@tangle-network/agent-eval 0.222.1 → 1.0.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +191 -0
- package/dist/adapters/http.d.ts +2 -2
- package/dist/{agent-profile-CSEQRC9l.d.ts → agent-profile-CoD1Dotf.d.ts} +2 -2
- package/dist/{agent-profile-CSEQRC9l.d.ts.map → agent-profile-CoD1Dotf.d.ts.map} +1 -1
- package/dist/analyst/index.d.ts +11 -11
- package/dist/analyst/index.d.ts.map +1 -1
- package/dist/analyst/index.js +1 -1
- package/dist/{analyze-runs-CGLRgZGd.d.ts → analyze-runs-C-pP1sK9.d.ts} +2 -2
- package/dist/{analyze-runs-CGLRgZGd.d.ts.map → analyze-runs-C-pP1sK9.d.ts.map} +1 -1
- package/dist/{analyze-runs-CTheASYV.js → analyze-runs-Ck68mbgj.js} +7 -9
- package/dist/analyze-runs-Ck68mbgj.js.map +1 -0
- package/dist/{benchmark-command-Cr7i7YSX.js → benchmark-command-DWkc5mzq.js} +32 -37
- package/dist/benchmark-command-DWkc5mzq.js.map +1 -0
- package/dist/benchmarks/index.d.ts +4 -4
- package/dist/benchmarks/index.js +3 -3
- package/dist/campaign/index.d.ts +6 -6
- package/dist/campaign/index.js +9 -9
- package/dist/{campaign-ddr1N3yD.js → campaign-DdOgQfp8.js} +14 -19
- package/dist/campaign-DdOgQfp8.js.map +1 -0
- package/dist/{campaign-evidence-DN8rWuBj.js → campaign-evidence-MjCjkb5a.js} +3 -3
- package/dist/{campaign-evidence-DN8rWuBj.js.map → campaign-evidence-MjCjkb5a.js.map} +1 -1
- package/dist/cli.js +4 -4
- package/dist/contract/index.d.ts +9 -9
- package/dist/contract/index.js +11 -10
- package/dist/contract/index.js.map +1 -1
- package/dist/{default-registry-1f4ukAIX.d.ts → default-registry-B1DstPYg.d.ts} +7 -7
- package/dist/{default-registry-1f4ukAIX.d.ts.map → default-registry-B1DstPYg.d.ts.map} +1 -1
- package/dist/{define-agent-eval-sUYh9BFA.d.ts → define-agent-eval-CAffiFbN.d.ts} +7 -7
- package/dist/{define-agent-eval-sUYh9BFA.d.ts.map → define-agent-eval-CAffiFbN.d.ts.map} +1 -1
- package/dist/{define-agent-eval-BLoP4Lf9.js → define-agent-eval-Cc15qGik.js} +5 -5
- package/dist/{define-agent-eval-BLoP4Lf9.js.map → define-agent-eval-Cc15qGik.js.map} +1 -1
- package/dist/diagnosis/index.d.ts +2 -2
- package/dist/diagnosis/index.js +1 -1
- package/dist/{engine-BITiErg-.d.ts → engine-CaHaCc3W.d.ts} +2 -2
- package/dist/{engine-BITiErg-.d.ts.map → engine-CaHaCc3W.d.ts.map} +1 -1
- package/dist/{eval-campaign-Cgixrvf7.js → eval-campaign-C_mNDshv.js} +2 -2
- package/dist/{eval-campaign-Cgixrvf7.js.map → eval-campaign-C_mNDshv.js.map} +1 -1
- package/dist/{evaluation-Djq7-Ybj.d.ts → evaluation-BPqUvZ66.d.ts} +3 -3
- package/dist/{evaluation-Djq7-Ybj.d.ts.map → evaluation-BPqUvZ66.d.ts.map} +1 -1
- package/dist/evaluation.d.ts +1 -1
- package/dist/{exact-types-CndDT_h7.d.ts → exact-types-D98BKpxo.d.ts} +2 -2
- package/dist/{exact-types-CndDT_h7.d.ts.map → exact-types-D98BKpxo.d.ts.map} +1 -1
- package/dist/experiment/index.d.ts +1 -1
- package/dist/experiment/index.js +7 -7
- package/dist/{feedback-trajectory-DsHR9krG.d.ts → feedback-trajectory-SBGncrCP.d.ts} +3 -3
- package/dist/{feedback-trajectory-DsHR9krG.d.ts.map → feedback-trajectory-SBGncrCP.d.ts.map} +1 -1
- package/dist/{index-Bdcpx2s8.d.ts → index-B7v2QHlD.d.ts} +6 -6
- package/dist/{index-Bdcpx2s8.d.ts.map → index-B7v2QHlD.d.ts.map} +1 -1
- package/dist/{index-IzWTz5EP.d.ts → index-CUjeTDPv.d.ts} +6 -6
- package/dist/{index-IzWTz5EP.d.ts.map → index-CUjeTDPv.d.ts.map} +1 -1
- package/dist/index-DAk2quYu.d.ts.map +1 -1
- package/dist/{index-BdKzYYkC.d.ts → index-DUn0GNiZ.d.ts} +2 -2
- package/dist/{index-BdKzYYkC.d.ts.map → index-DUn0GNiZ.d.ts.map} +1 -1
- package/dist/{index-Cu12qxJb.d.ts → index-aDPVYBIM.d.ts} +8 -8
- package/dist/{index-Cu12qxJb.d.ts.map → index-aDPVYBIM.d.ts.map} +1 -1
- package/dist/index.d.ts +17 -17
- package/dist/index.js +13 -12
- package/dist/index.js.map +1 -1
- package/dist/{judge-gate-C2BCOt-M.js → judge-gate-Di_6bifh.js} +4 -4
- package/dist/judge-gate-Di_6bifh.js.map +1 -0
- package/dist/{llm-judge-cgbhTKQp.js → llm-judge-D6JtIqiU.js} +9 -9
- package/dist/{llm-judge-cgbhTKQp.js.map → llm-judge-D6JtIqiU.js.map} +1 -1
- package/dist/{matrix-8X6i5kJ6.d.ts → matrix-CBUzIZs4.d.ts} +2 -2
- package/dist/{matrix-8X6i5kJ6.d.ts.map → matrix-CBUzIZs4.d.ts.map} +1 -1
- package/dist/mean-BPuUNznQ.js +16 -0
- package/dist/mean-BPuUNznQ.js.map +1 -0
- package/dist/meta-eval/index.js +1 -1
- package/dist/multishot/golden/index.d.ts +1 -1
- package/dist/multishot/index.d.ts +2 -2
- package/dist/openapi.json +1 -1
- package/dist/{paired-arms-D-54hO-k.js → paired-arms-D5H_58e8.js} +2 -2
- package/dist/{paired-arms-D-54hO-k.js.map → paired-arms-D5H_58e8.js.map} +1 -1
- package/dist/{paired-promotion-decision-FGvHA4p_.js → paired-promotion-decision-BiVumgvM.js} +2 -2
- package/dist/{paired-promotion-decision-FGvHA4p_.js.map → paired-promotion-decision-BiVumgvM.js.map} +1 -1
- package/dist/{mean-CdDYG-us.js → paired-tests-C2zUqgUW.js} +2 -7
- package/dist/paired-tests-C2zUqgUW.js.map +1 -0
- package/dist/{pareto-B9udAt2i.js → pareto-7_LEmYGG.js} +4 -3
- package/dist/{pareto-B9udAt2i.js.map → pareto-7_LEmYGG.js.map} +1 -1
- package/dist/{power-and-mde-ClI6RWw4.js → power-and-mde-CHvyDO_H.js} +2 -2
- package/dist/{power-and-mde-ClI6RWw4.js.map → power-and-mde-CHvyDO_H.js.map} +1 -1
- package/dist/{prime-protocol-BpQlTowq.d.ts → prime-protocol-COFkxas6.d.ts} +2 -2
- package/dist/{prime-protocol-BpQlTowq.d.ts.map → prime-protocol-COFkxas6.d.ts.map} +1 -1
- package/dist/{promotion-policy-CBYeD4zi.js → promotion-policy-D7B4_jEZ.js} +4 -4
- package/dist/{promotion-policy-CBYeD4zi.js.map → promotion-policy-D7B4_jEZ.js.map} +1 -1
- package/dist/{registry-D9A7Tybn.d.ts → registry-dhdDoAV8.d.ts} +3 -3
- package/dist/{registry-D9A7Tybn.d.ts.map → registry-dhdDoAV8.d.ts.map} +1 -1
- package/dist/reporting.js +3 -3
- package/dist/{reward-hacking-DZzHZlS1.js → reward-hacking-Bjk3MAjh.js} +7 -9
- package/dist/{reward-hacking-DZzHZlS1.js.map → reward-hacking-Bjk3MAjh.js.map} +1 -1
- package/dist/reward-hacking-C9hPVT0Q.d.ts.map +1 -1
- package/dist/rl.d.ts +3 -3
- package/dist/rl.js +6 -6
- package/dist/rollout/index.d.ts +2 -2
- package/dist/{run-record-CZy4-TnZ.js → run-record-BUm-Jggo.js} +2 -2
- package/dist/{run-record-CZy4-TnZ.js.map → run-record-BUm-Jggo.js.map} +1 -1
- package/dist/search/index.d.ts +2 -2
