@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
@@ -1,5 +1,5 @@
1
1
  import { p as CostProvenance } from "./cost-ledger-DbQdN3nO.js";
2
- import { w as JudgeScore } from "./types-CS0qk_Yp.js";
2
+ import { w as JudgeScore } from "./types-BTRcIqFr.js";
3
3
  import { o as MatrixResult } from "./index-DNgf5gyG.js";
4
4
  import { AgentProfile } from "@tangle-network/agent-interface";
5
5
  //#region src/multishot/types.d.ts
@@ -398,4 +398,4 @@ interface RunMultishotMatrixResult {
398
398
  declare function runMultishotMatrix<TPersona extends MultishotPersona>(opts: RunMultishotMatrixOptions<TPersona>): Promise<RunMultishotMatrixResult>;
399
399
  //#endregion
400
400
  export { MultishotToolExecutor as A, MultishotFatalToolError as C, MultishotShape as D, MultishotResult as E, assertMultishotShotResult as F, MultishotTransportRequest as M, MultishotTransportResponse as N, MultishotShotResultError as O, MultishotTransportToolCall as P, MultishotDriverEmptyError as S, MultishotPersona as T, JudgeRunResult as _, MultishotCellOutput as a, runJudge as b, RunMultishotMatrixResult as c, MultishotShot as d, RunMultishotOptions as f, JudgeDimension as g, JudgeConfig as h, ConversationJudgeInput as i, MultishotTransport as j, MultishotToolDefinition as k, computeCellComposite as l, DEFAULT_JUDGE_MODEL as m, CellCompositeInput as n, MultishotJudges as o, runMultishot as p, CellCompositeScore as r, RunMultishotMatrixOptions as s, ArtifactJudgeInput as t, runMultishotMatrix as u, renderDimensions as v, MultishotMessage as w, MultishotArtifact as x, renderJsonFooter as y };
401
- //# sourceMappingURL=matrix-CyhW-vgJ.d.ts.map
401
+ //# sourceMappingURL=matrix-CPKsRY8Y.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"matrix-CyhW-vgJ.d.ts","names":[],"sources":["../src/multishot/types.ts","../src/multishot/judges.ts","../src/multishot/multishot.ts","../src/multishot/matrix.ts"],"mappings":";;;;;UAIiB;EACf;EACA;EACA;EACA,YAAY;IAAQ;IAAY;IAAc,MAAM;;;UAGrC;EACf;EACA;EACA;IAAc;IAAc,MAAM;;EAClC;;UAGe;EACf,YAAY;EACZ,WAAW;EACX;EACA;;;EAGA;;;;;;;;EAQA,iBAAiB;;UAGF;EACf;EACA;IACE;IACA;IACA,YAAY;;;;;;UAOC;EACf;EACA,UAAU,MAAM;EAChB,QAAQ;EACR;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA;IAAY;IAAc;;;UAGX;EACf;IAAW;IAAyB,aAAa;;EACjD;IAAU;IAAwB;;;;EAGlC;;;;EAIA;;;;;;;;KASU,sBACV,KAAK,8BACF,QAAQ;KAED,yBACV,MAAM,yBACN;;EAEE,WAAW;EACX,SAAS;MAER;EAAU;EAAiB;;UAEf;;EAEf;;GAEC;;;;;;;;UASc,eAAe,iBAAiB;;EAE/C,eAAe,SAAS;;;EAGxB,2BAA2B,SAAS;;cAGzB,kCAAkC;WACjB;EAA5B,YAA4B;;cAMjB,gCAAgC;EAC3C,YAAY;;cAMD,iCAAiC;EAC5C,YAAY;;;;;;;;;;;;;;;;;;;iBAyBE,0BAA0B,yBAAyB,SAAS;;;cCvI/D;UAEI;;EAEf;;EAEA;;UAGe,YAAY;;EAE3B;;;EAGA,WAAW;;EAEX;;EAEA,YAAY;;EAEZ;;;EAGA,cAAc,OAAO;;EAErB;;UAGe;;EAEf,OAAO;;EAEP,MAAM;;iBAGc,SAAS,QAC7B,OAAO,YAAY,SACnB,OAAO,SACN,QAAQ;;;iBAsIK,iBAAiB,eAAe;;iBAKhC,iBAAiB,eAAe;;;UCxK/B,oBAAoB,iBAAiB;EACpD,SAAS;EACT,SAAS;;;EAGT,QAAQ,eAAe;;EAEvB,QAAQ;;EAER,gBAAgB,eAAe;;;;EAI/B,mBAAmB;EACnB;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,gBAAgB;;;;EAIhB,iBAAiB;;;EAGjB,gBAAgB;EAChB,SAAS;;;;;;;;;;;KAYC,cAAc,iBAAiB,qBACzC,MAAM,oBAAoB,cACvB,QAAQ;;;;;;;;;;;;;;;;;iBA2BS,aAAa,iBAAiB,kBAClD,MAAM,oBAAoB,YACzB,QAAQ;;;UClFM,uBAAuB,iBAAiB;EACvD,YAAY;EACZ,SAAS;;UAGM,mBAAmB,iBAAiB;EACnD,UAAU;EACV,SAAS;;UAGM,gBAAgB,iBAAiB;;EAEhD,cAAc,YAAY,uBAAuB;;EAEjD,aAAa,YAAY,mBAAmB;;EAE5C,iBAAiB,YAAY,mBAAmB;;EAEhD;;EAEA;;UAGe;EACf;EACA,cAAc;EACd;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;EAEF;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;;UAIa,0BAA0B,iBAAiB;;EAE1D,UAAU;IAAQ;IAAY,OAAO;;;EAErC,UAAU;;;EAGV,QAAQ,eAAe;;EAEvB,QAAQ,gBAAgB;;EAExB,QAAQ;;EAER,gBAAgB,eAAe;;EAE/B,mBAAmB;;EAEnB;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,gBAAgB;;EAEhB,iBAAiB;;;EAGjB,gBAAgB;;;;;;;;;;;;;;;EAehB,UAAU,cAAc;;;;;UAMT;EACf;EACA;EACA;;UAoBe;EACf,cAAc;;EAEd,cAAc,cAAc;;EAE5B,iBAAiB,cAAc;;;;;;;iBAQjB,qBAAqB,OAAO;EAC1C;EACA;EACA;EACA;;UAuBe;EACf,QAAQ,aAAa;;iBAGD,mBAAmB,iBAAiB,kBACxD,MAAM,0BAA0B,YAC/B,QAAQ"}
1
+ {"version":3,"file":"matrix-CPKsRY8Y.d.ts","names":[],"sources":["../src/multishot/types.ts","../src/multishot/judges.ts","../src/multishot/multishot.ts","../src/multishot/matrix.ts"],"mappings":";;;;;UAIiB;EACf;EACA;EACA;EACA,YAAY;IAAQ;IAAY;IAAc,MAAM;;;UAGrC;EACf;EACA;EACA;IAAc;IAAc,MAAM;;EAClC;;UAGe;EACf,YAAY;EACZ,WAAW;EACX;EACA;;;EAGA;;;;;;;;EAQA,iBAAiB;;UAGF;EACf;EACA;IACE;IACA;IACA,YAAY;;;;;;UAOC;EACf;EACA,UAAU,MAAM;EAChB,QAAQ;EACR;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA;IAAY;IAAc;;;UAGX;EACf;IAAW;IAAyB,aAAa;;EACjD;IAAU;IAAwB;;;;EAGlC;;;;EAIA;;;;;;;;KASU,sBACV,KAAK,8BACF,QAAQ;KAED,yBACV,MAAM,yBACN;;EAEE,WAAW;EACX,SAAS;MAER;EAAU;EAAiB;;UAEf;;EAEf;;GAEC;;;;;;;;UASc,eAAe,iBAAiB;;EAE/C,eAAe,SAAS;;;EAGxB,2BAA2B,SAAS;;cAGzB,kCAAkC;WACjB;EAA5B,YAA4B;;cAMjB,gCAAgC;EAC3C,YAAY;;cAMD,iCAAiC;EAC5C,YAAY;;;;;;;;;;;;;;;;;;;iBAyBE,0BAA0B,yBAAyB,SAAS;;;cCvI/D;UAEI;;EAEf;;EAEA;;UAGe,YAAY;;EAE3B;;;EAGA,WAAW;;EAEX;;EAEA,YAAY;;EAEZ;;;EAGA,cAAc,OAAO;;EAErB;;UAGe;;EAEf,OAAO;;EAEP,MAAM;;iBAGc,SAAS,QAC7B,OAAO,YAAY,SACnB,OAAO,SACN,QAAQ;;;iBAsIK,iBAAiB,eAAe;;iBAKhC,iBAAiB,eAAe;;;UCxK/B,oBAAoB,iBAAiB;EACpD,SAAS;EACT,SAAS;;;EAGT,QAAQ,eAAe;;EAEvB,QAAQ;;EAER,gBAAgB,eAAe;;;;EAI/B,mBAAmB;EACnB;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,gBAAgB;;;;EAIhB,iBAAiB;;;EAGjB,gBAAgB;EAChB,SAAS;;;;;;;;;;;KAYC,cAAc,iBAAiB,qBACzC,MAAM,oBAAoB,cACvB,QAAQ;;;;;;;;;;;;;;;;;iBA2BS,aAAa,iBAAiB,kBAClD,MAAM,oBAAoB,YACzB,QAAQ;;;UClFM,uBAAuB,iBAAiB;EACvD,YAAY;EACZ,SAAS;;UAGM,mBAAmB,iBAAiB;EACnD,UAAU;EACV,SAAS;;UAGM,gBAAgB,iBAAiB;;EAEhD,cAAc,YAAY,uBAAuB;;EAEjD,aAAa,YAAY,mBAAmB;;EAE5C,iBAAiB,YAAY,mBAAmB;;EAEhD;;EAEA;;UAGe;EACf;EACA,cAAc;EACd;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;EAEF;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;;UAIa,0BAA0B,iBAAiB;;EAE1D,UAAU;IAAQ;IAAY,OAAO;;;EAErC,UAAU;;;EAGV,QAAQ,eAAe;;EAEvB,QAAQ,gBAAgB;;EAExB,QAAQ;;EAER,gBAAgB,eAAe;;EAE/B,mBAAmB;;EAEnB;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,gBAAgB;;EAEhB,iBAAiB;;;EAGjB,gBAAgB;;;;;;;;;;;;;;;EAehB,UAAU,cAAc;;;;;UAMT;EACf;EACA;EACA;;UAoBe;EACf,cAAc;;EAEd,cAAc,cAAc;;EAE5B,iBAAiB,cAAc;;;;;;;iBAQjB,qBAAqB,OAAO;EAC1C;EACA;EACA;EACA;;UAuBe;EACf,QAAQ,aAAa;;iBAGD,mBAAmB,iBAAiB,kBACxD,MAAM,0BAA0B,YAC/B,QAAQ"}
@@ -5,7 +5,7 @@ import { r as LedgerHash } from "../canonical-CFpojCN5.js";
