@tangle-network/agent-eval 0.144.6 → 0.144.8
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 +30 -0
- package/README.md +2 -0
- package/dist/{benchmark-J9Qe6j2_.d.ts → agent-profile-CgDTo40f.d.ts} +91 -4
- package/dist/agent-profile-CgDTo40f.d.ts.map +1 -0
- package/dist/analyst/index.d.ts +470 -88
- package/dist/analyst/index.d.ts.map +1 -1
- package/dist/analyst/index.js +24 -5
- package/dist/analyst/index.js.map +1 -1
- package/dist/{analyze-runs-DWIvOAGk.js → analyze-runs-BNNK7irB.js} +3 -3
- package/dist/{analyze-runs-DWIvOAGk.js.map → analyze-runs-BNNK7irB.js.map} +1 -1
- package/dist/{baseline-D_fT6277.d.ts → baseline-CavEbRyH.d.ts} +2 -23
- package/dist/baseline-CavEbRyH.d.ts.map +1 -0
- package/dist/{benchmark-command-CQd78YHt.js → benchmark-command-BKENp2s5.js} +662 -605
- package/dist/benchmark-command-BKENp2s5.js.map +1 -0
- package/dist/benchmarks/index.d.ts +1 -1
- package/dist/benchmarks/index.js +1 -1
- package/dist/{benchmarks-BEOkuvIg.js → benchmarks-BlPmjd88.js} +4 -4
- package/dist/{benchmarks-BEOkuvIg.js.map → benchmarks-BlPmjd88.js.map} +1 -1
- package/dist/campaign/index.d.ts +7 -5
- package/dist/campaign/index.js +5 -3
- package/dist/{campaign-CXsdyym7.js → campaign--HVSuvV0.js} +17 -301
- package/dist/campaign--HVSuvV0.js.map +1 -0
- package/dist/cli.js +2 -2
- package/dist/{client-0JI64ovJ.d.ts → client-DjXROWpx.d.ts} +3 -3
- package/dist/{client-0JI64ovJ.d.ts.map → client-DjXROWpx.d.ts.map} +1 -1
- package/dist/{completion-verifier-CBiee74w.d.ts → completion-verifier-foUCLif_.d.ts} +5 -5
- package/dist/{completion-verifier-CBiee74w.d.ts.map → completion-verifier-foUCLif_.d.ts.map} +1 -1
- package/dist/contract/index.d.ts +9 -8
- package/dist/contract/index.d.ts.map +1 -1
- package/dist/contract/index.js +7 -6
- package/dist/contract/index.js.map +1 -1
- package/dist/control.d.ts +2 -2
- package/dist/control.js +1 -1
- package/dist/counterfactual-CWPTrMH7.js +126 -0
- package/dist/counterfactual-CWPTrMH7.js.map +1 -0
- package/dist/counterfactual-CxmxAONP.d.ts +72 -0
- package/dist/counterfactual-CxmxAONP.d.ts.map +1 -0
- package/dist/{default-registry-J9m-_tya.d.ts → default-registry-BZhStdjl.d.ts} +6 -5
- package/dist/default-registry-BZhStdjl.d.ts.map +1 -0
- package/dist/{default-registry-Dta70shL.js → default-registry-BaQXW1Ow.js} +2 -2
- package/dist/{default-registry-Dta70shL.js.map → default-registry-BaQXW1Ow.js.map} +1 -1
- package/dist/{dspy-rlm-engine-19FQEMBK.js → dspy-rlm-engine-BiN49gK6.js} +4 -3
- package/dist/{dspy-rlm-engine-19FQEMBK.js.map → dspy-rlm-engine-BiN49gK6.js.map} +1 -1
- package/dist/{tool-groups-CK0JCkqO.d.ts → engine-nB64f48I.d.ts} +18 -31
- package/dist/engine-nB64f48I.d.ts.map +1 -0
- package/dist/{eval-campaign-CfLQQs9B.js → eval-campaign-DNjCvAm-.js} +7 -6
- package/dist/{eval-campaign-CfLQQs9B.js.map → eval-campaign-DNjCvAm-.js.map} +1 -1
- package/dist/{exact-types-CBYF5MGd.d.ts → exact-types-Djvzosly.d.ts} +2 -2
- package/dist/{exact-types-CBYF5MGd.d.ts.map → exact-types-Djvzosly.d.ts.map} +1 -1
- package/dist/exec-BLtYZdWo.js +49 -0
- package/dist/exec-BLtYZdWo.js.map +1 -0
- package/dist/experiment/index.d.ts +802 -0
- package/dist/experiment/index.d.ts.map +1 -0
- package/dist/experiment/index.js +1108 -0
