;;;;;iBAc1B,aAAa,GAAG,qBAAqB,QAAQ;;;;;iBAS7C,eAAe,GAAG,iBAAiB;;;;;iBAYnC,mBACpB,UAAU,gBACV;EAAY;EAAW;EAAgB;IACtC,QAAQ;;;KCrFC;UAEK;EACf,UAAU;;EAEV;;EAEA;;;EAGA;;UAGe;;;EAGf;;;;EAIA;;;EAGA;;EAEA;;;;EAIA;;;EAGA;;;;EAIA,8BAA8B;;;EAG9B,kBAAkB;;EAElB;;;;;;;EAOA,SAAS;;;;;;KAOC,wBAAwB;EAAkB,UAAU;;UAE/C,qBAAqB,qBAAqB,kBAAkB,WAAW,kBAC9E,KAAK,WAAW;;;;;;;;;;EAUxB,QAAQ,gBAAgB;;;EAGxB,SAAS;;;;;;;;;;;;;;;;;iBAuPK,qBAAqB,qBAAqB,kBAAkB,WAAW,UACrF,SAAS,8BACR,qBAAqB,WAAW;UA6GlB;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;GACd;IAAQ,SAAS;;IAAyB;IAAe;;;EAE1D,SAAS;;;;;;;;;;;;;;;;;;;;;;;;;iBA0BK,iBAAiB,UAAS,0BAA+B;;;UC9dxD;;EAEf;;EAEA;;EAEA;;;;;;;;;;;;;iBA4Ec,YACd,WAAW,YAAY,eAAe,cACtC,UAAU,YAAY,eAAe,cACrC,aAAa,aACb,SAAS,GAAG,oCACX;UAkEc;EACf,QAAQ;;;;;;;;;;;;EAYR,WAAW;;;;EAIX,iBAAiB;;;;;;;EAOjB,UAAU;;EAEV,mBAAmB;;EAEnB,SAAS;;;;EAIT;;EAEA;;EAEA;EACA;;EAEA;;EAEA;;EAEA,gBAAgB;;EAEhB;;;;EAIA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;;EAEA;EACA;;EAEA,8BAA8B;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA6BhB,oBACd,QAAQ,eACR,OAAM,6BACL;UA2Ec;EACf;;;EAGA,WAAW;;;;EAIX;;EAEA;IAAM;IAAa;;;EAEnB,mBAAmB;;EAEnB,SAAS;;EAET;;;;EAIA;EACA;EACA;EACA;EACA;;EAEA;EACA;;EAEA;;EAEA;;;;;iBAMc,YAAY;;;;;;;;;iBAYZ,qBACd,WAAW,YAAY,eAAe,cACtC,UAAU,YAAY,eAAe,cACrC,aAAa,aACb,8BACA;EACE;EACA;EACA;EACA;;;EAGA;EACA,8BAA8B;EAC9B;IAED"}
1
+ {"version":3,"file":"statistical-heldout-Bu4oNOJh.d.ts","names":[],"sources":["../src/statistics/paired-binary.ts","../src/statistics/sequential-eprocess.ts","../src/experiment/claim.ts","../src/experiment/final-evidence.ts","../src/attestation.ts","../src/experiment/evidence-receipt.ts","../src/experiment/campaign-evidence.ts","../src/pareto.ts","../src/campaign/gates/promotion-policy.ts","../src/campaign/gates/power-preflight.ts","../src/paired-arms.ts","../src/pre-registration.ts","../src/campaign/gates/sequential.ts","../src/campaign/gates/statistical-heldout.ts"],"mappings":";;;;;;;UAgBiB;;EAEf;;EAEA;;EAEA;;;;;;;;;iBAUc,OAAO,mBAAmB,WAAW,sBAAoB;;;;;;;;;;;;;;;;;;;;;;;;;;iBA2CzD,sBAAsB,QAAQ;;UAU7B;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;iBAec,QACd,SAAS,6BACT,WAAW,8BACV;;UAmBc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;iBAkBc,qBACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;;UAkCc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA+Bc,0BACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;UAyEc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA8Dc,0BACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;;;;;;;;;;;;;;;;;;iBAiFa,kBACd,QAAQ,mBACR,OAAO;;;;;;;;;;;iBA0BO,QAAQ,WAAW,WAAW;;;UCrgB7B;;;EAGf;;;;EAIA;;;;EAIA;;;;;;;;EAQA,SAAS;;UAGM;;EAEf;;EAEA;;EAEA;;;;;;;;;;UAWe,sBAAsB;EACrC;EACA;EACA;;EAEA;;EAEA;;EAEA;;;EAGA;;UAGe;;;EAGf,OAAO,YAAY;EACnB,SAAS;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA8BK,SAAS,OAAM,kBAAuB;;;cClFhD,aAAW,EAAA;;;;;;;;;;;;;;;;;GASN,EAAA,KAAA;;KAGC,kBAAkB,EAAE,aAAa;;iBAG7B,sBAAsB,OAAO,kBAAkB;UAO9C;EACf;EACA;EACA,OAAO;IAAQ;IAAY;;;;iBAIb,yBACd,OAAO,iBACP,0BACC;;;cCjCG,mBAAiB,EAAA;;;;;;GAQZ,EAAA,KAAA;;KAGC,2BAA2B,EAAE,aAAa;cAEhD,mBAAiB,EAAA;;;GAKZ,EAAA,KAAA;KAEC,2BAA2B,EAAE,aAAa;UAgCrC;EACf,aAAa;EACb,iBAAiB;EACjB;IAAY,aAAa;IAA0B,WAAW;;;KAGpD,qBAAqB;EAC3B;EAAiB,OAAO;;EACxB;EAAkB;IAAS;IAA8C;;;UAE9D;EACf,QACE,OAAO,2BACN,QAAQ;IAAuB,QAAQ;IAAqB;;EAC/D,OACE,mBACA,aAAa,2BACZ,QAAQ;IAAuB,QAAQ;IAAqB;;EAC/D,QAAQ,QAAQ,qBAAqB;;;cAI1B,2BAA2B;WAE3B;EADX,YACW,8CACT;;;cAQS,mCAAmC;EAC9C,YAAY;;;iBA6HE,wBAAwB;EAAW;IAAiB;;;;;;;;;;;;;;;;;;;;;;;cCrMvD;UAEI;;EAEf,eAAe;;EAEf;;;EAGA;;EAEA;;EAEA;;;EAGA;;UAGe;;EAEf;EACA,YAAY;EACZ,kBAAkB;;EAElB;;UAGe;EACf;;EAEA;;;;;;;;iBAiBc,OAAO,iBAAiB,YAAY,wBAAwB;;;;;;;;;;iBAoB5D,kBACd,iBACA,UAAU,iBACT;;;cC1EU;;cAGA;KAQD,gCAAgC;cAE/B;UAOI;WACN,MAAM;;WAEN;;UAGM;WACN,sBAAsB;;WAEtB;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;WACA,WAAW;;WAEX;;WAEA;;UAGM;WACN,SAAS;WACT,aAAa;;UAGP;WACN;WACA;;UAGM,mCAAmC,KAAK;;;;;;;iBAQzC,sBACd,OAAO,4BACP,YAAY,wBACX;;;;;iBA4Ba,sBAAsB,SAAS,kBAAkB;;;;;;iBAkCjD,sBAAsB,SAAS;;;;KCjJnC,0BAA0B,KACpC;EAOI,YAAY,KAAK;;;iBAGP,8BAA8B,UAAU,UAAU,GAAG;EACnE,UAAU,eAAe,GAAG;EAC5B,SAAS;EACT,SAAS;IAAA;;;;;;;;;;;;;;;;;;KCPC;UAEK,UAAU;;EAEzB;EACA,WAAW;EACX,QAAQ,WAAW;;UAGJ,aAAa;EAC5B,UAAU;EACV,WAAW;;EAEX,cAAc;IAAQ,WAAW;IAAG,WAAW;;;;iBAIjC,UAAU,GAAG,GAAG,GAAG,GAAG,GAAG,YAAY,UAAU;;;;;;iBAmB/C,eAAe,GAAG,YAAY,KAAK,YAAY,UAAU,OAAO,aAAa;;;;;KCxBjF;EAAoB;;EAAwB;EAAmB;;UAE1D;;EAEf;EACA,QAAQ;;;EAGR,WAAW;;;;EAIX;;;;EAIA;;;;;EAKA;;;;KAKU;UAEK;EACf;EACA,QAAQ;EACR,WAAW;;;;;;;;EAQX,WAAW;;;;;EAKX;;;EAGA;IAAM;IAAa;;;EAEnB,mBAAmB;;EAEnB,SAAS;;;EAGT;;EAEA;EACA;EACA,gBAAgB;EAChB;EACA;EACA,SAAS;;UAGM;;EAEf,MAAM;;;EAGN;;;EAGA;IAAQ;IAAmB;;;;;;KAMjB,mBAAmB,IAAI,mBAAmB;UAErC;;;;EAIf;;EAEA;;EAEA;;EAEA;;;EAGA;;;;;;;;iBASc,oBAAoB,WAAW,kBAAkB,UAC/D,KAAK,YAAY,WAAW,YAC5B,YAAY,sBACZ,OAAM,6BACL;;;;;;cAkIU,cAAc;UA+EV,sCAAsC;;EAErD,YAAY;;;EAGZ,SAAS;;EAET;;;;;;;iBAQc,uBAAuB,qBAAqB,kBAAkB,WAAW,UACvF,SAAS,gCACR,KAAK,WAAW;;;;;;;;;;;;;;;;;UCzVF;;EAEf;;;EAGA;;EAEA;;EAEA;;;;EAIA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA;;;EAGA;EACA;EACA;;EAEA;;EAEA;;;iBAYc,eAAe,MAAM,wBAAwB;;;;;UCzB5C;;;EAGf;;;;;EAKA;;EAEA;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;;EAEA;;;UAIe;EACf;;;;EAIA;EACA,UAAU;EACV,WAAW;;UAGI;;EAEf,OAAO;;;EAGP,kBAAkB;;EAElB,mBAAmB;;;;;;;;;;;;;;;;;;;;iBAqBL,SAAS,eAAe,gBAAgB,MAAM,kBAAkB;;UA8G/D;;EAEf;;EAEA;;EAEA,SAAS;;EAET,gBAAgB;;;UAID;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,aAAa;;EAEb;IAAY;IAAW;;;UAGR,iCAAiC;;;;EAIhD;;EAEA,YAAY;;UAGG;EACf;EACA;EACA;;;EAGA,aAAa;EACb,cAAc;;;;;;;;;;;;;;;iBAgBA,kBACd,eAAe,gBACf,MAAM,2BACL;UAmEc;EACf;EACA;EACA,UAAU;EACV,WAAW;;UAGI;EACf,OAAO;EACP,kBAAkB;EAClB,mBAAmB;;;;;;;;;iBAeL,eACd,uBAAuB,aACvB,wBAAwB,cACvB;;;;;;;;;;;;;;;;UCrWc;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;;;KAIU;UAEK,uBAAuB;;EAEtC;;EAEA,MAAM;;UAGS;EACf,UAAU;EACV;EACA;EACA;;;EAGA;;EAEA,kBAAkB;EAGlB;;;;;;;;;;;;;;;;;iBAkBoB,SAAS,GAAG,KAAK,IAAI;;;;;iBAQ3B,sBAAsB,UAAU;;;;;;iBAc1B,aAAa,GAAG,qBAAqB,QAAQ;;;;;iBAS7C,eAAe,GAAG,iBAAiB;;;;;iBAYnC,mBACpB,UAAU,gBACV;EAAY;EAAW;EAAgB;IACtC,QAAQ;;;KCrFC;UAEK;EACf,UAAU;;EAEV;;EAEA;;;EAGA;;UAGe;;;EAGf;;;;EAIA;;;EAGA;;EAEA;;;;EAIA;;;EAGA;;;;EAIA,8BAA8B;;;EAG9B,kBAAkB;;EAElB;;;;;;;EAOA,SAAS;;;;;;KAOC,wBAAwB;EAAkB,UAAU;;UAE/C,qBAAqB,qBAAqB,kBAAkB,WAAW,kBAC9E,KAAK,WAAW;;;;;;;;;;EAUxB,QAAQ,gBAAgB;;;EAGxB,SAAS;;;;;;;;;;;;;;;;;iBAuPK,qBAAqB,qBAAqB,kBAAkB,WAAW,UACrF,SAAS,8BACR,qBAAqB,WAAW;UA6GlB;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;GACd;IAAQ,SAAS;;IAAyB;IAAe;;;EAE1D,SAAS;;;;;;;;;;;;;;;;;;;;;;;;;iBA0BK,iBAAiB,UAAS,0BAA+B;;;UC9dxD;;EAEf;;EAEA;;EAEA;;;;;;;;;;;;;iBA4Ec,YACd,WAAW,YAAY,eAAe,cACtC,UAAU,YAAY,eAAe,cACrC,aAAa,aACb,SAAS,GAAG,oCACX;UAkEc;EACf,QAAQ;;;;;;;;;;;;EAYR,WAAW;;;;EAIX,iBAAiB;;;;;;;EAOjB,UAAU;;EAEV,mBAAmB;;EAEnB,SAAS;;;;EAIT;;EAEA;;EAEA;EACA;;EAEA;;EAEA;;EAEA,gBAAgB;;EAEhB;;;;EAIA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;;EAEA;EACA;;EAEA,8BAA8B;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA6BhB,oBACd,QAAQ,eACR,OAAM,6BACL;UA2Ec;EACf;;;EAGA,WAAW;;;;EAIX;;EAEA;IAAM;IAAa;;;EAEnB,mBAAmB;;EAEnB,SAAS;;EAET;;;;EAIA;EACA;EACA;EACA;EACA;;EAEA;EACA;;EAEA;;EAEA;;;;;iBAMc,YAAY;;;;;;;;;iBAYZ,qBACd,WAAW,YAAY,eAAe,cACtC,UAAU,YAAY,eAAe,cACrC,aAAa,aACb,8BACA;EACE;EACA;EACA;EACA;;;EAGA;EACA,8BAA8B;EAC9B;IAED"}
@@ -0,0 +1,256 @@
1
+ import { r as pairedBootstrap } from "./paired-tests-C8iCsioC.js";
2
+ import { t as decidePairedPromotion } from "./paired-promotion-decision-ClJh_9nK.js";
3
+ //#region src/campaign/gates/statistical-heldout.ts
4
+ /**
5
+ * Held-out inference pairs execution cells and judge identities before taking
6
+ * means within registered independent units. Repetitions can improve a unit's
7
+ * precision without increasing n. Ungrouped inference concerns independently
8
+ * sampled execution cells conditional on a fixed scenario roster.
9
+ *
10
+ * The shared paired decision rule selects the estimator and statistical test.
11
+ * Dimension reports retain missing coverage so required safety checks cannot
12
+ * pass through absent evidence. Thresholds use the judge's native score scale.
13
+ */
14
+ /** Tie fraction at/above which a gate annotates its verdict with the tie share.
