@tangle-network/agent-eval 1.0.24 → 1.0.26

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 (124) hide show
  1. package/CHANGELOG.md +88 -0
  2. package/dist/adapters/http.d.ts +2 -2
  3. package/dist/{agent-profile-DC_v4p1f.d.ts → agent-profile-BF1WLMco.d.ts} +3 -3
  4. package/dist/{agent-profile-DC_v4p1f.d.ts.map → agent-profile-BF1WLMco.d.ts.map} +1 -1
  5. package/dist/analyst/index.d.ts +14 -14
  6. package/dist/analyst/index.js +1 -1
  7. package/dist/{analyze-runs-DGr7mpC3.d.ts → analyze-runs-BaziIw93.d.ts} +4 -4
  8. package/dist/{analyze-runs-DGr7mpC3.d.ts.map → analyze-runs-BaziIw93.d.ts.map} +1 -1
  9. package/dist/app-grade/index.d.ts +4 -4
  10. package/dist/benchmarks/index.d.ts +30 -5
  11. package/dist/benchmarks/index.d.ts.map +1 -1
  12. package/dist/benchmarks/index.js +103 -15
  13. package/dist/benchmarks/index.js.map +1 -1
  14. package/dist/campaign/index.d.ts +7 -7
  15. package/dist/campaign/index.js +1 -1
  16. package/dist/{campaign-C9QC27C2.js → campaign-Cqz2vk2p.js} +2 -2
  17. package/dist/{campaign-C9QC27C2.js.map → campaign-Cqz2vk2p.js.map} +1 -1
  18. package/dist/{chat-client-9XLgWrQ7.js → chat-client-PJEzoot9.js} +3 -3
  19. package/dist/chat-client-PJEzoot9.js.map +1 -0
  20. package/dist/cli.js +2 -2
  21. package/dist/{cluster-CkzP4h3T.d.ts → cluster-CSd0LE_M.d.ts} +3 -3
  22. package/dist/{cluster-CkzP4h3T.d.ts.map → cluster-CSd0LE_M.d.ts.map} +1 -1
  23. package/dist/{completion-verifier-BYLOcQt7.js → completion-verifier-DVZx0x6x.js} +3 -3
  24. package/dist/{completion-verifier-BYLOcQt7.js.map → completion-verifier-DVZx0x6x.js.map} +1 -1
  25. package/dist/{completion-verifier-CJVDPv2T.d.ts → completion-verifier-oAgSwdFe.d.ts} +3 -3
  26. package/dist/{completion-verifier-CJVDPv2T.d.ts.map → completion-verifier-oAgSwdFe.d.ts.map} +1 -1
  27. package/dist/contract/index.d.ts +10 -10
  28. package/dist/contract/index.js +1 -1
  29. package/dist/{default-registry-C_Akmpui.d.ts → default-registry-DoIBhuS-.d.ts} +7 -7
  30. package/dist/{default-registry-C_Akmpui.d.ts.map → default-registry-DoIBhuS-.d.ts.map} +1 -1
  31. package/dist/{define-agent-eval-nNPC7cD5.d.ts → define-agent-eval-Be2Ser8j.d.ts} +7 -7
  32. package/dist/{define-agent-eval-nNPC7cD5.d.ts.map → define-agent-eval-Be2Ser8j.d.ts.map} +1 -1
  33. package/dist/diagnosis/index.d.ts +2 -2
  34. package/dist/{engine-B0gr15-l.d.ts → engine-CPBhCLJR.d.ts} +2 -2
  35. package/dist/{engine-B0gr15-l.d.ts.map → engine-CPBhCLJR.d.ts.map} +1 -1
  36. package/dist/{evaluation-mk3u_ZJC.d.ts → evaluation-BjRgUyst.d.ts} +4 -4
  37. package/dist/{evaluation-mk3u_ZJC.d.ts.map → evaluation-BjRgUyst.d.ts.map} +1 -1
  38. package/dist/evaluation.d.ts +1 -1
  39. package/dist/{exact-types-Blq7YvsK.d.ts → exact-types-DEnrf7NX.d.ts} +2 -2
  40. package/dist/{exact-types-Blq7YvsK.d.ts.map → exact-types-DEnrf7NX.d.ts.map} +1 -1
  41. package/dist/experiment/index.d.ts +3 -3
  42. package/dist/{feedback-trajectory-D3oISeGl.d.ts → feedback-trajectory-DxkG5EPi.d.ts} +3 -3
  43. package/dist/{feedback-trajectory-D3oISeGl.d.ts.map → feedback-trajectory-DxkG5EPi.d.ts.map} +1 -1
  44. package/dist/{index-CxuK3sxA.d.ts → index-AjfEHHA8.d.ts} +8 -8
  45. package/dist/{index-CxuK3sxA.d.ts.map → index-AjfEHHA8.d.ts.map} +1 -1
  46. package/dist/{index-cp5gkylN.d.ts → index-BI9ecONy.d.ts} +10 -10
  47. package/dist/{index-cp5gkylN.d.ts.map → index-BI9ecONy.d.ts.map} +1 -1
  48. package/dist/{index-BdFlQP54.d.ts → index-BriU7rfL.d.ts} +3 -3
  49. package/dist/{index-BdFlQP54.d.ts.map → index-BriU7rfL.d.ts.map} +1 -1
  50. package/dist/{index-Dodw93oW.d.ts → index-C8a9Dyan.d.ts} +23 -23
  51. package/dist/{index-Dodw93oW.d.ts.map → index-C8a9Dyan.d.ts.map} +1 -1
  52. package/dist/{index-IE8O0HNr.d.ts → index-D-0mg-eu.d.ts} +8 -8
  53. package/dist/{index-IE8O0HNr.d.ts.map → index-D-0mg-eu.d.ts.map} +1 -1
  54. package/dist/{index-BN6htPVi.d.ts → index-DHWFYUVW.d.ts} +2 -2
  55. package/dist/{index-BN6htPVi.d.ts.map → index-DHWFYUVW.d.ts.map} +1 -1
  56. package/dist/{index-KAYaJtMA.d.ts → index-VSN1IZQp.d.ts} +3 -3
  57. package/dist/{index-KAYaJtMA.d.ts.map → index-VSN1IZQp.d.ts.map} +1 -1
  58. package/dist/index.d.ts +27 -27
  59. package/dist/index.js +5 -5
  60. package/dist/{integrity-zeWMeTQt.d.ts → integrity-BMyCpnZ8.d.ts} +2 -2
  61. package/dist/{integrity-zeWMeTQt.d.ts.map → integrity-BMyCpnZ8.d.ts.map} +1 -1
  62. package/dist/{matrix-Diw3yJCG.d.ts → matrix-DdniPG_M.d.ts} +2 -2
  63. package/dist/{matrix-Diw3yJCG.d.ts.map → matrix-DdniPG_M.d.ts.map} +1 -1
  64. package/dist/meta-eval/index.d.ts +2 -2
  65. package/dist/multishot/golden/index.d.ts +1 -1
  66. package/dist/multishot/index.d.ts +3 -3
  67. package/dist/multishot/index.js +2 -2
  68. package/dist/openapi.json +1 -1
  69. package/dist/{paired-promotion-decision-COeOgNHG.d.ts → paired-promotion-decision-DawN2o0G.d.ts} +2 -2
  70. package/dist/{paired-promotion-decision-COeOgNHG.d.ts.map → paired-promotion-decision-DawN2o0G.d.ts.map} +1 -1
  71. package/dist/{prime-protocol-bRbh6dqd.d.ts → prime-protocol-RUPtoYiI.d.ts} +2 -2
  72. package/dist/{prime-protocol-bRbh6dqd.d.ts.map → prime-protocol-RUPtoYiI.d.ts.map} +1 -1
  73. package/dist/{promotion-gate-XglbDCNo.d.ts → promotion-gate-DG_PSJx_.d.ts} +2 -2
  74. package/dist/{promotion-gate-XglbDCNo.d.ts.map → promotion-gate-DG_PSJx_.d.ts.map} +1 -1
  75. package/dist/{registry-Pfeqsr5k.d.ts → registry-BqoX13FD.d.ts} +4 -4
  76. package/dist/{registry-Pfeqsr5k.d.ts.map → registry-BqoX13FD.d.ts.map} +1 -1
  77. package/dist/reporting.d.ts +4 -4
  78. package/dist/{researcher-DlyjP5uq.d.ts → researcher-BuojS14L.d.ts} +4 -4
  79. package/dist/{researcher-DlyjP5uq.d.ts.map → researcher-BuojS14L.d.ts.map} +1 -1
  80. package/dist/rl.d.ts +7 -7
  81. package/dist/rollout/index.d.ts +2 -2
  82. package/dist/rollout/index.js +1 -1
  83. package/dist/{rollout-Dh5q6CPA.js → rollout-B86tqluQ.js} +2 -2
  84. package/dist/{rollout-Dh5q6CPA.js.map → rollout-B86tqluQ.js.map} +1 -1
  85. package/dist/search/index.d.ts +2 -2
  86. package/dist/{search-kernel-6P1dRpJ4.d.ts → search-kernel-C4XnLKIW.d.ts} +3 -3
  87. package/dist/{search-kernel-6P1dRpJ4.d.ts.map → search-kernel-C4XnLKIW.d.ts.map} +1 -1
  88. package/dist/{sequential-OeSwJ2uG.d.ts → sequential-1-T2oi1_.d.ts} +2 -2
  89. package/dist/{sequential-OeSwJ2uG.d.ts.map → sequential-1-T2oi1_.d.ts.map} +1 -1
  90. package/dist/{server-BorE0Y3O.js → server-BQbd__Hr.js} +3 -3
  91. package/dist/{server-BorE0Y3O.js.map → server-BQbd__Hr.js.map} +1 -1
  92. package/dist/{statistical-heldout-vjrs8iyO.d.ts → statistical-heldout-Bx1QUDNy.d.ts} +5 -5
  93. package/dist/{statistical-heldout-vjrs8iyO.d.ts.map → statistical-heldout-Bx1QUDNy.d.ts.map} +1 -1
  94. package/dist/statistics.d.ts +2 -2
  95. package/dist/{store-tool-spans-BemsQEhd.d.ts → store-tool-spans-kCDPEwjL.d.ts} +5 -5
  96. package/dist/{store-tool-spans-BemsQEhd.d.ts.map → store-tool-spans-kCDPEwjL.d.ts.map} +1 -1
  97. package/dist/{summary-report-vdhE20Zf.d.ts → summary-report-9kwI6jFQ.d.ts} +2 -2
  98. package/dist/{summary-report-vdhE20Zf.d.ts.map → summary-report-9kwI6jFQ.d.ts.map} +1 -1
  99. package/dist/supervisor-run/index.d.ts +1 -1
  100. package/dist/{systemone-DFUN9DL1.d.ts → systemone-CSwi1hp-.d.ts} +7 -7
  101. package/dist/{systemone-DFUN9DL1.d.ts.map → systemone-CSwi1hp-.d.ts.map} +1 -1
  102. package/dist/{systemone-review-4CKmHvr5.d.ts → systemone-review-WpWosn3a.d.ts} +2 -2
  103. package/dist/{systemone-review-4CKmHvr5.d.ts.map → systemone-review-WpWosn3a.d.ts.map} +1 -1
  104. package/dist/systemone.d.ts +3 -3
  105. package/dist/{tool-groups-Dpl3picW.d.ts → tool-groups-Bcasvncd.d.ts} +3 -3
  106. package/dist/tool-groups-Bcasvncd.d.ts.map +1 -0
  107. package/dist/{trace-questions-oT2fZWJ8.d.ts → trace-questions-DI42jXPj.d.ts} +3 -3
  108. package/dist/{trace-questions-oT2fZWJ8.d.ts.map → trace-questions-DI42jXPj.d.ts.map} +1 -1
  109. package/dist/trace-repair/index.d.ts +1 -1
  110. package/dist/traces.d.ts +5 -5
  111. package/dist/{types-C9yEqDjQ.d.ts → types-ByrbgALT.d.ts} +4 -4
  112. package/dist/{types-C9yEqDjQ.d.ts.map → types-ByrbgALT.d.ts.map} +1 -1
  113. package/dist/{types-CP3rkpir.d.ts → types-CEgIBy0W.d.ts} +5 -4
  114. package/dist/{types-CP3rkpir.d.ts.map → types-CEgIBy0W.d.ts.map} +1 -1
  115. package/dist/{types-CwHU7U9J.d.ts → types-CLGrB4Sv.d.ts} +2 -2
  116. package/dist/{types-CwHU7U9J.d.ts.map → types-CLGrB4Sv.d.ts.map} +1 -1
  117. package/dist/{types-6GuPL-Vt.d.ts → types-DbGSC2No.d.ts} +2 -2
  118. package/dist/{types-6GuPL-Vt.d.ts.map → types-DbGSC2No.d.ts.map} +1 -1
  119. package/dist/wire/index.d.ts +4 -4
  120. package/dist/wire/index.js +1 -1
  121. package/docs/wire-protocol.md +1 -1
  122. package/package.json +1 -1
  123. package/dist/chat-client-9XLgWrQ7.js.map +0 -1
  124. package/dist/tool-groups-Dpl3picW.d.ts.map +0 -1
@@ -1 +1 @@
1
- {"version":3,"file":"index-BN6htPVi.d.ts","names":[],"sources":["../src/judge-calibration.ts","../src/statistics/agreement-irr.ts","../src/statistics/comparison-effects.ts","../src/statistics/descriptive.ts","../src/statistics/distribution-summary.ts","../src/statistics/effect-sizes.ts","../src/statistics/inference.ts","../src/statistics/multiplicity.ts","../src/statistics/paired-binary.ts","../src/statistics/paired-tests.ts","../src/statistics/power-and-mde.ts","../src/statistics/random.ts","../src/statistics/rank-tests.ts","../src/statistics/samples.ts","../src/statistics/sequential-eprocess.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;UAwBiB;EACf;EACA;;EAEA;;UAGe;EACf;EACA;;EAEA;;UAGe;EACf;EACA;;EAEA;;EAEA;;EAEA,YAAY;IAAQ;IAAgB;IAAe;IAAe;;;;;;iBAMpD,eACd,QAAQ,cACR,WAAW,mBACV;UA0Bc;;;;;EAKf;EACA;;;;;;iBAOc,eAAe,QAAQ,mBAAmB;UAgBzC;;EAEf;EACA;;iBAGc,cACd,SAAS;EAAQ;EAAmB;KACnC;UAYc;;EAEf;EACA;EACA;EACA;;;;;;;iBAQc,eACd,SAAS;EAAQ;EAAe;KAC/B;UA8Ec;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;IACE;IACA;;;EAGF;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;;;;;iBAUc,oBACd,oBACA,OAAM,6BACL;UAoDc,oCAAoC;;EAEnD;;EAEA;EACA;EACA;IACE;IACA;;;;;;;iBAQY,yBACd,QAAQ,cACR,WAAW,kBACX,OAAM,6BACL;;;;;;;;;;;;;;;;;;iBCtTa,sBAAsB,aAAa;UA2DlC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe,oCAAoC;EACnD;;EAEA;;EAEA;;UAGe;;EAEf,cAAc;;EAEd;;EAEA;;EAEA;;EAEA;;UAGe,+BAA+B;;EAE9C;;EAEA;;;;;;;;;;;;;;iBAec,0BACd,SAAS,qBACT,OAAM,yBACL;;;;;;;;;iBAgHa,yCACd,aAAa;EAAQ;EAAgB,QAAQ;IAC7C,OAAM,yBACL;;UAuBc;;EAEf;;EAEA;;;;;;;EAOA;;EAEA,SAAS;;UAGM;;EAEf;;EAEA;EACA;;;;;;iBAOc,iBACd,OAAO,0CACP,UAAS,0BACR;;iBC3Sa,OAAO,kBAAkB;;;;;;iBCEzB,aAAa;EAAU;EAAe;;;;;;;;;;iBAoBtC,mBACd,kBACA,qBACA;EAAQ;EAAe;;EACpB;EAAc;EAAe;;;;;;;iBAqClB,iBAAiB,YAAY;;UAW5B;WACN;;WAEA;;WAEA;WACA;WACA;;WAEA;;;;;;;;;iBAUK,uBACd,4BACA,sBACC;;iBAkCa,cAAc,iBAAiB;;UAM9B;WACN;WACA;WACA;WACA;WACA;WACA;WACA,cAAc,SAAS;WACvB;aACE;aACA;aACA;aACA;aACA;aACA;aACA;;;UAII;;EAEf;;EAEA;IAAmB;;;;;;;;;;iBAUL,sBACd,2BACA,UAAU,4BACT;;;;;iBA6Ea,MAAM;;;;;;iBAmBN,SAAS,aAAa;;;;;iBAuBtB,UAAU,aAAa;UAKtB;;EAEf,MAAM;;;;;EAKN,SAAS;;EAET;;UAGe;EACf;EACA;;;;;;;;;;;iBAYc,kBAAkB,OAAO,yBAAyB;;;;;;;;;UC/TjD;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;;;;;;;;iBAkBc,sBACd,kBACA;iBAQc,sBACd,kBACA,UAAS,6BACR;;;iBCtDa,QAAQ,aAAa;;iBAcrB,eAAe,kBAAkB;KAoBrC;;iBAGI,YAAY,kBAAkB;;iBAc9B,gBAAgB,gBAAgB;;iBClDhC,qBAAqB;;iBAMrB,uBAAuB,WAAW;;iBAsBlC,WAAW,WAAW,WAAW,WAAW;;iBAc5C,iBAAiB;;iBAQjB,oBAAoB,WAAW,WAAW,WAAW;;iBAqBrD,eAAe,WAAW,WAAW,WAAW;UAQ/C;EACf;EACA;EACA;;;iBAIc,sBACd;EAAmB;EAAW;MAC7B;;;;;;;;;;;;;iBC/Ea,WACd,4BACA;EACG;EAAoB;;;;;;;;;;iBAgBT,KACd,4BACA;EACG;EAAoB;;;;;;;;;iBA4BT,kBACd,4BACA;EACG;EAAmB;;;UChEP;;EAEf;;EAEA;;EAEA;;;;;;;;;;;;iBAac,OAAO,mBAAmB,WAAW,sBAAoB;UAwBxD;EACf,QAAQ;IAAuB;IAAmB;;EAClD,OAAO;IAAuB;IAAmB;;;EAEjD;EACA;EACA;EACA;;;;;iBAMc,0BACd,yBACA,iBACA,wBACA,gBACA,sBACC;;UAmCc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;iBAec,QACd,SAAS,6BACT,WAAW,8BACV;;UAmBc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;;iBAmBc,qBACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;;UAkCc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAgCc,0BACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;UAyEc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA8Dc,0BACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;iBAgEa,kBACd,QAAQ,mBACR,OAAO;;;;;;;;;;;iBA0BO,QAAQ,WAAW,WAAW;;;;;;UChgB7B;;EAEf;EACA;;EAEA;;;;;;;;;;;;;;;;iBAiBc,YAAY,kBAAkB,kBAAkB;UA2B/C;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;cAaW;UAEI;;EAEf;;EAEA;;EAEA;;;EAGA;;;;;;;;;;;;iBAac,gBACd,kBACA,iBACA,OAAM,yBACL;;KA0DS;;UAGK;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,aAAa;;EAEb;;;;;;;;;;;;;;;iBAgBc,eACd,gCACA,aAAa,sBACZ;;;iBAqCa,uBACd,QAAQ,mBACR,OAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;cA4CI;;;;;;;iBC1RG,mBAAmB;EACjC;EACA;EACA;EACA;;;;;;;;;;;iBAsBc,yBAAyB;EACvC;EACA;EACA;EACA;;;;;;;;iBAkBc,UAAU;EACxB;EACA;EACA;EACA;;;;;;;;;;;;;;;;iBAyBc,iBAAiB;EAC/B;EACA;EACA;EACA;EACA;;;;;;;iBAyBc,aAAa;EAC3B;EACA;EACA;EACA;EACA;;;;;;iBCrHc,WAAW;KCMf;;KAGA;UAEK;;EAEf,SAAS;;EAET;;;EAGA;;UAYe;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER;;;;;;;;;;;;;;iBAec,aACd,aACA,aACA,OAAM,kBACL;UAuFc;;;EAGf;;EAEA;;EAEA,QAAQ;;EAER;;EAEA;;;;;;;;;;;;;;;iBAgBc,mBACd,kBACA,iBACA,OAAM,kBACL;UCnLc;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;IAAkB;IAAa;IAAc;IAAoB;;;;iBAInD,aAAa;;iBAMb,gBACd,2BACA,UAAS,uBACR;UCnCc;;;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"}
1