- package/dist/experiment/index.js.map +1 -0
- package/dist/experiment-tracker-CnRICnMl.js +500 -0
- package/dist/experiment-tracker-CnRICnMl.js.map +1 -0
- package/dist/experiment-tracker-IMntXr6J.d.ts +299 -0
- package/dist/experiment-tracker-IMntXr6J.d.ts.map +1 -0
- package/dist/{external-optimizer-contracts-Q9c0L3eY.d.ts → external-optimizer-contracts-CZuJNcT5.d.ts} +2 -2
- package/dist/{external-optimizer-contracts-Q9c0L3eY.d.ts.map → external-optimizer-contracts-CZuJNcT5.d.ts.map} +1 -1
- package/dist/{feedback-trajectory-GgoS0-MK.d.ts → feedback-trajectory-Rh280oXo.d.ts} +3 -3
- package/dist/{feedback-trajectory-GgoS0-MK.d.ts.map → feedback-trajectory-Rh280oXo.d.ts.map} +1 -1
- package/dist/fuzz.d.ts +1 -1
- package/dist/hosted/index.d.ts +3 -3
- package/dist/{index-BZUe-ODI.d.ts → index-BZ3-y4YL.d.ts} +3 -3
- package/dist/{index-BZUe-ODI.d.ts.map → index-BZ3-y4YL.d.ts.map} +1 -1
- package/dist/{index-DSC51roc.d.ts → index-C3ssXVLv.d.ts} +2 -2
- package/dist/{index-DSC51roc.d.ts.map → index-C3ssXVLv.d.ts.map} +1 -1
- package/dist/{index-4XwggC10.d.ts → index-C5HOo4ZF2.d.ts} +4 -4
- package/dist/index-C5HOo4ZF2.d.ts.map +1 -0
- package/dist/{index-B6-B0zTB.d.ts → index-CvXXlyz7.d.ts} +2 -2
- package/dist/{index-B6-B0zTB.d.ts.map → index-CvXXlyz7.d.ts.map} +1 -1
- package/dist/{index-Dx1kF3Ez.d.ts → index-CwDrUMe0.d.ts} +2 -2
- package/dist/{index-Dx1kF3Ez.d.ts.map → index-CwDrUMe0.d.ts.map} +1 -1
- package/dist/{index-BIL5vxxt.d.ts → index-Sh2I0DRc.d.ts} +11 -645
- package/dist/index-Sh2I0DRc.d.ts.map +1 -0
- package/dist/index.d.ts +214 -404
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +233 -649
- package/dist/index.js.map +1 -1
- package/dist/{insight-report-DqEsugpr.d.ts → insight-report-C6h6F_4L.d.ts} +3 -3
- package/dist/{insight-report-DqEsugpr.d.ts.map → insight-report-C6h6F_4L.d.ts.map} +1 -1
- package/dist/{integrity-CNGUaGBY.d.ts → integrity-BuqEKu-x.d.ts} +2 -2
- package/dist/{integrity-CNGUaGBY.d.ts.map → integrity-BuqEKu-x.d.ts.map} +1 -1
- package/dist/integrity-MLzHOfV9.js +141 -0
- package/dist/integrity-MLzHOfV9.js.map +1 -0
- package/dist/kind-factory-BHIgPmzS.js.map +1 -1
- package/dist/{llm-client-Dv5BiKLE.js → llm-client-DzvMUsS_.js} +24 -6
- package/dist/llm-client-DzvMUsS_.js.map +1 -0
- package/dist/matrix/index.d.ts +2 -2
- package/dist/meta-eval/index.d.ts +1 -1
- package/dist/{mint-DD-0oQTA.js → mint-CGEkzPLf.js} +2 -2
- package/dist/{mint-DD-0oQTA.js.map → mint-CGEkzPLf.js.map} +1 -1
- package/dist/{multi-layer-verifier-BHY1gWAc.d.ts → multi-layer-verifier-DnAqwl0h.d.ts} +2 -2
- package/dist/{multi-layer-verifier-BHY1gWAc.d.ts.map → multi-layer-verifier-DnAqwl0h.d.ts.map} +1 -1
- package/dist/multishot/index.d.ts +2 -2
- package/dist/openapi.json +1 -1
- package/dist/paired-promotion-decision-B6zJ3gYM.d.ts +114 -0
- package/dist/paired-promotion-decision-B6zJ3gYM.d.ts.map +1 -0
- package/dist/pipelines/index.d.ts +2 -1
- package/dist/pipelines/index.d.ts.map +1 -1
- package/dist/pre-registration-DakwTRXk.js +96 -0
- package/dist/pre-registration-DakwTRXk.js.map +1 -0
- package/dist/prime-bridge-transport-6feEglLf.d.ts +49 -0
- package/dist/prime-bridge-transport-6feEglLf.d.ts.map +1 -0
- package/dist/prime-protocol-BfSalTfR.js +453 -0