- package/dist/search/index.js +2 -2
- package/dist/{search-kernel-B77eE7Ac.d.ts → search-kernel-CBdwFZ05.d.ts} +2 -2
- package/dist/{search-kernel-B77eE7Ac.d.ts.map → search-kernel-CBdwFZ05.d.ts.map} +1 -1
- package/dist/{search-summary-Ct7SrmVc.js → search-summary-Bt3uuK53.js} +4 -4
- package/dist/{search-summary-Ct7SrmVc.js.map → search-summary-Bt3uuK53.js.map} +1 -1
- package/dist/{sequential-B1I0tRfL.js → sequential-De7BTAlV.js} +2 -2
- package/dist/{sequential-B1I0tRfL.js.map → sequential-De7BTAlV.js.map} +1 -1
- package/dist/{skillopt-optimization-method-qB1M554K.js → skillopt-optimization-method-BKFbJEyz.js} +9 -12
- package/dist/skillopt-optimization-method-BKFbJEyz.js.map +1 -0
- package/dist/{statistical-heldout-TP6ndL0B.js → statistical-heldout-C9hz_AF9.js} +3 -3
- package/dist/{statistical-heldout-TP6ndL0B.js.map → statistical-heldout-C9hz_AF9.js.map} +1 -1
- package/dist/{statistical-heldout-BYxHe3KO.d.ts → statistical-heldout-Cei036nE.d.ts} +3 -3
- package/dist/{statistical-heldout-BYxHe3KO.d.ts.map → statistical-heldout-Cei036nE.d.ts.map} +1 -1
- package/dist/{store-tool-spans-DMM6wMwB.d.ts → store-tool-spans-B3i3fl0A.d.ts} +4 -4
- package/dist/{store-tool-spans-DMM6wMwB.d.ts.map → store-tool-spans-B3i3fl0A.d.ts.map} +1 -1
- package/dist/storyboard/index.d.ts +17 -0
- package/dist/storyboard/index.d.ts.map +1 -1
- package/dist/storyboard/index.js +154 -107
- package/dist/storyboard/index.js.map +1 -1
- package/dist/{summary-report-CPfsrLac.js → summary-report-BLnTyQKH.js} +5 -4
- package/dist/{summary-report-CPfsrLac.js.map → summary-report-BLnTyQKH.js.map} +1 -1
- package/dist/{systemone-UHbKD6LH.d.ts → systemone-Cw2txK5m.d.ts} +5 -5
- package/dist/{systemone-UHbKD6LH.d.ts.map → systemone-Cw2txK5m.d.ts.map} +1 -1
- package/dist/{systemone-review-Cn9Uj0xG.d.ts → systemone-review-BLgxt-4Q.d.ts} +3 -3
- package/dist/{systemone-review-Cn9Uj0xG.d.ts.map → systemone-review-BLgxt-4Q.d.ts.map} +1 -1
- package/dist/systemone.d.ts +3 -3
- package/dist/{tool-groups-CFWmAGCV.d.ts → tool-groups-C3K-d3gd.d.ts} +3 -3
- package/dist/tool-groups-C3K-d3gd.d.ts.map +1 -0
- package/dist/{trace-questions-BFCQAzoT.d.ts → trace-questions-ej5AEJ5X.d.ts} +3 -3
- package/dist/{trace-questions-BFCQAzoT.d.ts.map → trace-questions-ej5AEJ5X.d.ts.map} +1 -1
- package/dist/trace-repair/index.d.ts +1 -1
- package/dist/trace-repair/index.js +2 -1
- package/dist/trace-repair/index.js.map +1 -1
- package/dist/traces.d.ts +3 -3
- package/dist/{tree-DrIChF4a.js → tree-DnBkWXun.js} +5 -4
- package/dist/{tree-DrIChF4a.js.map → tree-DnBkWXun.js.map} +1 -1
- package/dist/{types-DUR0dqET.d.ts → types-0oMg7zee.d.ts} +2 -2
- package/dist/{types-DUR0dqET.d.ts.map → types-0oMg7zee.d.ts.map} +1 -1
- package/dist/{types-e3HGaI7P.d.ts → types-C54Uk_a1.d.ts} +2 -2
- package/dist/{types-e3HGaI7P.d.ts.map → types-C54Uk_a1.d.ts.map} +1 -1
- package/dist/{types-BEvjDziM.d.ts → types-DzCO-S1m.d.ts} +8 -8
- package/dist/{types-BEvjDziM.d.ts.map → types-DzCO-S1m.d.ts.map} +1 -1
- package/dist/wire/index.d.ts +1 -1
- package/package.json +5 -5
- package/dist/analyze-runs-CTheASYV.js.map +0 -1
- package/dist/benchmark-command-Cr7i7YSX.js.map +0 -1
- package/dist/campaign-ddr1N3yD.js.map +0 -1
- package/dist/judge-gate-C2BCOt-M.js.map +0 -1
- package/dist/mean-CdDYG-us.js.map +0 -1
- package/dist/skillopt-optimization-method-qB1M554K.js.map +0 -1
- package/dist/tool-groups-CFWmAGCV.d.ts.map +0 -1
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index-DAk2quYu.d.ts","names":[],"sources":["../src/statistics/agreement-irr.ts","../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":";;;;;;;;;;;;;;;;;;;;;;;;;;iBAyBgB,sBAAsB,aAAa;UA0ElC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe,oCAAoC;EACnD;;EAEA;;EAEA;;UAGe;;EAEf,cAAc;;EAEd;;EAEA;;EAEA;;EAEA;;UAGe,+BAA+B;;;;;;EAM9C;;;;;;EAMA;;;;;;;;;;;;;;;;;;iBAmBc,0BACd,SAAS,qBACT,OAAM,yBACL;;;;;;;;;iBA0Ha,yCACd,aAAa;EAAQ;EAAgB,QAAQ;IAC7C,OAAM,yBACL;UC9Qc;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;;UC1N1C;EACf;EACA;;UAGe;;EAEf;;EAEA;;EAEA;;UAGe;EACf;EACA;;EAEA;EACA;EACA;EACA;EACA;;;;;;;;;;;;iBAac,qBACd,uBAAuB,wBACvB,UAAS,0BACR;cCjCU,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;
|
|
1
|
+
{"version":3,"file":"index-DAk2quYu.d.ts","names":[],"sources":["../src/statistics/agreement-irr.ts","../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":";;;;;;;;;;;;;;;;;;;;;;;;;;iBAyBgB,sBAAsB,aAAa;UA0ElC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe,oCAAoC;EACnD;;EAEA;;EAEA;;UAGe;;EAEf,cAAc;;EAEd;;EAEA;;EAEA;;EAEA;;UAGe,+BAA+B;;;;;;EAM9C;;;;;;EAMA;;;;;;;;;;;;;;;;;;iBAmBc,0BACd,SAAS,qBACT,OAAM,yBACL;;;;;;;;;iBA0Ha,yCACd,aAAa;EAAQ;EAAgB,QAAQ;IAC7C,OAAM,yBACL;UC9Qc;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;;UC1N1C;EACf;EACA;;UAGe;;EAEf;;EAEA;;EAEA;;UAGe;EACf;EACA;;EAEA;EACA;EACA;EACA;EACA;;;;;;;;;;;;iBAac,qBACd,uBAAuB,wBACvB,UAAS,0BACR;cCjCU,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"}
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
import { a as MintedRolloutLine, b as RolloutSplit, g as RolloutOutcome, h as RolloutLine, i as GatedEvidence, t as ChatMessage, w as ToolDef } from "./schema-DkFJr3Y8.js";
|
|
2
|
-
import { l as CodeSurface } from "./types-
|
|
2
|
+
import { l as CodeSurface } from "./types-C54Uk_a1.js";
|
|
3
3
|
import "./reward-ST4PXNXF.js";
|
|
4
4
|
import { n as CompletionVerdict } from "./completion-verifier-7KEfroqx.js";
|
|
5
5
|
import { o as TrainingExportOptions } from "./exporters-B5Ze_RKh.js";
|
|
@@ -733,4 +733,4 @@ interface TinkerExportOptions extends Pick<TrainingExportOptions, 'allowHeldOutT
|
|
|
733
733
|
*/
|
|
734
734
|
declare function toTinkerDatumRows(lines: readonly MintedRolloutLine[], options: TinkerExportOptions): TinkerDatumRow[];
|
|
735
735
|