+ {"version":3,"file":"index-DHWFYUVW.d.ts","names":[],"sources":["../src/judge-calibration.ts","../src/statistics/agreement-irr.ts","../src/statistics/comparison-effects.ts","../src/statistics/descriptive.ts","../src/statistics/distribution-summary.ts","../src/statistics/effect-sizes.ts","../src/statistics/inference.ts","../src/statistics/multiplicity.ts","../src/statistics/paired-binary.ts","../src/statistics/paired-tests.ts","../src/statistics/power-and-mde.ts","../src/statistics/random.ts","../src/statistics/rank-tests.ts","../src/statistics/samples.ts","../src/statistics/sequential-eprocess.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;UAwBiB;EACf;EACA;;EAEA;;UAGe;EACf;EACA;;EAEA;;UAGe;EACf;EACA;;EAEA;;EAEA;;EAEA,YAAY;IAAQ;IAAgB;IAAe;IAAe;;;;;;iBAMpD,eACd,QAAQ,cACR,WAAW,mBACV;UA0Bc;;;;;EAKf;EACA;;;;;;iBAOc,eAAe,QAAQ,mBAAmB;UAgBzC;;EAEf;EACA;;iBAGc,cACd,SAAS;EAAQ;EAAmB;KACnC;UAYc;;EAEf;EACA;EACA;EACA;;;;;;;iBAQc,eACd,SAAS;EAAQ;EAAe;KAC/B;UA8Ec;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;IACE;IACA;;;EAGF;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;;;;;iBAUc,oBACd,oBACA,OAAM,6BACL;UAoDc,oCAAoC;;EAEnD;;EAEA;EACA;EACA;IACE;IACA;;;;;;;iBAQY,yBACd,QAAQ,cACR,WAAW,kBACX,OAAM,6BACL;;;;;;;;;;;;;;;;;;iBCtTa,sBAAsB,aAAa;UA2DlC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe,oCAAoC;EACnD;;EAEA;;EAEA;;UAGe;;EAEf,cAAc;;EAEd;;EAEA;;EAEA;;EAEA;;UAGe,+BAA+B;;EAE9C;;EAEA;;;;;;;;;;;;;;iBAec,0BACd,SAAS,qBACT,OAAM,yBACL;;;;;;;;;iBAgHa,yCACd,aAAa;EAAQ;EAAgB,QAAQ;IAC7C,OAAM,yBACL;;UAuBc;;EAEf;;EAEA;;;;;;;EAOA;;EAEA,SAAS;;UAGM;;EAEf;;EAEA;EACA;;;;;;iBAOc,iBACd,OAAO,0CACP,UAAS,0BACR;;iBC3Sa,OAAO,kBAAkB;;;;;;iBCEzB,aAAa;EAAU;EAAe;;;;;;;;;;iBAoBtC,mBACd,kBACA,qBACA;EAAQ;EAAe;;EACpB;EAAc;EAAe;;;;;;;iBAqClB,iBAAiB,YAAY;;UAW5B;WACN;;WAEA;;WAEA;WACA;WACA;;WAEA;;;;;;;;;iBAUK,uBACd,4BACA,sBACC;;iBAkCa,cAAc,iBAAiB;;UAM9B;WACN;WACA;WACA;WACA;WACA;WACA;WACA,cAAc,SAAS;WACvB;aACE;aACA;aACA;aACA;aACA;aACA;aACA;;;UAII;;EAEf;;EAEA;IAAmB;;;;;;;;;;iBAUL,sBACd,2BACA,UAAU,4BACT;;;;;iBA6Ea,MAAM;;;;;;iBAmBN,SAAS,aAAa;;;;;iBAuBtB,UAAU,aAAa;UAKtB;;EAEf,MAAM;;;;;EAKN,SAAS;;EAET;;UAGe;EACf;EACA;;;;;;;;;;;iBAYc,kBAAkB,OAAO,yBAAyB;;;;;;;;;UC/TjD;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;;;;;;;;iBAkBc,sBACd,kBACA;iBAQc,sBACd,kBACA,UAAS,6BACR;;;iBCtDa,QAAQ,aAAa;;iBAcrB,eAAe,kBAAkB;KAoBrC;;iBAGI,YAAY,kBAAkB;;iBAc9B,gBAAgB,gBAAgB;;iBClDhC,qBAAqB;;iBAMrB,uBAAuB,WAAW;;iBAsBlC,WAAW,WAAW,WAAW,WAAW;;iBAc5C,iBAAiB;;iBAQjB,oBAAoB,WAAW,WAAW,WAAW;;iBAqBrD,eAAe,WAAW,WAAW,WAAW;UAQ/C;EACf;EACA;EACA;;;iBAIc,sBACd;EAAmB;EAAW;MAC7B;;;;;;;;;;;;;iBC/Ea,WACd,4BACA;EACG;EAAoB;;;;;;;;;;iBAgBT,KACd,4BACA;EACG;EAAoB;;;;;;;;;iBA4BT,kBACd,4BACA;EACG;EAAmB;;;UChEP;;EAEf;;EAEA;;EAEA;;;;;;;;;;;;iBAac,OAAO,mBAAmB,WAAW,sBAAoB;UAwBxD;EACf,QAAQ;IAAuB;IAAmB;;EAClD,OAAO;IAAuB;IAAmB;;;EAEjD;EACA;EACA;EACA;;;;;iBAMc,0BACd,yBACA,iBACA,wBACA,gBACA,sBACC;;UAmCc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;iBAec,QACd,SAAS,6BACT,WAAW,8BACV;;UAmBc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;;iBAmBc,qBACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;;UAkCc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAgCc,0BACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;UAyEc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA8Dc,0BACd,SAAS,6BACT,WAAW,6BACX,sBACC;;;iBAgEa,kBACd,QAAQ,mBACR,OAAO;;;;;;;;;;;iBA0BO,QAAQ,WAAW,WAAW;;;;;;UChgB7B;;EAEf;EACA;;EAEA;;;;;;;;;;;;;;;;iBAiBc,YAAY,kBAAkB,kBAAkB;UA2B/C;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;cAaW;UAEI;;EAEf;;EAEA;;EAEA;;;EAGA;;;;;;;;;;;;iBAac,gBACd,kBACA,iBACA,OAAM,yBACL;;KA0DS;;UAGK;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,aAAa;;EAEb;;;;;;;;;;;;;;;iBAgBc,eACd,gCACA,aAAa,sBACZ;;;iBAqCa,uBACd,QAAQ,mBACR,OAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;cA4CI;;;;;;;iBC1RG,mBAAmB;EACjC;EACA;EACA;EACA;;;;;;;;;;;iBAsBc,yBAAyB;EACvC;EACA;EACA;EACA;;;;;;;;iBAkBc,UAAU;EACxB;EACA;EACA;EACA;;;;;;;;;;;;;;;;iBAyBc,iBAAiB;EAC/B;EACA;EACA;EACA;EACA;;;;;;;iBAyBc,aAAa;EAC3B;EACA;EACA;EACA;EACA;;;;;;iBCrHc,WAAW;KCMf;;KAGA;UAEK;;EAEf,SAAS;;EAET;;;EAGA;;UAYe;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER;;;;;;;;;;;;;;iBAec,aACd,aACA,aACA,OAAM,kBACL;UAuFc;;;EAGf;;EAEA;;EAEA,QAAQ;;EAER;;EAEA;;;;;;;;;;;;;;;iBAgBc,mBACd,kBACA,iBACA,OAAM,kBACL;UCnLc;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;IAAkB;IAAa;IAAc;IAAoB;;;;iBAInD,aAAa;;iBAMb,gBACd,2BACA,UAAS,uBACR;UCnCc;;;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"}
@@ -1,7 +1,7 @@
1
1
  import { a as JudgeError } from "./errors-E9I2hldP.js";
2
2
  import { f as Run } from "./schema-CkIvnmLU.js";
3
- import { U as LlmClientOptions } from "./cluster-CkzP4h3T.js";
4
- import { Bt as CalibrationResult, Gt as GoldenItem, Ht as ContinuousAgreement, Mt as CorpusAgreementReport, Vt as CandidateScore, Wt as ContinuousCalibrationResult } from "./index-BN6htPVi.js";
3
+ import { U as LlmClientOptions } from "./cluster-CSd0LE_M.js";
4
+ import { Bt as CalibrationResult, Gt as GoldenItem, Ht as ContinuousAgreement, Mt as CorpusAgreementReport, Vt as CandidateScore, Wt as ContinuousCalibrationResult } from "./index-DHWFYUVW.js";
5
5
  import { r as LedgerHash } from "./canonical-gLeHiqzd.js";
6
6
  import { s as TraceStore } from "./store-CMs5wGXP.js";
7
7
  import { h as OutcomeStore, m as OutcomeFilter } from "./rubric-predictive-validity-Bbi88Aut.js";
@@ -1078,4 +1078,4 @@ interface CatchRateReport {
1078
1078
  */
1079
1079
  declare function catchRate(results: readonly PlantOutcome[], manifest: PlantManifest): CatchRateReport;
1080
1080
  export { measureJudgeAgreement as $, CorrelationStudyOptions as $t, JudgePanelOutput as A, ProbabilityPolicyAuditInput as At, JudgeConfusion as B, clusterBootstrapMean as Bt, JudgeCalibrationCellVerdict as C, EvaluatorAdmissionPolicy as Ct, JudgeCalibrationGoldItem as D, EvaluatorAuditPlan as Dt, JudgeCalibrationGateConfig as E, EvaluatorAuditObservation as Et, CalibrationExample as F, evaluatorAuditObservationSchema as Ft, JudgeGateRefusedError as G, CategoricalForecast as Gt, JudgeGateDecision as H, CalibrationOptions as Ht, CalibrationMerge as I, planEvaluatorAudit as It, JudgeRunMeta as J, calibrationCurve as Jt, JudgeGateStatus as K, OrdinalForecast as Kt, DEFAULT_JUDGE_GATE_POLICY as L, ClusterBootstrapOptions as Lt, RubricDimensionGuide as M, auditEvaluator as Mt, calibrateJudgePanel as N, auditProbabilityPolicy as Nt, JudgeDimensionCalibration as O, EvaluatorAuditPlanInput as Ot, gateJudgeCalibration as P, evaluatorAdmissionPolicySchema as Pt, judgeGateDecision as Q, CorrelationResult as Qt, JudgeAgreement as R, ClusterBootstrapResult as Rt, CalibrateJudgePanelInput as S, analyzeSeries as St, JudgeCalibrationGateCell as T, EvaluatorAuditInput as Tt, JudgeGateInput as U, CalibrationPair as Ut, JudgeExampleAgreement as V, CalibrationBin as Vt, JudgeGatePolicy as W, CalibrationReport as Wt, addCalibrationVerdicts as X, multiclassBrierScore as Xt, OwnerVerdict as Y, calibrationFromPairs as Yt, assertJudgeMayGate as Z, rankedProbabilityScore as Zt, catchRate as _, snapshotFromCalibration as _t, PerturbedField as a, SentinelMetrics as at, plantByPerturbation as b, SeriesConvergenceOptions as bt, PlantEvidence as c, SentinelSnapshot as ct, PlantKindCounts as d, SentinelTrend as dt, CorrelationStudyResult as en, registerCalibrationSet as et, PlantManifest as f, SnapshotMeta as ft, UnseededRejection as g, snapshotFromAgreement as gt, SeededGradingSet as h, judgeSentinelReport as ht, PerturbationKind as i, SentinelMetricName as it, JudgePanelPrompt as j, ProbabilityPolicyAuditReport as jt, JudgePanelCalibration as k, EvaluatorErrorRate as kt, PlantExpectation as l, SentinelStore as lt, SeedPlantsOptions as m, fileSentinelStore as mt, CatchRateStatus as n, OutcomePair as nn, EvalHealthStamp as nt, PerturbedPlant as o, SentinelReport as ot, PlantOutcome as p, evalHealthStamp as pt, JudgeRun as q, brierScore as qt, EvidencePerturbation as r, correlationStudy as rn, JudgeSentinelOptions as rt, Plant as s, SentinelSetOptions as st, CatchRateReport as t, EvalMetricSpec as tn, snapshotFromJudgeAgreement as tt, PlantKind as u, SentinelThresholds as ut, definePlant as v, snapshotFromSentinelSet as vt, JudgeCalibrationGate as w, EvaluatorAdmissionReport as wt, seedPlants as x, SeriesConvergenceResult as xt, perturbEvidence as y, validateSentinelSnapshot as yt, JudgeCalibrationSet as z, ClusteredObservation as zt };
1081
- //# sourceMappingURL=index-KAYaJtMA.d.ts.map
1081
+ //# sourceMappingURL=index-VSN1IZQp.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"index-KAYaJtMA.d.ts","names":[],"sources":["../src/meta-eval/correlation-study.ts","../src/meta-eval/calibration.ts","../src/meta-eval/cluster-bootstrap.ts","../src/meta-eval/evaluator-admission.ts","../src/series-convergence.ts","../src/meta-eval/sentinel.ts","../src/meta-eval/judge-gate.ts","../src/meta-eval/judge-panel-calibration.ts","../src/meta-eval/plants.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;;;;;;iBA+Da,WAAW,gBAAgB;;UAkB1B;;EAEf,eAAe,SAAS;EACxB;;;;;;;;;iBAUc,qBAAqB,oBAAoB;;UAaxC;EACf;;EAEA;;;;;;;;iBASc,uBAAuB,oBAAoB;;UCzN1C;EACf;EACA;EACA;;UAGe;;EAEf;;EAEA;;EAEA;;UAGe;EACf;EACA;;EAEA;EACA;EACA;EACA;EACA;;;iBAIc,qBACd,uBAAuB,wBACvB,UAAS,0BACR;cC1BU,gCAA8B,EAAA;;;;GAMhC,EAAA,KAAA;;KAGC,2BAA2B,EAAE,aAAa;cAEzC,iCAA+B,EAAA;;;;;;;;;;;;;;;;;GASjC,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;;cAkGZ,gBAAc,EAAA;;;;;;;;;;;;;;;;;;GAOT,EAAA,KAAA;KAEC,0BAA0B,EAAE,aAAa;;;;;;iBA0BrC,mBAAmB,OAAO;;;;;;;;;;;;;;;;;KAiC9B,qBAAqB,kBAAkB;;;;;;iBAOnC,eAAe,OAAO,sBAAsB;cAoFtD,wBAAsB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAG1B,EAAA,KAAA;;KAGU,8BAA8B,EAAE,aAAa;UAExC,qCAAqC;EACpD;IACE,uBAAuB;IACvB,YAAY;;IAEZ,cAAc;MAAQ;MAAY;;;;;;;;;iBAStB,uBACd,OAAO,8BACN;;;;;;;;;;;UC7Wc;;EAEf;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;iBAGc,cACd,kBACA,UAAS,2BACR;cCoBG;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;;;;;;;;iBASrB,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;KCha7C;;UAGK;;EAEf;EACA,SAAS;;EAET;;EAEA;;EAEA;;EAEA;EACA;;UAGe;EACf;;EAEA;EACA,mBAAmB;;;;;;iBAuDL,uBAAuB,KAAK,sBAAsB;UAwBjD;EACf,KAAK;;EAEL;;EAEA;;EAEA;;;;;;;;iBASc,uBACd,KAAK,qBACL,mBAAmB,uBAClB;;UAkCc;EACf;;EAEA,SAAS;;EAET,mBAAmB;;UAGJ;EACf;;EAEA;EACA;EACA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;EACf;EACA,SAAS;EACT;;EAEA;;EAEA;;UAGe,uBAAuB;EACtC;;EAEA;;EAEA;;EAEA,iBAAiB,OAAO;EACxB;;EAEA;;;;;;EAMA;IAAmB;IAAgB;IAAgB,UAAU;;;;;;EAK7D;;EAEA;;EAEA;EACA,WAAW;EACX;;IAEE;;IAEA;;IAEA;;;EAGF,eAAe;;EAEf;;;;;;;iBAyBc,sBACd,KAAK,qBACL,eAAe,YACf,MAAM,eACL;UA+Gc;;;;;EAKf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,QAAQ;;cAGG,2BAA2B,SAAS;KAUrC;UAEK;EACf;EACA;EACA;;EAEA;EACA,QAAQ;;EAER;EACA,WAAW;EACX,QAAQ;;UAGO;EACf;;EAEA;;EAEA;;EAEA,WAAW;EACX;;EAEA,mBAAmB;EACnB,SAAS,QAAQ;;;iBAUH,kBAAkB,OAAO,iBAAiB;;cA2G7C,8BAA8B;WAChC,UAAU;EAEP,YAAA,UAAU;;;iBASR,mBAAmB,UAAU;;;;;;;;iBAW7B,2BACd,WAAW,gBACX;EAAW,UAAU;IACpB;;UCtkBc;EACf;EACA;;UAGe,yBAAyB;EACxC;;EAEA;;EAEA;;EAEA;;;EAGA,YAAY,QAAQ,OAAO;;UAGZ;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;UAGe,yBAAyB;EACxC,QAAQ,SAAS,OAAO,GAAG;;EAE3B,sBAAsB;EACtB,gBAAgB,yBAAyB;EACzC;;EAEA,KAAK;EACL,QAAQ;;EAER;;EAEA;;EAEA;;EAEA,OAAO;;;;UAKQ;EACf,aAAa;EACb,QAAQ;EACR,QAAQ;;;KAIE,iBAAiB,oBAAoB,QAAQ,OAAO;EAAgB;;UAE/D,sBAAsB;EACrC;EACA,YAAY;EACZ,UAAU,eAAe,OAAO,GAAG;;EAEnC,WAAW,eAAe,OAAO,GAAG;;EAEpC,SAAS,eAAe,eAAe,iBAAiB;;EAExD,eAAe,OAAO;;EAEtB,UAAU;IAAQ;IAAe;IAAgB;;EACjD;;;iBAiDoB,oBAAoB,kBACxC,OAAO,yBAAyB,KAC/B,QAAQ,sBAAsB;;UA0GhB;;EAEf;;EAEA;;EAEA;;EAEA;;KAGU;UAEK;EACf;EACA;EACA;EACA;EACA;EACA,SAAS;EACT;;UAGe;EACf;EACA,QAAQ;EACR;EACA;EACA;EACA,OAAO;;;iBAIO,qBACd,UAAU,eAAe;EAAiB,aAAa;KACvD,QAAQ,6BACP;;;;;;KC1OS;;;;;;;;;;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"}
1
+ {"version":3,"file":"index-VSN1IZQp.d.ts","names":[],"sources":["../src/meta-eval/correlation-study.ts","../src/meta-eval/calibration.ts","../src/meta-eval/cluster-bootstrap.ts","../src/meta-eval/evaluator-admission.ts","../src/series-convergence.ts","../src/meta-eval/sentinel.ts","../src/meta-eval/judge-gate.ts","../src/meta-eval/judge-panel-calibration.ts","../src/meta-eval/plants.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;;;;;;iBA+Da,WAAW,gBAAgB;;UAkB1B;;EAEf,eAAe,SAAS;EACxB;;;;;;;;;iBAUc,qBAAqB,oBAAoB;;UAaxC;EACf;;EAEA;;;;;;;;iBASc,uBAAuB,oBAAoB;;UCzN1C;EACf;EACA;EACA;;UAGe;;EAEf;;EAEA;;EAEA;;UAGe;EACf;EACA;;EAEA;EACA;EACA;EACA;EACA;;;iBAIc,qBACd,uBAAuB,wBACvB,UAAS,0BACR;cC1BU,gCAA8B,EAAA;;;;GAMhC,EAAA,KAAA;;KAGC,2BAA2B,EAAE,aAAa;cAEzC,iCAA+B,EAAA;;;;;;;;;;;;;;;;;GASjC,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;;cAkGZ,gBAAc,EAAA;;;;;;;;;;;;;;;;;;GAOT,EAAA,KAAA;KAEC,0BAA0B,EAAE,aAAa;;;;;;iBA0BrC,mBAAmB,OAAO;;;;;;;;;;;;;;;;;KAiC9B,qBAAqB,kBAAkB;;;;;;iBAOnC,eAAe,OAAO,sBAAsB;cAoFtD,wBAAsB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAG1B,EAAA,KAAA;;KAGU,8BAA8B,EAAE,aAAa;UAExC,qCAAqC;EACpD;IACE,uBAAuB;IACvB,YAAY;;IAEZ,cAAc;MAAQ;MAAY;;;;;;;;;iBAStB,uBACd,OAAO,8BACN;;;;;;;;;;;UC7Wc;;EAEf;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;iBAGc,cACd,kBACA,UAAS,2BACR;cCoBG;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;;;;;;;;iBASrB,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;KCha7C;;UAGK;;EAEf;EACA,SAAS;;EAET;;EAEA;;EAEA;;EAEA;EACA;;UAGe;EACf;;EAEA;EACA,mBAAmB;;;;;;iBAuDL,uBAAuB,KAAK,sBAAsB;UAwBjD;EACf,KAAK;;EAEL;;EAEA;;EAEA;;;;;;;;iBASc,uBACd,KAAK,qBACL,mBAAmB,uBAClB;;UAkCc;EACf;;EAEA,SAAS;;EAET,mBAAmB;;UAGJ;EACf;;EAEA;EACA;EACA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;EACf;EACA,SAAS;EACT;;EAEA;;EAEA;;UAGe,uBAAuB;EACtC;;EAEA;;EAEA;;EAEA,iBAAiB,OAAO;EACxB;;EAEA;;;;;;EAMA;IAAmB;IAAgB;IAAgB,UAAU;;;;;;EAK7D;;EAEA;;EAEA;EACA,WAAW;EACX;;IAEE;;IAEA;;IAEA;;;EAGF,eAAe;;EAEf;;;;;;;iBAyBc,sBACd,KAAK,qBACL,eAAe,YACf,MAAM,eACL;UA+Gc;;;;;EAKf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,QAAQ;;cAGG,2BAA2B,SAAS;KAUrC;UAEK;EACf;EACA;EACA;;EAEA;EACA,QAAQ;;EAER;EACA,WAAW;EACX,QAAQ;;UAGO;EACf;;EAEA;;EAEA;;EAEA,WAAW;EACX;;EAEA,mBAAmB;EACnB,SAAS,QAAQ;;;iBAUH,kBAAkB,OAAO,iBAAiB;;cA2G7C,8BAA8B;WAChC,UAAU;EAEP,YAAA,UAAU;;;iBASR,mBAAmB,UAAU;;;;;;;;iBAW7B,2BACd,WAAW,gBACX;EAAW,UAAU;IACpB;;UCtkBc;EACf;EACA;;UAGe,yBAAyB;EACxC;;EAEA;;EAEA;;EAEA;;;EAGA,YAAY,QAAQ,OAAO;;UAGZ;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;UAGe,yBAAyB;EACxC,QAAQ,SAAS,OAAO,GAAG;;EAE3B,sBAAsB;EACtB,gBAAgB,yBAAyB;EACzC;;EAEA,KAAK;EACL,QAAQ;;EAER;;EAEA;;EAEA;;EAEA,OAAO;;;;UAKQ;EACf,aAAa;EACb,QAAQ;EACR,QAAQ;;;KAIE,iBAAiB,oBAAoB,QAAQ,OAAO;EAAgB;;UAE/D,sBAAsB;EACrC;EACA,YAAY;EACZ,UAAU,eAAe,OAAO,GAAG;;EAEnC,WAAW,eAAe,OAAO,GAAG;;EAEpC,SAAS,eAAe,eAAe,iBAAiB;;EAExD,eAAe,OAAO;;EAEtB,UAAU;IAAQ;IAAe;IAAgB;;EACjD;;;iBAiDoB,oBAAoB,kBACxC,OAAO,yBAAyB,KAC/B,QAAQ,sBAAsB;;UA0GhB;;EAEf;;EAEA;;EAEA;;EAEA;;KAGU;UAEK;EACf;EACA;EACA;EACA;EACA;EACA,SAAS;EACT;;UAGe;EACf;EACA,QAAQ;EACR;EACA;EACA;EACA,OAAO;;;iBAIO,qBACd,UAAU,eAAe;EAAiB,aAAa;KACvD,QAAQ,6BACP;;;;;;KC1OS;;;;;;;;;;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"}
package/dist/index.d.ts CHANGED
@@ -1,45 +1,45 @@
1
1
  import { _ as PendingCostCall, a as CostLedgerFilter, c as CostLedgerPersistence, d as CostReceipt, f as CostReceiptInput, g as PaidCallResult, h as MaximumCharge, i as CostChannel, l as CostLedgerSummary, m as CustomTokenPricing, n as ChannelRollup, o as CostLedgerHandle, p as CostUsage, s as CostLedgerOptions, t as CostLedger, u as CostProvenance, v as PendingCostCallView, y as RunPaidCallInput } from "./cost-ledger-Czro9wmo.js";
2
2
  import { a as JudgeError, c as ValidationError, i as ConfigError, n as AgentEvalErrorCode, s as NotFoundError, t as AgentEvalError } from "./errors-E9I2hldP.js";
3