- package/dist/prime-protocol-BfSalTfR.js.map +1 -0
- package/dist/profile-cell.js +242 -1
- package/dist/profile-cell.js.map +1 -0
- package/dist/promotion-policy-ChWhTDBH.d.ts +289 -0
- package/dist/promotion-policy-ChWhTDBH.d.ts.map +1 -0
- package/dist/promotion-policy-CrLrmys8.js +682 -0
- package/dist/promotion-policy-CrLrmys8.js.map +1 -0
- package/dist/{propose-review-control-BciUCZoh.js → propose-review-control-qgWLA6E9.js} +2 -2
- package/dist/{propose-review-control-BciUCZoh.js.map → propose-review-control-qgWLA6E9.js.map} +1 -1
- package/dist/{release-report-ChOgpIoQ.d.ts → release-report-CI8uisI1.d.ts} +2 -2
- package/dist/{release-report-ChOgpIoQ.d.ts.map → release-report-CI8uisI1.d.ts.map} +1 -1
- package/dist/{release-report-Sl0xfkFv.js → release-report-Dy39cbFF.js} +3 -3
- package/dist/{release-report-Sl0xfkFv.js.map → release-report-Dy39cbFF.js.map} +1 -1
- package/dist/{replay-Krvb114g.d.ts → replay-DFf-teiC.d.ts} +5 -4
- package/dist/replay-DFf-teiC.d.ts.map +1 -0
- package/dist/{replay-GW61ezMW.js → replay-DyBLaKFc.js} +3 -3
- package/dist/{replay-GW61ezMW.js.map → replay-DyBLaKFc.js.map} +1 -1
- package/dist/reporting.d.ts +3 -3
- package/dist/reporting.js +2 -2
- package/dist/{researcher-xLeNcpKX.d.ts → researcher-BoaxeCzP.d.ts} +4 -4
- package/dist/{researcher-xLeNcpKX.d.ts.map → researcher-BoaxeCzP.d.ts.map} +1 -1
- package/dist/{reward-hacking-CyuzxKly.js → reward-hacking-B2tPL9a3.js} +12 -3
- package/dist/reward-hacking-B2tPL9a3.js.map +1 -0
- package/dist/{reward-hacking-RZgnGWlx.d.ts → reward-hacking-Cf1PtEOz.d.ts} +33 -3
- package/dist/reward-hacking-Cf1PtEOz.d.ts.map +1 -0
- package/dist/rl.d.ts +17 -7
- package/dist/rl.d.ts.map +1 -1
- package/dist/rl.js +16 -7
- package/dist/rl.js.map +1 -1
- package/dist/rollout/index.d.ts +2 -2
- package/dist/rollout/index.js +2 -2
- package/dist/{rollout-C2fD1cf4.js → rollout-2ECTXb2N.js} +2 -2
- package/dist/{rollout-C2fD1cf4.js.map → rollout-2ECTXb2N.js.map} +1 -1
- package/dist/{run-evidence-C6G41MSI.d.ts → run-evidence-BDFFai9R.d.ts} +2 -2
- package/dist/{run-evidence-C6G41MSI.d.ts.map → run-evidence-BDFFai9R.d.ts.map} +1 -1
- package/dist/{run-record-CWN8-VsV.js → run-record-DqOw5X6_.js} +2 -2
- package/dist/{run-record-CWN8-VsV.js.map → run-record-DqOw5X6_.js.map} +1 -1
- package/dist/{schema-Cef2cFmb.d.ts → schema-Cef2cFmb2.d.ts} +1 -1
- package/dist/schema-Cef2cFmb2.d.ts.map +1 -0
- package/dist/{semantic-concept-judge-DwF6n05O.js → semantic-concept-judge-Bmrq6yqU.js} +2 -2
- package/dist/{semantic-concept-judge-DwF6n05O.js.map → semantic-concept-judge-Bmrq6yqU.js.map} +1 -1
- package/dist/sequential-D-BLJBKU.js +299 -0
- package/dist/sequential-D-BLJBKU.js.map +1 -0
- package/dist/{server-D6XJQHw7.js → server-iu0ede49.js} +2 -2
- package/dist/{server-D6XJQHw7.js.map → server-iu0ede49.js.map} +1 -1
- package/dist/{single-run-lock-D5iN0Xzb.js → single-run-lock-BMQEv1wG.js} +7 -161
- package/dist/single-run-lock-BMQEv1wG.js.map +1 -0
- package/dist/{skill-usage-GlOphAhX.d.ts → skill-usage-CJlWEUFt.d.ts} +10 -10
- package/dist/{skill-usage-GlOphAhX.d.ts.map → skill-usage-CJlWEUFt.d.ts.map} +1 -1
- package/dist/{skillopt-optimization-method-7S43rbDB.d.ts → skillopt-optimization-method-BO7NAl3b.d.ts} +8 -294