5
5
  import { s as TraceStore } from "../store-BErPvYBr.js";
6
6
  import { a as OutcomeFilter, i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, s as OutcomeStoreError, t as DeploymentOutcome } from "../outcome-store-CNt4iZ67.js";
7
7
  import { a as ComputedInterval } from "../ast-hI-vjW6J.js";
8
- import { a as RubricPredictiveValidityReport, c as CorrelationInterval, i as RubricPredictiveValidityInput, l as OutcomeReduction, n as RubricOutcomeExclusion, o as RubricRanking, r as RubricOutcomePair, s as rubricPredictiveValidity, t as OutcomeMetricSpec } from "../rubric-predictive-validity-Bmj2_cll.js";
8
+ import { a as RubricPredictiveValidityReport, c as CorrelationInterval, i as RubricPredictiveValidityInput, l as OutcomeReduction, n as RubricOutcomeExclusion, o as RubricRanking, r as RubricOutcomePair, s as rubricPredictiveValidity, t as OutcomeMetricSpec } from "../rubric-predictive-validity-CTU8y8Cv.js";
9
9
  import { z } from "zod";
10
10
  //#region src/meta-eval/correlation-study.d.ts
11
11
  interface EvalMetricSpec {
@@ -216,6 +216,58 @@ interface EvaluatorAdmissionReport {
216
216
  * The execution owner enforces auditor separation, fresh sampling, and evidence authenticity.
217
217
  */
218
218
  declare function auditEvaluator(input: EvaluatorAuditInput): EvaluatorAdmissionReport;
219
+ declare const probabilityAuditSchema: z.ZodObject<{
220
+ evaluatorDigest: z.ZodString;
221
+ population: z.ZodString;
222
+ samplingFrame: z.ZodString;
223
+ authority: z.ZodObject<{
224
+ evaluatorAuthorId: z.ZodString;
225
+ auditorId: z.ZodString;
226
+ independenceEvidenceRef: z.ZodString;
227
+ }, z.core.$strict>;
228
+ policy: z.ZodObject<{
229
+ confidence: z.ZodNumber;
230
+ maxFalseAcceptanceRate: z.ZodNumber;
231
+ maxFalseRejectionRate: z.ZodNumber;
232
+ }, z.core.$strict>;
233
+ thresholds: z.ZodObject<{
234
+ rejectAtOrBelow: z.ZodNumber;
235
+ acceptAtOrAbove: z.ZodNumber;
236
+ }, z.core.$strict>;
237
+ observations: z.ZodArray<z.ZodObject<{
238
+ id: z.ZodString;
239
+ independentUnitId: z.ZodString;
240
+ evidenceRef: z.ZodString;
241
+ expected: z.ZodEnum<{
242
+ accept: "accept";
243
+ reject: "reject";
244
+ }>;
245
+ exposure: z.ZodEnum<{
246
+ development: "development";
247
+ fresh: "fresh";
248
+ }>;
249
+ acceptProbability: z.ZodNullable<z.ZodNumber>;
250
+ }, z.core.$strict>>;
251
+ }, z.core.$strict>;
252
+ /** A frozen caller-owned rule over P(accept), not the provider's entropy/confidence statistic. */
253
+ type ProbabilityPolicyAuditInput = z.infer<typeof probabilityAuditSchema>;
254
+ interface ProbabilityPolicyAuditReport extends EvaluatorAdmissionReport {
255
+ probabilityPolicy: {
256
+ sourceEvaluatorDigest: LedgerHash;
257
+ thresholds: ProbabilityPolicyAuditInput['thresholds'];
258
+ /** Retain probabilities by observation id so the reported mapping can be reproduced. */
259
+ observations: Array<{
260
+ id: string;
261
+ acceptProbability: number | null;
262
+ }>;
263
+ };
264
+ }
265
+ /**
266
+ * Audit a fixed probability-to-decision rule with the existing independent-unit error bounds.