- import { $ as EnsembleAggregate, $i as makePolicyEditCandidateRecord, $n as ReviewInput, $r as pairedDeltaTest, $t as TraceContract, A as contentHash, Ai as AssertCapabilityHeadroomOptions, An as PromptRegistry, Ar as RuntimeTrajectoryHookEvent, At as checkCanaries, B as evaluateOracles, Bi as PolicyEditAdmissionOptions, Bn as ReflectionContext, Br as ProductBenchmarkExportResult, Bt as evaluateTraceContract, C as runAgentControlLoop, Ci as errorStreakDetector, Cn as aggregateRunScore, Cr as runBoundedProcess, Ct as DeployRunner, D as index_d_exports, Di as CostSummary, Dn as ReferenceReplayRun, Dr as ProjectRuntimeTrajectoryEvidenceOptions, Dt as viteDeployRunner, E as BudgetGuard, Ei as CostEntry, En as ReferenceReplayCaseRun, Er as proposeSynthesisTargets, Et as deployGateLayer, F as gateTreatmentApplied, Fi as assertCapabilityHeadroom, Fn as ModelEndpointCheck, Fr as ProductBenchmarkSplit, Ft as TraceContractSpec, G as KeywordCoverageResult, Gi as PolicyEditGainDirection, Gn as LlmReviewerConfig, Gr as HeldOutPartition, Gt as ContractRuleExecution, H as KeywordConceptSpec, Hi as PolicyEditCandidateRecord, Hn as TrialTrace, Hr as exportProductBenchmark, Ht as ArgumentCheck, I as Oracle, Ii as capabilityHeadroom, In as ModelEndpointRequest, Ir as ProductBenchmarkValidationReport, It as compileTraceContractSpec, J as EnsembleJudgeOptions, Ji as PolicyEditRisk, Jn as ProposeOutput, Jr as TrainSelectionTestPartition, Jt as ContractStatus, K as runKeywordCoverageJudge, Ki as PolicyEditGainUnit, Kn as ProposeFn, Kr as PartitionHeldOutOptions, Kt as ContractRun, L as OracleObservation, Li as FindingToPolicyEditOptions, Ln as preflightModels, Lr as assertProductBenchmarkRun, Lt as lintTraceContractSpec, M as TreatmentGate, Mi as CapabilityHeadroomResult, Mn as ModelSeats, Mr as RuntimeTrajectoryRunRecord, Mt as ActionPolicyDecision, N as TreatmentGateInput, Ni as HeadroomInput, Nn as resolveSeat, Nr as projectRuntimeTrajectoryEvidence, Nt as evaluateActionPolicy, O as VerdictCacheStore, Oi as CostTracker, On as ReferenceReplaySplit, Or as RuntimeTrajectoryEvidenceProjection, Ot as wranglerDeployRunner, P as TreatmentGateOptions, Pi as TaskHeadroom, Pn as seatPresets, Pr as ProductBenchmarkManifest, Pt as ContractLintFinding, Q as withJudgeRetry, Qi as PolicyEditTargetSurface, Qn as ReviewFn, Qr as minimumPairsForPairedDeltaTest, Qt as SpanPredicate, R as OracleReport, Ri as PolicyEdit, Rn as RunRecordBackend, Rr as readProductBenchmarkManifest, Rt as explainTraceContract, S as subjectiveEval, Si as StreamingDetector, Sn as RunScoreWeights, Sr as DEFAULT_BOUNDED_PROCESS_TIMEOUT_MS, St as DeployRunResult, T as BudgetBreachError, Ti as repeatedActionDetector, Tn as SteeringBundle, Tr as SynthesisTarget, Tt as WranglerDeployRunnerInput, U as KeywordCoverageFinding, Ui as PolicyEditChange, Un as buildReflectionPrompt, Ur as exportProductBenchmarkRuns, Ut as ContractCheckResult, V as jsonShape, Vi as PolicyEditAxis, Vn as ReflectionProposal, Vr as ProductBenchmarkSingleRunExportOptions, Vt as traceContract, W as KeywordCoverageOptions, Wi as PolicyEditExpectedGain, Wn as parseReflectionResponse, Wr as productBenchmarkRepoIdentity, Wt as ContractRule, X as JudgeRetryOutcome, Xi as PolicyEditSource, Xn as ProposeReviewReport, Xr as PairedDeltaTestOptions, Xt as OrderMode, Y as ensembleJudge, Yi as PolicyEditSchemaVersion, Yn as ProposeReviewConfig, Yr as partitionHeldOut, Yt as ContractVerdict, Z as JudgeRetryPolicy, Zi as PolicyEditTarget, Zn as Review, Zr as PairedDeltaTestResult, Zt as RetryWrite, _ as blockingKnowledgeEval, _i as extractErrorCount, _n as CounterfactualRunner, _r as RunCommandInput, _t as NoLeakOptions, a as assertHttpOk, aa as costReceiptFromLlmError, ai as KnowledgeBundle, an as ScorecardCellDiff, ar as inMemoryReviewStore, at as runIntentMatchJudge, b as userQuestionsForKnowledgeGaps, bi as ErrorStreakOptions, bn as SteeringOptimizationRow, br as BoundedProcessInput, bt as blendHeldout, c as assertStatus, ca as costForTokenPricing, ci as KnowledgeReadinessReport, cn as ScorecardLogLine, cr as ExperimentRep, ct as SingleBackendDivergence, d as MODEL_PRICING, da as CostCeilingReachedError, di as KnowledgeSensitivity, dn as loadScorecard, dr as ImprovementVerdictResult, dt as BlendWeights, ea as isTransientLlmError, ei as LeaderboardOptions, en as CellVerdict, er as ReviewMemoryEntry, et as JudgeVerdict, f as estimateCost, fi as UserQuestion, fn as recordRuns, fr as DiscoverPersonasOptions, ft as FieldDestination, g as acquisitionPlansForKnowledgeGaps, gi as ExtractResult, gn as CounterfactualResult, gr as DirEntry, gt as JudgeScoreInput, h as resolveModelPricing, hi as ExtractOptions, hn as CounterfactualMutation, hr as CommandRunner, ht as HiddenLeak, i as assertHasField, ia as costReceiptFromLlm, ii as KnowledgeAcquisitionMode, in as ScorecardCell, ir as createLlmReviewer, it as IntentMatchResult, j as fileVerdictCache, ji as CapabilityHeadroomOptions, jn as hashContent, jr as RuntimeTrajectoryRecord, jt as ActionExecutionPolicy, k as canonicalJson, ki as ScenarioCost, kn as PromptHandle, kr as RuntimeTrajectoryEvidenceSummary, kt as CanaryLeak, l as analyzeAntiSlop, la as costForUsage, li as KnowledgeRequirement, ln as diffScorecard, lr as ExperimentStats, lt as SingleBackendReport, m as isModelPriced, mi as ErrorCountPattern, mn as CounterfactualContext, mr as discoverPersonas, mt as HiddenGradeResult, n as assertEventCount, na as LlmCallError, ni as leaderboard, nn as RecordRunsOptions, nr as Verification, nt as IntentMatchInput, o as assertLatencyUnder, oa as maximumChargeForLlmRequest, oi as KnowledgeFreshness, on as ScorecardDiff, or as jsonlReviewStore, ot as AssertSingleBackendOptions, p as estimateTokens, pi as ERROR_COUNT_PATTERNS, pn as recordRunsToScorecard, pr as DiscoveredPersona, pt as HiddenCriteriaGrader, q as runKeywordCoverageJudgeUrl, qi as PolicyEditInit, qn as ProposeInput, qr as TrainSelectionTestOptions, qt as ContractSpan, r as assertExists, ra as LlmResponseError, ri as DataAcquisitionPlan, rn as Scorecard, rr as VerifyFn, rt as IntentMatchOptions, s as assertNotContains, sa as CostResult, si as KnowledgeImportance, sn as ScorecardEntry, sr as runProposeReview, st as BackendDescriptor, t as assertContains, ta as stripFencedJson, ti as LeaderboardRow, tn as DiffScorecardOptions, tr as ReviewMemoryStore, tt as aggregateJudgeVerdicts, u as createAntiSlopJudge, ua as CostAccountingIncompleteError, ui as KnowledgeRequirementCategory, un as formatScorecardDiff, ur as ImprovementThresholds, ut as assertSingleBackend, v as knowledgeReadinessTracePayload, vi as DetectorEvent, vn as PairwiseSteeringOptimizer, vr as RunCommandResult, vt as RoutedField, w as ProductClient, wi as observeAll, wn as clamp01, wr as ActiveLearningOptions, wt as ViteDeployRunnerInput, x as objectiveEval, xi as RepeatedActionOptions, xn as RunScore, xr as BoundedProcessResult, xt as DeployGateLayerInput, y as scoreKnowledgeReadiness, yi as DetectorSignal, yn as SteeringOptimizationResult, yr as localCommandRunner, yt as assertNoHiddenLeak, z as OracleResult, zi as PolicyEditAdmission, zn as jsonlRunRecordBackend, zr as ProductBenchmarkExportOptions, zt as checkTraceContracts } from "./index-Dodw93oW.js";
3
+ import { $ as EnsembleAggregate, $i as makePolicyEditCandidateRecord, $n as ReviewInput, $r as pairedDeltaTest, $t as TraceContract, A as contentHash, Ai as AssertCapabilityHeadroomOptions, An as PromptRegistry, Ar as RuntimeTrajectoryHookEvent, At as checkCanaries, B as evaluateOracles, Bi as PolicyEditAdmissionOptions, Bn as ReflectionContext, Br as ProductBenchmarkExportResult, Bt as evaluateTraceContract, C as runAgentControlLoop, Ci as errorStreakDetector, Cn as aggregateRunScore, Cr as runBoundedProcess, Ct as DeployRunner, D as index_d_exports, Di as CostSummary, Dn as ReferenceReplayRun, Dr as ProjectRuntimeTrajectoryEvidenceOptions, Dt as viteDeployRunner, E as BudgetGuard, Ei as CostEntry, En as ReferenceReplayCaseRun, Er as proposeSynthesisTargets, Et as deployGateLayer, F as gateTreatmentApplied, Fi as assertCapabilityHeadroom, Fn as ModelEndpointCheck, Fr as ProductBenchmarkSplit, Ft as TraceContractSpec, G as KeywordCoverageResult, Gi as PolicyEditGainDirection, Gn as LlmReviewerConfig, Gr as HeldOutPartition, Gt as ContractRuleExecution, H as KeywordConceptSpec, Hi as PolicyEditCandidateRecord, Hn as TrialTrace, Hr as exportProductBenchmark, Ht as ArgumentCheck, I as Oracle, Ii as capabilityHeadroom, In as ModelEndpointRequest, Ir as ProductBenchmarkValidationReport, It as compileTraceContractSpec, J as EnsembleJudgeOptions, Ji as PolicyEditRisk, Jn as ProposeOutput, Jr as TrainSelectionTestPartition, Jt as ContractStatus, K as runKeywordCoverageJudge, Ki as PolicyEditGainUnit, Kn as ProposeFn, Kr as PartitionHeldOutOptions, Kt as ContractRun, L as OracleObservation, Li as FindingToPolicyEditOptions, Ln as preflightModels, Lr as assertProductBenchmarkRun, Lt as lintTraceContractSpec, M as TreatmentGate, Mi as CapabilityHeadroomResult, Mn as ModelSeats, Mr as RuntimeTrajectoryRunRecord, Mt as ActionPolicyDecision, N as TreatmentGateInput, Ni as HeadroomInput, Nn as resolveSeat, Nr as projectRuntimeTrajectoryEvidence, Nt as evaluateActionPolicy, O as VerdictCacheStore, Oi as CostTracker, On as ReferenceReplaySplit, Or as RuntimeTrajectoryEvidenceProjection, Ot as wranglerDeployRunner, P as TreatmentGateOptions, Pi as TaskHeadroom, Pn as seatPresets, Pr as ProductBenchmarkManifest, Pt as ContractLintFinding, Q as withJudgeRetry, Qi as PolicyEditTargetSurface, Qn as ReviewFn, Qr as minimumPairsForPairedDeltaTest, Qt as SpanPredicate, R as OracleReport, Ri as PolicyEdit, Rn as RunRecordBackend, Rr as readProductBenchmarkManifest, Rt as explainTraceContract, S as subjectiveEval, Si as StreamingDetector, Sn as RunScoreWeights, Sr as DEFAULT_BOUNDED_PROCESS_TIMEOUT_MS, St as DeployRunResult, T as BudgetBreachError, Ti as repeatedActionDetector, Tn as SteeringBundle, Tr as SynthesisTarget, Tt as WranglerDeployRunnerInput, U as KeywordCoverageFinding, Ui as PolicyEditChange, Un as buildReflectionPrompt, Ur as exportProductBenchmarkRuns, Ut as ContractCheckResult, V as jsonShape, Vi as PolicyEditAxis, Vn as ReflectionProposal, Vr as ProductBenchmarkSingleRunExportOptions, Vt as traceContract, W as KeywordCoverageOptions, Wi as PolicyEditExpectedGain, Wn as parseReflectionResponse, Wr as productBenchmarkRepoIdentity, Wt as ContractRule, X as JudgeRetryOutcome, Xi as PolicyEditSource, Xn as ProposeReviewReport, Xr as PairedDeltaTestOptions, Xt as OrderMode, Y as ensembleJudge, Yi as PolicyEditSchemaVersion, Yn as ProposeReviewConfig, Yr as partitionHeldOut, Yt as ContractVerdict, Z as JudgeRetryPolicy, Zi as PolicyEditTarget, Zn as Review, Zr as PairedDeltaTestResult, Zt as RetryWrite, _ as blockingKnowledgeEval, _i as extractErrorCount, _n as CounterfactualRunner, _r as RunCommandInput, _t as NoLeakOptions, a as assertHttpOk, aa as costReceiptFromLlmError, ai as KnowledgeBundle, an as ScorecardCellDiff, ar as inMemoryReviewStore, at as runIntentMatchJudge, b as userQuestionsForKnowledgeGaps, bi as ErrorStreakOptions, bn as SteeringOptimizationRow, br as BoundedProcessInput, bt as blendHeldout, c as assertStatus, ca as costForTokenPricing, ci as KnowledgeReadinessReport, cn as ScorecardLogLine, cr as ExperimentRep, ct as SingleBackendDivergence, d as MODEL_PRICING, da as CostCeilingReachedError, di as KnowledgeSensitivity, dn as loadScorecard, dr as ImprovementVerdictResult, dt as BlendWeights, ea as isTransientLlmError, ei as LeaderboardOptions, en as CellVerdict, er as ReviewMemoryEntry, et as JudgeVerdict, f as estimateCost, fi as UserQuestion, fn as recordRuns, fr as DiscoverPersonasOptions, ft as FieldDestination, g as acquisitionPlansForKnowledgeGaps, gi as ExtractResult, gn as CounterfactualResult, gr as DirEntry, gt as JudgeScoreInput, h as resolveModelPricing, hi as ExtractOptions, hn as CounterfactualMutation, hr as CommandRunner, ht as HiddenLeak, i as assertHasField, ia as costReceiptFromLlm, ii as KnowledgeAcquisitionMode, in as ScorecardCell, ir as createLlmReviewer, it as IntentMatchResult, j as fileVerdictCache, ji as CapabilityHeadroomOptions, jn as hashContent, jr as RuntimeTrajectoryRecord, jt as ActionExecutionPolicy, k as canonicalJson, ki as ScenarioCost, kn as PromptHandle, kr as RuntimeTrajectoryEvidenceSummary, kt as CanaryLeak, l as analyzeAntiSlop, la as costForUsage, li as KnowledgeRequirement, ln as diffScorecard, lr as ExperimentStats, lt as SingleBackendReport, m as isModelPriced, mi as ErrorCountPattern, mn as CounterfactualContext, mr as discoverPersonas, mt as HiddenGradeResult, n as assertEventCount, na as LlmCallError, ni as leaderboard, nn as RecordRunsOptions, nr as Verification, nt as IntentMatchInput, o as assertLatencyUnder, oa as maximumChargeForLlmRequest, oi as KnowledgeFreshness, on as ScorecardDiff, or as jsonlReviewStore, ot as AssertSingleBackendOptions, p as estimateTokens, pi as ERROR_COUNT_PATTERNS, pn as recordRunsToScorecard, pr as DiscoveredPersona, pt as HiddenCriteriaGrader, q as runKeywordCoverageJudgeUrl, qi as PolicyEditInit, qn as ProposeInput, qr as TrainSelectionTestOptions, qt as ContractSpan, r as assertExists, ra as LlmResponseError, ri as DataAcquisitionPlan, rn as Scorecard, rr as VerifyFn, rt as IntentMatchOptions, s as assertNotContains, sa as CostResult, si as KnowledgeImportance, sn as ScorecardEntry, sr as runProposeReview, st as BackendDescriptor, t as assertContains, ta as stripFencedJson, ti as LeaderboardRow, tn as DiffScorecardOptions, tr as ReviewMemoryStore, tt as aggregateJudgeVerdicts, u as createAntiSlopJudge, ua as CostAccountingIncompleteError, ui as KnowledgeRequirementCategory, un as formatScorecardDiff, ur as ImprovementThresholds, ut as assertSingleBackend, v as knowledgeReadinessTracePayload, vi as DetectorEvent, vn as PairwiseSteeringOptimizer, vr as RunCommandResult, vt as RoutedField, w as ProductClient, wi as observeAll, wn as clamp01, wr as ActiveLearningOptions, wt as ViteDeployRunnerInput, x as objectiveEval, xi as RepeatedActionOptions, xn as RunScore, xr as BoundedProcessResult, xt as DeployGateLayerInput, y as scoreKnowledgeReadiness, yi as DetectorSignal, yn as SteeringOptimizationResult, yr as localCommandRunner, yt as assertNoHiddenLeak, z as OracleResult, zi as PolicyEditAdmission, zn as jsonlRunRecordBackend, zr as ProductBenchmarkExportOptions, zt as checkTraceContracts } from "./index-C8a9Dyan.js";
4
4
  import { _ as toAgentProfileJson, g as groupRunsByAgentProfileCell, h as buildAgentProfileCell, i as AgentProfileCellInput, l as AgentProfileJson, m as agentProfileCellKey, p as AgentProfileSourceInput, r as AgentProfileCell, s as AgentProfileDimensionValue, t as AGENT_PROFILE_KINDS, y as verifyAgentProfileCell } from "./agent-profile-cell-3kYxUsK2.js";
5
5
  import { E as isToolSpan, S as TraceEvent, T as isLlmSpan, _ as Span, a as FAILURE_CLASSES, c as JudgeSpan, f as Run, h as RunStatus, l as LlmSpan, n as BudgetLedgerEntry, o as FailureClass, p as RunCaptureReport, r as BudgetSpec, s as GenericSpan, t as Artifact, w as isJudgeSpan, x as ToolSpan } from "./schema-CkIvnmLU.js";
6
6
  import { _ as runTaskScore, a as RunRecord, c as RunSplitTag, d as RunTokenUsage, f as RunTraceRef, g as parseRunRecordSafe, h as modelHasSnapshot, i as RunOutcome, l as RunTaskFailure, m as isRunRecord, n as RunCostProvenance, o as RunRecordValidationError, p as UNKNOWN_MODEL, r as RunJudgeMetadata, s as RunSearchCoordinates, t as JudgeScoresRecord, u as RunTerminalOutcome, v as validateRunRecord } from "./run-record-pBlhZizT.js";
7
- import { $ as ErrorCluster, Q as DatasetOverview, S as makeProposalFinding, U as createBoundedTraceAnalysisStore, Y as TraceAnalysisStore, Z as DEFAULT_TRACE_ANALYST_BUDGETS, at as TraceAnalystFilters, b as computeFindingId, c as AnalystRunEvent, ct as TraceAnalystSpanStatus, d as AnalystRunSummary, dt as ViewTraceOversized, et as QueryTracesPage, f as AnalystSeverity, ft as ViewTraceResult, i as AnalystFinding, it as TraceAnalystByteBudgets, l as AnalystRunInputs, lt as TraceAnalystTraceSummary, m as EvidenceRef, n as AnalystContext, nt as SearchTraceResult, ot as TraceAnalystSpan, p as AnalystUsageReceipt, rt as SpanMatchRecord, st as TraceAnalystSpanKind, t as Analyst, tt as SearchSpanResult, u as AnalystRunResult, ut as ViewSpansResult, v as ProposalFinding, x as makeFinding } from "./types-6GuPL-Vt.js";