export { HarborMetrics as $, harnessSessionRolloutLine as A, ReadClaudeTranscriptOptions as B, GateReport as C, GateEntryPoint as Ct, readOpencodeSessionMessages as D, DEFAULT_OPENCODE_DB as E, gatedEvidenceOf as Et, readCodexRollout as F, readRolloutJournal as G, parseClaudeTranscript as H, ClaudeTranscript as I, FromHarborOptions as J, readRolloutLedger as K, ClaudeTranscriptRef as L, gateReceiptRolloutLines as M, CodexRollout as N, HarnessSessionImportOptions as O, parseCodexRollout as P, HarborImageSource as Q, ClaudeUsageTotals as R, GateDisposition as S, GateCheckedOutcome as St, assertGateReport as T, gateErrors as Tt, readClaudeTranscript as U, findClaudeTranscripts as V, appendRolloutLines as W, HarborContentPart as X, HarborAgent as Y, HarborFinalMetrics as Z, privacyHit as _, GATE_CHECK_IDS as _t, BuildOptions as a, HarborToolCall as at, EmittedEvidence as b, GateCheckDisposition as bt, ScrubReport as c, relabelImportedSplit as ct, ReleaseFormat as d, HARBOR_TASK_CHECKS as dt, HarborObservation as et, buildDatasetCard as f, HarborTaskFile as ft, defaultRolloutScrubber as g, GATE_CHECKS as gt, ScrubRule as h, writeHarborTask as ht, toTinkerDatumRows as i, HarborSubagentTrajectoryRef as it, harnessSessionRolloutLines as j, harnessSessionRefs as k, buildHfDataset as l, toHarborTrajectories as lt, ScrubCounts as m, WriteHarborTaskOptions as mt, TinkerDatumRow as n, HarborStep as nt, BuildSummary as o, HarborTrajectory as ot, PRIVACY_RULE_NAMES as p, HarborTaskVerification as pt, writeRolloutLedger as q, TinkerExportOptions as r, HarborStepSource as rt, RolloutReleaseCliArgs as s, fromHarborTrajectory as st, TinkerDatum as t, HarborObservationResult as tt, DatasetCardInputs as u, toHarborTrajectory as ut, scrubLines as v, GATE_POLICIES as vt, ReleaseRowRef as w, GatePolicy as wt, FormatGateCounts as x, GateCheckId as xt, scrubText as y, GateCheck as yt, DEFAULT_CLAUDE_PROJECTS_DIR as z };
|
|
736
|
-
//# sourceMappingURL=index-
|
|
736
|
+
//# sourceMappingURL=index-DUn0GNiZ.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index-
|
|
1
|
+
{"version":3,"file":"index-DUn0GNiZ.d.ts","names":[],"sources":["../src/rollout/gate-checks.ts","../src/rollout/harbor-task.ts","../src/rollout/interchange/harbor.ts","../src/rollout/ledger.ts","../src/rollout/readers/claude-jsonl.ts","../src/rollout/readers/codex-rollout.ts","../src/rollout/readers/gate-receipts.ts","../src/rollout/readers/harness-session-lines.ts","../src/rollout/readers/opencode-sqlite.ts","../src/rollout/release/gate-report.ts","../src/rollout/release/scrub.ts","../src/rollout/release/card.ts","../src/rollout/release/hf-dataset.ts","../src/rollout/tinker.ts"],"mappings":";;;;;;;;;;;;;cA2Ca;KAOD,sBAAsB;;;;;;;;;KAUtB,qBAAqB,QAAQ,kBAAkB,SAAS;;;;;;;;;;;;;;;;;UAkB1D;EACR,SAAS;;EAET;;UAOe;EACf,IAAI;;EAEJ;;EAEA,SAAS,SAAS;;;;;;;;;;;;EAYlB,WAAW;;;;;;;;;;;;;iBA6LG,gBAAgB,SAAS,cAAc;;;;;;cA6I1C,yBAAyB,KAAK,cAAc;;;;;;;;;KAe7C;WACG;;;;WAEA;WAAyB;;;;WAEzB;WAAuB;;;KAU1B,yBAAyB,KAAK,cAAc;;;;;;;;;cAU3C;;;;;;;;eAxBQ;;;;;;;;;;;;;eAAA;;;;;eAAA;;;;;;;;;;;eAAA;;;eAAA;;;eAAA;;;eAAA;;;;;;;;;;;eAAA;;;eAAA;;;eAAA;;;eAAA;;;;;KA+FT,8BAA8B;;;;;;iBAO1B,WAAW,SAAS,aAAa,QAAQ;;cCphB5C;UAQI;;WAEN;;WAEA;WACA;WACA;;UAGM;;WAEN;WACA;WACA;WACA,MAAM;WACN,QAAQ,KAAK;WACb;WACA,gBAAgB;;UAGV;;WAEN;WACA,SAAS;WACT,SAAS;WACT;;WAEA;WACA;;WAEA;;WAEA,YAAY;WACZ;aACE;aACA;aACA;;WAEF;WACA;;;;;;;;;;WAUA,SACP,SAAS,wBACT,SAAS,gBACN,QAAQ;WACJ,SAAS;;;;;;;iBAgDE,gBAAgB,SAAS,yBAAyB;KCrB5D;UAEK;EACf;EACA;;UAGe;EACf;EACA;EACA,SAAS;;UAGM;EACf;EACA;;EAEA,WAAW;EACX,QAAQ;;UAGO;EACf;EACA;;EAEA;EACA,QAAQ;;UAGO;EACf;EACA,mBAAmB;EACnB,0BAA0B;EAC1B,QAAQ;;UAGO;EACf,SAAS;;UAGM;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA,QAAQ;;UAGO;;EAEf;EACA;EACA,QAAQ;EACR;EACA;EACA,kBAAkB;EAClB;EACA,aAAa;EACb,cAAc;EACd,UAAU;EACV;EACA;EACA,QAAQ;;UAGO;EACf;EACA;EACA;;EAEA,mBAAmB;EACnB,QAAQ;;UAGO;EACf;EACA;EACA;EACA;EACA;EACA,QAAQ;;UAGO;EACf;EACA;;EAEA;EACA,OAAO;EACP,OAAO;EACP;EACA,gBAAgB;EAChB;EACA,wBAAwB;EACxB,QAAQ;;;;;;;;;;;;;;;;;;iBAgXM,mBAAmB,OAAO,gBAAgB;;iBAa1C,qBAAqB,OAAO,gBAAgB;UA2O3C;;EAEf,YAAY;;;;;;;;;;;iBAwGE,qBACd,YAAY,kBACZ,UAAS,oBACR;;;;;;;;;;;;;;;iBAiCa,qBACd,gBAAgB,eAChB,OAAO,eACN;;iBC76BmB,mBAAmB,cAAc,OAAO,gBAAgB;;iBAOxD,mBAAmB,cAAc,OAAO,gBAAgB;;;;;;;;;;;iBAiBxD,kBAAkB,eAAe,QAAQ;;;;;;iBASzC,mBAAmB,eAAe,QAAQ;cCnCnD;UAOI;EACf;EACA;;;iBAIoB,sBACpB,aACA,uBACC,QAAQ;UASM;EACf;EACA;EACA;EACA;;UAGe;EACf,UAAU;EACV,OAAO;;EAEP;EACA;;EAEA;;EAEA;;UAGe;;;;;WAKN;;;iBAqDW,qBACpB,cACA,UAAS,8BACR,QAAQ;;iBAQK,sBACd,aACA,UAAS,8BACR;;KC3HS,eAAe,KACzB;EAUA,UAAU;;EAEV;EACA;;;iBA2BoB,iBACpB,cACA,UAAS,cACR,QAAQ;;iBAKK,kBACd,aACA,SAAQ,kBACR,UAAS,KAAK,6BACb;iBCxDmB,wBACpB,cACA,sBACC;EAAU,OAAO;EAAe;;UCSlB;;EAEf;;EAEA;EACA;;;iBAWoB,mBACpB,gBACA,UAAS,8BACR,QAAQ;;iBAwCK,0BACd,SAAS,gBACT,KAAK,YACL,sBACC;;iBA4EoB,2BACrB,eAAe,cACf,sBACC;EAAiB,KAAK;EAAY,OAAO;EAAa;;cCtJ5C;;iBAeS,4BACpB,mBACA,cACC,QAAQ;;KCoBC;;UAUK;EACf;EACA;;;;;;;;;;;;EAYA;;;;;;;;EAQA;;;;;;;;;;;;EAYA;;;UAIe;;EAEf;EACA;;UAGe;;EAEf;;EAEA;;;;;;EAMA;;EAEA;;;;;;;;;;;;;;EAcA,oBAAoB;;;;;;;;;;;;EAYpB;;EAEA;;;;;;;;;;EAUA,wBAAwB;;UAGT;;EAEf;EACA,UAAU,QAAQ,OAAO,eAAe;;;;;;;;;;;;;;;;;iBA6K1B,iBAAiB,QAAQ;UCrUxB;EACf;EACA,SAAS;;EAET,UAAU,eAAe;;;KAqEf,cAAc;iBAWV,UAAU,cAAc,QAAQ;iBAkChC,WAAW,OAAO;EAChC,OAAO;EACP,QAAQ;;;;;;;;;iBAaM,uBAAuB;cAsF1B;;;;;;iBAkBG,WAAW,gBAAgB;cC7O9B;KACD,wBAAwB;UAkBnB;;EAEf,OAAO;EACP,SAAS;EACT;;EAEA;EACA,aAAa;EACb;IAAY;IAAmB;IAAkB;;EACjD,cAAc,QAAQ,OAAO;;;;;;;EAO7B,MAAM;;iBA4GQ,iBAAiB,QAAQ;UClHxB;EACf;EACA,SAAS;EACT;;EAEA;;UAGe;;EAEf,OAAO,eAAe;EACtB,QAAQ;EACR;IAAY;IAAmB;IAAkB;;;EAEjD,SAAS;;UAGM;EACf;EACA;EACA;EACA,OAAO;EACP,cAAc,QAAQ,OAAO;;EAE7B,MAAM;EACN;;iBAGoB,eACpB,kBACA,SAAS,eACR,QAAQ;UAkIM,8BAA8B;EAC7C;EACA;EACA;;;UCxMe;EACf;IAAe,QAAQ;MAAQ;MAAsB;;;EACrD;IACE;MAAiB;MAAgB;MAAgB;;IACjD;MAAY;MAAgB;MAAkB;;IAC9C;MAAc;MAAgB;MAAkB;;;;UAInC;;EAEf,OAAO;EACP;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe,4BACP,KAAK;;EAEb;;EAEA,eAAe;;EAEf,cAAc,MAAM,mBAAmB;;;;;;;;iBAsFzB,kBACd,gBAAgB,qBAChB,SAAS,sBACR"}
|
|
@@ -1,13 +1,13 @@
|
|
|
1
1
|
import { o as CostLedgerHandle } from "./cost-ledger-Czro9wmo.js";
|
|
2
2
|
import { a as RunRecord, c as RunSplitTag, n as RunCostProvenance } from "./run-record-D2BFMPGl.js";
|
|
3
|
-
import "./types-
|
|
4
|
-
import "./statistical-heldout-
|
|
5
|
-
import "./index-
|
|
6
|
-
import "./types-
|
|
3
|
+
import "./types-DzCO-S1m.js";
|
|
4
|
+
import "./statistical-heldout-Cei036nE.js";
|
|
5
|
+
import "./index-B7v2QHlD.js";
|
|
6
|
+
import "./types-C54Uk_a1.js";
|
|
7
7
|
import { f as TraceSpanEvent } from "./client-BIFwUg7I.js";
|
|
8
|
-
import { d as InsightReport, t as AnalyzeRunsOptions } from "./analyze-runs-
|
|
9
|
-
import "./define-agent-eval-