- package/dist/skillopt-optimization-method-BO7NAl3b.d.ts.map +1 -0
- package/dist/{skillopt-optimization-method-Bfb-vBKe.js → skillopt-optimization-method-CkaI2ly4.js} +19 -654
- package/dist/skillopt-optimization-method-CkaI2ly4.js.map +1 -0
- package/dist/statistical-heldout-Dn9ruizm.d.ts +555 -0
- package/dist/statistical-heldout-Dn9ruizm.d.ts.map +1 -0
- package/dist/{statistics-C-dm-J6H.d.ts → statistics-D6Uebe_4.d.ts} +2 -2
- package/dist/{statistics-C-dm-J6H.d.ts.map → statistics-D6Uebe_4.d.ts.map} +1 -1
- package/dist/steps-BArUxhna.d.ts +51 -0
- package/dist/steps-BArUxhna.d.ts.map +1 -0
- package/dist/{integrity-fdt8XPAv.js → store-DNe_Uv1Q.js} +2 -140
- package/dist/store-DNe_Uv1Q.js.map +1 -0
- package/dist/{summary-report-B0cAyA7N.d.ts → summary-report-DuUS_i7W.d.ts} +3 -114
- package/dist/summary-report-DuUS_i7W.d.ts.map +1 -0
- package/dist/{summary-report-9A5y7EsK.js → summary-report-Lf-5I7xh.js} +2 -2
- package/dist/{summary-report-9A5y7EsK.js.map → summary-report-Lf-5I7xh.js.map} +1 -1
- package/dist/supervisor-run/index.d.ts +2 -2
- package/dist/supervisor-run/index.js +1 -1
- package/dist/{supervisor-run-DiyQVczd.js → supervisor-run-D_sokXcO.js} +32 -8
- package/dist/{supervisor-run-DiyQVczd.js.map → supervisor-run-D_sokXcO.js.map} +1 -1
- package/dist/tool-groups-DIVBnJyl.d.ts +28 -0
- package/dist/tool-groups-DIVBnJyl.d.ts.map +1 -0
- package/dist/trace-repair/index.d.ts +2102 -0
- package/dist/trace-repair/index.d.ts.map +1 -0
- package/dist/trace-repair/index.js +3878 -0
- package/dist/trace-repair/index.js.map +1 -0
- package/dist/traces.d.ts +6 -6
- package/dist/traces.js +3 -2
- package/dist/trajectory-YC15QDYQ.d.ts +24 -0
- package/dist/trajectory-YC15QDYQ.d.ts.map +1 -0
- package/dist/trajectory-replay/index.d.ts +781 -0
- package/dist/trajectory-replay/index.d.ts.map +1 -0
- package/dist/trajectory-replay/index.js +2103 -0
- package/dist/trajectory-replay/index.js.map +1 -0
- package/dist/{types-XMVEdrE_.d.ts → types-D216SgwM.d.ts} +24 -6
- package/dist/types-D216SgwM.d.ts.map +1 -0
- package/dist/{types-yLK8gXE9.d.ts → types-D4mog56g.d.ts} +2 -2
- package/dist/{types-yLK8gXE9.d.ts.map → types-D4mog56g.d.ts.map} +1 -1
- package/dist/{types-BhP9q0Fq.d.ts → types-DF_Udrp-.d.ts} +52 -3
- package/dist/{types-BhP9q0Fq.d.ts.map → types-DF_Udrp-.d.ts.map} +1 -1
- package/dist/{types-DOZyvsFU.d.ts → types-DYuNHo9R.d.ts} +3 -3
- package/dist/{types-DOZyvsFU.d.ts.map → types-DYuNHo9R.d.ts.map} +1 -1
- package/dist/usage-receipt-EVI8B8Xu.js.map +1 -1
- package/dist/verdict-DExhxfgR.d.ts +201 -0
- package/dist/verdict-DExhxfgR.d.ts.map +1 -0
- package/dist/verdict-cache-BCcOh0kF.js +159 -0
- package/dist/verdict-cache-BCcOh0kF.js.map +1 -0
- package/dist/wire/index.d.ts +2 -2
- package/dist/wire/index.js +1 -1
- package/docs/campaign-proposers.md +5 -0
- package/docs/charter.md +112 -0
- package/docs/experiment.md +104 -0
- package/docs/prime-analyst.md +1 -0
- package/docs/trace-analysis.md +26 -0
- package/docs/trace-repair-admission.md +194 -0
- package/docs/trace-repair-analyst-arms.md +121 -0
- package/docs/trace-repair-continuation.md +107 -0
- package/docs/trace-repair-grader.md +163 -0
- package/docs/trajectory-replay.md +110 -0