15
+ * Tie-domination of the median bites structurally at >= 0.5 (the median is then
16
+ * 0 by construction); 0.4 is a softer warn threshold that flags a run APPROACHING
17
+ * that regime, so an operator sees it before the median goes fully blind. */
18
+ const TIE_WARN_FRACTION = .4;
19
+ /** Campaign cell IDs append a numeric repetition after the scenario's full ID. */
20
+ function scenarioIdFromCellId(cellId) {
21
+ const separator = cellId.lastIndexOf(":");
22
+ const repetition = cellId.slice(separator + 1);
23
+ if (separator < 1 || cellId.trim() !== cellId || !/^(0|[1-9]\d*)$/.test(repetition) || !Number.isSafeInteger(Number(repetition))) throw new Error(`pairHoldout: malformed cellId '${cellId}'; expected scenarioId:rep`);
24
+ return cellId.slice(0, separator);
25
+ }
26
+ /** Preserve cell pairing before taking equal-weight independent-unit means. */
27
+ function aggregatePairedHoldout(paired, independentUnitByScenarioId) {
28
+ if (paired.before.length !== paired.after.length || paired.before.length !== paired.cellIds.length) throw new Error("aggregatePairedHoldout: scores and cellIds must have the same length");
29
+ if (new Set(paired.cellIds).size !== paired.cellIds.length) throw new Error("aggregatePairedHoldout: duplicate cellIds cannot count as new observations");
30
+ if (paired.before.some((value) => !Number.isFinite(value)) || paired.after.some((value) => !Number.isFinite(value))) throw new Error("aggregatePairedHoldout: paired scores must be finite");
31
+ if (independentUnitByScenarioId === void 0) return {
32
+ before: [...paired.before],
33
+ after: [...paired.after],
34
+ unitIds: [...paired.cellIds]
35
+ };
36
+ const scenarioIds = paired.cellIds.map(scenarioIdFromCellId);
37
+ const groups = /* @__PURE__ */ new Map();
38
+ for (let i = 0; i < paired.cellIds.length; i++) {
39
+ const scenarioId = scenarioIds[i];
40
+ const unitId = independentUnitByScenarioId.get(scenarioId);
41
+ if (typeof unitId !== "string" || unitId.length === 0 || unitId.trim() !== unitId) throw new Error(`aggregatePairedHoldout: missing independent unit for scenario '${scenarioId}'`);
42
+ const group = groups.get(unitId) ?? {
43
+ before: 0,
44
+ after: 0,
45
+ n: 0
46
+ };
47
+ group.before += paired.before[i];
48
+ group.after += paired.after[i];
49
+ group.n += 1;
50
+ groups.set(unitId, group);
51
+ }
52
+ const unitIds = [...groups.keys()].sort();
53
+ return {
54
+ before: unitIds.map((id) => groups.get(id).before / groups.get(id).n),
55
+ after: unitIds.map((id) => groups.get(id).after / groups.get(id).n),
56
+ unitIds
57
+ };
58
+ }
59
+ /**
60
+ * Pair candidate vs baseline holdout observations by FULL cellId. `select`
61
+ * pulls the scalar from a cell's judge reports (composite, or a named
62
+ * dimension); a cell contributes the mean of `select` across its judges. Cells
63
+ * whose scenario is not in `scenarioIds`, or where `select` is undefined for
64
+ * every judge on both sides, are skipped. The selected judge IDs must agree
65
+ * within each pair. Throws when the two maps disagree on holdout cell IDs — a
66
+ * load-bearing invariant: the baseline + winner holdout campaigns run the same
67
+ * scenarios with the same seed base, so their cellIds MUST align; a mismatch
68
+ * means a silent pairing bug, not a soft fallback.
69
+ */
70
+ function pairHoldout(candidate, baseline, scenarioIds, select) {
71
+ const cellValues = (byCell, cellId) => {
72
+ const scores = byCell.get(cellId);
73
+ const values = /* @__PURE__ */ new Map();
74
+ if (!scores) return values;
75
+ for (const [judgeId, s] of Object.entries(scores)) {
76
+ if (s.failed === true) throw new Error(`pairHoldout: cell '${cellId}' contains a failed judge score`);
77
+ const v = select(s);
78
+ if (typeof v === "number" && !Number.isFinite(v)) throw new Error(`pairHoldout: cell '${cellId}' contains a non-finite selected score`);
79
+ if (typeof v === "number") values.set(judgeId, v);
80
+ }
81
+ return values;
82
+ };
83
+ const inScope = (cellId) => scenarioIds.has(scenarioIdFromCellId(cellId));
84
+ const candCells = [...candidate.keys()].filter(inScope).sort();
85
+ const baseCells = [...baseline.keys()].filter(inScope).sort();
86
+ if (candCells.length !== baseCells.length || candCells.some((c, i) => c !== baseCells[i])) throw new Error(`pairHoldout: candidate/baseline holdout cells do not align — candidate=[${candCells.join(",")}] baseline=[${baseCells.join(",")}]. Both holdout campaigns must run the same scenarios with the same seed base.`);
87
+ const before = [];
88
+ const after = [];
89
+ const cellIds = [];
90
+ for (const cellId of candCells) {
91
+ const b = cellValues(baseline, cellId);
92
+ const a = cellValues(candidate, cellId);
93
+ if (b.size === 0 && a.size === 0) continue;
94
+ if (b.size === 0 || a.size === 0) throw new Error(`pairHoldout: cell '${cellId}' has a selected score on only one arm`);
95
+ if (b.size !== a.size || [...b.keys()].some((id) => !a.has(id))) throw new Error(`pairHoldout: cell '${cellId}' selected judge IDs do not align`);
96
+ const judgeIds = [...b.keys()].sort();
97
+ before.push(meanSelectedScores(judgeIds.map((id) => b.get(id))));
98
+ after.push(meanSelectedScores(judgeIds.map((id) => a.get(id))));
99
+ cellIds.push(cellId);
100
+ }
101
+ return {
102
+ before,
103
+ after,
104
+ cellIds
105
+ };
106
+ }
107
+ function meanSelectedScores(values) {
108
+ const first = values[0];
109
+ return values.every((value) => value === first) ? first : values.reduce((sum, value) => sum + value, 0) / values.length;
110
+ }
111
+ /**
112
+ * Significance of the held-out composite lift: ship only when the lower bound
113
+ * of the interval the outcome's shape ADMITS exceeds `deltaThreshold` (default
114
+ * 0 ⇒ "confidently positive"). Interpret `deltaThreshold` in the judge's native
115
+ * scale.
116
+ *
117
+ * The decision is delegated whole to {@link decidePairedPromotion}, the one
118
+ * copy of the rule (`src/paired-promotion-decision.ts`), which `HeldOutGate`
119
+ * also calls. That module's header carries the measurements; the short version
120
+ * is three guards a bare `bootstrap.low > threshold` does not have:
121
+ *
122
+ * - a two-point (pass/fail) outcome decides on Tango's SCORE interval, the
123
+ * only paired-binary construction that stays valid at a nonzero margin;
124
+ * - McNemar's exact test VETOES at any non-negative threshold;
125
+ * - a ZERO-WIDTH interval is refused rather than promoted, in either
126
+ * direction — [0,0] clears every negative threshold and [g,g] clears every
127
+ * threshold below g, and both are an absence of evidence, not a result.
128
+ *
129
+ * Measured on this function before those guards landed, at a nominal 5 %:
130
+ * 14.60 % false promotion at n = 40 on a paired-binary noninferiority boundary,
131
+ * and 88.50 % at n = 6 under a bounded asymmetric null whose true mean paired
132
+ * delta is exactly 0.
133
+ *
134
+ * Continuous mean targets require bootstrap eligibility. Explicit median
135
+ * targets can use the exact sign test at its confidence-dependent minimum.
136
+ */
137
+ function heldoutSignificance(paired, opts = {}) {
138
+ const deltaThreshold = opts.deltaThreshold ?? 0;
139
+ const confidence = opts.confidence ?? .95;
140
+ const resamples = opts.resamples ?? 2e3;
141
+ const seed = opts.seed ?? 1337;
142
+ const statistic = opts.statistic ?? "mean";
143
+ const observations = aggregatePairedHoldout(paired, opts.independentUnitByScenarioId);
144
+ const decision = decidePairedPromotion(observations.before, observations.after, {
145
+ confidence,
146
+ resamples,
147
+ statistic,
148
+ seed,
149
+ threshold: deltaThreshold,
150
+ minPairs: opts.minProductiveRuns
151
+ });
152
+ const bootstrap = decision.bootstrap ?? pairedBootstrap(observations.before, observations.after, {
153
+ confidence,
154
+ resamples,
155
+ statistic,
156
+ seed
157
+ });
158
+ const medianBootstrap = statistic === "median" ? bootstrap : pairedBootstrap(observations.before, observations.after, {
159
+ confidence,
160
+ resamples,
161
+ statistic: "median",
162
+ seed
163
+ });
164
+ const n = observations.before.length;
165
+ let ties = 0;
166
+ for (let i = 0; i < n; i += 1) {
167
+ const after = observations.after[i];
168
+ const before = observations.before[i];
169
+ if (Math.abs(after - before) < 1e-9) ties += 1;
170
+ }
171
+ const tieFraction = n === 0 ? 0 : ties / n;
172
+ return {
173
+ paired,
174
+ bootstrap,
175
+ medianBootstrap,
176
+ decision,
177
+ decisionStatistic: decision.statistic,
178
+ mcnemar: decision.mcnemar,
179
+ tieFraction,
180
+ n,
181
+ pairedCellN: paired.cellIds.length,
182
+ observationUnit: opts.independentUnitByScenarioId === void 0 ? "cell" : "registered",
183
+ unitIds: observations.unitIds,
184
+ minimumRequired: decision.minimumPairs,
185
+ decisionMethod: decision.method,
186
+ pValue: decision.pValue,
187
+ significant: decision.promote,
188
+ fewRuns: !decision.sufficient
189
+ };
190
+ }
191
+ /** Detect the native scale of a set of scores: 0-100 when any magnitude clears
192
+ * 1.5, else [0,1]. Used to auto-scale the regression tolerance so a default
193
+ * expressed for [0,1] is not silently a no-op on a 0-100 dimension. */
194
+ function detectScale(values) {
195
+ return values.some((v) => Math.abs(v) > 1.5) ? 100 : 1;
196
+ }
197
+ /**
198
+ * Report required-dimension evidence after full pairing and optional unit means.