|
|
10
|
-
import "./default-registry-
|
|
8
|
+
import { d as InsightReport, t as AnalyzeRunsOptions } from "./analyze-runs-C-pP1sK9.js";
|
|
9
|
+
import "./define-agent-eval-CAffiFbN.js";
|
|
10
|
+
import "./default-registry-B1DstPYg.js";
|
|
11
11
|
import { AgentCandidateBenchmarkCellRef, AgentCandidateBenchmarkSuiteInputs, AgentCandidateBenchmarkTask, AgentCandidateBenchmarkTaskMaterial, AgentCandidateBundle, AgentCandidateEvaluationPolicy, AgentCandidateExperiment, AgentCandidateExperimentMaterial, AgentCandidateExperimentMeasurement, AgentImprovementCost, AgentImprovementMeasuredComparison, AgentProfileImprovementExperiment, AgentProfileImprovementExperimentMaterial, AgentProfileImprovementMeasuredComparison, AgentProfileImprovementMeasurement, AgentProfileImprovementRunCell, AgentProfileImprovementRunReceipt, AgentProfileImprovementSuiteInputs, AgentProfileImprovementTask, AgentProfileImprovementTaskMaterial, CandidateExecutionEvidence, Sha256Digest } from "@tangle-network/agent-interface";
|
|
12
12
|
interface SealCandidateBenchmarkSuiteOptions {
|
|
13
13
|
tasks: [AgentCandidateBenchmarkTask, ...AgentCandidateBenchmarkTask[]];
|
|
@@ -650,4 +650,4 @@ interface FromOtelSpansOptions {
|
|
|
650
650
|
}
|
|
651
651
|
declare function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[];
|
|
652
652
|
export { AgentProfileImprovementExperimentExecutionInput as $, CodeAgentSessionActionSurface as A, AgentTraceIndex as B, CodeAgentSessionIntakeOptions as C, PairedMeasurementAdapter as Ct, CodeAgentSessionAction as D, ParsedCodeAgentJsonl as E, SealCandidateBenchmarkSuiteOptions as Et, observeCodeAgentSession as F, parseAgentTrace as G, AgentTraceRecord as H, AgentTraceContributor as I, EvalReportingSuiteOptions as J, partitionRunsByAuthoringModel as K, AgentTraceContributorType as L, CodeAgentSessionObservation as M, CodeAgentSessionSource as N, CodeAgentSessionActionKind as O, CodeAgentSessionTerminalStatus as P, RunRecordRejection as Q, AgentTraceConversation as R, CodeAgentSessionDiagnostic as S, PairedMeasurement as St, CodeAgentSessionMetrics as T, RunCandidateExperimentOptions as Tt, AuthoringProvenance as U, AgentTraceRange as V, PartitionByAuthoringModelResult as W, FromRunRecordDirOptions as X, EvalReportingSuiteResult as Y, FromRunRecordDirResult as Z, fromPiSession as _, verifyCandidateExperimentComparison as _t, FromFeedbackTableOptions as a, runAgentProfileImprovementExperiment as at, streamCodeAgentJsonlFile as b, CompareCandidateExperimentOptions as bt, CodeAgentStoreObservation as c, sealAgentProfileImprovementTask as ct, CodeAgentStoreScope as d, measuredComparisonFromCandidateExperiment as dt, AgentProfileImprovementExperimentRun as et, CodeAgentStoreSession as f, runCandidateExperiment as ft, fromOpenCodeSession as g, verifyCandidateExperiment as gt, fromKimiCodeSession as h, sealCandidateExperiment as ht, FeedbackTableRow as i, measuredComparisonFromAgentProfileImprovementExperiment as it, CodeAgentSessionExecutionReceipt as j, CodeAgentSessionActionStatus as k, CodeAgentStoreRejection as l, verifyAgentProfileImprovementExperimentComparison as lt, fromCodexSession as m, sealCandidateBenchmarkTask as mt, fromOtelSpans as n, RunAgentProfileImprovementExperimentOptions as nt, FromFeedbackTableResult as o, sealAgentProfileImprovementExperiment as ot, fromClaudeCodeSession as p, sealCandidateBenchmarkSuite as pt, EvalReportingSuiteInput as q, FeedbackTableMeta as r, SealAgentProfileImprovementSuiteOptions as rt, fromFeedbackTable as s, sealAgentProfileImprovementSuite as st, FromOtelSpansOptions as t, CompareAgentProfileImprovementExperimentOptions as tt, CodeAgentStoreRejections as u, evaluatePairedMeasurements as ut, parseCodeAgentJsonl as v, CandidateExperimentExecutionInput as vt, CodeAgentSessionIntakeResult as w, PairedMeasurementEvaluation as wt, CodeAgentJsonlLine as x, EvaluatePairedMeasurementsOptions as xt, parseCodeAgentJsonlFile as y, CandidateExperimentRun as yt, AgentTraceFile as z };
|
|
653
|
-
//# sourceMappingURL=index-
|
|
653
|
+
//# sourceMappingURL=index-aDPVYBIM.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index-
|
|
1
|
+
{"version":3,"file":"index-aDPVYBIM.d.ts","names":[],"sources":["../src/contract/measured-comparison-types.ts","../src/contract/candidate-experiment.ts","../src/contract/paired-measurement-evaluation.ts","../src/contract/profile-measured-comparison.ts","../src/contract/intake/run-record-dir.ts","../src/contract/eval-reporting-suite.ts","../src/contract/intake/agent-trace.ts","../src/contract/intake/code-agent-observation.ts","../src/contract/intake/code-agent-session-types.ts","../src/contract/intake/code-agent-jsonl.ts","../src/contract/intake/code-agent-session.ts","../src/contract/intake/code-agent-store.ts","../src/contract/intake/feedback-table.ts","../src/contract/intake/otel-spans.ts"],"mappings":";;;;;;;;;;;UAeiB;EACf,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,QAAQ;EACR,MAAM;EACN,eAAe;EACf;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QAAQ,OAAO,oCAAoC,QAAQ;;EAE3D;;EAEA,aAAa;EACb,SAAS;;UAGM;EACf,cAAc;EACd;IACE;IACA,MAAM;;;UAIO;EACf,YAAY;EACZ,cAAc;EACd;IACE;IACA,MAAM;;EAER,aAAa;EACb;EACA,YAAY;EACZ;EACA,WAAW;;;UAII,kBAAkB;EACjC;EACA,UAAU;EACV,WAAW;;;UAII,yBAAyB;EACxC,MAAM,KAAK;EACX,WAAW,KAAK;IAAkB;IAAc;;EAChD,QAAQ,KAAK;EACb,eAAe,KAAK,OAAO;EAC3B,UAAU,KAAK;EACf,UAAU,KAAK;EACf,OAAO,KAAK;;UAGG,kCAAkC;EACjD,uBAAuB,kBAAkB;EACzC,QAAQ;EACR,SAAS,yBAAyB;;EAElC;;EAEA,kBAAkB;;EAElB,kBAAkB;;;KAIR,8BAA8B,KACxC;EAGA,iBAAiB;EACjB,WAAW;EACX;;;iBC9Dc,2BACd,UAAU,sCACT;;iBAQa,4BACd,SAAS,qCACR;;iBAiBa,wBACd,UAAU,mCACT;iBAQa,0BAA0B,iBAAiB;;iBAarC,uBACpB,SAAS,gCACR,QAAQ;;iBA4DK,0CACd,SAAS,oCACR;;iBA2Da,oCACd,iBACC;;;;;;;;iBC1La,2BAA2B,MACzC,SAAS,kCAAkC,QAC1C;UCFc;EACf,aAAa;EACb,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,aAAa;EACb,MAAM;EACN,SAAS;EACT;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QACE,OAAO,kDACN,QAAQ;;EAEX;;EAEA,aAAa;EACb,SAAS;;UAGM;EACf,cAAc;EACd;IACE;IACA,MAAM;;;UAIO;EACf,YAAY;EACZ,cAAc;EACd;IACE;IACA,MAAM;;EAER,aAAa;EACb;EACA,YAAY;EACZ;EACA,WAAW;;;iBAIG,gCACd,UAAU,sCACT;;iBAQa,iCACd,SAAS,0CACR;;iBAqBa,sCACd,UAAU,4CACT;;;;;;iBAsBmB,qCACpB,SAAS,8CACR,QAAQ;;iBA0CK,wDACd,SAAS