267
+ * No threshold search, model calls, promotion, or assumption of calibrated probabilities.
268
+ * Null scores and the open interval between thresholds remain unknown, never correct by default.
269
+ */
270
+ declare function auditProbabilityPolicy(input: ProbabilityPolicyAuditInput): ProbabilityPolicyAuditReport;
219
271
  //#endregion
220
272
  //#region src/meta-eval/plants.d.ts
221
273
  /**
@@ -601,5 +653,5 @@ interface EvalHealthStamp {
601
653
  */
602
654
  declare function evalHealthStamp(report: SentinelReport): EvalHealthStamp;
603
655
  //#endregion
604
- export { CalibrationBin, CalibrationOptions, CalibrationPair, CalibrationReport, type CalibrationResult, type CandidateScore, CatchRateReport, CatchRateStatus, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type CorrelationInterval, CorrelationResult, CorrelationStudyOptions, CorrelationStudyResult, DeploymentOutcome, EvalHealthStamp, EvalMetricSpec, type EvaluatorAdmissionPolicy, type EvaluatorAdmissionReport, type EvaluatorAuditInput, type EvaluatorAuditObservation, type EvaluatorErrorRate, EvidencePerturbation, FileSystemOutcomeStore, FileSystemOutcomeStoreOptions, type GoldenItem, InMemoryOutcomeStore, JudgeSentinelOptions, OutcomeFilter, OutcomeMetricSpec, OutcomePair, type OutcomeReduction, OutcomeStore, OutcomeStoreError, PerturbationKind, PerturbedField, PerturbedPlant, Plant, PlantEvidence, PlantExpectation, PlantKind, PlantKindCounts, PlantManifest, PlantOutcome, type PositionalBiasResult, RubricOutcomeExclusion, RubricOutcomePair, RubricPredictiveValidityInput, RubricPredictiveValidityReport, RubricRanking, SeedPlantsOptions, SeededGradingSet, type SelfPreferenceResult, SentinelMetricName, SentinelMetrics, SentinelReport, SentinelSetOptions, SentinelSnapshot, SentinelStore, SentinelThresholds, SentinelTrend, SnapshotMeta, UnseededRejection, type VerbosityBiasResult, auditEvaluator, calibrateJudge, calibrateJudgeContinuous, calibrationCurve, calibrationFromPairs, catchRate, continuousAgreement, correlationStudy, definePlant, evalHealthStamp, fileSentinelStore, inMemorySentinelStore, judgeSentinelReport, perturbEvidence, plantByPerturbation, positionalBias, rubricPredictiveValidity, seedPlants, selfPreference, snapshotFromAgreement, snapshotFromCalibration, snapshotFromSentinelSet, validateSentinelSnapshot, verbosityBias };
656
+ export { CalibrationBin, CalibrationOptions, CalibrationPair, CalibrationReport, type CalibrationResult, type CandidateScore, CatchRateReport, CatchRateStatus, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type CorrelationInterval, CorrelationResult, CorrelationStudyOptions, CorrelationStudyResult, DeploymentOutcome, EvalHealthStamp, EvalMetricSpec, type EvaluatorAdmissionPolicy, type EvaluatorAdmissionReport, type EvaluatorAuditInput, type EvaluatorAuditObservation, type EvaluatorErrorRate, EvidencePerturbation, FileSystemOutcomeStore, FileSystemOutcomeStoreOptions, type GoldenItem, InMemoryOutcomeStore, JudgeSentinelOptions, OutcomeFilter, OutcomeMetricSpec, OutcomePair, type