199
+ * A bootstrap floor breach or a shared paired test supporting a drop marks
200
+ * regression. These two criteria are distinct; `ci` records the shared
201
+ * estimator and `bootstrap` records the floor interval. Missing observations
202
+ * and insufficient n remain explicit for the caller's evidence policy.
203
+ * The default tolerance is 0.05 on [0,1] and 5 on a detected 0-100 scale.
204
+ */
205
+ function dimensionRegressions(candidate, baseline, scenarioIds, criticalDimensions, opts = {}) {
206
+ const out = [];
207
+ const expectedCellIds = [...baseline.keys()].filter((cellId) => scenarioIds.has(scenarioIdFromCellId(cellId))).sort();
208
+ for (const dim of criticalDimensions) {
209
+ const paired = pairHoldout(candidate, baseline, scenarioIds, (s) => s.dimensions[dim]);
210
+ if (paired.before.length === 0) continue;
211
+ const observations = aggregatePairedHoldout(paired, opts.independentUnitByScenarioId);
212
+ const measuredCells = new Set(paired.cellIds);
213
+ const measuredScenarios = new Set(paired.cellIds.map(scenarioIdFromCellId));
214
+ const tolerance = opts.tolerance ?? .05 * detectScale([...paired.before, ...paired.after]);
215
+ const bootstrapStatistic = opts.statistic ?? "mean";
216
+ const shared = {
217
+ confidence: opts.confidence ?? .95,
218
+ resamples: opts.resamples ?? 2e3,
219
+ statistic: bootstrapStatistic,
220
+ seed: opts.seed ?? 1337,
221
+ minPairs: opts.minProductiveRuns
222
+ };
223
+ const guard = decidePairedPromotion(observations.before, observations.after, shared);
224
+ const regression = decidePairedPromotion(observations.after, observations.before, {
225
+ ...shared,
226
+ threshold: tolerance
227
+ });
228
+ const bootstrap = guard.bootstrap ?? pairedBootstrap(observations.before, observations.after, shared);
229
+ out.push({
230
+ dimension: dim,
231
+ bootstrap,
232
+ bootstrapStatistic,
233
+ ci: {
234
+ low: guard.low,
235
+ high: guard.high
236
+ },
237
+ decisionStatistic: guard.statistic,
238
+ mcnemar: guard.mcnemar,
239
+ indeterminate: guard.indeterminate,
240
+ regressed: bootstrap.low < -tolerance || regression.promote,
241
+ tolerance,
242
+ n: observations.before.length,
243
+ pairedCellN: paired.cellIds.length,
244
+ observationUnit: opts.independentUnitByScenarioId === void 0 ? "cell" : "registered",
245
+ minimumRequired: guard.minimumPairs,
246
+ fewRuns: !guard.sufficient,
247
+ missingCellIds: expectedCellIds.filter((cellId) => !measuredCells.has(cellId)),
248
+ missingScenarioIds: [...scenarioIds].filter((id) => !measuredScenarios.has(id)).sort()
249
+ });
250
+ }
251
+ return out;
252
+ }
253
+ //#endregion
254
+ export { heldoutSignificance as a, dimensionRegressions as i, aggregatePairedHoldout as n, pairHoldout as o, detectScale as r, TIE_WARN_FRACTION as t };
255
+
256
+ //# sourceMappingURL=statistical-heldout-CeWT_W7y.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"statistical-heldout-CeWT_W7y.js","names":[],"sources":["../src/campaign/gates/statistical-heldout.ts"],"sourcesContent":["/**\n * Held-out inference pairs execution cells and judge identities before taking\n * means within registered independent units. Repetitions can improve a unit's\n * precision without increasing n. Ungrouped inference concerns independently\n * sampled execution cells conditional on a fixed scenario roster.\n *\n * The shared paired decision rule selects the estimator and statistical test.\n * Dimension reports retain missing coverage so required safety checks cannot\n * pass through absent evidence. Thresholds use the judge's native score scale.\n */\n\nimport {\n decidePairedPromotion,\n type PairedDecisionMethod,\n type PairedDecisionStatistic,\n type PairedMcNemarEvidence,\n type PairedPromotionDecision,\n} from '../../paired-promotion-decision'\nimport {\n DECISION_PAIRED_DELTA_STATISTIC,\n type PairedBootstrapResult,\n pairedBootstrap,\n} from '../../statistics'\nimport type { JudgeScore } from '../types'\n\n/** Tie fraction at/above which a gate annotates its verdict with the tie share.\n * Tie-domination of the median bites structurally at >= 0.5 (the median is then\n * 0 by construction); 0.4 is a softer warn threshold that flags a run APPROACHING\n * that regime, so an operator sees it before the median goes fully blind. */\nexport const TIE_WARN_FRACTION = 0.4\n\nexport interface PairedHoldout {\n /** Baseline scalar per paired cell (same order as `after`/`cellIds`). */\n before: number[]\n /** Candidate scalar per paired cell. */\n after: number[]\n /** The full cellIds (`scenario:rep`) that paired, in order. */\n cellIds: string[]\n}\n\n/** Campaign cell IDs append a numeric repetition after the scenario's full ID. */\nexport function scenarioIdFromCellId(cellId: string): string {\n const separator = cellId.lastIndexOf(':')\n const repetition = cellId.slice(separator + 1)\n if (\n separator < 1 ||\n cellId.trim() !== cellId ||\n !/^(0|[1-9]\\d*)$/.test(repetition) ||\n !Number.isSafeInteger(Number(repetition))\n ) {\n throw new Error(`pairHoldout: malformed cellId '${cellId}'; expected scenarioId:rep`)\n }\n return cellId.slice(0, separator)\n}\n\n/** Preserve cell pairing before taking equal-weight independent-unit means. */\nexport function aggregatePairedHoldout(\n paired: PairedHoldout,\n independentUnitByScenarioId?: ReadonlyMap<string, string>,\n): { before: number[]; after: number[]; unitIds: string[] } {\n if (\n paired.before.length !== paired.after.length ||\n paired.before.length !== paired.cellIds.length\n ) {\n throw new Error('aggregatePairedHoldout: scores and cellIds must have the same length')\n }\n if (new Set(paired.cellIds).size !== paired.cellIds.length) {\n throw new Error('aggregatePairedHoldout: duplicate cellIds cannot count as new observations')\n }\n if (\n paired.before.some((value) => !Number.isFinite(value)) ||\n paired.after.some((value) => !Number.isFinite(value))\n ) {\n throw new Error('aggregatePairedHoldout: paired scores must be finite')\n }\n if (independentUnitByScenarioId === undefined) {\n return { before: [...paired.before], after: [...paired.after], unitIds: [...paired.cellIds] }\n }\n const scenarioIds = paired.cellIds.map(scenarioIdFromCellId)\n const groups = new Map<string, { before: number; after: number; n: number }>()\n for (let i = 0; i < paired.cellIds.length; i++) {\n const scenarioId = scenarioIds[i]!