,kDACR;;iBA2Da,kDACd,iBACC;;UClOc;;EAEf;;EAEA;;EAEA;;UAGe;;;;;;EAMf;;;;;;;EAOA,WAAW;;;;;;EAMX;;UAGe;;EAEf,MAAM;;EAEN,UAAU;;EAEV;;;;KCxBU,0BAA0B;UAErB;;;;;EAKf,UAAU,KAAK;;EAEf,OAAO;;;;;;;;;;EAUP;;;;UAKe;;;EAGf,QAAQ;;EAER;;IAEE;;IAEA;;;IAGA;;IAEA;;;IAGA,UAAU;;;EAGZ;;KCnDU;UAEK;EACf,MAAM;;EAEN;;UAGe;EACf;EACA;EACA;;;EAGA,cAAc;;UAGC;EACf;EACA,cAAc;EACd,QAAQ;;UAGO;EACf;EACA,eAAe;;UAGA;EACf;EACA;EACA;EACA;IAAQ;IAAc;;EACtB;IAAS;IAAe;;EACxB,OAAO;;;;UAOQ;EACf;;EAEA;;EAEA;EACA;EACA;;EAEA;;EAEA;;KAGU,kBAAkB,YAAY;;;;;;iBAW1B,gBAAgB,SAAS,qBAAqB;UAmE7C;;;;EAIf,SAAS,YAAY;;;EAGrB,cAAc;;;;;;;;iBASA,8BACd,MAAM,aACN,OAAO,kBACN;;;;;;KCjKS;KAQA;KAEA;KAEA;KASA;UAEK;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA,MAAM;EACN,SAAS;EACT;EACA,QAAQ;EACR;EACA;EACA,UAAU;;UAGK;EACf,QAAQ;EACR;EACA;EACA;IACE,QAAQ;IACR;;EAEF,SAAS;;UAGD;EACR,QAAQ;EACR;EACA;EACA,YAAY;;;;;;;iBAeE,wBACd,SAAS,iCACR;UClFc;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf,QAAQ;EACR;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,WAAW;EACX;;UAGe;EACf,MAAM;EACN,aAAa;EACb,SAAS;EACT,cAAc;;UAGC;EACf;EACA;EACA;EACA;EACA;EACA;EACA,WAAW;EACX;EACA;EACA;EACA;EACA;EACA;;;EAGA,iBAAiB;;;EAGjB,YAAY;;;;KAKF;EACN;EAAe;EAAoB;;EACnC;EAAmB;;iBCvFT,oBAAoB,gBAAgB;;;;;;;;;;iBAuB7B,yBAAyB,eAAe,eAAe;;;;;iBA6BxD,wBAAwB,eAAe,QAAQ;iBCDrD,iBACd,SAAS,gCACR;iBAIa,sBACd,SAAS,gCACR;iBAIa,oBACd,SAAS,gCACR;iBAIa,oBACd,SAAS,gCACR;;iBAKa,cACd,SAAS,gCACR;;UCrDc;;;;;WAKN;;WAEA,QAAQ;;;;;WAKR;;;;;WAKA;aACE;;aAEA;;;;KAKD;;UAUK;WACN;WACA;WACA;WACA;WACA,aAAa;;;UAIP;WACN;WACA,UAAU,SAAS,QAAQ,OAAO;;WAElC,QAAQ;aACN;aACA,QAAQ;;;;;;;;;KAUT;WAEG;WACA,OAAO;WACP;WACA,mBAAmB;;WAEnB;;WAEA;;WAEA;WACA,WAAW,SAAS,QAAQ,OAAO;WACnC,YAAY;;WAGZ;WACA,OAAO;WACP;;WAEA;WACA,YAAY;;UCvGV;;;EAGf;;EAEA;;;EAGA;;;EAGA,WAAW;;UAGI;EACf;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,WAAW;;;EAGX,SAAS;;UAGM;;EAEf,SAAS;;;EAGT,OAAO;;;;EAIP;IAAU;IAAa;;;;;EAIvB;;UAGe;EACf,MAAM;;;EAGN,aAAa;IAAQ;IAAe;IAAe;;;iBAGrC,kBAAkB,MAAM,2BAA2B;UC3BlD;EACf,OAAO;;EAEP,eAAe;;EAEf;;;;;;EAMA,eAAe,eAAe,gBAAgB;;iBAGhC,cAAc,MAAM,uBAAuB"}
|
package/dist/index.d.ts
CHANGED
|
@@ -3,46 +3,46 @@ import { a as JudgeError, c as ValidationError, i as ConfigError, n as AgentEval
|
|
|
3
3
|
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";
|
|
4
4
|
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";
|
|
5
5
|
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-D2BFMPGl.js";
|
|
6
|
-
import { $ as QueryTracesPage, H as createBoundedTraceAnalysisStore, J as TraceAnalysisStore, Q as ErrorCluster, S as makeProposalFinding, X as DEFAULT_TRACE_ANALYST_BUDGETS, Z as DatasetOverview, at as TraceAnalystSpan, b as computeFindingId, c as AnalystRunEvent, ct as TraceAnalystTraceSummary, d as AnalystRunSummary, dt as ViewTraceResult, et as SearchSpanResult, f as AnalystSeverity, i as AnalystFinding, it as TraceAnalystFilters, l as AnalystRunInputs, lt as ViewSpansResult, m as EvidenceRef, n as AnalystContext, nt as SpanMatchRecord, ot as TraceAnalystSpanKind, p as AnalystUsageReceipt, rt as TraceAnalystByteBudgets, st as TraceAnalystSpanStatus, t as Analyst, tt as SearchTraceResult, u as AnalystRunResult, ut as ViewTraceOversized, v as ProposalFinding, x as makeFinding } from "./types-
|
|
6
|
+
import { $ as QueryTracesPage, H as createBoundedTraceAnalysisStore, J as TraceAnalysisStore, Q as ErrorCluster, S as makeProposalFinding, X as DEFAULT_TRACE_ANALYST_BUDGETS, Z as DatasetOverview, at as TraceAnalystSpan, b as computeFindingId, c as AnalystRunEvent, ct as TraceAnalystTraceSummary, d as AnalystRunSummary, dt as ViewTraceResult, et as SearchSpanResult, f as AnalystSeverity, i as AnalystFinding, it as TraceAnalystFilters, l as AnalystRunInputs, lt as ViewSpansResult, m as EvidenceRef, n as AnalystContext, nt as SpanMatchRecord, ot as TraceAnalystSpanKind, p as AnalystUsageReceipt, rt as TraceAnalystByteBudgets, st as TraceAnalystSpanStatus, t as Analyst, tt as SearchTraceResult, u as AnalystRunResult, ut as ViewTraceOversized, v as ProposalFinding, x as makeFinding } from "./types-DzCO-S1m.js";
|
|
7
7
|
import { $ as AssertCrossFamilyServedOptions, A as createChatClient, B as LlmToolChoice, C as ChatResponse, D as MockTransportOpts, E as CustomTransportOpts, G as InMemoryRawProviderSink, H as LlmUsage, I as LlmMessage, M as LlmCallRequest, N as LlmCallResult, O as OpenAiCompatibleTransportOpts, P as LlmChargeBounds, R as LlmTokenLogprob, S as ChatRequest, T as CreateChatClientOpts, U as FileSystemRawProviderSink, V as LlmToolDefinition, X as RawProviderEvent, Z as RawProviderSink, _ as ScenarioResult, a as CompletionCriterion, at as assertCrossFamilyServed, b as ChatCallOpts, c as JudgeFn, ct as checkServedModel, d as PersonaConfig, dt as CrossFamilyError, et as AssertServedModelOptions, f as ProductClientConfig, ft as JudgeFamily, g as ScenarioFile, h as RubricDimension, i as CollectedArtifacts, it as ServedModelVerdict, j as LlmCallMetadata, k as SandboxSdkTransportOpts, l as JudgeInput, lt as describeServedModelCheck, m as RouteMap, mt as judgeFamily, n as ArtifactResult, nt as ServedModelCheck, o as DimensionJudgeScore, ot as assertServedModel, p as ProductScenario, pt as assertCrossFamily, q as NoopRawProviderSink, r as CheckResult, rt as ServedModelPolicy, s as DriverState, st as assertServedModels, t as ArtifactCheck, tt as ModelSubstitutionError, u as JudgeRubric, ut as AssertCrossFamilyOptions, v as Turn, w as ChatTransport, x as ChatClient, y as TurnResult, z as LlmToolCall } from "./types-0KglnQ9o.js";
|
|
8
8
|
import { a as ContinuousCalibrationResult, d as calibrateJudgeContinuous, f as continuousAgreement, h as verbosityBias, i as ContinuousAgreementOptions, l as VerbosityBiasResult, n as CandidateScore, o as GoldenItem, r as ContinuousAgreement, t as CalibrationResult, u as calibrateJudge } from "./judge-calibration-DnrbQaXI.js";
|
|
9
9
|
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, an as CorpusAgreementPerDimension, bt as SeriesConvergenceOptions, cn as corpusInterRaterAgreement, et as registerCalibrationSet, in as CorpusAgreementOptions, ln as corpusInterRaterAgreementFromJudgeScores, on as CorpusAgreementReport, q as JudgeRun, sn as CorpusScoreRecord, tt as snapshotFromJudgeAgreement, un as interRaterReliability, xt as SeriesConvergenceResult, z as JudgeCalibrationSet } from "./index-DAk2quYu.js";