OutcomeReduction, OutcomeStore, OutcomeStoreError, PerturbationKind, PerturbedField, PerturbedPlant, Plant, PlantEvidence, PlantExpectation, PlantKind, PlantKindCounts, PlantManifest, PlantOutcome, type PositionalBiasResult, type ProbabilityPolicyAuditInput, type ProbabilityPolicyAuditReport, RubricOutcomeExclusion, RubricOutcomePair, RubricPredictiveValidityInput, RubricPredictiveValidityReport, RubricRanking, SeedPlantsOptions, SeededGradingSet, type SelfPreferenceResult, SentinelMetricName, SentinelMetrics, SentinelReport, SentinelSetOptions, SentinelSnapshot, SentinelStore, SentinelThresholds, SentinelTrend, SnapshotMeta, UnseededRejection, type VerbosityBiasResult, auditEvaluator, auditProbabilityPolicy, calibrateJudge, calibrateJudgeContinuous, calibrationCurve, calibrationFromPairs, catchRate, continuousAgreement, correlationStudy, definePlant, evalHealthStamp, fileSentinelStore, inMemorySentinelStore, judgeSentinelReport, perturbEvidence, plantByPerturbation, positionalBias, rubricPredictiveValidity, seedPlants, selfPreference, snapshotFromAgreement, snapshotFromCalibration, snapshotFromSentinelSet, validateSentinelSnapshot, verbosityBias };
605
657
  //# sourceMappingURL=index.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"index.d.ts","names":[],"sources":["../../src/meta-eval/correlation-study.ts","../../src/meta-eval/calibration.ts","../../src/meta-eval/evaluator-admission.ts","../../src/meta-eval/plants.ts","../../src/meta-eval/sentinel.ts"],"mappings":";;;;;;;;;;UAoBiB;EACf;;;EAGA,WAAW,KAAK,KAAK,OAAO,eAAe;;UAG5B;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA;IAAe;IAAe;;;EAE9B;IAAgB;IAAe;;;EAE/B;;UAGe;EACf,OAAO;;EAEP,eAAe,MACb;IACE;IACA;;EAGJ;EACA;;UAGe;;EAEf;;EAEA,gBAAgB;;EAEhB;;EAEA;;;EAGA;;iBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,aAAa,kBACb,8BACA,UAAS,0BACR,QAAQ;;;UChEM;EACf;EACA;EACA;EACA;EACA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,MAAM;;EAEN;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;IAAU;IAAY;;;UAGP;EACf;EACA;;iBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,YAAY,gBACZ,uBACA,UAAS,qBACR,QAAQ;;iBA0CK,qBACd,qBAAqB,mBACrB,oBACA,uBACA,UAAS,qBACR;;;cC9FG,cAAY,EAAA;;;;GAMP,EAAA,KAAA;;KAGC,2BAA2B,EAAE,aAAa;cAEhD,mBAAiB,EAAA;;;;;;;;;;;;;;;;;GASZ,EAAA,KAAA;;KAGC,4BAA4B,EAAE,aAAa;cAEjD,aAAW,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAeN,EAAA,KAAA;KAEC,sBAAsB,EAAE,aAAa;UAEhC;;EAEf;EACA;;EAEA;;EAEA;EACA;;EAEA;;EAEA,UAAU;EACV;EACA;;UAGe;EACf,iBAAiB;EACjB,cAAc;EACd,aAAa;EACb,cAAc;EACd;EACA;EACA,WAAW;EACX,QAAQ;;EAER;EACA;EACA;EACA;EACA,cAAc;EACd;IACE;IACA;IACA;IACA;IACA;IACA;;EAEF,YAAY;IAAQ;IAAY;IAA2B;;EAC3D,iBAAiB;EACjB,gBAAgB;;;;;;;iBAoEF,eAAe,OAAO,sBAAsB;;;;;;;;KCrGhD;;;;;;;;;;KAkBA;UAOK;;EAEf;EACA,MAAM;;EAEN,MAAM;;EAEN,iBAAiB;;;;;;;;;;;;;;iBAeH,YAAY;EAC1B;EACA,MAAM;EACN,MAAM;EACN,iBAAiB;IACf;;;;;UAqDa;;EAEf;;EAEA;;;KAIU;;KAGA;UAEK;;EAEf,UAAU;;EAEV,OAAO;;EAEP,KAAK;;;;;;EAML;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAsJc,gBAAgB,UAAU,gBAAgB;UA8BzC;EACf,OAAO;;EAEP,UAAU;EACV,OAAO;EACP,KAAK;EACL;EACA;;;;;;;;;;;;;;;;;;;;;;iBAuBc,oBAAoB;EAClC;EACA;EACA,UAAU;;EAEV;EACA;IACE;UA0Ba;;;;;;EAMf,MAAM;;EAEN;;EAEA;;EAEA;;EAEA,QAAQ;;UAGO;;;;;;EAMf,OAAO;EACP,UAAU;;UAGK;;;;;EAKf;;;;;;EAMA;;;;;;;;;;;iBAYc,WACd,kBAAkB,cAClB,iBAAiB,SACjB,UAAS,oBACR;;;;;;UAqGc;EACf;EACA;;;;;;KAOU;UAEK;EACf;EACA;EACA;EACA;;EAEA;;;;;;;;UASe;EACf;EACA;EACA;;EAEA;;UAGe;EACf,QAAQ;;EAER;EACA;EACA;EACA;;;;;EAKA;;EAEA;;EAEA,QAAQ,QAAQ,OAAO,WAAW;;EAElC;;EAEA;EACA,UAAU;;;;;;;;;;iBAWI,UACd,kBAAkB,gBAClB,UAAU,gBACT;;;cC/lBG;KAEM,6BAA6B;UAExB;;EAEf;;EAEA;;EAEA;;UAGe;;EAEf;EACA;;EAEA;;;;;;;EAOA,SAAS;;;UAIM;EACf;EACA;EACA;;;iBAcc,yBAAyB,UAAU,kBAAkB;;;;;;iBAiCrD,wBACd,QAAQ,oBAAoB,6BAC5B,MAAM,eACL;;;;;;;;iBAmBa,sBACd,QAAQ,wBAAwB,qBAChC,MAAM,eACL;UAYc;;EAEf;;;;;;;;;iBAUc,wBACd,QAAQ,kBACR,QAAQ,cACR,MAAM,cACN,UAAS,qBACR;UA4Cc;EACf,OAAO,UAAU,mBAAmB;;EAEpC,QAAQ,mBAAmB,QAAQ;;iBAGrB,sBAAsB,UAAS,qBAA0B;;;;;;;iBAqBzD,kBAAkB,eAAe;UA0ChC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;;EAEf;EACA,aAAa;;EAEb,cAAc;;UAGC;EACf;EACA,QAAQ;;EAER,OAAO;;EAEP;;EAEA;;EAEA;EACA;EACA;;UAGe;EACf,UAAU;EACV;EACA;;;;;;;EAOA;;iBAKc,oBACd,SAAS,oBACT,MAAM,uBACL;UA6Gc;EACf;EACA;;;;;;;;;;iBAWc,gBAAgB,QAAQ,iBAAiB"}
1
+ {"version":3,"file":"index.d.ts","names":[],"sources":["../../src/meta-eval/correlation-study.ts","../../src/meta-eval/calibration.ts","../../src/meta-eval/evaluator-admission.ts","../../src/meta-eval/plants.ts","../../src/meta-eval/sentinel.ts"],"mappings":";;;;;;;;;;UAoBiB;EACf;;;EAGA,WAAW,KAAK,KAAK,OAAO,eAAe;;UAG5B;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA;IAAe;IAAe;;;EAE9B;IAAgB;IAAe;;;EAE/B;;UAGe;EACf,OAAO;;EAEP,eAAe,MACb;IACE;IACA;;EAGJ;EACA;;UAGe;;EAEf;;EAEA,gBAAgB;;EAEhB;;EAEA;;;EAGA;;iBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,aAAa,kBACb,8BACA,UAAS,0BACR,QAAQ;;;UChEM;EACf;EACA;EACA;EACA;EACA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,MAAM;;EAEN;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;IAAU;IAAY;;;UAGP;EACf;EACA;;iBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,YAAY,gBACZ,uBACA,UAAS,qBACR,QAAQ;;iBA0CK,qBACd,qBAAqB,mBACrB,oBACA,uBACA,UAAS,qBACR;;;cC9FG,cAAY,EAAA;;;;GAMP,EAAA,KAAA;;KAGC,2BAA2B,EAAE,aAAa;cAEhD,mBAAiB,EAAA;;;;;;;;;;;;;;;;;GASZ,EAAA,KAAA;;KAGC,4BAA4B,EAAE,aAAa;cAEjD,aAAW,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAeN,EAAA,KAAA;KAEC,sBAAsB,EAAE,aAAa;UAEhC;;EAEf;EACA;;EAEA;;EAEA;EACA;;EAEA;;EAEA,UAAU;EACV;EACA;;UAGe;EACf,iBAAiB;EACjB,cAAc;EACd,aAAa;EACb,cAAc;EACd;EACA;EACA,WAAW;EACX,QAAQ;;EAER;EACA;EACA;EACA;EACA,cAAc;EACd;IACE;IACA;IACA;IACA;IACA;IACA;;EAEF,YAAY;IAAQ;IAAY;IAA2B;;EAC3D,iBAAiB;EACjB,gBAAgB;;;;;;;iBAoEF,eAAe,OAAO,sBAAsB;cA2GtD,wBAAsB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAG1B,EAAA,KAAA;;KAGU,8BAA8B,EAAE,aAAa;UAExC,qCAAqC;EACpD;IACE,uBAAuB;IACvB,YAAY;;IAEZ,cAAc;MAAQ;MAAY;;;;;;;;;iBAStB,uBACd,OAAO,8BACN;;;;;;;;KCxOS;;;;;;;;;;KAkBA;UAOK;;EAEf;EACA,MAAM;;EAEN,MAAM;;EAEN,iBAAiB;;;;;;;;;;;;;;iBAeH,YAAY;EAC1B;EACA,MAAM;EACN,MAAM;EACN,iBAAiB;IACf;;;;;UAqDa;;EAEf;;EAEA;;;KAIU;;KAGA;UAEK;;EAEf,UAAU;;EAEV,OAAO;;EAEP,KAAK;;;;;;EAML;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAsJc,gBAAgB,UAAU,gBAAgB;UA8BzC;EACf,OAAO;;EAEP,UAAU;EACV,OAAO;EACP,KAAK;EACL;EACA;;;;;;;;;;;;;;;;;;;;;;iBAuBc,oBAAoB;EAClC;EACA;EACA,UAAU;;EAEV;EACA;IACE;UA0Ba;;;;;;EAMf,MAAM;;EAEN;;EAEA;;EAEA;;EAEA,QAAQ;;UAGO;;;;;;EAMf,OAAO;EACP,UAAU;;UAGK;;;;;EAKf;;;;;;EAMA;;;;;;;;;;;iBAYc,WACd,kBAAkB,cAClB,iBAAiB,SACjB,UAAS,oBACR;;;;;;UAqGc;EACf;EACA;;;;;;KAOU;UAEK;EACf;EACA;EACA;EACA;;EAEA;;;;;;;;UASe;EACf;EACA;EACA;;EAEA;;UAGe;EACf,QAAQ;;EAER;EACA;EACA;EACA;;;;;EAKA;;EAEA;;EAEA,QAAQ,QAAQ,OAAO,WAAW;;EAElC;;EAEA;EACA,UAAU;;;;;;;;;;iBAWI,UACd,kBAAkB,gBAClB,UAAU,gBACT;;;cC/lBG;KAEM,6BAA6B;UAExB;;EAEf;;EAEA;;EAEA;;UAGe;;EAEf;EACA;;EAEA;;;;;;;EAOA,SAAS;;;UAIM;EACf;EACA;EACA;;;iBAcc,yBAAyB,UAAU,kBAAkB;;;;;;iBAiCrD,wBACd,QAAQ,oBAAoB,6BAC5B,MAAM,eACL;;;;;;;;iBAmBa,sBACd,QAAQ,wBAAwB,qBAChC,MAAM,eACL;UAYc;;EAEf;;;;;;;;;iBAUc,wBACd,QAAQ,kBACR,QAAQ,cACR,MAAM,cACN,UAAS,qBACR;UA4Cc;EACf,OAAO,UAAU,mBAAmB;;EAEpC,QAAQ,mBAAmB,QAAQ;;iBAGrB,sBAAsB,UAAS,qBAA0B;;;;;;;iBAqBzD,kBAAkB,eAAe;UA0ChC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;;EAEf;EACA,aAAa;;EAEb,cAAc;;UAGC;EACf;EACA,QAAQ;;EAER,OAAO;;EAEP;;EAEA;;EAEA;EACA;EACA;;UAGe;EACf,UAAU;EACV;EACA;;;;;;;EAOA;;iBAKc,oBACd,SAAS,oBACT,MAAM,uBACL;UA6Gc;EACf;EACA;;;;;;;;;;iBAWc,gBAAgB,QAAQ,iBAAiB"}
@@ -1,8 +1,8 @@
1
1
  import { n as CaptureIntegrityError, s as ValidationError } from "../errors-Dngq5h35.js";
2
- import { a as hashCanonical, i as compareCodeUnits } from "../canonical-DPyQ_rpt.js";
2
+ import { a as hashCanonical, i as compareCodeUnits } from "../canonical-BsZVgBhW.js";
3
3
  import { t as mulberry32 } from "../random-Dn5fPWkt.js";
4
4
  import { a as selfPreference, i as positionalBias, n as calibrateJudgeContinuous, o as verbosityBias, r as continuousAgreement, t as calibrateJudge } from "../judge-calibration-DYmaBtJr.js";
5
- import { u as runMetricExtractor } from "../query-D1nLIKt7.js";
5
+ import { u as runMetricExtractor } from "../query-DlV0pdx4.js";
6
6
  import { r as computeInterval } from "../ast-CP9ae9B0.js";
7
7
  import { t as analyzeSeries } from "../series-convergence-CjO2QdRW.js";
8
8
  import { a as reduceOutcomeMetric, i as hasVariation, n as assertUniqueObservationIds, o as validateObservationOptions, r as correlationSummary, t as rubricPredictiveValidity } from "../rubric-predictive-validity-CCK-1B7w.js";
@@ -375,6 +375,55 @@ function auditEvaluator(input) {
375
375
  reportDigest: hashCanonical(body)
376
376
  };
377
377
  }
378
+ const probabilityThresholdsSchema = z.object({
379
+ rejectAtOrBelow: z.number().finite().min(0).max(1),
380
+ acceptAtOrAbove: z.number().finite().min(0).max(1)
381
+ }).strict().refine((value) => value.rejectAtOrBelow < value.acceptAtOrAbove, { message: "rejection and acceptance thresholds must not overlap" });
382
+ const probabilityObservationSchema = observationSchema.omit({ observed: true }).extend({ acceptProbability: z.number().finite().min(0).max(1).nullable() }).strict();
383
+ const probabilityAuditSchema = inputSchema.extend({
384
+ thresholds: probabilityThresholdsSchema,
385
+ observations: z.array(probabilityObservationSchema)
386
+ });
387
+ /**
388
+ * Audit a fixed probability-to-decision rule with the existing independent-unit error bounds.