\n const unitId = independentUnitByScenarioId.get(scenarioId)\n if (typeof unitId !== 'string' || unitId.length === 0 || unitId.trim() !== unitId) {\n throw new Error(\n `aggregatePairedHoldout: missing independent unit for scenario '${scenarioId}'`,\n )\n }\n const group = groups.get(unitId) ?? { before: 0, after: 0, n: 0 }\n group.before += paired.before[i]!\n group.after += paired.after[i]!\n group.n += 1\n groups.set(unitId, group)\n }\n const unitIds = [...groups.keys()].sort()\n return {\n before: unitIds.map((id) => groups.get(id)!.before / groups.get(id)!.n),\n after: unitIds.map((id) => groups.get(id)!.after / groups.get(id)!.n),\n unitIds,\n }\n}\n\n/**\n * Pair candidate vs baseline holdout observations by FULL cellId. `select`\n * pulls the scalar from a cell's judge reports (composite, or a named\n * dimension); a cell contributes the mean of `select` across its judges. Cells\n * whose scenario is not in `scenarioIds`, or where `select` is undefined for\n * every judge on both sides, are skipped. The selected judge IDs must agree\n * within each pair. Throws when the two maps disagree on holdout cell IDs — a\n * load-bearing invariant: the baseline + winner holdout campaigns run the same\n * scenarios with the same seed base, so their cellIds MUST align; a mismatch\n * means a silent pairing bug, not a soft fallback.\n */\nexport function pairHoldout(\n candidate: Map<string, Record<string, JudgeScore>>,\n baseline: Map<string, Record<string, JudgeScore>>,\n scenarioIds: Set<string>,\n select: (s: JudgeScore) => number | undefined,\n): PairedHoldout {\n const cellValues = (\n byCell: Map<string, Record<string, JudgeScore>>,\n cellId: string,\n ): Map<string, number> => {\n const scores = byCell.get(cellId)\n const values = new Map<string, number>()\n if (!scores) return values\n for (const [judgeId, s] of Object.entries(scores)) {\n if (s.failed === true) {\n throw new Error(`pairHoldout: cell '${cellId}' contains a failed judge score`)\n }\n const v = select(s)\n if (typeof v === 'number' && !Number.isFinite(v)) {\n throw new Error(`pairHoldout: cell '${cellId}' contains a non-finite selected score`)\n }\n if (typeof v === 'number') values.set(judgeId, v)\n }\n return values\n }\n\n const inScope = (cellId: string) => scenarioIds.has(scenarioIdFromCellId(cellId))\n const candCells = [...candidate.keys()].filter(inScope).sort()\n const baseCells = [...baseline.keys()].filter(inScope).sort()\n // Alignment invariant — the holdout campaigns share scenarios + seed, so the\n // cell sets must be identical. Differ ⇒ a real pairing bug; fail loud.\n if (candCells.length !== baseCells.length || candCells.some((c, i) => c !== baseCells[i])) {\n throw new Error(\n `pairHoldout: candidate/baseline holdout cells do not align — ` +\n `candidate=[${candCells.join(',')}] baseline=[${baseCells.join(',')}]. ` +\n `Both holdout campaigns must run the same scenarios with the same seed base.`,\n )\n }\n\n const before: number[] = []\n const after: number[] = []\n const cellIds: string[] = []\n for (const cellId of candCells) {\n const b = cellValues(baseline, cellId)\n const a = cellValues(candidate, cellId)\n // A scalar absent on both sides means that dimension was not scored. A\n // one-sided absence is asymmetric evidence loss, never a row to discard.\n if (b.size === 0 && a.size === 0) continue\n if (b.size === 0 || a.size === 0) {\n throw new Error(`pairHoldout: cell '${cellId}' has a selected score on only one arm`)\n }\n if (b.size !== a.size || [...b.keys()].some((id) => !a.has(id))) {\n throw new Error(`pairHoldout: cell '${cellId}' selected judge IDs do not align`)\n }\n const judgeIds = [...b.keys()].sort()\n before.push(meanSelectedScores(judgeIds.map((id) => b.get(id)!)))\n after.push(meanSelectedScores(judgeIds.map((id) => a.get(id)!)))\n cellIds.push(cellId)\n }\n return { before, after, cellIds }\n}\n\nfunction meanSelectedScores(values: number[]): number {\n const first = values[0]!\n // Summing identical fractional scores can round their mean outside the\n // declared binary support. An agreeing judge set preserves its exact value.\n return values.every((value) => value === first)\n ? first\n : values.reduce((sum, value) => sum + value, 0) / values.length\n}\n\nexport interface HeldoutSignificance {\n paired: PairedHoldout\n /**\n * The paired bootstrap on the requested statistic (MEAN by default — see the\n * tie note on `heldoutSignificance`).\n *\n * DIAGNOSTIC, not necessarily the interval the verdict keyed on. On a\n * two-point (pass/fail) outcome the decision routes to Tango's score interval\n * instead, because a percentile bootstrap of the mean over a three-atom\n * lattice is not a valid interval at a nonzero margin. Read\n * `decision.low`/`decision.high` for the interval that actually decided, and\n * `decisionStatistic` for which one it is.\n */\n bootstrap: PairedBootstrapResult\n /** The MEDIAN paired-delta bootstrap, reported as a diagnostic. When many\n * scenarios are tied (both sides solve them), the median is pinned near 0\n * regardless of the mean lift — comparing the two exposes tie-domination. */\n medianBootstrap: PairedBootstrapResult\n /**\n * The full promotion decision: which estimator the outcome's shape admits,\n * the interval it produced, McNemar's exact veto on the two-point path, and\n * whether the interval was zero-width (no evidence in either direction). The\n * single source of `significant`.