|
|
10
10
|
import { B as WeightedCompositeResult, G as spearmanR, H as partialCredit, J as weightedMean, K as summarizeNumberSeries, L as SeriesDistribution, R as SeriesDistributionOptions, U as pearsonR, V as confidenceInterval, W as ranks, a as MintedRolloutLine, q as weightedComposite, z as WeightedCompositeInput } from "./schema-DkFJr3Y8.js";
|
|
11
11
|
import { n as bonferroni, r as holm, t as benjaminiHochberg } from "./multiplicity-D_SEsc5j.js";
|
|
12
|
-
import { A as PairRunRecordsResult, At as EProcess, B as ArmPopulationSummary, Bt as mcnemar, D as MatchedRunRecordPair, E as MatchedPair, F as comparePairedArms, Ft as ExactRiskDifferenceResult, Gt as wilson, Ht as pairedRiskDifferenceExact, I as pairArms, It as McNemarResult, L as pairRunRecords, Lt as ProportionInterval, M as PairedArmsComparison, Mt as EProcessState, N as PairedCorrectness, Nt as EProcessStep, O as PairArmsOptions, P as PairedMetricDelta, Pt as eProcess, R as ArmAssignment, Rt as RiskDifferenceResult, S as manifestContentDigest, T as ComparePairedArmsOptions, Ut as pairedRiskDifferenceScore, Vt as pairedRiskDifference, Wt as passAtK, _t as AttestationProvenance, bt as attest, gt as ATTESTATION_ALGORITHM, it as paretoFrontier, j as PairedArmRow, jt as EProcessOptions, 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, zt as ScoreRiskDifferenceResult } from "./statistical-heldout-
|
|
12
|
+
import { A as PairRunRecordsResult, At as EProcess, B as ArmPopulationSummary, Bt as mcnemar, D as MatchedRunRecordPair, E as MatchedPair, F as comparePairedArms, Ft as ExactRiskDifferenceResult, Gt as wilson, Ht as pairedRiskDifferenceExact, I as pairArms, It as McNemarResult, L as pairRunRecords, Lt as ProportionInterval, M as PairedArmsComparison, Mt as EProcessState, N as PairedCorrectness, Nt as EProcessStep, O as PairArmsOptions, P as PairedMetricDelta, Pt as eProcess, R as ArmAssignment, Rt as RiskDifferenceResult, S as manifestContentDigest, T as ComparePairedArmsOptions, Ut as pairedRiskDifferenceScore, Vt as pairedRiskDifference, Wt as passAtK, _t as AttestationProvenance, bt as attest, gt as ATTESTATION_ALGORITHM, it as paretoFrontier, j as PairedArmRow, jt as EProcessOptions, 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, zt as ScoreRiskDifferenceResult } from "./statistical-heldout-Cei036nE.js";
|
|
13
13
|
import { a as PairedPromotionDecision, c as BOOTSTRAP_GATE_MIN_N, d as PairedSignTestResult, f as PairedTTestResult, g as pairedTTest, h as pairedSignTest, i as PairedMcNemarEvidence, l as PairedBootstrapOptions, m as pairedBootstrap, n as PairedDecisionShape, o as PairedPromotionDecisionOptions, p as SignTestAlternative, r as PairedDecisionStatistic, s as decidePairedPromotion, t as PairedDecisionMethod, u as PairedBootstrapResult } from "./paired-promotion-decision-BDnqepZ1.js";
|
|
14
14
|
import { a as requiredPairedSampleSize, i as pairedMde, n as mcnemarPower, o as requiredSampleSize, r as mcnemarRequiredN, t as mulberry32 } from "./random-BN150Q_X.js";
|
|
15
15
|
import { C as mannWhitneyU, S as WilcoxonSignedRankResult, _ as evaluateReleaseConfidence, a as bootstrapCi, b as RankTestMethodRequest, 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, v as MannWhitneyResult, w as wilcoxonSignedRank, x as RankTestOptions, y as RankTestMethod } from "./promotion-gate-CixshGI8.js";
|
|
16
|
-
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-
|
|
17
|
-
import { $ as MintRolloutResult, Q as MintRolloutOptions, R as Scenario, S as JudgeConfig, _ as GateDecision, a as CampaignResult, nt as mintRolloutRows, tt as RolloutSearchLineage, w as JudgeScore } from "./types-
|
|
16
|
+
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-B7v2QHlD.js";
|
|
17
|
+
import { $ as MintRolloutResult, Q as MintRolloutOptions, R as Scenario, S as JudgeConfig, _ as GateDecision, a as CampaignResult, nt as mintRolloutRows, tt as RolloutSearchLineage, w as JudgeScore } from "./types-C54Uk_a1.js";
|
|
18
18
|
import { a as RunFilter, i as InMemoryTraceStore, n as FileSystemTraceStore, o as SpanFilter, s as TraceStore, t as EventFilter } from "./store-CMs5wGXP.js";
|
|
19
19
|
import { n as ScorePreference } from "./reward-ST4PXNXF.js";
|
|
20
|
-
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-
|
|
20
|
+
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-CoD1Dotf.js";
|
|
21
21
|
import { i as DatasetSplit, n as DatasetManifest, r as DatasetScenario, t as Dataset } from "./dataset-DThC1S6G.js";
|
|
22
22
|
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";
|
|
23
23
|
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";
|
|
24
24
|
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-0La5vejY.js";
|
|
25
25
|
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";
|
|
26
|
-
import { d as InsightReport, n as analyzeRuns, t as AnalyzeRunsOptions } from "./analyze-runs-
|
|
26
|
+
import { d as InsightReport, n as analyzeRuns, t as AnalyzeRunsOptions } from "./analyze-runs-C-pP1sK9.js";
|
|
27
27
|
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-7KEfroqx.js";
|
|
28
|
-
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-
|
|
29
|
-
import { B as FindingsDiff, H as PersistedFinding, U as diffFindings, V as FindingsStore, z as DiffPolicy } from "./systemone-
|
|
30
|
-
import { i as ExactAnalystRunEvent, l as ExactCapableAnalyst, o as ExactAnalystRunResult } from "./exact-types-
|
|
31
|
-
import { a as createTraceAnalyst, i as TraceAnalystDefinition, n as buildDefaultAnalystRegistry, r as CreateTraceAnalystOptions, t as DefaultAnalystRegistryOptions } from "./default-registry-
|
|
32
|
-
import { r as TraceAnalysisEngineResult, s as RawAnalystFinding, t as TraceAnalysisEngine } from "./engine-
|
|
33
|
-
import { a as BudgetPolicy, c as RegistryRunOpts, i as AnalystRegistryOptions, s as ExactRegistryRunOpts, t as AnalystRegistry } from "./registry-
|
|
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-
|
|
35
|
-
import { i as BenchmarkEvaluation } from "./types-
|
|
28
|
+
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-CAffiFbN.js";
|
|
29
|
+