- package/docs/verification-strategies.md +103 -0
- package/package.json +19 -2
- package/dist/agent-profile-cell-CbfBm2g6.js +0 -335
- package/dist/agent-profile-cell-CbfBm2g6.js.map +0 -1
- package/dist/baseline-D_fT6277.d.ts.map +0 -1
- package/dist/benchmark-J9Qe6j2_.d.ts.map +0 -1
- package/dist/benchmark-command-CQd78YHt.js.map +0 -1
- package/dist/campaign-CXsdyym7.js.map +0 -1
- package/dist/default-registry-J9m-_tya.d.ts.map +0 -1
- package/dist/index-4XwggC10.d.ts.map +0 -1
- package/dist/index-BIL5vxxt.d.ts.map +0 -1
- package/dist/integrity-fdt8XPAv.js.map +0 -1
- package/dist/llm-client-Dv5BiKLE.js.map +0 -1
- package/dist/replay-Krvb114g.d.ts.map +0 -1
- package/dist/reward-hacking-CyuzxKly.js.map +0 -1
- package/dist/reward-hacking-RZgnGWlx.d.ts.map +0 -1
- package/dist/schema-Cef2cFmb.d.ts.map +0 -1
- package/dist/single-run-lock-D5iN0Xzb.js.map +0 -1
- package/dist/skillopt-optimization-method-7S43rbDB.d.ts.map +0 -1
- package/dist/skillopt-optimization-method-Bfb-vBKe.js.map +0 -1
- package/dist/summary-report-B0cAyA7N.d.ts.map +0 -1
- package/dist/tool-groups-CK0JCkqO.d.ts.map +0 -1
- package/dist/types-XMVEdrE_.d.ts.map +0 -1
- package/dist/verdict-Dps8_okt.d.ts +0 -37
- package/dist/verdict-Dps8_okt.d.ts.map +0 -1
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"default-registry-J9m-_tya.d.ts","names":[],"sources":["../src/trace-analyst/behavioral-metrics.ts","../src/analyst/behavioral-analyst.ts","../src/analyst/kind-factory.ts","../src/analyst/default-registry.ts"],"mappings":";;;;;KAoBY;UAMK;EACf,MAAM;EACN;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;EACA;;EAEA,gBAAgB;;EAEhB;EACA;EACA,eAAe;EACf;EACA;;EAEA;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA,sBAAsB;EACtB,uBAAuB;;;;;;iBAoFT,oBAAoB,gBAAgB,qBAAqB;;;UC9GxD;;EAEf;;EAEA;;;;;;iBA+Dc,yBACd,SAAS,mBACT;EAAQ;EAAoB;IAC3B;;iBAkCa,kBACd,UAAS,2BACR,oBAAoB;;;;UCrHN;EACf;EACA;EACA;EACA;;EAEA,sBAAsB,SAAS;;EAE/B;;EAEA,WAAW;;EAEX,kBACE,OAAO,oBACP,SAAS,wCACe;EAC1B,SAAS,QAAQ;EACjB,eAAe,KAAK,mBAAmB,SAAS,mBAAmB;EACnE;EACA;;UAGe;EACf,QAAQ;;EAER;EACA;;;iBAIoB,gBAAgB;EACpC,YAAY;EACZ,QAAQ;EACR,OAAO;EACP,SAAS;EACT;IACE,QAAQ;;iBAgFI,mBACd,YAAY,wBACZ,SAAS,4BACR,oBAAoB;iBA+QP,oBAAoB,OAAO;iBAe3B,uBAAuB,UAAU;;;UCnahC;;;;;;EAMf,SAAS;EACT,uBAAuB;;EAEvB;EACA,aAAa;EACb,WAAW;;iBAGG,4BACd,UAAS,gCACR"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"index-4XwggC10.d.ts","names":[],"sources":["../src/benchmarks/types.ts","../src/benchmarks/calibration.ts","../src/benchmarks/routing/dataset.ts","../src/benchmarks/routing/index.ts","../src/benchmarks/runner.ts","../src/benchmarks/standard-formats.ts","../src/benchmarks/index.ts"],"mappings":";;;;;KASY;KAQA;UAgBK,qBAAqB;;;EAGpC;;EAEA,SAAS;;EAET,QAAQ;;EAER,SAAS;;EAET,WAAW;;EAEX;;EAEA,SAAS;EACT,WAAW;;UAGI;;;;EAIf;;EAEA;;EAEA,aAAa;;;EAGb,KAAK;EACL;;UAGe;EACf;EACA;EACA;EACA;EACA;;;UAQe,iBAAiB,kBAAkB,oBAAoB;;EAEtE;EACA,SAAS;EACT,WAAW;EACX;EACA,SAAS;EACT;;;;;EAKA,YAAY,OAAO,cAAc,QAAQ,qBAAqB;;EAE9D,SAAS,MAAM,qBAAqB,WAAW,UAAU,YAAY,QAAQ;;;;EAI7E,YAAY,iBAAiB;;UAGd,kBAAkB,4BAA4B;EAC7D;EACA;EACA,QAAQ;EACR,UAAU;EACV,UAAU;EACV,MAAM,qBAAqB;;KAGjB,mBAAmB,oBAAoB,uBAAuB;EACxE,UAAU,kBAAkB;EAC5B,MAAM,qBAAqB;EAC3B,SAAS;MACL,QAAQ,aAAa;;;cAoBd;;;;;;;;;iBAUG,mBACd,gBACA,gBACC;;;UCjJc,kCAAkC,oBAAoB;EACrE,SAAS,iBAAiB,qBAAqB,WAAW,UAAU;EACpE,MAAM,qBAAqB;EAC3B,cAAc;EACd,gBAAgB;EAChB;EACA;EACA;;UAGe;EACf;EACA,MAAM;EACN,QAAQ;EACR;EACA;EACA;EACA;;iBAGoB,yBAAyB,oBAAoB,oBACjE,SAAS,kCAAkC,UAAU,aACpD,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;UCFM;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;;cAGW,iBAAiB;;;;KChBlB,iBAAiB;KACjB,qBAAqB,qBAAqB;cAEhD,0BAA0B,iBAAiB,oBAAoB;WAC1D;WACA;WACA;WACA;WACA;EAEH,YAAY,OAAO,cAAc,QAAQ;EAMzC,SAAS,MAAM,oBAAoB,mBAAmB,QAAQ;EAqBpE,YAAY,iBAAiB;;;;;;;;iBAef,mBAAmB;cAOtB,cAAW,OAjDG,gBAAc,QAAQ;cAkDpC,WAAQ,MA5CE,oBAAkB,qBAAqB,QAAQ;cA6CzD,cAAW,mBAxBO;;;UClCd,oBAAoB,oBAAoB;EACvD,SAAS,iBAAiB,qBAAqB,WAAW,UAAU;EACpE,SAAS,mBAAmB,UAAU;EACtC,kBAAkB;EAClB;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,UAAU;EACV,YAAY;;UAGG;EACf;EACA,QAAQ;EACR,UAAU;EACV,SAAS;EACT;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,QAAQ,eAAe;EACvB,MAAM,eAAe;EACrB,YAAY,eAAe;EAC3B,OAAO;EACP,SAAS;EACT,WAAW;;UAGI;EACf;EACA;EACA;EACA,OAAO;EACP,SAAS;EACT,WAAW;;UAGI;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe,mBAAmB,oBAAoB;EACtD,WAAW,MAAM,kBAAkB;EACnC,UAAU,eAAe,WAAW,kBAAkB;EACtD,QAAQ;EACR;EACA;;iBAGoB,oBAAoB,oBAAoB,oBAC5D,SAAS,oBAAoB,UAAU,aACtC,QAAQ,mBAAmB,UAAU;iBAmGxB,2BAA2B,UAAU,WAAW;EAC9D,SAAS,iBAAiB,qBAAqB,WAAW,UAAU;EACpE,WAAW,MAAM,kBAAkB;EACnC,UAAU,eAAe,WAAW,kBAAkB;IACpD;iBA6CY,8BAA8B,QAAQ;;;UCpOrC;EACf;EACA;EACA;EACA,WAAW;;UAGI;EACf;EACA;EACA,WAAW;;UAGI;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA,gBAAgB;EAChB,SAAS,eAAe;EACxB,WAAW;;UAGI;EACf;EACA,QAAQ;EACR,kBAAkB;EAClB,gBAAgB;EAChB,kBAAkB;EAClB;EACA,SAAS;EACT;EACA,WAAW,oBAAoB;;UAGhB,kCAAkC;EACjD,oBAAoB;EACpB;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;;KAGU,kEAGC;EAEP;EACA;EACA,mBAAmB;;UAGR;EACf,oBAAoB;EACpB;EACA;EACA;EACA;;iBAGc,eAAe,aAAa,eAAe;iBAc3C,aAAa;iBASb,WAAW,eAAe;iBAgB1B,qBAAqB,eAAe;iBAgBpC,sBAAsB,eAAe;iBAerC,4BACd,SAAS,qCACR,MAAM,qBAAqB;iBAuCd,kCACd,SAAS,4BACR,iBACD,qBAAqB,2BACrB,0BACA;iBAqBc,0BACd,SAAS,0BACT,UAAU,2BACV,UAAS;;;;;;;;;;iBAsCK,8BACd,UAAU,2BACV,oBAAmB;iBA0BL,yBAAyB;EACvC;EACA,gBAAgB;EAChB;IACE"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"index-BIL5vxxt.d.ts","names":[],"sources":["../src/campaign/analyst-surface.ts","../src/paired-arms.ts","../src/campaign/cross-surface-types.ts","../src/campaign/cross-surface-interaction.ts","../src/campaign/fixtures.ts","../src/campaign/gates/neutralization-gate.ts","../src/pre-registration.ts","../src/campaign/gates/sequential.ts","../src/campaign/gates/statistical-heldout.ts","../src/campaign/grounded-reflection.ts","../src/campaign/labeled-store/fs-adapter.ts","../src/campaign/neutralize.ts","../src/produced-state.ts","../src/agent-profile.ts","../src/integrity/backend-integrity.ts","../src/campaign/presets/run-profile-matrix.ts","../src/campaign/presets/playback.ts","../src/campaign/run-dir.ts","../src/campaign/scenario-selection.ts","../src/campaign/score-utils.ts","../src/campaign/search-ledger-errors.ts","../src/campaign/search-ledger.ts","../src/campaign/single-run-lock.ts","../src/campaign/surface-identity.ts","../src/campaign/transient-failure.ts","../src/campaign/upstream-evaluators.ts","../src/campaign/worktree/index.ts"],"mappings":";;;;;;;;;;;;;UAWiB,6BAA6B;EAC5C;EACA,YAAY;EACZ,YAAY;EACZ,yBAAyB;EACzB,kBAAkB;;UAGH;