8
- import { $ as InMemoryRawProviderSink, A as ChatResponse, B as LlmCallRequest, C as ScenarioFile, D as ChatCallOpts, E as TurnResult, F as MockTransportOpts, H as LlmChargeBounds, I as OpenAiCompatibleTransportOpts, J as LlmToolChoice, K as LlmTokenLogprob, L as SandboxSdkTransportOpts, M as CreateChatClientOpts, N as CustomTransportOpts, O as ChatClient, R as createChatClient, S as RubricDimension, St as judgeFamily, T as Turn, V as LlmCallResult, W as LlmMessage, X as LlmUsage, Y as LlmToolDefinition, Z as FileSystemRawProviderSink, _ as JudgeRubric, _t as describeServedModelCheck, at as RawProviderSink, b as ProductScenario, bt as JudgeFamily, c as ArtifactCheck, ct as AssertServedModelOptions, d as CollectedArtifacts, dt as ServedModelPolicy, f as CompletionCriterion, ft as ServedModelVerdict, g as JudgeInput, gt as checkServedModel, h as JudgeFn, ht as assertServedModels, it as RawProviderEvent, j as ChatTransport, k as ChatRequest, l as ArtifactResult, lt as ModelSubstitutionError, m as DriverState, mt as assertServedModel, p as DimensionJudgeScore, pt as assertCrossFamilyServed, q as LlmToolCall, st as AssertCrossFamilyServedOptions, tt as NoopRawProviderSink, u as CheckResult, ut as ServedModelCheck, v as PersonaConfig, vt as AssertCrossFamilyOptions, w as ScenarioResult, x as RouteMap, xt as assertCrossFamily, y as ProductClientConfig, yt as CrossFamilyError, z as LlmCallMetadata } from "./cluster-CkzP4h3T.js";
9
- import { $ as chiSquaredPValue, A as pairedBootstrap, At as CorpusAgreementOptions, Bt as CalibrationResult, C as BOOTSTRAP_GATE_MIN_N, Ct as ranks, D as PairedSignTestResult, Dt as weightedComposite, E as PairedBootstrapResult, Et as summarizeNumberSeries, G as pairedRiskDifferenceScore, Gt as GoldenItem, Ht as ContinuousAgreement, I as McNemarResult, It as corpusInterRaterAgreement, J as benjaminiHochberg, Jt as VerbosityBiasResult, K as passAtK, L as ProportionInterval, Lt as corpusInterRaterAgreementFromJudgeScores, M as pairedSignTest, Mt as CorpusAgreementReport, N as pairedTTest, Nt as CorpusScoreRecord, O as PairedTTestResult, Ot as weightedMean, P as ExactRiskDifferenceResult, Q as chiSquared, R as RiskDifferenceResult, Rt as interRaterReliability, S as requiredSampleSize, St as pearsonR, T as PairedBootstrapOptions, Tt as studentTCritical, U as pairedRiskDifference, Ut as ContinuousAgreementOptions, V as mcnemar, Vt as CandidateScore, W as pairedRiskDifferenceExact, Wt as ContinuousCalibrationResult, X as holm, Xt as calibrateJudgeContinuous, Y as bonferroni, Yt as calibrateJudge, Z as LinearRegressionResult, Zt as continuousAgreement, _ as mulberry32, _t as WeightedCompositeInput, a as eProcess, at as CliffsMagnitude, b as pairedMde, bt as meanConfidenceInterval, c as sampleMedian, ct as interpretCliffs, d as RankTestMethod, dt as SampleDistribution, en as verbosityBias, et as fisherExactTwoSided, f as RankTestMethodRequest, ft as percentileNearestRank, g as wilcoxonSignedRank, gt as SeriesDistributionOptions, h as mannWhitneyU, ht as SeriesDistribution, i as EProcessStep, it as twoByTwoPValue, jt as CorpusAgreementPerDimension, k as SignTestAlternative, kt as rocAuc, l as summarizeSample, lt as pairedCohensDz, m as WilcoxonSignedRankResult, mt as MeanConfidenceInterval, n as EProcessOptions, nt as normalTwoSidedPValue, o as SampleSummary, ot as cliffsDelta, p as RankTestOptions, pt as summarizeDistribution, q as wilson, r as EProcessState, rt as studentTTwoSidedPValue, s as SampleSummaryOptions, st as cohensD, t as EProcess, tt as linearRegressionSlope, u as MannWhitneyResult, v as mcnemarPower, vt as WeightedCompositeResult, wt as spearmanR, x as requiredPairedSampleSize, xt as partialCredit, y as mcnemarRequiredN, yt as confidenceInterval, z as ScoreRiskDifferenceResult } from "./index-BN6htPVi.js";
10
- import { $ as summarizeBackendIntegrity, $n as RedTeamCategory, Di as RunCampaignOptions, Ei as CampaignCellRetryPolicy, Oi as runCampaign, Q as assertRealBackend, Qn as RedTeamCase, Ti as CampaignCellFailureReceipt, X as BackendIntegrityError, Z as BackendIntegrityReport, Zn as DEFAULT_RED_TEAM_CORPUS, ar as CanaryAlert, ci as LlmJudgeOptions, cr as CanaryOptions, er as RedTeamFinding, fi as transientDispatchFailure, ft as ArtifactEventLike, gt as extractProducedState, ht as ToolCallEventLike, ir as scoreRedTeamOutput, li as llmJudge, lr as CanaryReport, mt as RuntimeEventLike, nr as redTeamDataset, or as CanaryEvaluation, pt as ProposalEventLike, rr as redTeamReport, sr as CanaryKind, tr as RedTeamReport, ur as runCanaries } from "./index-cp5gkylN.js";
11
- import { $ as MintRolloutResult, Q as MintRolloutOptions, R as Scenario, _ as GateDecision, a as CampaignResult, nt as mintRolloutRows, tt as RolloutSearchLineage, w as JudgeScore } from "./types-C9yEqDjQ.js";
7
+ import { $ as ErrorCluster, Q as DatasetOverview, S as makeProposalFinding, U as createBoundedTraceAnalysisStore, Y as TraceAnalysisStore, Z as DEFAULT_TRACE_ANALYST_BUDGETS, at as TraceAnalystFilters, b as computeFindingId, c as AnalystRunEvent, ct as TraceAnalystSpanStatus, d as AnalystRunSummary, dt as ViewTraceOversized, et as QueryTracesPage, f as AnalystSeverity, ft as ViewTraceResult, i as AnalystFinding, it as TraceAnalystByteBudgets, l as AnalystRunInputs, lt as TraceAnalystTraceSummary, m as EvidenceRef, n as AnalystContext, nt as SearchTraceResult, ot as TraceAnalystSpan, p as AnalystUsageReceipt, rt as SpanMatchRecord, st as TraceAnalystSpanKind, t as Analyst, tt as SearchSpanResult, u as AnalystRunResult, ut as ViewSpansResult, v as ProposalFinding, x as makeFinding } from "./types-DbGSC2No.js";
8
+ import { $ as InMemoryRawProviderSink, A as ChatResponse, B as LlmCallRequest, C as ScenarioFile, D as ChatCallOpts, E as TurnResult, F as MockTransportOpts, H as LlmChargeBounds, I as OpenAiCompatibleTransportOpts, J as LlmToolChoice, K as LlmTokenLogprob, L as SandboxSdkTransportOpts, M as CreateChatClientOpts, N as CustomTransportOpts, O as ChatClient, R as createChatClient, S as RubricDimension, St as judgeFamily, T as Turn, V as LlmCallResult, W as LlmMessage, X as LlmUsage, Y as LlmToolDefinition, Z as FileSystemRawProviderSink, _ as JudgeRubric, _t as describeServedModelCheck, at as RawProviderSink, b as ProductScenario, bt as JudgeFamily, c as ArtifactCheck, ct as AssertServedModelOptions, d as CollectedArtifacts, dt as ServedModelPolicy, f as CompletionCriterion, ft as ServedModelVerdict, g as JudgeInput, gt as checkServedModel, h as JudgeFn, ht as assertServedModels, it as RawProviderEvent, j as ChatTransport, k as ChatRequest, l as ArtifactResult, lt as ModelSubstitutionError, m as DriverState, mt as assertServedModel, p as DimensionJudgeScore, pt as assertCrossFamilyServed, q as LlmToolCall, st as AssertCrossFamilyServedOptions, tt as NoopRawProviderSink, u as CheckResult, ut as ServedModelCheck, v as PersonaConfig, vt as AssertCrossFamilyOptions, w as ScenarioResult, x as RouteMap, xt as assertCrossFamily, y as ProductClientConfig, yt as CrossFamilyError, z as LlmCallMetadata } from "./cluster-CSd0LE_M.js";
9
+ import { $ as chiSquaredPValue, A as pairedBootstrap, At as CorpusAgreementOptions, Bt as CalibrationResult, C as BOOTSTRAP_GATE_MIN_N, Ct as ranks, D as PairedSignTestResult, Dt as weightedComposite, E as PairedBootstrapResult, Et as summarizeNumberSeries, G as pairedRiskDifferenceScore, Gt as GoldenItem, Ht as ContinuousAgreement, I as McNemarResult, It as corpusInterRaterAgreement, J as benjaminiHochberg, Jt as VerbosityBiasResult, K as passAtK, L as ProportionInterval, Lt as corpusInterRaterAgreementFromJudgeScores, M as pairedSignTest, Mt as CorpusAgreementReport, N as pairedTTest, Nt as CorpusScoreRecord, O as PairedTTestResult, Ot as weightedMean, P as ExactRiskDifferenceResult, Q as chiSquared, R as RiskDifferenceResult, Rt as interRaterReliability, S as requiredSampleSize, St as pearsonR, T as PairedBootstrapOptions, Tt as studentTCritical, U as pairedRiskDifference, Ut as ContinuousAgreementOptions, V as mcnemar, Vt as CandidateScore, W as pairedRiskDifferenceExact, Wt as ContinuousCalibrationResult, X as holm, Xt as calibrateJudgeContinuous, Y as bonferroni, Yt as calibrateJudge, Z as LinearRegressionResult, Zt as continuousAgreement, _ as mulberry32, _t as WeightedCompositeInput, a as eProcess, at as CliffsMagnitude, b as pairedMde, bt as meanConfidenceInterval, c as sampleMedian, ct as interpretCliffs, d as RankTestMethod, dt as SampleDistribution, en as verbosityBias, et as fisherExactTwoSided, f as RankTestMethodRequest, ft as percentileNearestRank, g as wilcoxonSignedRank, gt as SeriesDistributionOptions, h as mannWhitneyU, ht as SeriesDistribution, i as EProcessStep, it as twoByTwoPValue, jt as CorpusAgreementPerDimension, k as SignTestAlternative, kt as rocAuc, l as summarizeSample, lt as pairedCohensDz, m as WilcoxonSignedRankResult, mt as MeanConfidenceInterval, n as EProcessOptions, nt as normalTwoSidedPValue, o as SampleSummary, ot as cliffsDelta, p as RankTestOptions, pt as summarizeDistribution, q as wilson, r as EProcessState, rt as studentTTwoSidedPValue, s as SampleSummaryOptions, st as cohensD, t as EProcess, tt as linearRegressionSlope, u as MannWhitneyResult, v as mcnemarPower, vt as WeightedCompositeResult, wt as spearmanR, x as requiredPairedSampleSize, xt as partialCredit, y as mcnemarRequiredN, yt as confidenceInterval, z as ScoreRiskDifferenceResult } from "./index-DHWFYUVW.js";
10
+ import { $ as summarizeBackendIntegrity, $n as RedTeamCategory, Di as RunCampaignOptions, Ei as CampaignCellRetryPolicy, Oi as runCampaign, Q as assertRealBackend, Qn as RedTeamCase, Ti as CampaignCellFailureReceipt, X as BackendIntegrityError, Z as BackendIntegrityReport, Zn as DEFAULT_RED_TEAM_CORPUS, ar as CanaryAlert, ci as LlmJudgeOptions, cr as CanaryOptions, er as RedTeamFinding, fi as transientDispatchFailure, ft as ArtifactEventLike, gt as extractProducedState, ht as ToolCallEventLike, ir as scoreRedTeamOutput, li as llmJudge, lr as CanaryReport, mt as RuntimeEventLike, nr as redTeamDataset, or as CanaryEvaluation, pt as ProposalEventLike, rr as redTeamReport, sr as CanaryKind, tr as RedTeamReport, ur as runCanaries } from "./index-BI9ecONy.js";
11
+ import { $ as MintRolloutResult, Q as MintRolloutOptions, R as Scenario, _ as GateDecision, a as CampaignResult, nt as mintRolloutRows, tt as RolloutSearchLineage, w as JudgeScore } from "./types-ByrbgALT.js";
12
12
  import { a as RunFilter, i as InMemoryTraceStore, n as FileSystemTraceStore, o as SpanFilter, s as TraceStore, t as EventFilter } from "./store-CMs5wGXP.js";
13
13
  import { a as MintedRolloutLine } from "./schema-PDKQ_iNE.js";
14
14
  import { n as ScorePreference } from "./reward-ap_R9mCH.js";
15
- import { A as PairRunRecordsResult, B as ArmPopulationSummary, D as MatchedRunRecordPair, E as MatchedPair, F as comparePairedArms, I as pairArms, L as pairRunRecords, M as PairedArmsComparison, N as PairedCorrectness, O as PairArmsOptions, P as PairedMetricDelta, R as ArmAssignment, S as manifestContentDigest, T as ComparePairedArmsOptions, _t as AttestationProvenance, bt as attest, gt as ATTESTATION_ALGORITHM, it as paretoFrontier, j as PairedArmRow, k as PairArmsResult, nt as ParetoResult, rt as dominates, tt as Objective, vt as AttestationVerification, x as hashJson, xt as verifyAttestation, yt as AttestedReport, z as ArmPopulationReading } from "./statistical-heldout-vjrs8iyO.js";
16
- import { a as ProfileAxisSpec, c as expandProfileAxes, i as HarnessType, l as harnessAxisOf, n as CODING_HARNESSES, o as agentProfileHash, r as HARNESS_NATIVE_MODEL, s as agentProfileId, t as AgentProfile } from "./agent-profile-DC_v4p1f.js";
17
- import { a as PairedPromotionDecision, i as PairedMcNemarEvidence, n as PairedDecisionShape, o as PairedPromotionDecisionOptions, r as PairedDecisionStatistic, s as decidePairedPromotion, t as PairedDecisionMethod } from "./paired-promotion-decision-COeOgNHG.js";
15
+ import { A as PairRunRecordsResult, B as ArmPopulationSummary, D as MatchedRunRecordPair, E as MatchedPair, F as comparePairedArms, I as pairArms, L as pairRunRecords, M as PairedArmsComparison, N as PairedCorrectness, O as PairArmsOptions, P as PairedMetricDelta, R as ArmAssignment, S as manifestContentDigest, T as ComparePairedArmsOptions, _t as AttestationProvenance, bt as attest, gt as ATTESTATION_ALGORITHM, it as paretoFrontier, j as PairedArmRow, k as PairArmsResult, nt as ParetoResult, rt as dominates, tt as Objective, vt as AttestationVerification, x as hashJson, xt as verifyAttestation, yt as AttestedReport, z as ArmPopulationReading } from "./statistical-heldout-Bx1QUDNy.js";
16
+ import { a as ProfileAxisSpec, c as expandProfileAxes, i as HarnessType, l as harnessAxisOf, n as CODING_HARNESSES, o as agentProfileHash, r as HARNESS_NATIVE_MODEL, s as agentProfileId, t as AgentProfile } from "./agent-profile-BF1WLMco.js";
17
+ import { a as PairedPromotionDecision, i as PairedMcNemarEvidence, n as PairedDecisionShape, o as PairedPromotionDecisionOptions, r as PairedDecisionStatistic, s as decidePairedPromotion, t as PairedDecisionMethod } from "./paired-promotion-decision-DawN2o0G.js";
18
18
  import { i as DatasetSplit, n as DatasetManifest, r as DatasetScenario } from "./dataset-DThC1S6G.js";
19
19
  import { C as runEquivalenceCheck, S as defineEquivalenceCheck, _ as StrategyChecker, a as CheckerIdentity, b as VerificationStrategyProfile, c as EquivalenceCheckDefinition, d as EquivalenceCheckerInput, f as EquivalenceCheckerResult, g as EquivalenceRecord, h as EquivalenceProtocolError, i as equivalenceVerdict, l as EquivalenceCheckSpec, m as EquivalenceObligationStatus, n as VerdictCertification, o as CheckerOutcome, p as EquivalenceObligation, r as certificationEvidenceDigest, s as EquivalenceArm, t as DefaultVerdict, u as EquivalenceChecker, v as VERIFICATION_STRATEGIES, x as VerificationStrategySource, y as VERIFICATION_STRATEGY_SOURCES } from "./verdict-D_qiBQ6_.js";
20
20
  import { a as MultiLayerVerifier, c as VerifyOptions, i as LayerStatus, l as gradeSemanticStatus, n as Layer, o as Severity, r as LayerResult, s as VerificationReport, t as Finding } from "./multi-layer-verifier-lM4aVN78.js";
21
- import { C as HeldOutGateRejectionCode, S as HeldOutGateDecision, _ as gainHistogram, a as ResearchReportOptions, b as HeldOutGate, d as GainDistributionOptions, g as SummaryTableRow, h as SummaryTableOptions, l as GainDistributionBin, m as SummaryTable, p as ParetoPoint, t as ResearchReport, u as GainDistributionFigureSpec, v as summaryTable, w as SplitCoverage, x as HeldOutGateConfig, y as GateEvidence } from "./summary-report-vdhE20Zf.js";
21
+ import { C as HeldOutGateRejectionCode, S as HeldOutGateDecision, _ as gainHistogram, a as ResearchReportOptions, b as HeldOutGate, d as GainDistributionOptions, g as SummaryTableRow, h as SummaryTableOptions, l as GainDistributionBin, m as SummaryTable, p as ParetoPoint, t as ResearchReport, u as GainDistributionFigureSpec, v as summaryTable, w as SplitCoverage, x as HeldOutGateConfig, y as GateEvidence } from "./summary-report-9kwI6jFQ.js";