import { B as FindingsDiff, H as PersistedFinding, U as diffFindings, V as FindingsStore, z as DiffPolicy } from "./systemone-Cw2txK5m.js";
|
|
30
|
+
import { i as ExactAnalystRunEvent, l as ExactCapableAnalyst, o as ExactAnalystRunResult } from "./exact-types-D98BKpxo.js";
|
|
31
|
+
import { a as createTraceAnalyst, i as TraceAnalystDefinition, n as buildDefaultAnalystRegistry, r as CreateTraceAnalystOptions, t as DefaultAnalystRegistryOptions } from "./default-registry-B1DstPYg.js";
|
|
32
|
+
import { r as TraceAnalysisEngineResult, s as RawAnalystFinding, t as TraceAnalysisEngine } from "./engine-CaHaCc3W.js";
|
|
33
|
+
import { a as BudgetPolicy, c as RegistryRunOpts, i as AnalystRegistryOptions, s as ExactRegistryRunOpts, t as AnalystRegistry } from "./registry-dhdDoAV8.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-CUjeTDPv.js";
|
|
35
|
+
import { i as BenchmarkEvaluation } from "./types-0oMg7zee.js";
|
|
36
36
|
import { a as SpanHandle, n as CaptureWrite, o as TraceEmitter, s as TraceEmitterOptions, t as CaptureErrorHandler } from "./emitter-mAXC5PSN.js";
|
|
37
37
|
import { n as SandboxDriver } from "./sandbox-harness-B3Kx1fOA.js";
|
|
38
38
|
import { _ as shareAllowed, a as RedactionProfile, d as assessShareSafety, g as redactText, h as redactForShare, l as ShareSafetyStatus, m as redact, n as REDACTION_VERSION, o as RedactionReport, r as RedactOptions, u as ShareSafetyVerdict } from "./redact-C_wzC4c_.js";
|
|
39
39
|
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-Bb_pdTqG.js";
|
|
40
40
|
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-BFG1vwTV.js";
|
|
41
|
-
import { A as AnalystFindingDigest, B as ControlRuntimeConfig, C as createFeedbackTrajectory, D as renderPreferenceMemoryMarkdown, E as feedbackTrajectoryToOptimizerRow, F as ControlBudget, H as ControlSeverity, 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-
|
|
42
|
-
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-
|
|
41
|
+
import { A as AnalystFindingDigest, B as ControlRuntimeConfig, C as createFeedbackTrajectory, D as renderPreferenceMemoryMarkdown, E as feedbackTrajectoryToOptimizerRow, F as ControlBudget, H as ControlSeverity, 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-SBGncrCP.js";
|
|
42
|
+
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-B3i3fl0A.js";
|
|
43
43
|
import { a as RunIntegrityReport, o as assertRunCaptured, t as RunIntegrityError } from "./integrity-BAdVzT8u.js";
|
|
44
44
|
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-CLnNJ7Or.js";
|
|
45
|
-
import "./index-
|
|
45
|
+
import "./index-DUn0GNiZ.js";
|
|
46
46
|
import { y as OUTPUT_VALUE } from "./attribute-vocabulary-5oo8ykIx.js";
|
|
47
47
|
import { i as judgeSpans, n as argHash, s as runsForScenario } from "./query-UD0B08Sb.js";
|
|
48
48
|
import { SpanKind } from "@tangle-network/agent-trace-contract";
|
package/dist/index.js
CHANGED
|
@@ -6,33 +6,34 @@ import { a as groupRunsByAgentProfileCell, c as verifyAgentProfileCell, i as bui
|
|
|
6
6
|
import { r as observedSplitScore, t as isRealnessGated } from "./reward-P0VD_Aft.js";
|
|
7
7
|
import { i as isToolSpan, n as isJudgeSpan, r as isLlmSpan, t as FAILURE_CLASSES } from "./schema-D-xu3MU_.js";
|
|
8
8
|
import { a as parseRunRecordSafe, c as validateRunRecord, i as modelHasSnapshot, n as UNKNOWN_MODEL, o as runTaskScore, r as isRunRecord, t as RunRecordValidationError } from "./run-record-DEjx43cI.js";
|
|
9
|
-
import { a as cliffsDelta, c as pairedCohensDz, i as summaryTable, n as gainHistogram, o as cohensD, s as interpretCliffs } from "./summary-report-
|
|
9
|
+
import { a as cliffsDelta, c as pairedCohensDz, i as summaryTable, n as gainHistogram, o as cohensD, s as interpretCliffs } from "./summary-report-BLnTyQKH.js";
|
|
10
10
|
import { t as interpolatedQuantile } from "./quantile-CncXP-Ss.js";
|
|
11
|
-
import { a as wilcoxonSignedRank, i as mannWhitneyU, n as pairArms, r as pairRunRecords, t as comparePairedArms } from "./paired-arms-
|
|
11
|
+
import { a as wilcoxonSignedRank, i as mannWhitneyU, n as pairArms, r as pairRunRecords, t as comparePairedArms } from "./paired-arms-D5H_58e8.js";
|
|
12
12
|
import { a as medianInPlace, i as makeRng, t as assertFiniteSample } from "./internal-v6XfOeJW.js";
|
|
13
13
|
import { t as mulberry32 } from "./random-DrzcgC7F.js";
|
|
14
14
|
import { a as spearmanR, c as weightedMean, i as ranks, n as partialCredit, o as summarizeNumberSeries, r as pearsonR, s as weightedComposite, t as confidenceInterval } from "./descriptive-cPwITVfy.js";
|
|
15
15
|
import { n as calibrateJudgeContinuous, o as verbosityBias, r as continuousAgreement, t as calibrateJudge } from "./judge-calibration-DGqlfWW8.js";
|
|
16
16
|
import { n as corpusInterRaterAgreementFromJudgeScores, r as interRaterReliability, t as corpusInterRaterAgreement } from "./agreement-irr-BTjDN0i6.js";
|
|
17
|
-
import { a as welchsTTest, c as fisherExactTwoSided, d as studentTTwoSidedPValue, f as twoByTwoPValue, i as iqr, l as linearRegressionSlope, n as checkCanaries, o as chiSquared, s as chiSquaredPValue, t as analyzeRuns, u as normalTwoSidedPValue } from "./analyze-runs-
|
|
17
|
+
import { a as welchsTTest, c as fisherExactTwoSided, d as studentTTwoSidedPValue, f as twoByTwoPValue, i as iqr, l as linearRegressionSlope, n as checkCanaries, o as chiSquared, s as chiSquaredPValue, t as analyzeRuns, u as normalTwoSidedPValue } from "./analyze-runs-Ck68mbgj.js";
|
|
18
18
|
import { n as bonferroni, r as holm, t as benjaminiHochberg } from "./multiplicity-DAd4xaDC.js";
|
|
19
19
|
import { a as pairedRiskDifferenceScore, i as pairedRiskDifferenceExact, o as passAtK, r as pairedRiskDifference, s as wilson, t as mcnemar } from "./paired-binary-BR0nwHOA.js";
|
|
20
|
-
import {
|
|
21
|
-
import { a as requiredSampleSize, i as requiredPairedSampleSize, n as mcnemarRequiredN, r as pairedMde, t as mcnemarPower } from "./power-and-mde-
|
|
20
|
+
import { a as pairedSignTest, o as pairedTTest, r as pairedBootstrap, t as BOOTSTRAP_GATE_MIN_N } from "./paired-tests-C2zUqgUW.js";
|
|
21
|
+
import { a as requiredSampleSize, i as requiredPairedSampleSize, n as mcnemarRequiredN, r as pairedMde, t as mcnemarPower } from "./power-and-mde-CHvyDO_H.js";