EACf,mBAAmB;EACnB,QAAQ;EACR,MAAM;;UAGS;EACf,QAAQ;IACN;IACA,YAAY;IACZ;IACA,QAAQ;MACN,QAAQ;;iBAGE,iCACd,SAAS,2CAET,SAAS,gBACT,UAAU,sBACV,SAAS,oBACN,QAAQ;iBAgBG,4BAA4B,YAC1C,sBACA;;;;;UClBe;;;EAGf;;;;;EAKA;;EAEA;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;;EAEA;;;UAIe;EACf;;;;EAIA;EACA,UAAU;EACV,WAAW;;UAGI;;EAEf,OAAO;;;EAGP,kBAAkB;;EAElB,mBAAmB;;;;;;;;;;;;;;;;;;;;iBAqBL,SAAS,eAAe,gBAAgB,MAAM,kBAAkB;;UA8G/D;;EAEf;;EAEA;;EAEA,SAAS;;EAET,gBAAgB;;;UAID;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,aAAa;;EAEb;IAAY;IAAW;;;UAGR,iCAAiC;;;;EAIhD;;EAEA,YAAY;;UAGG;EACf;EACA;EACA;;;EAGA,aAAa;EACb,cAAc;;;;;;;;;;;;;;;iBAgBA,kBACd,eAAe,gBACf,MAAM,2BACL;UAmEc;EACf;EACA;EACA,UAAU;EACV,WAAW;;UAGI;EACf,OAAO;EACP,kBAAkB;EAClB,mBAAmB;;;;;;;;;iBAeL,eACd,uBAAuB,aACvB,wBAAwB,cACvB;;;;KCjXS;;UAGK;EACf;EACA;;EAEA;;;UAIe;EACf;EACA;EACA;EACA;;;UAIe;EACf;;EAEA;;EAEA;;;;;;;UAQe;EACf;EACA;;EAEA;EACA,cAAc;EACd;EACA;;;;;;EAMA,MAAM;EACN,mBAAmB;;EAEnB;;UAGe;EACf;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;;EAEA,kCAAkC;;EAElC;;UAGe,qCACf,aAAa,sBAAsB;EAEnC,qBAAqB;EACrB,qBAAqB;EACrB,eAAe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;EACA,WAAW;EACX,WAAW;;UAGI;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;;UAGe;EACf,aAAa;EACb;EACA;EACA;EACA;;KAGU;UAWK;EACf;EACA,SAAS;;UAGM;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf,WAAW;EACX,SAAS;EACT,OAAO;EACP,OAAO,eAAe;EACtB,QAAQ;EACR,QAAQ;;EAER,sBAAsB;;EAEtB,aAAa;;UAGE;EACf;EACA;EACA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA,QAAQ;EACR,cAAc,eAAe;;UAGd;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;EACA;;UAGe;EACf,SAAS;EACT;EACA;EACA;EACA;EACA,eAAe;EACf,gBAAgB;;KAGN;UAYK;EACf;EACA,SAAS;EACT;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA,4BAA4B,iCAAiC;EAC7D,wBAAwB,eAAe;EACvC,QAAQ;EACR,QAAQ;EACR,aAAa;EACb,eAAe;;UAGA;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA,SAAS;;UAGM;;EAEf;EACA;;KAGU;UAaK;EACf;EACA;EACA;EACA;EACA;EACA,uBAAuB;EACvB;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA,YAAY;EACZ;;;UAIe;EACf;EACA;EACA;EACA;EACA,OAAO;;UAGQ;;EAEf;;EAEA;EACA;;EAEA;EACA;;EAEA,gBAAgB;;EAEhB,OAAO;;UAGQ;EACf,YAAY;EACZ,YAAY;EACZ,kBAAkB;;UAGH,8BACf,aAAa,sBAAsB;EAEnC;EACA;EACA;EACA;;EAEA,MAAM;EACN,iBAAiB;EACjB,iBAAiB;EACjB,YAAY;EACZ,UAAU;EACV,YAAY;;;;;;;;;iBCnRE,gCAAgC,aAAa,qBAC3D,OAAO,qCAAqC,QAC3C,8BAA8B;;;KC7CrB;UAEK;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;EACP;;UAGe,4BAA4B;EAC3C;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;;EAEf,aAAa;;EAEb;;UAGe,wCAAwC;EACvD;;UAGe,0BAA0B,6BACjC,KACN,uBAAuB,qBAAqB;EAG9C;EACA,aAAa;EACb;EACA;EACA,UAAU;;KAGA,qBAAqB;EAC/B,UAAU,MAAM,KAAK;;;iBAIP,qBAAqB;;;;;iBA4BrB,gBACd,kBACA,cACA,UAAS,yBACR;;iBA4Ca,yBACd,kBACA,UAAS,kCACR;;;;;iBAuBa,mBAAmB,qBACjC,SAAS,0BAA0B,aAClC;;;UChJc,0BAA0B,kBAAkB,WAAW;EACtE,WAAW;;;;;EAKX;;;;;;iBAuBc,mBAAmB,WAAW,kBAAkB,UAC9D,SAAS,0BAA0B,aAClC,KAAK,WAAW;;;;;;;;;;;;;UC9CF;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;;;;;;;;;;KAWU;UAEK,uBAAuB;;EAEtC;;;;;;;;;EASA,OAAO;;UAGQ;EACf,UAAU;EACV;EACA;EACA;;;EAGA;;EAEA,kBAAkB;EAGlB;;;;;;;;;;;iBAYc,aAAa;;;;;;;;;;;;;;;;;;;;;;iBA8BP,SAAS,GAAG,KAAK,IAAI;;;;;;;;;;iBAkBrB,aAAa,GAAG,qBAAqB,QAAQ;;;;;;;iBAW7C,eAAe,GAAG,iBAAiB;;;;;iBAWnC,mBACpB,UAAU,gBACV;EAAY;EAAW;EAAgB;IACtC,QAAQ;;;KCpHC;UAEK;