22
22
  import { a as FailureClassification, c as FailureRule, d as failureBlame, i as FailureBlame, l as classifyFailure, o as FailureContext, r as failureClusterView, s as FailureReasonRule, u as classifyFailureReason } from "./failure-cluster-DOoa7CBg.js";
23
- import { d as InsightReport, n as analyzeRuns, t as AnalyzeRunsOptions } from "./analyze-runs-DGr7mpC3.js";
23
+ import { d as InsightReport, n as analyzeRuns, t as AnalyzeRunsOptions } from "./analyze-runs-BaziIw93.js";
24
24
  import { a as SpanHandle, n as CaptureWrite, o as TraceEmitter, s as TraceEmitterOptions, t as CaptureErrorHandler } from "./emitter-mAXC5PSN.js";
25
25
  import { c as SandboxDriver } from "./sandbox-harness-K3LuyORm.js";
26
- import { a as ProducedProposal, c as SatisfiedBy, d as createLlmCorrectnessChecker, f as verifyCompletion, i as LlmCorrectnessCheckerOpts, l as TaskGold, n as CompletionVerdict, o as ProducedState, r as CorrectnessChecker, s as RequirementCheck, t as CompletionRequirement, u as completionVerdict } from "./completion-verifier-CJVDPv2T.js";
27
- import { _ as SelfImproveMethodResult, a as DefinedAgentEval, c as SelfImproveMethodOptions, d as SelfImproveProposerOptions, f as SelfImproveProposerResult, g as SelfImproveMethodProvenance, h as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveOptions, n as AgentEvalEvaluateOptions, o as defineAgentEval, p as SelfImproveResult, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveProgressEvent } from "./define-agent-eval-nNPC7cD5.js";
28
- import { a as createTraceAnalyst, i as TraceAnalystDefinition, n as buildDefaultAnalystRegistry, r as CreateTraceAnalystOptions, t as DefaultAnalystRegistryOptions } from "./default-registry-C_Akmpui.js";
29
- import { i as ExactAnalystRunEvent, l as ExactCapableAnalyst, o as ExactAnalystRunResult } from "./exact-types-Blq7YvsK.js";
30
- import { $ as measureJudgeAgreement, B as JudgeConfusion, F as CalibrationExample, G as JudgeGateRefusedError, H as JudgeGateDecision, I as CalibrationMerge, J as JudgeRunMeta, K as JudgeGateStatus, L as DEFAULT_JUDGE_GATE_POLICY, Q as judgeGateDecision, R as JudgeAgreement, St as analyzeSeries, U as JudgeGateInput, V as JudgeExampleAgreement, W as JudgeGatePolicy, X as addCalibrationVerdicts, Y as OwnerVerdict, Z as assertJudgeMayGate, bt as SeriesConvergenceOptions, et as registerCalibrationSet, q as JudgeRun, tt as snapshotFromJudgeAgreement, xt as SeriesConvergenceResult, z as JudgeCalibrationSet } from "./index-KAYaJtMA.js";
31
- import { B as DiffPolicy, H as FindingsStore, U as PersistedFinding, V as FindingsDiff, W as diffFindings } from "./systemone-DFUN9DL1.js";
32
- import { r as TraceAnalysisEngineResult, s as RawAnalystFinding, t as TraceAnalysisEngine } from "./engine-B0gr15-l.js";
33
- import { a as BudgetPolicy, c as RegistryRunOpts, i as AnalystRegistryOptions, s as ExactRegistryRunOpts, t as AnalystRegistry } from "./registry-Pfeqsr5k.js";
34
- import { C as SEMANTIC_CONCEPT_JUDGE_VERSION, D as runSemanticConceptJudge, E as SemanticConceptJudgeResult, S as ConceptSpec, T as SemanticConceptJudgeOptions, _ as DspyRlmTraceEngineOptions, a as FAILURE_MODE_KIND_SPEC, b as createChatTraceEngine, d as FindingSubject, f as FindingSubjectKind, t as DEFAULT_TRACE_ANALYST_KINDS, v as createDspyRlmTraceEngine, w as SemanticConceptJudgeInput, x as ConceptFinding, y as ChatTraceEngineOptions } from "./index-CxuK3sxA.js";
26
+ import { a as ProducedProposal, c as SatisfiedBy, d as createLlmCorrectnessChecker, f as verifyCompletion, i as LlmCorrectnessCheckerOpts, l as TaskGold, n as CompletionVerdict, o as ProducedState, r as CorrectnessChecker, s as RequirementCheck, t as CompletionRequirement, u as completionVerdict } from "./completion-verifier-oAgSwdFe.js";
27
+ import { _ as SelfImproveMethodResult, a as DefinedAgentEval, c as SelfImproveMethodOptions, d as SelfImproveProposerOptions, f as SelfImproveProposerResult, g as SelfImproveMethodProvenance, h as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveOptions, n as AgentEvalEvaluateOptions, o as defineAgentEval, p as SelfImproveResult, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveProgressEvent } from "./define-agent-eval-Be2Ser8j.js";
28
+ import { a as createTraceAnalyst, i as TraceAnalystDefinition, n as buildDefaultAnalystRegistry, r as CreateTraceAnalystOptions, t as DefaultAnalystRegistryOptions } from "./default-registry-DoIBhuS-.js";
29
+ import { i as ExactAnalystRunEvent, l as ExactCapableAnalyst, o as ExactAnalystRunResult } from "./exact-types-DEnrf7NX.js";
30
+ import { $ as measureJudgeAgreement, B as JudgeConfusion, F as CalibrationExample, G as JudgeGateRefusedError, H as JudgeGateDecision, I as CalibrationMerge, J as JudgeRunMeta, K as JudgeGateStatus, L as DEFAULT_JUDGE_GATE_POLICY, Q as judgeGateDecision, R as JudgeAgreement, St as analyzeSeries, U as JudgeGateInput, V as JudgeExampleAgreement, W as JudgeGatePolicy, X as addCalibrationVerdicts, Y as OwnerVerdict, Z as assertJudgeMayGate, bt as SeriesConvergenceOptions, et as registerCalibrationSet, q as JudgeRun, tt as snapshotFromJudgeAgreement, xt as SeriesConvergenceResult, z as JudgeCalibrationSet } from "./index-VSN1IZQp.js";
31
+ import { B as DiffPolicy, H as FindingsStore, U as PersistedFinding, V as FindingsDiff, W as diffFindings } from "./systemone-CSwi1hp-.js";
32
+ import { r as TraceAnalysisEngineResult, s as RawAnalystFinding, t as TraceAnalysisEngine } from "./engine-CPBhCLJR.js";
33
+ import { a as BudgetPolicy, c as RegistryRunOpts, i as AnalystRegistryOptions, s as ExactRegistryRunOpts, t as AnalystRegistry } from "./registry-BqoX13FD.js";
34
+ import { C as SEMANTIC_CONCEPT_JUDGE_VERSION, D as runSemanticConceptJudge, E as SemanticConceptJudgeResult, S as ConceptSpec, T as SemanticConceptJudgeOptions, _ as DspyRlmTraceEngineOptions, a as FAILURE_MODE_KIND_SPEC, b as createChatTraceEngine, d as FindingSubject, f as FindingSubjectKind, t as DEFAULT_TRACE_ANALYST_KINDS, v as createDspyRlmTraceEngine, w as SemanticConceptJudgeInput, x as ConceptFinding, y as ChatTraceEngineOptions } from "./index-AjfEHHA8.js";
35
35
  import { S as welchsTTest, _ as computeToolUseMetrics, b as WelchTestStatus, d as Trajectory, f as TrajectoryStep, g as ToolUseOptions, h as ToolUseMetrics, i as toolWasteView, m as ToolStats, p as buildTrajectory, s as judgeAgreementView, u as budgetBreachView, x as iqr, y as WelchTestResult } from "./tool-waste-CUASoTtk.js";
36
- import { i as BenchmarkEvaluation } from "./types-CP3rkpir.js";
37
- import { A as AnalystFindingDigest, B as ControlRuntimeConfig, C as createFeedbackTrajectory, D as renderPreferenceMemoryMarkdown, E as feedbackTrajectoryToOptimizerRow, F as ControlBudget, I as ControlContext, L as ControlDecision, M as AnalystRunDigest, N as analystRunDigest, O as summarizePreferenceMemory, P as ControlActionOutcome, R as ControlEvalResult, S as controlRunToFeedbackTrajectory, T as feedbackTrajectoriesToOptimizerRows, U as ControlStep, V as ControlRuntimeError, W as StopDecision, _ as PreferenceMemoryEntry, a as FeedbackLabel, b as analystRunToReviewRequests, c as FeedbackOptimizerRow, d as FeedbackTask, f as FeedbackTrajectory, g as InMemoryFeedbackTrajectoryStore, h as FileSystemFeedbackTrajectoryStore, i as FeedbackAttempt, j as AnalystReviewDecision, k as withAssignedFeedbackSplit, l as FeedbackOutcome, m as FeedbackTrajectoryStore, n as AnalystReviewRequest, o as FeedbackLabelKind, p as FeedbackTrajectoryFilter, r as FeedbackArtifactType, s as FeedbackLabelSource, t as AnalystFeedbackTrajectoryOptions, u as FeedbackSplitPolicy, v as ProposedSideEffect, w as feedbackTrajectoriesToDatasetScenarios, x as assignFeedbackSplit, y as analystRunToFeedbackTrajectory, z as ControlRunResult } from "./feedback-trajectory-D3oISeGl.js";
38
- import { A as TraceInsightReadiness, F as describeTraceInsightScope, G as AnalyzeTracesResult, H as tokenizeDomainWords, I as domainEvidencePattern, K as analyzeTraces, L as inferDomainKeywords, N as buildTraceInsightContext, P as buildTraceInsightPrompt, V as scoreTraceInsightReadiness, a as captureFetchToRawSink, b as OtlpSpan, c as ExtractedUsage, d as extractUsageFromSse, f as OtlpFileTraceStore, g as spanRecordsToTraceAnalysisStore, h as otlpTextToTraceAnalysisStore, j as TraceInsightSuite, m as SpanRecordsToTraceAnalysisStoreOptions, n as toolSpansToTraceAnalysisStore, o as OtlpFlatLine, u as extractUsage, v as OtlpExport, x as exportRunAsOtlp } from "./store-tool-spans-BemsQEhd.js";
39
- import { a as RunIntegrityReport, o as assertRunCaptured, t as RunIntegrityError } from "./integrity-zeWMeTQt.js";
40
- import { a as SteeringChange, c as CampaignRunContext, d as CampaignVariant, f as EvalCampaignOptions, h as runEvalCampaign, i as Researcher, l as CampaignRunOutcome, m as FailedRun, n as ExperimentResult, o as CampaignFactoryParams, p as EvalCampaignResult, r as FailureMode, s as CampaignIntegrityPolicy, t as ExperimentPlan, u as CampaignScenario } from "./researcher-DlyjP5uq.js";
41
- import { C as requiredPairsForPairedPromotion, S as pairedPromotionPower, _ as PairedPromotionPowerOptions, a as PairedEvalueStep, b as RequiredPairsForPairedPromotionOptions, c as SequentialDecision, d as sequentialCrossingHorizon, f as sequentialPairedPower, g as PairedPromotionPowerCall, h as PairedPromotionCallRates, i as PairedEvalueSequence, l as evaluateInterimReleaseConfidence, m as PairedPromotionAlternativeCell, n as InterimReleaseConfidenceInput, p as PairedPromotionAlternative, r as PairedEvalueOptions, t as InterimReleaseConfidence, u as pairedEvalueSequence, v as PairedPromotionPowerResult, x as RequiredPairsForPairedPromotionResult, y as PairedValueLaw } from "./sequential-OeSwJ2uG.js";
42
- import { _ as evaluateReleaseConfidence, a as bootstrapCi, d as ReleaseConfidenceIssue, f as ReleaseConfidenceMetrics, g as ReleaseTraceEvidence, i as Verdict, n as BootstrapResult, p as ReleaseConfidenceScorecard, s as ActionableSideInfo, t as BootstrapOptions, u as ReleaseConfidenceInput } from "./promotion-gate-XglbDCNo.js";
36
+ import { i as BenchmarkEvaluation } from "./types-CEgIBy0W.js";
37
+ import { A as AnalystFindingDigest, B as ControlRuntimeConfig, C as createFeedbackTrajectory, D as renderPreferenceMemoryMarkdown, E as feedbackTrajectoryToOptimizerRow, F as ControlBudget, I as ControlContext, L as ControlDecision, M as AnalystRunDigest, N as analystRunDigest, O as summarizePreferenceMemory, P as ControlActionOutcome, R as ControlEvalResult, S as controlRunToFeedbackTrajectory, T as feedbackTrajectoriesToOptimizerRows, U as ControlStep, V as ControlRuntimeError, W as StopDecision, _ as PreferenceMemoryEntry, a as FeedbackLabel, b as analystRunToReviewRequests, c as FeedbackOptimizerRow, d as FeedbackTask, f as FeedbackTrajectory, g as InMemoryFeedbackTrajectoryStore, h as FileSystemFeedbackTrajectoryStore, i as FeedbackAttempt, j as AnalystReviewDecision, k as withAssignedFeedbackSplit, l as FeedbackOutcome, m as FeedbackTrajectoryStore, n as AnalystReviewRequest, o as FeedbackLabelKind, p as FeedbackTrajectoryFilter, r as FeedbackArtifactType, s as FeedbackLabelSource, t as AnalystFeedbackTrajectoryOptions, u as FeedbackSplitPolicy, v as ProposedSideEffect, w as feedbackTrajectoriesToDatasetScenarios, x as assignFeedbackSplit, y as analystRunToFeedbackTrajectory, z as ControlRunResult } from "./feedback-trajectory-DxkG5EPi.js";
38
+ import { A as TraceInsightReadiness, F as describeTraceInsightScope, G as AnalyzeTracesResult, H as tokenizeDomainWords, I as domainEvidencePattern, K as analyzeTraces, L as inferDomainKeywords, N as buildTraceInsightContext, P as buildTraceInsightPrompt, V as scoreTraceInsightReadiness, a as captureFetchToRawSink, b as OtlpSpan, c as ExtractedUsage, d as extractUsageFromSse, f as OtlpFileTraceStore, g as spanRecordsToTraceAnalysisStore, h as otlpTextToTraceAnalysisStore, j as TraceInsightSuite, m as SpanRecordsToTraceAnalysisStoreOptions, n as toolSpansToTraceAnalysisStore, o as OtlpFlatLine, u as extractUsage, v as OtlpExport, x as exportRunAsOtlp } from "./store-tool-spans-kCDPEwjL.js";
39
+ import { a as RunIntegrityReport, o as assertRunCaptured, t as RunIntegrityError } from "./integrity-BMyCpnZ8.js";
40
+ import { a as SteeringChange, c as CampaignRunContext, d as CampaignVariant, f as EvalCampaignOptions, h as runEvalCampaign, i as Researcher, l as CampaignRunOutcome, m as FailedRun, n as ExperimentResult, o as CampaignFactoryParams, p as EvalCampaignResult, r as FailureMode, s as CampaignIntegrityPolicy, t as ExperimentPlan, u as CampaignScenario } from "./researcher-BuojS14L.js";
41
+ import { C as requiredPairsForPairedPromotion, S as pairedPromotionPower, _ as PairedPromotionPowerOptions, a as PairedEvalueStep, b as RequiredPairsForPairedPromotionOptions, c as SequentialDecision, d as sequentialCrossingHorizon, f as sequentialPairedPower, g as PairedPromotionPowerCall, h as PairedPromotionCallRates, i as PairedEvalueSequence, l as evaluateInterimReleaseConfidence, m as PairedPromotionAlternativeCell, n as InterimReleaseConfidenceInput, p as PairedPromotionAlternative, r as PairedEvalueOptions, t as InterimReleaseConfidence, u as pairedEvalueSequence, v as PairedPromotionPowerResult, x as RequiredPairsForPairedPromotionResult, y as PairedValueLaw } from "./sequential-1-T2oi1_.js";
42
+ import { _ as evaluateReleaseConfidence, a as bootstrapCi, d as ReleaseConfidenceIssue, f as ReleaseConfidenceMetrics, g as ReleaseTraceEvidence, i as Verdict, n as BootstrapResult, p as ReleaseConfidenceScorecard, s as ActionableSideInfo, t as BootstrapOptions, u as ReleaseConfidenceInput } from "./promotion-gate-DG_PSJx_.js";
43
43
  import { _ as redactText, d as ShareSafetyVerdict, f as assessShareSafety, g as redactForShare, h as redact, i as RedactOptions, o as RedactionProfile, r as REDACTION_VERSION, s as RedactionReport, u as ShareSafetyStatus, v as shareAllowed } from "./redact-jrT0APTU.js";
44
44
  import { y as OUTPUT_VALUE } from "./attribute-vocabulary-5oo8ykIx.js";
45
45
  import { i as judgeSpans, n as argHash, s as runsForScenario } from "./query-UD0B08Sb.js";
package/dist/index.js CHANGED
@@ -41,9 +41,9 @@ import { i as failureBlame, n as classifyFailure, r as classifyFailureReason } f
41
41
  import { n as runBoundedProcess, t as DEFAULT_BOUNDED_PROCESS_TIMEOUT_MS } from "./bounded-process-BZrWXIKW.js";
42
42
  import { t as runEvalCampaign } from "./eval-campaign-DpkToY_w.js";
43
43
  import { _ as judgeFamily, a as maximumChargeForLlmRequest, c as ModelSubstitutionError, f as assertServedModels, g as assertCrossFamily, h as CrossFamilyError, i as costReceiptFromLlmError, l as assertCrossFamilyServed, m as describeServedModelCheck, p as checkServedModel, r as costReceiptFromLlm, s as isTransientLlmError, u as assertServedModel } from "./llm-client-Bj-TooRS.js";
44
- import { n as createChatClient } from "./chat-client-9XLgWrQ7.js";
45
- import { n as createLlmCorrectnessChecker, r as verifyCompletion, t as completionVerdict } from "./completion-verifier-BYLOcQt7.js";