|
|
22
22
|
import { t as eProcess } from "./sequential-eprocess-C1YLwFaz.js";
|
|
23
|
-
import {
|
|
23
|
+
import { t as mean$1 } from "./mean-BPuUNznQ.js";
|
|
24
|
+
import { i as pairedDeltaTest, n as pairedDecisionShape, r as minimumPairsForPairedDeltaTest, t as decidePairedPromotion } from "./paired-promotion-decision-BiVumgvM.js";
|
|
24
25
|
import { c as runsForScenario, i as judgeSpans, t as argHash } from "./query-CGXzO9d4.js";
|
|
25
26
|
import { n as contentHash, r as fileVerdictCache, t as canonicalJson } from "./verdict-cache-BVH-fl18.js";
|
|
26
|
-
import { B as DEFAULT_RED_TEAM_CORPUS, H as redTeamReport, U as scoreRedTeamOutput, V as redTeamDataset, W as runCanaries, Z as runCampaign, a as clamp01, i as aggregateRunScore, r as transientDispatchFailure, t as llmJudge } from "./llm-judge-
|
|
27
|
+
import { B as DEFAULT_RED_TEAM_CORPUS, H as redTeamReport, U as scoreRedTeamOutput, V as redTeamDataset, W as runCanaries, Z as runCampaign, a as clamp01, i as aggregateRunScore, r as transientDispatchFailure, t as llmJudge } from "./llm-judge-D6JtIqiU.js";
|
|
27
28
|
import { a as resolveModelPricing, i as isModelPriced, n as estimateCost, r as estimateTokens, t as MODEL_PRICING } from "./metrics-DNCAUtuB.js";
|
|
28
29
|
import { c as costForTokenPricing, i as CostCeilingReachedError, l as costForUsage, n as CostAccountingIncompleteError, t as CostLedger } from "./cost-ledger-CS8R3IZj.js";
|
|
29
|
-
import { C as assertRealBackend, S as BackendIntegrityError, _ as parseReflectionResponse, a as ATTESTATION_ALGORITHM, g as buildReflectionPrompt, o as attest, s as verifyAttestation, w as summarizeBackendIntegrity } from "./campaign-evidence-
|
|
30
|
+
import { C as assertRealBackend, S as BackendIntegrityError, _ as parseReflectionResponse, a as ATTESTATION_ALGORITHM, g as buildReflectionPrompt, o as attest, s as verifyAttestation, w as summarizeBackendIntegrity } from "./campaign-evidence-MjCjkb5a.js";
|
|
30
31
|
import { s as surfaceContentHash } from "./json-recovery-DK4MDm5r.js";
|
|
31
32
|
import { a as redact, c as shareAllowed, n as assessShareSafety, o as redactForShare, s as redactText, t as REDACTION_VERSION } from "./redact-D6APJS1d.js";
|
|
32
33
|
import { a as makeProposalFinding, i as makeFinding, r as computeFindingId } from "./types-BK0blWhe.js";
|
|
33
|
-
import { S as requiredPairsForPairedPromotion, n as paretoFrontier, t as dominates, x as pairedPromotionPower } from "./pareto-
|
|
34
|
+
import { S as requiredPairsForPairedPromotion, n as paretoFrontier, t as dominates, x as pairedPromotionPower } from "./pareto-7_LEmYGG.js";
|
|
34
35
|
import { a as LlmCallError, i as stripFencedJson, n as isRecord, o as LlmResponseError } from "./json-DtvZNAU9.js";
|
|
35
|
-
import { r as selfImprove, t as defineAgentEval } from "./define-agent-eval-
|
|
36
|
+
import { r as selfImprove, t as defineAgentEval } from "./define-agent-eval-Cc15qGik.js";
|
|
36
37
|
import { n as equivalenceVerdict, t as certificationEvidenceDigest } from "./verdict-C8TcdwfD.js";
|
|
37
38
|
import { c as createBoundedTraceAnalysisStore, p as DEFAULT_TRACE_ANALYST_BUDGETS, t as createTraceAnalyst } from "./kind-factory-VKysp1ds.js";
|
|
38
39
|
import { C as snapshotAnalystRun, S as readAnalystReview, b as assertUniqueFindingIds, c as FAILURE_MODE_KIND_SPEC, i as DEFAULT_TRACE_ANALYST_KINDS, n as AnalystRegistry, t as buildDefaultAnalystRegistry, v as analystFindingDigest, w as validateAnalystReviewDecisions, x as completedAnalystReviewQuality, y as analystRunDigest } from "./default-registry-C1tPEjZ4.js";
|
|
@@ -47,7 +48,7 @@ import { t as packageVersion$1 } from "./package-version-CJl_pY9W.js";
|
|
|
47
48
|
import { t as buildTrajectory } from "./trajectory-O3GDHmev.js";
|
|
48
49
|
import { i as failureBlame, n as classifyFailure, r as classifyFailureReason } from "./failure-taxonomy-9eQfXKoW.js";
|
|
49
50
|
import { n as runBoundedProcess, t as DEFAULT_BOUNDED_PROCESS_TIMEOUT_MS } from "./bounded-process-BZrWXIKW.js";
|
|
50
|
-
import { t as runEvalCampaign } from "./eval-campaign-
|
|
51
|
+
import { t as runEvalCampaign } from "./eval-campaign-C_mNDshv.js";
|
|
51
52
|
import { _ as assertCrossFamily, a as maximumChargeForLlmRequest, d as assertServedModel, g as CrossFamilyError, h as describeServedModelCheck, i as costReceiptFromLlmError, l as ModelSubstitutionError, m as checkServedModel, p as assertServedModels, r as costReceiptFromLlm, s as isTransientLlmError, u as assertCrossFamilyServed, v as judgeFamily } from "./llm-client-kJ9EOqjt.js";
|
|
52
53
|
import { n as createLlmCorrectnessChecker, r as verifyCompletion, t as completionVerdict } from "./completion-verifier-rqbMh5-4.js";
|
|
53
54
|
import "./rollout-BhkMpAJT.js";
|
|
@@ -57,7 +58,7 @@ import { n as paidJsonChat } from "./chat-json-call-8yUv4LFO.js";
|
|
|
57
58
|
import { t as createDspyRlmTraceEngine } from "./dspy-rlm-engine-jDiWkul6.js";
|
|
58
59
|
import { n as diffFindings, r as createChatTraceEngine, t as FindingsStore } from "./findings-store-B75Dspaq.js";
|
|
59
60
|
import { a as budgetBreachView, i as failureClusterView, n as computeToolUseMetrics, r as judgeAgreementView, t as toolWasteView } from "./tool-waste-CLJ46AbT.js";
|
|
60
|
-
import { _ as analyzeSeries, a as judgeGateDecision, c as snapshotFromJudgeAgreement, g as withJudgeRetry, i as assertJudgeMayGate, n as JudgeGateRefusedError, o as measureJudgeAgreement, r as addCalibrationVerdicts, s as registerCalibrationSet, t as DEFAULT_JUDGE_GATE_POLICY } from "./judge-gate-
|
|
61
|
+
import { _ as analyzeSeries, a as judgeGateDecision, c as snapshotFromJudgeAgreement, g as withJudgeRetry, i as assertJudgeMayGate, n as JudgeGateRefusedError, o as measureJudgeAgreement, r as addCalibrationVerdicts, s as registerCalibrationSet, t as DEFAULT_JUDGE_GATE_POLICY } from "./judge-gate-Di_6bifh.js";
|
|
61
62
|
import { a as spanRecordsToTraceAnalysisStore, i as otlpTextToTraceAnalysisStore, r as OtlpFileTraceStore } from "./store-otlp-DmLm65XW.js";
|
|
62
63
|
import { n as bootstrapCi, t as evaluateReleaseConfidence } from "./release-confidence-CTUJz5mM.js";
|
|
63
64
|
import { t as createChatClient } from "./chat-client-CO3uB33i.js";
|