\n */\n decision: PairedPromotionDecision\n /** Which paired estimator the verdict was decided on. */\n decisionStatistic: PairedDecisionStatistic\n /** McNemar's exact evidence on the two-point path; null otherwise. */\n mcnemar: PairedMcNemarEvidence | null\n /** Fraction of paired observations that are exact ties (|delta| < 1e-9). A\n * high tie fraction is WHY a median-based gate would have missed a real lift;\n * it is the observability the tie fix adds. */\n tieFraction: number\n /** Number of paired observation units, after configured aggregation. */\n n: number\n /** Original matched execution cells, before aggregation. */\n pairedCellN: number\n observationUnit: 'registered' | 'cell'\n /** Registered unit IDs, or cell IDs on the ungrouped path. */\n unitIds: string[]\n /** Effective minimum for the requested target and chosen estimator. */\n minimumRequired: number\n /** Statistical method that carried the decision. */\n decisionMethod: PairedDecisionMethod\n /** Exact one-sided p-value on the small-sample path; otherwise null. */\n pValue: number | null\n /** True iff n >= minimumRequired, the DECIDING interval has nonzero width,\n * its lower bound clears the threshold, and McNemar's exact test does not\n * veto at a non-negative threshold. */\n significant: boolean\n /** Set when n < minimumRequired — too little evidence to claim significance. */\n fewRuns: boolean\n}\n\nexport interface HeldoutSignificanceOptions {\n deltaThreshold?: number\n minProductiveRuns?: number\n confidence?: number\n resamples?: number\n /** Fixed by default for a deterministic, reproducible gate verdict. */\n seed?: number\n statistic?: 'mean' | 'median'\n /** Group full cell pairs into equal-weight independent units before inference. */\n independentUnitByScenarioId?: ReadonlyMap<string, string>\n}\n\n/**\n * Significance of the held-out composite lift: ship only when the lower bound\n * of the interval the outcome's shape ADMITS exceeds `deltaThreshold` (default\n * 0 ⇒ \"confidently positive\"). Interpret `deltaThreshold` in the judge's native\n * scale.\n *\n * The decision is delegated whole to {@link decidePairedPromotion}, the one\n * copy of the rule (`src/paired-promotion-decision.ts`), which `HeldOutGate`\n * also calls. That module's header carries the measurements; the short version\n * is three guards a bare `bootstrap.low > threshold` does not have:\n *\n * - a two-point (pass/fail) outcome decides on Tango's SCORE interval, the\n * only paired-binary construction that stays valid at a nonzero margin;\n * - McNemar's exact test VETOES at any non-negative threshold;\n * - a ZERO-WIDTH interval is refused rather than promoted, in either\n * direction — [0,0] clears every negative threshold and [g,g] clears every\n * threshold below g, and both are an absence of evidence, not a result.\n *\n * Measured on this function before those guards landed, at a nominal 5 %:\n * 14.60 % false promotion at n = 40 on a paired-binary noninferiority boundary,\n * and 88.50 % at n = 6 under a bounded asymmetric null whose true mean paired\n * delta is exactly 0.\n *\n * Continuous mean targets require bootstrap eligibility. Explicit median\n * targets can use the exact sign test at its confidence-dependent minimum.\n */\nexport function heldoutSignificance(\n paired: PairedHoldout,\n opts: HeldoutSignificanceOptions = {},\n): HeldoutSignificance {\n const deltaThreshold = opts.deltaThreshold ?? 0\n const confidence = opts.confidence ?? 0.95\n const resamples = opts.resamples ?? 2000\n const seed = opts.seed ?? 1337\n // DEFAULT to the MEAN paired delta, not the median. The median is destroyed by\n // TIES: whenever both baseline and candidate solve a holdout scenario (a common\n // case once the agent is decent — and INCREASINGLY common as you add holdout\n // scenarios for statistical power), that scenario contributes delta 0. When\n // >=50% of paired cells are ties the median is pinned at 0 regardless of a large,\n // consistent lift on the rest, and the gate holds a genuinely better candidate.\n // (Measured live, supervisor-lab run 6: 40 cells, 20 ties, MEAN +0.177, MEDIAN 0\n // → false hold. Doubling the holdout for \"power\" made it WORSE by adding ties.)\n // The mean equals the reported aggregate lift, ties correctly contribute 0\n // without dominating, and it is the textbook paired-comparison estimator; the\n // median is kept as a reported diagnostic. Callers wanting outlier-robustness at\n // the cost of tie-blindness can still pass `statistic: 'median'`.\n const statistic = opts.statistic ?? 'mean'\n const observations = aggregatePairedHoldout(paired, opts.independentUnitByScenarioId)\n const decision = decidePairedPromotion(observations.before, observations.after, {\n confidence,\n resamples,\n statistic,\n seed,\n threshold: deltaThreshold,\n minPairs: opts.minProductiveRuns,\n })\n // `decision.bootstrap` is null exactly when the score interval decided, so\n // the requested-statistic bootstrap is computed here for the diagnostic\n // field. Same two bootstraps as before on every path.\n const bootstrap =\n decision.bootstrap ??\n pairedBootstrap(observations.before, observations.after, {\n confidence,\n resamples,\n statistic,\n seed,\n })\n const medianBootstrap =\n statistic === 'median'\n ? bootstrap\n : pairedBootstrap(observations.before, observations.after, {\n confidence,\n resamples,\n statistic: 'median',\n seed,\n })\n const n = observations.before.length\n let ties = 0\n for (let i = 0; i < n; i += 1) {\n const after = observations.after[i]!\n const before = observations.before[i]!\n if (Math.abs(after - before) < 1e-9) ties += 1\n }\n const tieFraction = n === 0 ? 0 : ties / n\n return {\n paired,\n bootstrap,\n medianBootstrap,\n decision,\n decisionStatistic: decision.statistic,\n mcnemar: decision.mcnemar,\n tieFraction,\n n,\n pairedCellN: paired.cellIds.length,\n observationUnit: opts.independentUnitByScenarioId === undefined ? 'cell' : 'registered',\n unitIds: observations.unitIds,\n minimumRequired: decision.minimumPairs,\n decisionMethod: decision.method,\n pValue: decision.pValue,\n significant: decision.promote,\n fewRuns: !decision.sufficient,\n }\n}\n\nexport interface DimensionRegression {\n dimension: string\n /** Paired bootstrap on (candidate − baseline). DIAGNOSTIC on a pass/fail\n * dimension, where `ci` carries the interval that decided instead. */\n bootstrap: PairedBootstrapResult\n /** Which paired statistic `bootstrap.low` is the lower bound of. `'mean'`\n * unless the caller asked for the median. `bootstrap.median` still carries\n * the median point estimate either way. */\n bootstrapStatistic: 'median' | 'mean'\n /** The interval `regressed` was decided on, in the dimension's native units. */\n ci: { low: number; high: number }\n /** Which estimator produced `ci`. */\n decisionStatistic: PairedDecisionStatistic\n /** McNemar's exact evidence on a pass/fail dimension; null