46
- import "./rollout-Dh5q6CPA.js";
44
+ import { n as createChatClient } from "./chat-client-PJEzoot9.js";
45
+ import { n as createLlmCorrectnessChecker, r as verifyCompletion, t as completionVerdict } from "./completion-verifier-DVZx0x6x.js";
46
+ import "./rollout-B86tqluQ.js";
47
47
  import { t as mintRolloutRows } from "./mint-Ck92efNo.js";
48
48
  import { i as sequentialPairedPower, n as pairedEvalueSequence, r as sequentialCrossingHorizon, t as evaluateInterimReleaseConfidence } from "./sequential-Ko7T-zvu.js";
49
49
  import { n as paidJsonChat } from "./chat-json-call-CaNO3k0A.js";
@@ -7195,7 +7195,7 @@ async function runIntentMatchJudge(input, options) {
7195
7195
  const start = Date.now();
7196
7196
  const opts = {
7197
7197
  chat: options.chat,
7198
- model: options.model ?? options.chat.defaultModel ?? "gpt-6-luna",
7198
+ model: options.model ?? options.chat.defaultModel ?? "zai/glm-5.3",
7199
7199
  timeoutMs: options.timeoutMs ?? DEFAULT_TIMEOUT$1,
7200
7200
  maxTokens: options.maxTokens ?? DEFAULT_MAX_TOKENS$1,
7201
7201
  maxSourceChars: options.maxSourceChars ?? DEFAULT_MAX_SOURCE$1,
@@ -8087,7 +8087,7 @@ async function runSemanticConceptJudge(input, options) {
8087
8087
  };
8088
8088
  const opts = {
8089
8089
  chat: options.chat,
8090
- model: options.model ?? options.chat.defaultModel ?? "gpt-6-luna",
8090
+ model: options.model ?? options.chat.defaultModel ?? "zai/glm-5.3",
8091
8091
  timeoutMs: options.timeoutMs ?? DEFAULT_TIMEOUT,
8092
8092
  maxTokens: options.maxTokens ?? DEFAULT_MAX_TOKENS,
8093
8093
  maxSourceChars: options.maxSourceChars ?? DEFAULT_MAX_SOURCE,
@@ -1,5 +1,5 @@
1
1
  import { r as CaptureIntegrityError } from "./errors-E9I2hldP.js";
2
- import { at as RawProviderSink } from "./cluster-CkzP4h3T.js";
2
+ import { at as RawProviderSink } from "./cluster-CSd0LE_M.js";
3
3
  import { s as TraceStore } from "./store-CMs5wGXP.js";
4
4
  interface RunIntegrityExpectations {
5
5
  /** Minimum LLM span count. Default 0 (no requirement). */
@@ -54,4 +54,4 @@ declare class RunIntegrityError extends CaptureIntegrityError {
54
54
  }
55
55
  declare function assertRunCaptured(store: TraceStore, runId: string, expectations?: RunIntegrityExpectations): Promise<RunIntegrityReport>;
56
56
  export { RunIntegrityReport as a, RunIntegrityIssueCode as i, RunIntegrityExpectations as n, assertRunCaptured as o, RunIntegrityIssue as r, RunIntegrityError as t };
57
- //# sourceMappingURL=integrity-zeWMeTQt.d.ts.map
57
+ //# sourceMappingURL=integrity-BMyCpnZ8.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"integrity-zeWMeTQt.d.ts","names":[],"sources":["../src/trace/integrity.ts"],"mappings":";;;UAyBiB;;EAEf;;EAEA;;EAEA;;;;;EAKA,UAAU;;EAEV;;;;;;EAMA;;EAEA;;KAGU;UAWK;EACf,MAAM;EACN;EACA,SAAS;;UAGM;EACf;EACA;EACA;EACA;EACA;EACA;;;;;;EAMA;IAAmB;IAAiB;;EACpC,QAAQ;;cAGG,0BAA0B;WACT,QAAQ;EAAR,YAAA,QAAQ;;iBAOhB,kBACpB,OAAO,YACP,eACA,eAAc,2BACb,QAAQ"}
1
+ {"version":3,"file":"integrity-BMyCpnZ8.d.ts","names":[],"sources":["../src/trace/integrity.ts"],"mappings":";;;UAyBiB;;EAEf;;EAEA;;EAEA;;;;;EAKA,UAAU;;EAEV;;;;;;EAMA;;EAEA;;KAGU;UAWK;EACf,MAAM;EACN;EACA,SAAS;;UAGM;EACf;EACA;EACA;EACA;EACA;EACA;;;;;;EAMA;IAAmB;IAAiB;;EACpC,QAAQ;;cAGG,0BAA0B;WACT,QAAQ;EAAR,YAAA,QAAQ;;iBAOhB,kBACpB,OAAO,YACP,eACA,eAAc,2BACb,QAAQ"}
@@ -1,5 +1,5 @@
1
1
  import { u as CostProvenance } from "./cost-ledger-Czro9wmo.js";
2
- import { C as JudgeDimension, w as JudgeScore } from "./types-C9yEqDjQ.js";
2
+ import { C as JudgeDimension, w as JudgeScore } from "./types-ByrbgALT.js";
3
3
  import { o as MatrixResult } from "./index-DwUGehqX.js";
4
4
  import { AgentProfile } from "@tangle-network/agent-interface";
5
5
  interface MultishotMessage {
@@ -380,4 +380,4 @@ interface RunMultishotMatrixResult {
380
380
  }
381
381
  declare function runMultishotMatrix<TPersona extends MultishotPersona>(opts: RunMultishotMatrixOptions<TPersona>): Promise<RunMultishotMatrixResult>;
382
382
  export { MultishotTransportResponse as A, MultishotPersona as C, MultishotToolExecutor as D, MultishotToolDefinition as E, assertMultishotShotResult as M, MultishotTransport as O, MultishotMessage as S, MultishotShape as T, renderJsonFooter as _, MultishotCellOutput as a, MultishotDriverEmptyError as b, RunMultishotMatrixResult as c, MultishotShot as d, RunMultishotOptions as f, renderDimensions as g, MultishotJudgeConfig as h, ConversationJudgeInput as i, MultishotTransportToolCall as j, MultishotTransportRequest as k, computeCellComposite as l, JudgeRunResult as m, CellCompositeInput as n, MultishotJudges as o, runMultishot as p, CellCompositeScore as r, RunMultishotMatrixOptions as s, ArtifactJudgeInput as t, runMultishotMatrix as u, runJudge as v, MultishotResult as w, MultishotFatalToolError as x, MultishotArtifact as y };
383
- //# sourceMappingURL=matrix-Diw3yJCG.d.ts.map
383
+ //# sourceMappingURL=matrix-DdniPG_M.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"matrix-Diw3yJCG.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;EAAA,YAAA;;cAMjB,gCAAgC;EAC/B,YAAA;;;;;;;;;;;;;;;;;;;iBAgCE,0BAA0B,yBAAyB,SAAS;UCtI3D,qBAAqB;;EAEpC;;;EAGA,WAAW;;EAEX;;EAEA,YAAY;;EAEZ;;;EAGA,cAAc,OAAO;;EAErB;;UAGe;;EAEf,OAAO;;EAEP,MAAM;;iBAGc,SAAS,QAC7B,OAAO,qBAAqB,SAC5B,OAAO,SACN,QAAQ;;;iBAsIK,iBAAiB,eAAe;;iBAKhC,iBAAiB,eAAe;UCvK/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;;;;;;;;;;;;;;;;;iBAkBS,aAAa,iBAAiB,kBAClD,MAAM,oBAAoB,YACzB,QAAQ;UClEM,uBAAuB,iBAAiB;EACvD,YAAY;EACZ,SAAS;;UAGM,mBAAmB,iBAAiB;EACnD,UAAU;EACV,SAAS;;UAGM,gBAAgB,iBAAiB;;EAEhD,cAAc,qBAAqB,uBAAuB;;EAE1D,aAAa,qBAAqB,mBAAmB;;EAErD,iBAAiB,qBAAqB,mBAAmB;;EAEzD;;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-DdniPG_M.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;EAAA,YAAA;;cAMjB,gCAAgC;EAC/B,YAAA;;;;;;;;;;;;;;;;;;;iBAgCE,0BAA0B,yBAAyB,SAAS;UCtI3D,qBAAqB;;EAEpC;;;EAGA,WAAW;;EAEX;;EAEA,YAAY;;EAEZ;;;EAGA,cAAc,OAAO;;EAErB;;UAGe;;EAEf,OAAO;;EAEP,MAAM;;iBAGc,SAAS,QAC7B,OAAO,qBAAqB,SAC5B,OAAO,SACN,QAAQ;;;iBAsIK,iBAAiB,eAAe;;iBAKhC,iBAAiB,eAAe;UCvK/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;;;;;;;;;;;;;;;;;iBAkBS,aAAa,iBAAiB,kBAClD,MAAM,oBAAoB,YACzB,QAAQ;UClEM,uBAAuB,iBAAiB;EACvD,YAAY;EACZ,SAAS;;UAGM,mBAAmB,iBAAiB;EACnD,UAAU;EACV,SAAS;;UAGM,gBAAgB,iBAAiB;;EAEhD,cAAc,qBAAqB,uBAAuB;;EAE1D,aAAa,qBAAqB,mBAAmB;;EAErD,iBAAiB,qBAAqB,mBAAmB;;EAEzD;;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"}
@@ -1,4 +1,4 @@
1
- import { $t as selfPreference, Bt as CalibrationResult, Gt as GoldenItem, Ht as ContinuousAgreement, Jt as VerbosityBiasResult, Kt as PositionalBiasResult, Qt as positionalBias, Ut as ContinuousAgreementOptions, Vt as CandidateScore, Wt as ContinuousCalibrationResult, Xt as calibrateJudgeContinuous, Yt as calibrateJudge, Zt as continuousAgreement, en as verbosityBias, qt as SelfPreferenceResult } from "../index-BN6htPVi.js";
1
+ import { $t as selfPreference, Bt as CalibrationResult, Gt as GoldenItem, Ht as ContinuousAgreement, Jt as VerbosityBiasResult, Kt as PositionalBiasResult, Qt as positionalBias, Ut as ContinuousAgreementOptions, Vt as CandidateScore, Wt as ContinuousCalibrationResult, Xt as calibrateJudgeContinuous, Yt as calibrateJudge, Zt as continuousAgreement, en as verbosityBias, qt as SelfPreferenceResult } from "../index-DHWFYUVW.js";
2
2
  import { a as RubricPredictiveValidityReport, c as CorrelationInterval, d as FileSystemOutcomeStore, f as FileSystemOutcomeStoreOptions, g as OutcomeStoreError, h as OutcomeStore, i as RubricPredictiveValidityInput, l as OutcomeReduction, m as OutcomeFilter, n as RubricOutcomeExclusion, o as RubricRanking, p as InMemoryOutcomeStore, r as RubricOutcomePair, s as rubricPredictiveValidity, t as OutcomeMetricSpec, u as DeploymentOutcome } from "../rubric-predictive-validity-Bbi88Aut.js";
3
- import { $ as measureJudgeAgreement, $t as CorrelationStudyOptions, A as JudgePanelOutput, At as ProbabilityPolicyAuditInput, B as JudgeConfusion, Bt as clusterBootstrapMean, C as JudgeCalibrationCellVerdict, Ct as EvaluatorAdmissionPolicy, D as JudgeCalibrationGoldItem, Dt as EvaluatorAuditPlan, E as JudgeCalibrationGateConfig, Et as EvaluatorAuditObservation, F as CalibrationExample, Ft as evaluatorAuditObservationSchema, G as JudgeGateRefusedError, Gt as CategoricalForecast, H as JudgeGateDecision, Ht as CalibrationOptions, I as CalibrationMerge, It as planEvaluatorAudit, J as JudgeRunMeta, Jt as calibrationCurve, K as JudgeGateStatus, Kt as OrdinalForecast, L as DEFAULT_JUDGE_GATE_POLICY, Lt as ClusterBootstrapOptions, M as RubricDimensionGuide, Mt as auditEvaluator, N as calibrateJudgePanel, Nt as auditProbabilityPolicy, O as JudgeDimensionCalibration, Ot as EvaluatorAuditPlanInput, P as gateJudgeCalibration, Pt as evaluatorAdmissionPolicySchema, Q as judgeGateDecision, Qt as CorrelationResult, R as JudgeAgreement, Rt as ClusterBootstrapResult, S as CalibrateJudgePanelInput, T as JudgeCalibrationGateCell, Tt as EvaluatorAuditInput, U as JudgeGateInput, Ut as CalibrationPair, V as JudgeExampleAgreement, Vt as CalibrationBin, W as JudgeGatePolicy, Wt as CalibrationReport, X as addCalibrationVerdicts, Xt as multiclassBrierScore, Y as OwnerVerdict, Yt as calibrationFromPairs, Z as assertJudgeMayGate, Zt as rankedProbabilityScore, _ as catchRate, _t as snapshotFromCalibration, a as PerturbedField, at as SentinelMetrics, b as plantByPerturbation, c as PlantEvidence, ct as SentinelSnapshot, d as PlantKindCounts, dt as SentinelTrend, en as CorrelationStudyResult, et as registerCalibrationSet, f as PlantManifest, ft as SnapshotMeta, g as UnseededRejection, gt as snapshotFromAgreement, h as SeededGradingSet, ht as judgeSentinelReport, i as PerturbationKind, it as SentinelMetricName, j as JudgePanelPrompt, jt as ProbabilityPolicyAuditReport, k as JudgePanelCalibration, kt as EvaluatorErrorRate, l as PlantExpectation, lt as SentinelStore, m as SeedPlantsOptions, mt as fileSentinelStore, n as CatchRateStatus, nn as OutcomePair, nt as EvalHealthStamp, o as PerturbedPlant, ot as SentinelReport, p as PlantOutcome, pt as evalHealthStamp, q as JudgeRun, qt as brierScore, r as EvidencePerturbation, rn as correlationStudy, rt as JudgeSentinelOptions, s as Plant, st as SentinelSetOptions, t as CatchRateReport, tn as EvalMetricSpec, tt as snapshotFromJudgeAgreement, u as PlantKind, ut as SentinelThresholds, v as definePlant, vt as snapshotFromSentinelSet, w as JudgeCalibrationGate, wt as EvaluatorAdmissionReport, x as seedPlants, y as perturbEvidence, yt as validateSentinelSnapshot, z as JudgeCalibrationSet, zt as ClusteredObservation } from "../index-KAYaJtMA.js";
3
+ import { $ as measureJudgeAgreement, $t as CorrelationStudyOptions, A as JudgePanelOutput, At as ProbabilityPolicyAuditInput, B as JudgeConfusion, Bt as clusterBootstrapMean, C as JudgeCalibrationCellVerdict, Ct as EvaluatorAdmissionPolicy, D as JudgeCalibrationGoldItem, Dt as EvaluatorAuditPlan, E as JudgeCalibrationGateConfig, Et as EvaluatorAuditObservation, F as CalibrationExample, Ft as evaluatorAuditObservationSchema, G as JudgeGateRefusedError, Gt as CategoricalForecast, H as JudgeGateDecision, Ht as CalibrationOptions, I as CalibrationMerge, It as planEvaluatorAudit, J as JudgeRunMeta, Jt as calibrationCurve, K as JudgeGateStatus, Kt as OrdinalForecast, L as DEFAULT_JUDGE_GATE_POLICY, Lt as ClusterBootstrapOptions, M as RubricDimensionGuide, Mt as auditEvaluator, N as calibrateJudgePanel, Nt as auditProbabilityPolicy, O as JudgeDimensionCalibration, Ot as EvaluatorAuditPlanInput, P as gateJudgeCalibration, Pt as evaluatorAdmissionPolicySchema, Q as judgeGateDecision, Qt as CorrelationResult, R as JudgeAgreement, Rt as ClusterBootstrapResult, S as CalibrateJudgePanelInput, T as JudgeCalibrationGateCell, Tt as EvaluatorAuditInput, U as JudgeGateInput, Ut as CalibrationPair, V as JudgeExampleAgreement, Vt as CalibrationBin, W as JudgeGatePolicy, Wt as CalibrationReport, X as addCalibrationVerdicts, Xt as multiclassBrierScore, Y as OwnerVerdict, Yt as calibrationFromPairs, Z as assertJudgeMayGate, Zt as rankedProbabilityScore, _ as catchRate, _t as snapshotFromCalibration, a as PerturbedField, at as SentinelMetrics, b as plantByPerturbation, c as PlantEvidence, ct as SentinelSnapshot, d as PlantKindCounts, dt as SentinelTrend, en as CorrelationStudyResult, et as registerCalibrationSet, f as PlantManifest, ft as SnapshotMeta, g as UnseededRejection, gt as snapshotFromAgreement, h as SeededGradingSet, ht as judgeSentinelReport, i as PerturbationKind, it as SentinelMetricName, j as JudgePanelPrompt, jt as ProbabilityPolicyAuditReport, k as JudgePanelCalibration, kt as EvaluatorErrorRate, l as PlantExpectation, lt as SentinelStore, m as SeedPlantsOptions, mt as fileSentinelStore, n as CatchRateStatus, nn as OutcomePair, nt as EvalHealthStamp, o as PerturbedPlant, ot as SentinelReport, p as PlantOutcome, pt as evalHealthStamp, q as JudgeRun, qt as brierScore, r as EvidencePerturbation, rn as correlationStudy, rt as JudgeSentinelOptions, s as Plant, st as SentinelSetOptions, t as CatchRateReport, tn as EvalMetricSpec, tt as snapshotFromJudgeAgreement, u as PlantKind, ut as SentinelThresholds, v as definePlant, vt as snapshotFromSentinelSet, w as JudgeCalibrationGate, wt as EvaluatorAdmissionReport, x as seedPlants, y as perturbEvidence, yt as validateSentinelSnapshot, z as JudgeCalibrationSet, zt as ClusteredObservation } from "../index-VSN1IZQp.js";
4
4
  export { type CalibrateJudgePanelInput, CalibrationBin, CalibrationExample, CalibrationMerge, CalibrationOptions, CalibrationPair, CalibrationReport, type CalibrationResult, type CandidateScore, CatchRateReport, CatchRateStatus, CategoricalForecast, ClusterBootstrapOptions, ClusterBootstrapResult, ClusteredObservation, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type CorrelationInterval, CorrelationResult, CorrelationStudyOptions, CorrelationStudyResult, DEFAULT_JUDGE_GATE_POLICY, DeploymentOutcome, type EvalHealthStamp, EvalMetricSpec, type EvaluatorAdmissionPolicy, type EvaluatorAdmissionReport, type EvaluatorAuditInput, type EvaluatorAuditObservation, type EvaluatorAuditPlan, type EvaluatorAuditPlanInput, type EvaluatorErrorRate, EvidencePerturbation, FileSystemOutcomeStore, FileSystemOutcomeStoreOptions, type GoldenItem, InMemoryOutcomeStore, JudgeAgreement, type JudgeCalibrationCellVerdict, type JudgeCalibrationGate, type JudgeCalibrationGateCell, type JudgeCalibrationGateConfig, type JudgeCalibrationGoldItem, JudgeCalibrationSet, JudgeConfusion, type JudgeDimensionCalibration, JudgeExampleAgreement, JudgeGateDecision, JudgeGateInput, JudgeGatePolicy, JudgeGateRefusedError, JudgeGateStatus, type JudgePanelCalibration, type JudgePanelOutput, type JudgePanelPrompt, JudgeRun, JudgeRunMeta, type JudgeSentinelOptions, OrdinalForecast, OutcomeFilter, OutcomeMetricSpec, OutcomePair, type OutcomeReduction, OutcomeStore, OutcomeStoreError, OwnerVerdict, PerturbationKind, PerturbedField, PerturbedPlant, Plant, PlantEvidence, PlantExpectation, PlantKind, PlantKindCounts, PlantManifest, PlantOutcome, type PositionalBiasResult, type ProbabilityPolicyAuditInput, type ProbabilityPolicyAuditReport, type RubricDimensionGuide, RubricOutcomeExclusion, RubricOutcomePair, RubricPredictiveValidityInput, RubricPredictiveValidityReport, RubricRanking, SeedPlantsOptions, SeededGradingSet, type SelfPreferenceResult, type SentinelMetricName, type SentinelMetrics, type SentinelReport, type SentinelSetOptions, type SentinelSnapshot, type SentinelStore, type SentinelThresholds, type SentinelTrend, type SnapshotMeta, UnseededRejection, type VerbosityBiasResult, addCalibrationVerdicts, assertJudgeMayGate, auditEvaluator, auditProbabilityPolicy, brierScore, calibrateJudge, calibrateJudgeContinuous, calibrateJudgePanel, calibrationCurve, calibrationFromPairs, catchRate, clusterBootstrapMean, continuousAgreement, correlationStudy, definePlant, evalHealthStamp, evaluatorAdmissionPolicySchema, evaluatorAuditObservationSchema, fileSentinelStore, gateJudgeCalibration, judgeGateDecision, judgeSentinelReport, measureJudgeAgreement, multiclassBrierScore, perturbEvidence, planEvaluatorAudit, plantByPerturbation, positionalBias, rankedProbabilityScore, registerCalibrationSet, rubricPredictiveValidity, seedPlants, selfPreference, snapshotFromAgreement, snapshotFromCalibration, snapshotFromJudgeAgreement, snapshotFromSentinelSet, validateSentinelSnapshot, verbosityBias };