389
+ * No threshold search, model calls, promotion, or assumption of calibrated probabilities.
390
+ * Null scores and the open interval between thresholds remain unknown, never correct by default.
391
+ */
392
+ function auditProbabilityPolicy(input) {
393
+ const parsed = probabilityAuditSchema.safeParse(input);
394
+ if (!parsed.success) throw new ValidationError(`invalid probability policy audit: ${parsed.error.message}`);
395
+ const { thresholds, observations, ...source } = parsed.data;
396
+ observations.sort((a, b) => compareCodeUnits(a.id, b.id));
397
+ const probabilityPolicy = {
398
+ sourceEvaluatorDigest: source.evaluatorDigest,
399
+ thresholds,
400
+ observations: observations.map(({ id, acceptProbability }) => ({
401
+ id,
402
+ acceptProbability
403
+ }))
404
+ };
405
+ const { reportDigest: _baseDigest, ...baseReport } = auditEvaluator({
406
+ ...source,
407
+ evaluatorDigest: hashCanonical({
408
+ kind: "probability-policy-v1",
409
+ sourceEvaluatorDigest: source.evaluatorDigest,
410
+ thresholds
411
+ }),
412
+ observations: observations.map(({ acceptProbability, ...observation }) => ({
413
+ ...observation,
414
+ observed: acceptProbability === null ? "unknown" : acceptProbability >= thresholds.acceptAtOrAbove ? "accept" : acceptProbability <= thresholds.rejectAtOrBelow ? "reject" : "unknown"
415
+ }))
416
+ });
417
+ const body = {
418
+ ...baseReport,
419
+ inputDigest: hashCanonical(parsed.data),
420
+ probabilityPolicy
421
+ };
422
+ return {
423
+ ...body,
424
+ reportDigest: hashCanonical(body)
425
+ };
426
+ }
378
427
  //#endregion
379
428
  //#region src/meta-eval/plants.ts
380
429
  /**
@@ -1130,6 +1179,6 @@ function evalHealthStamp(report) {
1130
1179
  };
1131
1180
  }
1132
1181
  //#endregion
1133
- export { FileSystemOutcomeStore, InMemoryOutcomeStore, OutcomeStoreError, auditEvaluator, calibrateJudge, calibrateJudgeContinuous, calibrationCurve, calibrationFromPairs, catchRate, continuousAgreement, correlationStudy, definePlant, evalHealthStamp, fileSentinelStore, inMemorySentinelStore, judgeSentinelReport, perturbEvidence, plantByPerturbation, positionalBias, rubricPredictiveValidity, seedPlants, selfPreference, snapshotFromAgreement, snapshotFromCalibration, snapshotFromSentinelSet, validateSentinelSnapshot, verbosityBias };
1182
+ export { FileSystemOutcomeStore, InMemoryOutcomeStore, OutcomeStoreError, auditEvaluator, auditProbabilityPolicy, calibrateJudge, calibrateJudgeContinuous, calibrationCurve, calibrationFromPairs, catchRate, continuousAgreement, correlationStudy, definePlant, evalHealthStamp, fileSentinelStore, inMemorySentinelStore, judgeSentinelReport, perturbEvidence, plantByPerturbation, positionalBias, rubricPredictiveValidity, seedPlants, selfPreference, snapshotFromAgreement, snapshotFromCalibration, snapshotFromSentinelSet, validateSentinelSnapshot, verbosityBias };
1134
1183
 
1135
1184
  //# sourceMappingURL=index.js.map