otherwise. */\n mcnemar: PairedMcNemarEvidence | null\n /** `ci` has zero width — no evidence in either direction. */\n indeterminate: boolean\n /** The bootstrap lower bound is below negative tolerance, or the shared\n * paired test supports a drop exceeding tolerance. Missing coverage and\n * insufficient observations are reported separately. */\n regressed: boolean\n tolerance: number\n n: number\n pairedCellN: number\n observationUnit: 'registered' | 'cell'\n /** Statistical minimum for the configured independent observation unit. */\n minimumRequired: number\n fewRuns: boolean\n /** Both arms lack this dimension on these otherwise matched execution cells. */\n missingCellIds: string[]\n /** Configured scenarios with no paired dimension measurement at all. */\n missingScenarioIds: string[]\n}\n\n/** Detect the native scale of a set of scores: 0-100 when any magnitude clears\n * 1.5, else [0,1]. Used to auto-scale the regression tolerance so a default\n * expressed for [0,1] is not silently a no-op on a 0-100 dimension. */\nexport function detectScale(values: number[]): 1 | 100 {\n return values.some((v) => Math.abs(v) > 1.5) ? 100 : 1\n}\n\n/**\n * Report required-dimension evidence after full pairing and optional unit means.\n * A bootstrap floor breach or a shared paired test supporting a drop marks\n * regression. These two criteria are distinct; `ci` records the shared\n * estimator and `bootstrap` records the floor interval. Missing observations\n * and insufficient n remain explicit for the caller's evidence policy.\n * The default tolerance is 0.05 on [0,1] and 5 on a detected 0-100 scale.\n */\nexport function dimensionRegressions(\n candidate: Map<string, Record<string, JudgeScore>>,\n baseline: Map<string, Record<string, JudgeScore>>,\n scenarioIds: Set<string>,\n criticalDimensions: string[],\n opts: {\n tolerance?: number\n confidence?: number\n resamples?: number\n seed?: number\n /** Paired statistic the CI is computed on. Default `'mean'` — see\n * {@link DECISION_PAIRED_DELTA_STATISTIC} for why the median is not. */\n statistic?: 'mean' | 'median'\n independentUnitByScenarioId?: ReadonlyMap<string, string>\n minProductiveRuns?: number\n } = {},\n): DimensionRegression[] {\n const out: DimensionRegression[] = []\n const expectedCellIds = [...baseline.keys()]\n .filter((cellId) => scenarioIds.has(scenarioIdFromCellId(cellId)))\n .sort()\n for (const dim of criticalDimensions) {\n const paired = pairHoldout(candidate, baseline, scenarioIds, (s) => s.dimensions[dim])\n if (paired.before.length === 0) continue // dimension not scored on this judge\n const observations = aggregatePairedHoldout(paired, opts.independentUnitByScenarioId)\n const measuredCells = new Set(paired.cellIds)\n const measuredScenarios = new Set(paired.cellIds.map(scenarioIdFromCellId))\n const tolerance = opts.tolerance ?? 0.05 * detectScale([...paired.before, ...paired.after])\n const bootstrapStatistic = opts.statistic ?? DECISION_PAIRED_DELTA_STATISTIC\n const shared = {\n confidence: opts.confidence ?? 0.95,\n resamples: opts.resamples ?? 2000,\n statistic: bootstrapStatistic,\n seed: opts.seed ?? 1337,\n minPairs: opts.minProductiveRuns,\n }\n const guard = decidePairedPromotion(observations.before, observations.after, shared)\n const regression = decidePairedPromotion(observations.after, observations.before, {\n ...shared,\n threshold: tolerance,\n })\n const bootstrap =\n guard.bootstrap ?? pairedBootstrap(observations.before, observations.after, shared)\n out.push({\n dimension: dim,\n bootstrap,\n bootstrapStatistic,\n ci: { low: guard.low, high: guard.high },\n decisionStatistic: guard.statistic,\n mcnemar: guard.mcnemar,\n indeterminate: guard.indeterminate,\n // Fires on EITHER burden of proof — see `floorBreached` in\n // `promotion-policy.ts` for the same rule and the reasoning. The CI arm\n // is the contract this interface and `default-production-gate`'s reason\n // string both state (\"CI.low < -tolerance\"), read off the DECIDING\n // interval; the exact-test arm adds the small-sample path where the\n // bootstrap interval is descriptive only.\n // The credible-worst-case arm stays on the BOOTSTRAP — see the same\n // decision in `floorBreached` (`promotion-policy.ts`). On a pass/fail\n // dimension the score interval is ±z²/(n+z²) even when every pair is\n // concordant, so reading the floor off it would flag an unchanged safety\n // dimension as regressed at any realistic n. The PROVEN-drop arm does use\n // the shared rule, which is what closes the fail-open hole that mattered:\n // a genuine pass/fail regression is now judged on an interval valid at\n // the nonzero `tolerance`, instead of a bootstrap that is pinned wherever\n // ties dominate.\n regressed: bootstrap.low < -tolerance || regression.promote,\n tolerance,\n n: observations.before.length,\n pairedCellN: paired.cellIds.length,\n observationUnit: opts.independentUnitByScenarioId === undefined ? 'cell' : 'registered',\n minimumRequired: guard.minimumPairs,\n fewRuns: !guard.sufficient,\n missingCellIds: expectedCellIds.filter((cellId) => !measuredCells.has(cellId)),\n missingScenarioIds: [...scenarioIds].filter((id) => !measuredScenarios.has(id)).sort(),\n })\n }\n return out\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AA6BA,MAAa,oBAAoB;;AAYjC,SAAgB,qBAAqB,QAAwB;CAC3D,MAAM,YAAY,OAAO,YAAY,GAAG;CACxC,MAAM,aAAa,OAAO,MAAM,YAAY,CAAC;CAC7C,IACE,YAAY,KACZ,OAAO,KAAK,MAAM,UAClB,CAAC,iBAAiB,KAAK,UAAU,KACjC,CAAC,OAAO,cAAc,OAAO,UAAU,CAAC,GAExC,MAAM,IAAI,MAAM,kCAAkC,OAAO,2BAA2B;CAEtF,OAAO,OAAO,MAAM,GAAG,SAAS;AAClC;;AAGA,SAAgB,uBACd,QACA,6BAC0D;CAC1D,IACE,OAAO,OAAO,WAAW,OAAO,MAAM,UACtC,OAAO,OAAO,WAAW,OAAO,QAAQ,QAExC,MAAM,IAAI,MAAM,sEAAsE;CAExF,IAAI,IAAI,IAAI,OAAO,OAAO,CAAC,CAAC,SAAS,OAAO,QAAQ,QAClD,MAAM,IAAI,MAAM,4EAA4E;CAE9F,IACE,OAAO,OAAO,MAAM,UAAU,CAAC,OAAO,SAAS,KAAK,CAAC,KACrD,OAAO,MAAM,MAAM,UAAU,CAAC,OAAO,SAAS,KAAK,CAAC,GAEpD,MAAM,IAAI,MAAM,sDAAsD;CAExE,IAAI,gCAAgC,KAAA,GAClC,OAAO;EAAE,QAAQ,CAAC,GAAG,OAAO,MAAM;EAAG,OAAO,CAAC,GAAG,OAAO,KAAK;EAAG,SAAS,CAAC,GAAG,OAAO,OAAO;CAAE;CAE9F,MAAM,cAAc,OAAO,QAAQ,IAAI,oBAAoB;CAC3D,MAAM,yBAAS,IAAI,IAA0D;CAC7E,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,QAAQ,KAAK;EAC9C,MAAM,aAAa,YAAY;EAC/B,MAAM,SAAS,4BAA4B,IAAI,UAAU;EACzD,IAAI,OAAO,WAAW,YAAY,OAAO,WAAW,KAAK,OAAO,KAAK,MAAM,QACzE,MAAM,IAAI,MACR,kEAAkE,WAAW,EAC/E;EAEF,MAAM,QAAQ,OAAO,IAAI,MAAM,KAAK;GAAE,QAAQ;GAAG,OAAO;GAAG,GAAG;EAAE;EAChE,MAAM,UAAU,OAAO,OAAO;EAC9B,MAAM,SAAS,OAAO,MAAM;EAC5B,MAAM,KAAK;EACX,OAAO,IAAI,QAAQ,KAAK;CAC1B;CACA,MAAM,UAAU,CAAC,GAAG,OAAO,KAAK,CAAC,CAAC,CAAC,KAAK;CACxC,OAAO;EACL,QAAQ,QAAQ,KAAK,OAAO,OAAO,IAAI,EAAE,CAAC,CAAE,SAAS,OAAO,IAAI,EAAE,CAAC,CAAE,CAAC;EACtE,OAAO,QAAQ,KAAK,OAAO,OAAO,IAAI,EAAE,CAAC,CAAE,QAAQ,OAAO,IAAI,EAAE,CAAC,CAAE,CAAC;EACpE;CACF;AACF;;;;;;;;;;;;AAaA,SAAgB,YACd,WACA,UACA,aACA,QACe;CACf,MAAM,cACJ,QACA,WACwB;EACxB,MAAM,SAAS,OAAO,IAAI,MAAM;EAChC,MAAM,yBAAS,IAAI,IAAoB;EACvC,IAAI,CAAC,QAAQ,OAAO;EACpB,KAAK,MAAM,CAAC,SAAS,MAAM,OAAO,QAAQ,MAAM,GAAG;GACjD,IAAI,EAAE,WAAW,MACf,MAAM,IAAI,MAAM,sBAAsB,OAAO,gCAAgC;GAE/E,MAAM,IAAI,OAAO,CAAC;GAClB,IAAI,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,GAC7C,MAAM,IAAI,MAAM,sBAAsB,OAAO,uCAAuC;GAEtF,IAAI,OAAO,MAAM,UAAU,OAAO,IAAI,SAAS,CAAC;EAClD;EACA,OAAO;CACT;CAEA,MAAM,WAAW,WAAmB,YAAY,IAAI,qBAAqB,MAAM,CAAC;CAChF,MAAM,YAAY,CAAC,GAAG,UAAU,KAAK,CAAC,CAAC,CAAC,OAAO,OAAO,CAAC,CAAC,KAAK;CAC7D,MAAM,YAAY,CAAC,GAAG,SAAS,KAAK,CAAC,CAAC,CAAC,OAAO,OAAO,CAAC,CAAC,KAAK;CAG5D,IAAI,UAAU,WAAW,UAAU,UAAU,UAAU,MAAM,GAAG,MAAM,MAAM,UAAU,EAAE,GACtF,MAAM,IAAI,MACR,2EACgB,UAAU,KAAK,GAAG,EAAE,cAAc,UAAU,KAAK,GAAG,EAAE,+EAExE;CAGF,MAAM,SAAmB,CAAC;CAC1B,MAAM,QAAkB,CAAC;CACzB,MAAM,UAAoB,CAAC;CAC3B,KAAK,MAAM,UAAU,WAAW;EAC9B,MAAM,IAAI,WAAW,UAAU,MAAM;EACrC,MAAM,IAAI,WAAW,WAAW,MAAM;EAGtC,IAAI,EAAE,SAAS,KAAK,EAAE,SAAS,GAAG;EAClC,IAAI,EAAE,SAAS,KAAK,EAAE,SAAS,GAC7B,MAAM,IAAI,MAAM,sBAAsB,OAAO,uCAAuC;EAEtF,IAAI,EAAE,SAAS,EAAE,QAAQ,CAAC,GAAG,EAAE,KAAK,CAAC,CAAC,CAAC,MAAM,OAAO,CAAC,EAAE,IAAI,EAAE,CAAC,GAC5D,MAAM,IAAI,MAAM,sBAAsB,OAAO,kCAAkC;EAEjF,MAAM,WAAW,CAAC,GAAG,EAAE,KAAK,CAAC,CAAC,CAAC,KAAK;EACpC,OAAO,KAAK,mBAAmB,SAAS,KAAK,OAAO,EAAE,IAAI,EAAE,CAAE,CAAC,CAAC;EAChE,MAAM,KAAK,mBAAmB,SAAS,KAAK,OAAO,EAAE,IAAI,EAAE,CAAE,CAAC,CAAC;EAC/D,QAAQ,KAAK,MAAM;CACrB;CACA,OAAO;EAAE;EAAQ;EAAO;CAAQ;AAClC;AAEA,SAAS,mBAAmB,QAA0B;CACpD,MAAM,QAAQ,OAAO;CAGrB,OAAO,OAAO,OAAO,UAAU,UAAU,KAAK,IAC1C,QACA,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAC7D;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8FA,SAAgB,oBACd,QACA,OAAmC,CAAC,GACf;CACrB,MAAM,iBAAiB,KAAK,kBAAkB;CAC9C,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,OAAO,KAAK,QAAQ;CAa1B,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,eAAe,uBAAuB,QAAQ,KAAK,2BAA2B;CACpF,MAAM,WAAW,sBAAsB,aAAa,QAAQ,aAAa,OAAO;EAC9E;EACA;EACA;EACA;EACA,WAAW;EACX,UAAU,KAAK;CACjB,CAAC;CAID,MAAM,YACJ,SAAS,aACT,gBAAgB,aAAa,QAAQ,aAAa,OAAO;EACvD;EACA;EACA;EACA;CACF,CAAC;CACH,MAAM,kBACJ,cAAc,WACV,YACA,gBAAgB,aAAa,QAAQ,aAAa,OAAO;EACvD;EACA;EACA,WAAW;EACX;CACF,CAAC;CACP,MAAM,IAAI,aAAa,OAAO;CAC9B,IAAI,OAAO;CACX,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,GAAG;EAC7B,MAAM,QAAQ,aAAa,MAAM;EACjC,MAAM,SAAS,aAAa,OAAO;EACnC,IAAI,KAAK,IAAI,QAAQ,MAAM,IAAI,MAAM,QAAQ;CAC/C;CACA,MAAM,cAAc,MAAM,IAAI,IAAI,OAAO;CACzC,OAAO;EACL;EACA;EACA;EACA;EACA,mBAAmB,SAAS;EAC5B,SAAS,SAAS;EAClB;EACA;EACA,aAAa,OAAO,QAAQ;EAC5B,iBAAiB,KAAK,gCAAgC,KAAA,IAAY,SAAS;EAC3E,SAAS,aAAa;EACtB,iBAAiB,SAAS;EAC1B,gBAAgB,SAAS;EACzB,QAAQ,SAAS;EACjB,aAAa,SAAS;EACtB,SAAS,CAAC,SAAS;CACrB;AACF;;;;AAuCA,SAAgB,YAAY,QAA2B;CACrD,OAAO,OAAO,MAAM,MAAM,KAAK,IAAI,CAAC,IAAI,GAAG,IAAI,MAAM;AACvD;;;;;;;;;AAUA,SAAgB,qBACd,WACA,UACA,aACA,oBACA,OAUI,CAAC,GACkB;CACvB,MAAM,MAA6B,CAAC;CACpC,MAAM,kBAAkB,CAAC,GAAG,SAAS,KAAK,CAAC,CAAC,CACzC,QAAQ,WAAW,YAAY,IAAI,qBAAqB,MAAM,CAAC,CAAC,CAAC,CACjE,KAAK;CACR,KAAK,MAAM,OAAO,oBAAoB;EACpC,MAAM,SAAS,YAAY,WAAW,UAAU,cAAc,MAAM,EAAE,WAAW,IAAI;EACrF,IAAI,OAAO,OAAO,WAAW,GAAG;EAChC,MAAM,eAAe,uBAAuB,QAAQ,KAAK,2BAA2B;EACpF,MAAM,gBAAgB,IAAI,IAAI,OAAO,OAAO;EAC5C,MAAM,oBAAoB,IAAI,IAAI,OAAO,QAAQ,IAAI,oBAAoB,CAAC;EAC1E,MAAM,YAAY,KAAK,aAAa,MAAO,YAAY,CAAC,GAAG,OAAO,QAAQ,GAAG,OAAO,KAAK,CAAC;EAC1F,MAAM,qBAAqB,KAAK,aAAA;EAChC,MAAM,SAAS;GACb,YAAY,KAAK,cAAc;GAC/B,WAAW,KAAK,aAAa;GAC7B,WAAW;GACX,MAAM,KAAK,QAAQ;GACnB,UAAU,KAAK;EACjB;EACA,MAAM,QAAQ,sBAAsB,aAAa,QAAQ,aAAa,OAAO,MAAM;EACnF,MAAM,aAAa,sBAAsB,aAAa,OAAO,aAAa,QAAQ;GAChF,GAAG;GACH,WAAW;EACb,CAAC;EACD,MAAM,YACJ,MAAM,aAAa,gBAAgB,aAAa,QAAQ,aAAa,OAAO,MAAM;EACpF,IAAI,KAAK;GACP,WAAW;GACX;GACA;GACA,IAAI;IAAE,KAAK,MAAM;IAAK,MAAM,MAAM;GAAK;GACvC,mBAAmB,MAAM;GACzB,SAAS,MAAM;GACf,eAAe,MAAM;GAgBrB,WAAW,UAAU,MAAM,CAAC,aAAa,WAAW;GACpD;GACA,GAAG,aAAa,OAAO;GACvB,aAAa,OAAO,QAAQ;GAC5B,iBAAiB,KAAK,gCAAgC,KAAA,IAAY,SAAS;GAC3E,iBAAiB,MAAM;GACvB,SAAS,CAAC,MAAM;GAChB,gBAAgB,gBAAgB,QAAQ,WAAW,CAAC,cAAc,IAAI,MAAM,CAAC;GAC7E,oBAAoB,CAAC,GAAG,WAAW,CAAC,CAAC,QAAQ,OAAO,CAAC,kBAAkB,IAAI,EAAE,CAAC,CAAC,CAAC,KAAK;EACvF,CAAC;CACH;CACA,OAAO;AACT"}
@@ -1,5 +1,5 @@
1
1
  import { n as CaptureIntegrityError } from "./errors-Dngq5h35.js";
2
- import { C as TraceNotFoundError, E as extractOtlpAttributes, O as projectOtlpFlatLine, S as TraceFileTooLargeError, T as compareSpanTime, _ as TraceAnalysisLimitError, b as TraceFileMalformedError, c as bindSpanSourceReader, d as truncateForBudget, f as validateInteger, g as SpanNotFoundError, h as TRACE_ANALYSIS_LIMITS, k as spanEpochMillis, p as DEFAULT_TRACE_ANALYST_BUDGETS, u as compileSearchRegex, x as TraceFileMissingError, y as TraceAnalysisValidationError } from "./kind-factory-DxTzg32u.js";
2
+ import { C as TraceNotFoundError, E as extractOtlpAttributes, O as projectOtlpFlatLine, S as TraceFileTooLargeError, T as compareSpanTime, _ as TraceAnalysisLimitError, b as TraceFileMalformedError, c as bindSpanSourceReader, d as truncateForBudget, f as validateInteger, g as SpanNotFoundError, h as TRACE_ANALYSIS_LIMITS, k as spanEpochMillis, p as DEFAULT_TRACE_ANALYST_BUDGETS, u as compileSearchRegex, x as TraceFileMissingError, y as TraceAnalysisValidationError } from "./kind-factory-V79Hw0AU.js";
3
3
  import { readFile, stat } from "node:fs/promises";
4
4
  //#region src/trace-analyst/shared-abortable-task.ts
5
5
  function createSharedAbortableTask(factory) {
@@ -759,4 +759,4 @@ function finalizeErrorClusters(clusters, errorTraceCount) {
759
759
  //#endregion
760
760
  export { otlpTextToTraceAnalysisStore as i, OtlpFileTraceStore as n, createOtlpBufferTraceStore as r, DEFAULT_MAX_TRACE_FILE_BYTES as t };
761
761
 
762
- //# sourceMappingURL=store-otlp-DAaEbMLf.js.map
762
+ //# sourceMappingURL=store-otlp-2gYCy-v0.js.map