@@ -1,4 +1,4 @@
1
- import { C as MultishotPersona, E as MultishotToolDefinition, c as RunMultishotMatrixResult, f as RunMultishotOptions, k as MultishotTransportRequest, s as RunMultishotMatrixOptions, w as MultishotResult } from "../../matrix-Diw3yJCG.js";
1
+ import { C as MultishotPersona, E as MultishotToolDefinition, c as RunMultishotMatrixResult, f as RunMultishotOptions, k as MultishotTransportRequest, s as RunMultishotMatrixOptions, w as MultishotResult } from "../../matrix-DdniPG_M.js";
2
2
  interface CompareOptions {
3
3
  /** Stop after this many mismatches. A structural divergence high in the
4
4
  * tree would otherwise report every leaf below it. */
@@ -1,6 +1,6 @@
1
- import { P as DEFAULT_JUDGE_MODEL } from "../cluster-CkzP4h3T.js";
2
- import { C as JudgeDimension, w as JudgeScore } from "../types-C9yEqDjQ.js";
3
- import { A as MultishotTransportResponse, C as MultishotPersona, D as MultishotToolExecutor, E as MultishotToolDefinition, M as assertMultishotShotResult, O as MultishotTransport, S as MultishotMessage, T as MultishotShape, _ as renderJsonFooter, a as MultishotCellOutput, b as MultishotDriverEmptyError, c as RunMultishotMatrixResult, d as MultishotShot, f as RunMultishotOptions, g as renderDimensions, h as MultishotJudgeConfig, i as ConversationJudgeInput, j as MultishotTransportToolCall, k as MultishotTransportRequest, l as computeCellComposite, m as JudgeRunResult, n as CellCompositeInput, o as MultishotJudges, p as runMultishot, r as CellCompositeScore, s as RunMultishotMatrixOptions, t as ArtifactJudgeInput, u as runMultishotMatrix, v as runJudge, w as MultishotResult, x as MultishotFatalToolError, y as MultishotArtifact } from "../matrix-Diw3yJCG.js";
1
+ import { P as DEFAULT_JUDGE_MODEL } from "../cluster-CSd0LE_M.js";
2
+ import { C as JudgeDimension, w as JudgeScore } from "../types-ByrbgALT.js";
3
+ import { A as MultishotTransportResponse, C as MultishotPersona, D as MultishotToolExecutor, E as MultishotToolDefinition, M as assertMultishotShotResult, O as MultishotTransport, S as MultishotMessage, T as MultishotShape, _ as renderJsonFooter, a as MultishotCellOutput, b as MultishotDriverEmptyError, c as RunMultishotMatrixResult, d as MultishotShot, f as RunMultishotOptions, g as renderDimensions, h as MultishotJudgeConfig, i as ConversationJudgeInput, j as MultishotTransportToolCall, k as MultishotTransportRequest, l as computeCellComposite, m as JudgeRunResult, n as CellCompositeInput, o as MultishotJudges, p as runMultishot, r as CellCompositeScore, s as RunMultishotMatrixOptions, t as ArtifactJudgeInput, u as runMultishotMatrix, v as runJudge, w as MultishotResult, x as MultishotFatalToolError, y as MultishotArtifact } from "../matrix-DdniPG_M.js";
4
4
  import { AgentProfile } from "@tangle-network/agent-interface";
5
5
  /** Prompt and completion tokens one leg reported. */
6
6
  interface MultishotLegTokens {
@@ -1,5 +1,5 @@
1
1
  import { d as estimateCost, p as isModelPriced } from "../cost-ledger-1CwUnVOd.js";
2
- import { t as DEFAULT_JUDGE_MODEL } from "../chat-client-9XLgWrQ7.js";
2
+ import { t as DEFAULT_JUDGE_MODEL } from "../chat-client-PJEzoot9.js";
3
3
  import { n as withCellSpend } from "../cell-spend-zxmzg4GH.js";
4
4
  import { t as runAgentMatrix } from "../matrix-CgpX8wdm.js";
5
5
  import { n as MultishotFatalToolError, r as assertMultishotShotResult, t as MultishotDriverEmptyError } from "../types-DK4vNv5q.js";
@@ -433,7 +433,7 @@ function driverVisibleContent(msg) {
433
433
  }
434
434
  // src/multishot/judges.ts
435
435
  async function runJudge(judge, input) {
436
- const model = judge.model ?? "gpt-6-luna";
436
+ const model = judge.model ?? "zai/glm-5.3";
437
437
  const prompt = judge.buildPrompt(input);
438
438
  let raw = "";
439
439
  let llmCall;
package/dist/openapi.json CHANGED
@@ -2,7 +2,7 @@
2
2
  "openapi": "3.1.0",
3
3
  "info": {
4
4
  "title": "@tangle-network/agent-eval — wire protocol",
5
- "version": "1.0.24",
5
+ "version": "1.0.26",
6
6
  "description": "HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: 1.1.0. Bumps on breaking changes to request/response schemas.",
7
7
  "contact": {
8
8
  "name": "Tangle Network",
@@ -1,4 +1,4 @@
1
- import { E as PairedBootstrapResult } from "./index-BN6htPVi.js";
1
+ import { E as PairedBootstrapResult } from "./index-DHWFYUVW.js";
2
2
  /** Which paired estimator produced the deciding interval. */
3
3
  type PairedDecisionStatistic = 'paired_risk_difference' | 'mean_bootstrap' | 'median_bootstrap';
4
4
  /** Which test carried the decision, given the estimator and the sample size. */
@@ -120,4 +120,4 @@ interface PairedDecisionShape {
120
120
  */
121
121
  declare function decidePairedPromotion(before: number[], after: number[], options?: PairedPromotionDecisionOptions): PairedPromotionDecision;
122
122
  export { PairedPromotionDecision as a, PairedMcNemarEvidence as i, PairedDecisionShape as n, PairedPromotionDecisionOptions as o, PairedDecisionStatistic as r, decidePairedPromotion as s, PairedDecisionMethod as t };
123
- //# sourceMappingURL=paired-promotion-decision-COeOgNHG.d.ts.map
123
+ //# sourceMappingURL=paired-promotion-decision-DawN2o0G.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"paired-promotion-decision-COeOgNHG.d.ts","names":[],"sources":["../src/paired-promotion-decision.ts"],"mappings":";;KA2DY;;KAMA;;UAGK;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;;;EAGf;;EAEA;;EAEA;;EAEA;;;;EAIA;;;;;;;;EAQA;;;;;EAKA;;;;;;;;;;;;EAYA;;UAGe;;EAEf;;EAEA;EACA;EACA,WAAW;EACX,QAAQ;;;;;EAKR;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,WAAW;;EAEX,SAAS;;;EAGT;;EAEA;;EAEA;;;EAGA;;EAEA;;EAEA;;;EAGA;;EAEA;;EAEA;;EAEA;;;UAIe;EACf,WAAW;;EAEX;;EAEA;;;;;;;iBA2Dc,sBACd,kBACA,iBACA,UAAS,iCACR"}
1
+ {"version":3,"file":"paired-promotion-decision-DawN2o0G.d.ts","names":[],"sources":["../src/paired-promotion-decision.ts"],"mappings":";;KA2DY;;KAMA;;UAGK;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;;;EAGf;;EAEA;;EAEA;;EAEA;;;;EAIA;;;;;;;;EAQA;;;;;EAKA;;;;;;;;;;;;EAYA;;UAGe;;EAEf;;EAEA;EACA;EACA,WAAW;EACX,QAAQ;;;;;EAKR;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,WAAW;;EAEX,SAAS;;;EAGT;;EAEA;;EAEA;;;EAGA;;EAEA;;EAEA;;;EAGA;;EAEA;;EAEA;;EAEA;;;UAIe;EACf,WAAW;;EAEX;;EAEA;;;;;;;iBA2Dc,sBACd,kBACA,iBACA,UAAS,iCACR"}
@@ -1,5 +1,5 @@
1
1
  import { m as CustomTokenPricing } from "./cost-ledger-Czro9wmo.js";
2
- import { p as AnalystUsageReceipt } from "./types-6GuPL-Vt.js";
2
+ import { p as AnalystUsageReceipt } from "./types-DbGSC2No.js";
3
3
  import { t as PrimeBridgeTransport } from "./prime-bridge-transport-Cfqir_Im.js";
4
4
  /**
5
5
  * The reply grammar an analyst arm holds a model to, independent of transport.
@@ -306,4 +306,4 @@ interface PrimeProtocolIdentity {
306
306
  */
307
307
  declare function primeProtocolSha256(identity: PrimeProtocolIdentity): string;
308
308
  export { runPrimeExchange as C, ReplyRowDecoded as D, ReplyEnvelope as E, ReplyRowRejection as O, projectPrimeTrajectory as S, ReplyContract as T, buildPrimePrompt as _, PrimeProjectionOutcome as a, normalizePrimeUsage as b, PrimeProtocolIdentity as c, PrimeRepairPromptSpec as d, PrimeRepairState as f, analystUsageReceiptFromPrimeUsage as g, PrimeTurnRecord as h, PrimeProjectionDelivery as i, PrimeRawUsage as l, PrimeRowDecoded as m, PrimeExchangeOutcome as n, PrimeProjectionSource as o, PrimeReplyContract as p, PrimeFailure as r, PrimePromptSpec as s, PrimeExchangeOptions as t, PrimeRejectedRow as u, buildPrimeRepairPrompt as v, DecodedReply as w, primeProtocolSha256 as x, mergePrimeRawUsage as y };
309
- //# sourceMappingURL=prime-protocol-bRbh6dqd.d.ts.map
309
+ //# sourceMappingURL=prime-protocol-RUPtoYiI.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"prime-protocol-bRbh6dqd.d.ts","names":[],"sources":["../src/analyst/reply-contract.ts","../src/analyst/prime-protocol.ts"],"mappings":";;;;;;;;;;;;;KAaY,gBAAgB;EAAU;EAAU,KAAK;;EAAW;EAAW;;;UAG1D;EACf;;EAEA,QAAQ;;UAGO,cAAc;;EAE7B;;EAEA;;;;;EAKA;;;;;;EAMA,UAAU,cAAc,gBAAgB,gBAAgB;;;;;;EAMxD;;;;;;EAMA,eAAe,iBAAiB;;;;;;EAMhC;;;;;EAKA;;;UAIe;EACf;EACA;;UAGe,aAAa;EAC5B,MAAM;;EAEN,QAAQ;EACR,UAAU;;EAEV;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;KC3CU,gBAAgB,QAAQ,gBAAgB;KAExC,mBAAmB,QAAQ,cAAc;UAIpC;EACf;;EAEA;EACA;;EAEA;EACA;;EAEA;;iBAGc,iBAAiB,MAAM;UAatB;EACf;EACA;EACA;;;iBAIc,uBAAuB,MAAM;;;;;;UAoE5B;;;;;;EAMf;;EAEA;;EAEA;;;;;;;EAOA;;;iBAQc,oBAAoB,eAAe;;;;;;iBAgBnC,mBAAmB,GAAG,eAAe,GAAG,gBAAgB;;;;;;;;;;;;;iBA6BxD,kCACd,OAAO,eACP,SAAS,qBACR;;;;;;;;KAqCS;EAEN;EACA;;EAEA;EAAqB;EAAiB;EAAgB;;EACtD;EAAiB;EAAiB;;UAEvB;EACf;;EAEA;;;UAIe;EACf;EACA,OAAO;;EAEP;;UAGe;EACf;EACA;;KAGU,qBAAqB;EAE3B;;EAEA;EACA,MAAM;EACN,UAAU;;EAEV;;EAEA;EACA,OAAO;EACP,OAAO;EACP,QAAQ;EACR;;EAGA;EACA,SAAS;EACT,OAAO;EACP,OAAO;EACP,QAAQ;;EAER;;UAGW,qBAAqB;EACpC,UAAU,mBAAmB;EAC7B;EACA,WAAW;;EAEX;EACA;;EAEA;;EAEA;EACA,SAAS;;;;;;;iBAQW,iBAAiB,MACrC,SAAS,qBAAqB,QAC7B,QAAQ,qBAAqB;;;;;;UAyLf,sBAAsB;;EAErC,QAAQ,iBAAiB;;EAEzB,UAAU,iBAAiB;;EAE3B;;UAGe;EACf;EACA;EACA;;KAGU,uBAAuB;EAE7B;EACA,gBAAgB;EAChB;EACA,UAAU;;EAEV;EAAW;EAAgB;;;;;;;;;iBASX,uBAAuB,OAC3C,QAAQ,sBAAsB,QAC9B;EAAU;IACT,QAAQ,uBAAuB;UA8BjB;EACf;EACA;EACA;EACA;;EAEA,QAAQ,SAAS;;;;;;;;;;;;iBAaH,oBAAoB,UAAU"}
1
+ {"version":3,"file":"prime-protocol-RUPtoYiI.d.ts","names":[],"sources":["../src/analyst/reply-contract.ts","../src/analyst/prime-protocol.ts"],"mappings":";;;;;;;;;;;;;KAaY,gBAAgB;EAAU;EAAU,KAAK;;EAAW;EAAW;;;UAG1D;EACf;;EAEA,QAAQ;;UAGO,cAAc;;EAE7B;;EAEA;;;;;EAKA;;;;;;EAMA,UAAU,cAAc,gBAAgB,gBAAgB;;;;;;EAMxD;;;;;;EAMA,eAAe,iBAAiB;;;;;;EAMhC;;;;;EAKA;;;UAIe;EACf;EACA;;UAGe,aAAa;EAC5B,MAAM;;EAEN,QAAQ;EACR,UAAU;;EAEV;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;KC3CU,gBAAgB,QAAQ,gBAAgB;KAExC,mBAAmB,QAAQ,cAAc;UAIpC;EACf;;EAEA;EACA;;EAEA;EACA;;EAEA;;iBAGc,iBAAiB,MAAM;UAatB;EACf;EACA;EACA;;;iBAIc,uBAAuB,MAAM;;;;;;UAoE5B;;;;;;EAMf;;EAEA;;EAEA;;;;;;;EAOA;;;iBAQc,oBAAoB,eAAe;;;;;;iBAgBnC,mBAAmB,GAAG,eAAe,GAAG,gBAAgB;;;;;;;;;;;;;iBA6BxD,kCACd,OAAO,eACP,SAAS,qBACR;;;;;;;;KAqCS;EAEN;EACA;;EAEA;EAAqB;EAAiB;EAAgB;;EACtD;EAAiB;EAAiB;;UAEvB;EACf;;EAEA;;;UAIe;EACf;EACA,OAAO;;EAEP;;UAGe;EACf;EACA;;KAGU,qBAAqB;EAE3B;;EAEA;EACA,MAAM;EACN,UAAU;;EAEV;;EAEA;EACA,OAAO;EACP,OAAO;EACP,QAAQ;EACR;;EAGA;EACA,SAAS;EACT,OAAO;EACP,OAAO;EACP,QAAQ;;EAER;;UAGW,qBAAqB;EACpC,UAAU,mBAAmB;EAC7B;EACA,WAAW;;EAEX;EACA;;EAEA;;EAEA;EACA,SAAS;;;;;;;iBAQW,iBAAiB,MACrC,SAAS,qBAAqB,QAC7B,QAAQ,qBAAqB;;;;;;UAyLf,sBAAsB;;EAErC,QAAQ,iBAAiB;;EAEzB,UAAU,iBAAiB;;EAE3B;;UAGe;EACf;EACA;EACA;;KAGU,uBAAuB;EAE7B;EACA,gBAAgB;EAChB;EACA,UAAU;;EAEV;EAAW;EAAgB;;;;;;;;;iBASX,uBAAuB,OAC3C,QAAQ,sBAAsB,QAC9B;EAAU;IACT,QAAQ,uBAAuB;UA8BjB;EACf;EACA;EACA;EACA;;EAEA,QAAQ,SAAS;;;;;;;;;;;;iBAaH,oBAAoB,UAAU"}
@@ -1,7 +1,7 @@
1
1
  import { o as FailureClass } from "./schema-CkIvnmLU.js";
2
2
  import { a as RunRecord, c as RunSplitTag } from "./run-record-pBlhZizT.js";
3
3
  import { i as DatasetSplit, n as DatasetManifest, r as DatasetScenario } from "./dataset-DThC1S6G.js";
4
- import { S as HeldOutGateDecision } from "./summary-report-vdhE20Zf.js";
4
+ import { S as HeldOutGateDecision } from "./summary-report-9kwI6jFQ.js";
5
5
  /** Severity of an actionable finding attached to a run/trace. */
6
6
  type AsiSeverity = 'info' | 'warning' | 'error' | 'critical';
7
7
  /** Actionable side-info — a diagnosed finding the loop can act on. */
@@ -227,4 +227,4 @@ declare function judgeReplayGate<TOutput>(args: JudgeReplayGateArgs<TOutput>): P
227
227
  candidateSamples: number;
228
228
  }>;
229
229
  export { evaluateReleaseConfidence as _, bootstrapCi as a, ReleaseConfidenceAxis as c, ReleaseConfidenceIssue as d, ReleaseConfidenceMetrics as f, ReleaseTraceEvidence as g, ReleaseConfidenceThresholds as h, Verdict as i, ReleaseConfidenceAxisName as l, ReleaseConfidenceStatus as m, BootstrapResult as n, judgeReplayGate as o, ReleaseConfidenceScorecard as p, JudgeReplayGateArgs as r, ActionableSideInfo as s, BootstrapOptions as t, ReleaseConfidenceInput as u };
230
- //# sourceMappingURL=promotion-gate-XglbDCNo.d.ts.map
230
+ //# sourceMappingURL=promotion-gate-DG_PSJx_.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"promotion-gate-XglbDCNo.d.ts","names":[],"sources":["../src/release-confidence.ts","../src/promotion-gate.ts"],"mappings":";;;;;KAuBK;;UAGY;;EAEf;;EAEA;EACA,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;EACA,WAAW;;KAGD;KACA;UAQK;EACf;EACA;EACA,QAAQ;EACR;EACA;EACA;EACA;EACA;;EAEA,eAAe;EACf,MAAM;EACN,WAAW;;UAGI;;EAEf;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,UAAU;EACV,qBAAqB;EACrB,gBAAgB;EAChB,kBAAkB;EAClB,eAAe;EACf,aAAa;;UAGE;EACf,MAAM;EACN,QAAQ;EACR;EACA;;UAGe;EACf,MAAM;EACN;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,aAAa,OAAO;EACpB,cAAc;EACd,oBAAoB,QAAQ,OAAO;EACnC,0BAA0B;;;;;;;EAO1B;;UAGe;EACf;EACA;EACA;EACA,QAAQ;EACR;EACA,MAAM;EACN,QAAQ;EACR,SAAS;EACT,SAAS;EACT,cAAc;EACd;;iBAkBc,0BACd,OAAO,yBACN;;;;;;;;;;;;;;;;;;;;;;;;;KC3JS;UAEK;EACf;EACA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,SAAS;;UAGM;;EAEf;;EAEA;;;;;;EAMA;;;EAGA;;;;;;;;;iBAUc,YACd,oBACA,qBACA,UAAS,mBACR;;;;;;;;;;;;;;;;UA0Gc,oBAAoB;EACnC,iBAAiB;EACjB,kBAAkB;;EAElB,QAAQ,QAAQ,YAAY;EAC5B;EACA;;EAEA;;EAEA;;;;;iBAMoB,gBAAgB,SACpC,MAAM,oBAAoB,WACzB,QAAQ;EAAoB;EAAyB"}
1
+ {"version":3,"file":"promotion-gate-DG_PSJx_.d.ts","names":[],"sources":["../src/release-confidence.ts","../src/promotion-gate.ts"],"mappings":";;;;;KAuBK;;UAGY;;EAEf;;EAEA;EACA,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;EACA,WAAW;;KAGD;KACA;UAQK;EACf;EACA;EACA,QAAQ;EACR;EACA;EACA;EACA;EACA;;EAEA,eAAe;EACf,MAAM;EACN,WAAW;;UAGI;;EAEf;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,UAAU;EACV,qBAAqB;EACrB,gBAAgB;EAChB,kBAAkB;EAClB,eAAe;EACf,aAAa;;UAGE;EACf,MAAM;EACN,QAAQ;EACR;EACA;;UAGe;EACf,MAAM;EACN;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,aAAa,OAAO;EACpB,cAAc;EACd,oBAAoB,QAAQ,OAAO;EACnC,0BAA0B;;;;;;;EAO1B;;UAGe;EACf;EACA;EACA;EACA,QAAQ;EACR;EACA,MAAM;EACN,QAAQ;EACR,SAAS;EACT,SAAS;EACT,cAAc;EACd;;iBAkBc,0BACd,OAAO,yBACN;;;;;;;;;;;;;;;;;;;;;;;;;KC3JS;UAEK;EACf;EACA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,SAAS;;UAGM;;EAEf;;EAEA;;;;;;EAMA;;;EAGA;;;;;;;;;iBAUc,YACd,oBACA,qBACA,UAAS,mBACR;;;;;;;;;;;;;;;;UA0Gc,oBAAoB;EACnC,iBAAiB;EACjB,kBAAkB;;EAElB,QAAQ,QAAQ,YAAY;EAC5B;EACA;;EAEA;;EAEA;;;;;iBAMoB,gBAAgB,SACpC,MAAM,oBAAoB,WACzB,QAAQ;EAAoB;EAAyB"}
@@ -1,7 +1,7 @@
1
1
  import { o as CostLedgerHandle } from "./cost-ledger-Czro9wmo.js";
2
- import { c as AnalystRunEvent, d as AnalystRunSummary, i as AnalystFinding, l as AnalystRunInputs, n as AnalystContext, t as Analyst, u as AnalystRunResult } from "./types-6GuPL-Vt.js";
3
- import { O as ChatClient } from "./cluster-CkzP4h3T.js";
4
- import { i as ExactAnalystRunEvent, o as ExactAnalystRunResult, u as ExactExecutionComponentIdentity } from "./exact-types-Blq7YvsK.js";
2
+ import { c as AnalystRunEvent, d as AnalystRunSummary, i as AnalystFinding, l as AnalystRunInputs, n as AnalystContext, t as Analyst, u as AnalystRunResult } from "./types-DbGSC2No.js";
3
+ import { O as ChatClient } from "./cluster-CSd0LE_M.js";
4
+ import { i as ExactAnalystRunEvent, o as ExactAnalystRunResult, u as ExactExecutionComponentIdentity } from "./exact-types-DEnrf7NX.js";
5
5
  interface AnalystHooks {
6
6
  /** Legacy runs may mutate ctx; exact runs provide a frozen observational context. */
7
7
  onBeforeAnalyze?(args: {
@@ -165,4 +165,4 @@ declare class AnalystRegistry {
165
165
  runStream(runId: string, inputs: AnalystRunInputs, runOpts?: RegistryRunOpts): AsyncGenerator<AnalystRunEvent, void, void>;
166
166
  }
167
167
  export { BudgetPolicy as a, RegistryRunOpts as c, AnalystRegistryOptions as i, ExactAnalystRunExecutionError as n, ExactAnalystBudgetPolicy as o, AnalystHooks as r, ExactRegistryRunOpts as s, AnalystRegistry as t };
168
- //# sourceMappingURL=registry-Pfeqsr5k.d.ts.map
168
+ //# sourceMappingURL=registry-BqoX13FD.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"registry-Pfeqsr5k.d.ts","names":[],"sources":["../src/analyst/registry-plan.ts","../src/analyst/registry-execution.ts","../src/analyst/registry.ts"],"mappings":";;;;UAiBiB;;EAEf,iBAAiB;IACf,SAAS;IACT,KAAK;IACL;aACS;;EAEX,gBAAgB;IACd,SAAS;IACT,SAAS;IACT,UAAU;IACV;aACS;;;;;;EAMX,SAAS;IACP,SAAS;IACT,OAAO;IACP;MACE,+BAA+B,QAAQ;;EAE3C,YAAY;IAAQ,QAAQ;aAA4B;;UAGzC;;EAEf;;EAEA,UAAU;;;;;;;EAOV,YAAY;IACV,SAAS;IACT;IACA;IACA;;;UAIa;;EAEf,OAAO;;EAEP,OAAO,aAAa,SAAS;;EAE7B,QAAQ;;EAER,gBAAgB;;EAEhB,gBAAgB;;EAEhB,eAAe;;UAGA;;EAEf;;EAEA;;EAEA,SAAS;;EAET;;EAEA,SAAS;;EAET,aAAa;;EAEb;;EAEA,OAAO;;;;;;;;;EASP,gBAAgB,cAAc,kBAAkB,eAAe,cAAc;;;;;;EAM7E;;;KAIU;WAEG;WACA;;WAGA;WACA;;WAEA,SAAS,SAAS;;;;;;;;;;UAWhB;WACN;WACA,QAAQ;WACR;WACA,QAAQ;WACR,YAAY;WACZ,oBAAoB;WACpB;WACA,MAAM,SAAS;WACf,eACL,cAAc,kBACd,SAAS,eAAe,cAAc;WAEjC;WACA;WACA;WACA;;;cChHE,sCAAsC;WACxC;WACA,QAAQ;EAEL,YAAA,iBAAiB,QAAQ,uBAAuB,UAAU;;cCN3D;mBACM;mBACA;EAEL,YAAA,UAAS;EAIrB,SAAS,SAAS;EAsBlB,QAAQ;IACN;IACA;IACA;IACA,MAAM;;EAUF,IACJ,eACA,QAAQ,kBACR,UAAS,kBACR,QAAQ;;EAUL,SACJ,eACA,QAAQ,kBACR,SAAS,uBACR,QAAQ;;EAQJ,eACL,eACA,QAAQ,kBACR,SAAS,uBACR,eAAe;;;;;;;;;;;;EAmBX,UACL,eACA,QAAQ,kBACR,UAAS,kBACR,eAAe"}
1
+ {"version":3,"file":"registry-BqoX13FD.d.ts","names":[],"sources":["../src/analyst/registry-plan.ts","../src/analyst/registry-execution.ts","../src/analyst/registry.ts"],"mappings":";;;;UAiBiB;;EAEf,iBAAiB;IACf,SAAS;IACT,KAAK;IACL;aACS;;EAEX,gBAAgB;IACd,SAAS;IACT,SAAS;IACT,UAAU;IACV;aACS;;;;;;EAMX,SAAS;IACP,SAAS;IACT,OAAO;IACP;MACE,+BAA+B,QAAQ;;EAE3C,YAAY;IAAQ,QAAQ;aAA4B;;UAGzC;;EAEf;;EAEA,UAAU;;;;;;;EAOV,YAAY;IACV,SAAS;IACT;IACA;IACA;;;UAIa;;EAEf,OAAO;;EAEP,OAAO,aAAa,SAAS;;EAE7B,QAAQ;;EAER,gBAAgB;;EAEhB,gBAAgB;;EAEhB,eAAe;;UAGA;;EAEf;;EAEA;;EAEA,SAAS;;EAET;;EAEA,SAAS;;EAET,aAAa;;EAEb;;EAEA,OAAO;;;;;;;;;EASP,gBAAgB,cAAc,kBAAkB,eAAe,cAAc;;;;;;EAM7E;;;KAIU;WAEG;WACA;;WAGA;WACA;;WAEA,SAAS,SAAS;;;;;;;;;;UAWhB;WACN;WACA,QAAQ;WACR;WACA,QAAQ;WACR,YAAY;WACZ,oBAAoB;WACpB;WACA,MAAM,SAAS;WACf,eACL,cAAc,kBACd,SAAS,eAAe,cAAc;WAEjC;WACA;WACA;WACA;;;cChHE,sCAAsC;WACxC;WACA,QAAQ;EAEL,YAAA,iBAAiB,QAAQ,uBAAuB,UAAU;;cCN3D;mBACM;mBACA;EAEL,YAAA,UAAS;EAIrB,SAAS,SAAS;EAsBlB,QAAQ;IACN;IACA;IACA;IACA,MAAM;;EAUF,IACJ,eACA,QAAQ,kBACR,UAAS,kBACR,QAAQ;;EAUL,SACJ,eACA,QAAQ,kBACR,SAAS,uBACR,QAAQ;;EAQJ,eACL,eACA,QAAQ,kBACR,SAAS,uBACR,eAAe;;;;;;;;;;;;EAmBX,UACL,eACA,QAAQ,kBACR,UAAS,kBACR,eAAe"}
@@ -1,8 +1,8 @@
1
- import { A as pairedBootstrap, E as PairedBootstrapResult, J as benjaminiHochberg, T as PairedBootstrapOptions, g as wilcoxonSignedRank } from "./index-BN6htPVi.js";
1
+ import { A as pairedBootstrap, E as PairedBootstrapResult, J as benjaminiHochberg, T as PairedBootstrapOptions, g as wilcoxonSignedRank } from "./index-DHWFYUVW.js";
2
2
  import { a as RubricPredictiveValidityReport, i as RubricPredictiveValidityInput, o as RubricRanking, r as RubricOutcomePair, s as rubricPredictiveValidity } from "./rubric-predictive-validity-Bbi88Aut.js";
3
- import { _ as gainHistogram, a as ResearchReportOptions, c as RESEARCH_REPORT_HARD_PAIR_FLOOR, d as GainDistributionOptions, f as ParetoFigureSpec, g as SummaryTableRow, h as SummaryTableOptions, i as ResearchReportMethodology, l as GainDistributionBin, m as SummaryTable, n as ResearchReportCandidate, o as ResearchReportRecommendation, p as ParetoPoint, r as ResearchReportDecision, s as researchReport, t as ResearchReport, u as GainDistributionFigureSpec, v as summaryTable } from "./summary-report-vdhE20Zf.js";
4
- import { a as PairedEvalueStep, c as SequentialDecision, i as PairedEvalueSequence, l as evaluateInterimReleaseConfidence, n as InterimReleaseConfidenceInput, r as PairedEvalueOptions, t as InterimReleaseConfidence, u as pairedEvalueSequence } from "./sequential-OeSwJ2uG.js";
5
- import { _ as evaluateReleaseConfidence, a as bootstrapCi, c as ReleaseConfidenceAxis, d as ReleaseConfidenceIssue, f as ReleaseConfidenceMetrics, g as ReleaseTraceEvidence, h as ReleaseConfidenceThresholds, i as Verdict, l as ReleaseConfidenceAxisName, m as ReleaseConfidenceStatus, n as BootstrapResult, o as judgeReplayGate, p as ReleaseConfidenceScorecard, r as JudgeReplayGateArgs, t as BootstrapOptions, u as ReleaseConfidenceInput } from "./promotion-gate-XglbDCNo.js";
3
+ import { _ as gainHistogram, a as ResearchReportOptions, c as RESEARCH_REPORT_HARD_PAIR_FLOOR, d as GainDistributionOptions, f as ParetoFigureSpec, g as SummaryTableRow, h as SummaryTableOptions, i as ResearchReportMethodology, l as GainDistributionBin, m as SummaryTable, n as ResearchReportCandidate, o as ResearchReportRecommendation, p as ParetoPoint, r as ResearchReportDecision, s as researchReport, t as ResearchReport, u as GainDistributionFigureSpec, v as summaryTable } from "./summary-report-9kwI6jFQ.js";
4
+ import { a as PairedEvalueStep, c as SequentialDecision, i as PairedEvalueSequence, l as evaluateInterimReleaseConfidence, n as InterimReleaseConfidenceInput, r as PairedEvalueOptions, t as InterimReleaseConfidence, u as pairedEvalueSequence } from "./sequential-1-T2oi1_.js";
5
+ import { _ as evaluateReleaseConfidence, a as bootstrapCi, c as ReleaseConfidenceAxis, d as ReleaseConfidenceIssue, f as ReleaseConfidenceMetrics, g as ReleaseTraceEvidence, h as ReleaseConfidenceThresholds, i as Verdict, l as ReleaseConfidenceAxisName, m as ReleaseConfidenceStatus, n as BootstrapResult, o as judgeReplayGate, p as ReleaseConfidenceScorecard, r as JudgeReplayGateArgs, t as BootstrapOptions, u as ReleaseConfidenceInput } from "./promotion-gate-DG_PSJx_.js";
6
6
  /** A text report column; `none` leaves the cell unpadded. */
7
7
  export interface ReportTableColumn {
8
8
  header: string;
@@ -1,10 +1,10 @@
1
1
  import { i as AgentProfileCellInput, r as AgentProfileCell } from "./agent-profile-cell-3kYxUsK2.js";
2
2
  import { a as RunRecord, c as RunSplitTag, d as RunTokenUsage, l as RunTaskFailure, n as RunCostProvenance, r as RunJudgeMetadata, t as JudgeScoresRecord } from "./run-record-pBlhZizT.js";
3
- import { O as ChatClient, at as RawProviderSink } from "./cluster-CkzP4h3T.js";
3
+ import { O as ChatClient, at as RawProviderSink } from "./cluster-CSd0LE_M.js";
4
4
  import { s as TraceStore } from "./store-CMs5wGXP.js";
5
- import { S as HeldOutGateDecision, a as ResearchReportOptions, t as ResearchReport } from "./summary-report-vdhE20Zf.js";
5
+ import { S as HeldOutGateDecision, a as ResearchReportOptions, t as ResearchReport } from "./summary-report-9kwI6jFQ.js";
6
6
  import { o as TraceEmitter, r as RunCompleteHook } from "./emitter-mAXC5PSN.js";
7
- import { a as RunIntegrityReport, n as RunIntegrityExpectations } from "./integrity-zeWMeTQt.js";
7
+ import { a as RunIntegrityReport, n as RunIntegrityExpectations } from "./integrity-BMyCpnZ8.js";
8
8
  interface CampaignVariant<V> {
9
9
  id: string;
10
10
  payload: V;
@@ -286,4 +286,4 @@ interface Researcher {
286
286
  evaluateChange(plan: ExperimentPlan): Promise<ExperimentResult>;
287
287
  }
288
288
  export { SteeringChange as a, CampaignRunContext as c, CampaignVariant as d, EvalCampaignOptions as f, runEvalCampaign as h, Researcher as i, CampaignRunOutcome as l, FailedRun as m, ExperimentResult as n, CampaignFactoryParams as o, EvalCampaignResult as p, FailureMode as r, CampaignIntegrityPolicy as s, ExperimentPlan as t, CampaignScenario as u };
289
- //# sourceMappingURL=researcher-DlyjP5uq.d.ts.map
289
+ //# sourceMappingURL=researcher-BuojS14L.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"researcher-DlyjP5uq.d.ts","names":[],"sources":["../src/eval-campaign.ts","../src/researcher.ts"],"mappings":";;;;;;;UAuEiB,gBAAgB;EAC/B;EACA,SAAS;;UAGM;EACf;;EAEA,OAAO;;UAGQ,mBAAmB;;EAElC;;EAEA;EACA,SAAS;EACT;EACA;EACA,cAAc;EACd;EACA,UAAU;;;;;;;EAOV,SAAS;EACT,OAAO;EACP,SAAS;;;;;EAKT,MAAM;;;UAIS;;;;;;EAMf,SAAS;EACT;;UAGQ;;EAER;;EAEA;;EAEA;;EAEA,gBAAgB;EAChB,YAAY;;EAEZ;;EAEA;;EAEA;;EAEA,MAAM;;EAEN,gBAAgB;;;;;;EAMhB,cAAc;;;;;;EAMd,eAAe,mBAAmB;;;KAIxB,qBAAqB,2BAA2B;KAEhD,eAAe,MAAM,KAAK,mBAAmB,OAAO,QAAQ;KAE5D;UAEK,oBAAoB;;;;;EAKnC;EACA,UAAU,gBAAgB;EAC1B,WAAW;;EAEX;;EAEA,WAAW;;EAEX;;;;;;;;EAQA,cAAc,QAAQ,uBAAuB;;;;;;;EAO7C;;;;;;EAMA,eAAe,QAAQ,0BAA0B;;;;;;;EAOjD,kBAAkB,QAAQ,0BAA0B;;;;;EAKpD;;;;;EAKA,gBAAgB;;;;;;EAMhB,YAAY;;EAEZ,qBAAqB;;;;;EAKrB,QAAQ,eAAe;;;;;EAKvB;IAAW;MAAwB,KACjC;;;;;EAOF;;EAEA;;;;EAIA;;EAEA,SAAS,QAAQ;;;;;;;EAOjB,eACI,mBACA,0BAEE,QAAQ;IACN,SAAS;IACT,cAAc;QAGd,mBACA,wBACA,QAAQ,mBAAmB;;UAGpB;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;EACA;;EAEA,MAAM;;EAEN,kBAAkB;EAClB,YAAY;;EAEZ,SAAS;EACT;EACA;;iBAyDoB,gBAAgB,GACpC,MAAM,oBAAoB,KACzB,QAAQ;;UChUM;;;EAGf;;EAEA;EACA;;;IAGE;;IAEA;;;;UAKa;EACf;;;;EAIA;;;EAGA;;EAEA;;;UAIe;EACf;EACA;EACA,SAAS;;;EAGT;;EAEA;IAAU;IAAkB;;;;UAIb;EACf,MAAM;EACN,MAAM;EACN,cAAc;;;;;;;;;;;;;;;;;;;UAoBC;EACf,gBAAgB,MAAM,cAAc,QAAQ;EAC5C,cAAc,UAAU,gBAAgB,QAAQ;EAChD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAC1E,eAAe,MAAM,iBAAiB,QAAQ"}
1
+ {"version":3,"file":"researcher-BuojS14L.d.ts","names":[],"sources":["../src/eval-campaign.ts","../src/researcher.ts"],"mappings":";;;;;;;UAuEiB,gBAAgB;EAC/B;EACA,SAAS;;UAGM;EACf;;EAEA,OAAO;;UAGQ,mBAAmB;;EAElC;;EAEA;EACA,SAAS;EACT;EACA;EACA,cAAc;EACd;EACA,UAAU;;;;;;;EAOV,SAAS;EACT,OAAO;EACP,SAAS;;;;;EAKT,MAAM;;;UAIS;;;;;;EAMf,SAAS;EACT;;UAGQ;;EAER;;EAEA;;EAEA;;EAEA,gBAAgB;EAChB,YAAY;;EAEZ;;EAEA;;EAEA;;EAEA,MAAM;;EAEN,gBAAgB;;;;;;EAMhB,cAAc;;;;;;EAMd,eAAe,mBAAmB;;;KAIxB,qBAAqB,2BAA2B;KAEhD,eAAe,MAAM,KAAK,mBAAmB,OAAO,QAAQ;KAE5D;UAEK,oBAAoB;;;;;EAKnC;EACA,UAAU,gBAAgB;EAC1B,WAAW;;EAEX;;EAEA,WAAW;;EAEX;;;;;;;;EAQA,cAAc,QAAQ,uBAAuB;;;;;;;EAO7C;;;;;;EAMA,eAAe,QAAQ,0BAA0B;;;;;;;EAOjD,kBAAkB,QAAQ,0BAA0B;;;;;EAKpD;;;;;EAKA,gBAAgB;;;;;;EAMhB,YAAY;;EAEZ,qBAAqB;;;;;EAKrB,QAAQ,eAAe;;;;;EAKvB;IAAW;MAAwB,KACjC;;;;;EAOF;;EAEA;;;;EAIA;;EAEA,SAAS,QAAQ;;;;;;;EAOjB,eACI,mBACA,0BAEE,QAAQ;IACN,SAAS;IACT,cAAc;QAGd,mBACA,wBACA,QAAQ,mBAAmB;;UAGpB;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;EACA;;EAEA,MAAM;;EAEN,kBAAkB;EAClB,YAAY;;EAEZ,SAAS;EACT;EACA;;iBAyDoB,gBAAgB,GACpC,MAAM,oBAAoB,KACzB,QAAQ;;UChUM;;;EAGf;;EAEA;EACA;;;IAGE;;IAEA;;;;UAKa;EACf;;;;EAIA;;;EAGA;;EAEA;;;UAIe;EACf;EACA;EACA,SAAS;;;EAGT;;EAEA;IAAU;IAAkB;;;;UAIb;EACf,MAAM;EACN,MAAM;EACN,cAAc;;;;;;;;;;;;;;;;;;;UAoBC;EACf,gBAAgB,MAAM,cAAc,QAAQ;EAC5C,cAAc,UAAU,gBAAgB,QAAQ;EAChD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAC1E,eAAe,MAAM,iBAAiB,QAAQ"}
package/dist/rl.d.ts CHANGED
@@ -1,16 +1,16 @@
1
1
  import { _ as Span } from "./schema-CkIvnmLU.js";
2
2
  import { a as RunRecord, c as RunSplitTag } from "./run-record-pBlhZizT.js";
3
- import { E as PairedBootstrapResult } from "./index-BN6htPVi.js";
4
- import { Q as MintRolloutOptions, a as CampaignResult } from "./types-C9yEqDjQ.js";
3
+ import { E as PairedBootstrapResult } from "./index-DHWFYUVW.js";
4
+ import { Q as MintRolloutOptions, a as CampaignResult } from "./types-ByrbgALT.js";
5
5
  import { a as MintedRolloutLine, b as RolloutSplit, t as ChatMessage } from "./schema-PDKQ_iNE.js";
6
- import { a as PairedPromotionDecision } from "./paired-promotion-decision-COeOgNHG.js";
6
+ import { a as PairedPromotionDecision } from "./paired-promotion-decision-DawN2o0G.js";
7
7
  import { a as detectRewardHacking, c as VerifiableRewardSource, i as RewardHackingSignal, l as extractVerifiableReward, n as RewardHackingFinding, o as VerifiableReward, r as RewardHackingReport, s as VerifiableRewardExtractionOptions, t as DetectRewardHackingInput, u as extractVerifiableRewardsFromRecords } from "./reward-hacking-CIdVwVe4.js";
8
8
  import { a as RubricPredictiveValidityReport, d as FileSystemOutcomeStore, f as FileSystemOutcomeStoreOptions, h as OutcomeStore, p as InMemoryOutcomeStore, t as OutcomeMetricSpec, u as DeploymentOutcome } from "./rubric-predictive-validity-Bbi88Aut.js";
9
- import { a as SystemOneReviewReport, t as PreparedSystemOneReview } from "./systemone-review-4CKmHvr5.js";
10
- import { n as TraceQuestionOutcome, r as TraceQuestionSpec } from "./trace-questions-oT2fZWJ8.js";
11
- import { a as SteeringChange, f as EvalCampaignOptions, h as runEvalCampaign, i as Researcher, n as ExperimentResult, p as EvalCampaignResult, r as FailureMode, t as ExperimentPlan } from "./researcher-DlyjP5uq.js";
9
+ import { a as SystemOneReviewReport, t as PreparedSystemOneReview } from "./systemone-review-WpWosn3a.js";
10
+ import { n as TraceQuestionOutcome, r as TraceQuestionSpec } from "./trace-questions-DI42jXPj.js";
11
+ import { a as SteeringChange, f as EvalCampaignOptions, h as runEvalCampaign, i as Researcher, n as ExperimentResult, p as EvalCampaignResult, r as FailureMode, t as ExperimentPlan } from "./researcher-BuojS14L.js";
12
12
  import { o as TrainingExportOptions } from "./exporters-dQKYD_Zo.js";
13
- import { t as InterimReleaseConfidence } from "./sequential-OeSwJ2uG.js";
13
+ import { t as InterimReleaseConfidence } from "./sequential-1-T2oi1_.js";
14
14
  import { t as AdversarialMutation } from "./adversarial-DnQQFqTR.js";
15
15
  /**
16
16
  * Adaptive curriculum: active scenario selection by variance.