EACf,UAAU;;EAEV;;EAEA;;;EAGA;;UAGe;;;EAGf;;;;EAIA;;;EAGA;;EAEA;;;;EAIA;;;EAGA;;;EAGA,kBAAkB;;EAElB;;UAGe,qBAAqB,qBAAqB,kBAAkB,WAAW,kBAC9E,KAAK,WAAW;;;;;;;;EAQxB,QAAQ,gBAAgB;;EAExB,SAAS;IAAkB,UAAU;;;;;;;;;;;;;;;;;iBA0MvB,qBAAqB,qBAAqB,kBAAkB,WAAW,UACrF,SAAS,8BACR,qBAAqB,WAAW;UAsFlB;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;GACd;IAAQ,SAAS;;IAAyB;IAAe;;;EAE1D,SAAS;;;;;;;;;;;;;;;;;;;;;;;;;iBA0BK,iBAAiB,UAAS,0BAA+B;;;UCtXxD;;EAEf;;EAEA;;EAEA;;;;;;;;;;;;;iBAcc,YACd,WAAW,YAAY,eAAe,cACtC,UAAU,YAAY,eAAe,cACrC,aAAa,aACb,SAAS,GAAG,oCACX;UAsDc;EACf,QAAQ;;;;;;;;;;;;EAYR,WAAW;;;;EAIX,iBAAiB;;;;;;;EAOjB,UAAU;;EAEV,mBAAmB;;EAEnB,SAAS;;;;EAIT;;EAEA;;EAEA;;EAEA,gBAAgB;;EAEhB;;;;EAIA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;;EAEA;EACA;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA6Bc,oBACd,QAAQ,eACR,OAAM,6BACL;UAkEc;EACf;;;EAGA,WAAW;;;;EAIX;;EAEA;IAAM;IAAa;;;EAEnB,mBAAmB;;EAEnB,SAAS;;EAET;;;;;EAKA;EACA;EACA;;;;;iBAMc,YAAY;;;;;;;;;;;;;;;;;;;iBAsBZ,qBACd,WAAW,YAAY,eAAe,cACtC,UAAU,YAAY,eAAe,cACrC,aAAa,aACb,8BACA;EACE;EACA;EACA;EACA;;;EAGA;IAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC7Tc;WACN;WACA,MAAM,SAAS;;;UAIT;;WAEN;;WAEA;WACA,gBAAgB;;UAGV;;WAEN;;WAEA;;UAGM;;WAEN;;WAEA,eAAe;;WAEf,eAAe;;;;;;;;iBASV,oBACd,mBAAmB,iBACnB,OAAM,6BACL;UAsCc;;WAEN;;WAEA;;;;;;;;;iBAUK,2BACd,cACA,MAAM,KAAK,0DACV;;;UCjFc;;EAEf;;;EAGA;;EAEA;;;cAIW,kCAAkC;WAE3B;EADlB,YACkB,cAChB;;;;;;cAiBS,kCAAkC;mBAIhB;mBAHZ;mBACA;EAEjB,YAA6B,SAAS;EAKhC,QAAQ,OAAO,uBAAuB;EAYtC,OAAO,MAAM,4BAA4B,QAAQ;EA8CjD,QAAQ;IACZ;IACA;IACA,UAAU;IACV,SAAS,OAAO;;UAkCV;UAiCA;UAmBA;UAkBA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBCtNM,eAAe;;;;UCZd;EACf;EACA;;;;;;;UAQe;EACf;EACA;EACA;EACA;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;EAIA;;;;;;;;KASU,mBACR,oBACA,oBACA;EACE;;;;;;;;;;;;iBAmBU,qBAAqB,iBAAiB,qBAAqB;;;;;;;;;;cCjE9D,2BAA2B;UAOvB;;;EAGf,MAAM;;EAEN,qBAAqB;;;EAGrB;;;;;;EAMA;;;;;;;cAQW;;;;;;;;;;;;;;;;;;;iBAoBG,kBAAkB,MAAM,kBAAkB;;;;;;;;iBAwD1C,cACd,SAAS,KAAK;EACX,SAAS;EAAa;;;;;;;;;iBAiBX,eAAe,SAAS;;;;;iBASxB,oBAAoB,SAAS;;;;;;;;;;;iBA0C7B,iBAAiB,SAAS;;;UClKzB;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;cAQW,8BAA8B;WAGvB,QAAQ;EAF1B,YACE,iBACgB,QAAQ;;;;;;;iBAWZ,0BACd,SAAS,cAAc,aACtB;;iBAWa,+BACd,UAAU,cAAc,eACvB;;;;;;;;iBA0Ga,kBACd,SAAS,cAAc,YACvB;EAAQ;IACP;;iBAMa,wBACd,UAAU,cAAc,cACxB;EAAQ;IACP;;;;;;cC7HU,2BAA2B;EACtC,YAAY;;;;;KAQF,kBAAkB,kBAAkB,UAAU,cACxD,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;UAEI,wBAAwB,kBAAkB,UAAU;;EAEnE,UAAU;;EAEV,WAAW;;EAEX,UAAU,kBAAkB,WAAW;;EAEvC,SAAS,YAAY,WAAW;;;EAGhC;;;EAGA;;;EAGA;;;;EAIA;;EAEA,WAAW;;EAEX;;EAEA;;;;;;;;EAQA;;;EAGA;;EAEA;;;EAGA;;EAEA;;EAEA,eAAe;EACf,gBAAgB;;;EAGhB,UAAU;;EAEV,YAAY;;;EAGZ,aAAa,UAAU;;;;EAIvB;;;;;;;EAOA,cACE,UAAU,WACV,UAAU;IACL;IAAgB;;;UAGR;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;EACA,gBAAgB;;;EAGhB,WAAW;;UAGI;EACf;EACA;;UAGe,uBAAuB,WAAW,kBAAkB;EACnE;EACA;;;;EAIA,SAAS;EACT,WAAW,eAAe;EAC1B,YAAY,eAAe;;EAE3B,YAAY,eAAe;;EAE3B,WAAW;;EAEX,WAAW,eAAe,eAAe,WAAW;;;;;iBA2IhC,iBAAiB,kBAAkB,UAAU,WACjE,MAAM,wBAAwB,WAAW,aACxC,QAAQ,uBAAuB,WAAW;;;;;UCzT5B;;EAEf;;EAEA,UAAU;;;;;;;UAQK,kBAAkB;;EAEjC;;EAEA,OAAO;;EAEP,cAAc;;;UAIC,wBAAwB;EACvC,SAAS;;;;;;;;;;;UAYM,eAAe,eAAe,YAAY;EACzD,IAAI,OAAO,QAAQ,KAAK,kBAAkB,iBAAiB;;;;;;;iBAQ7C,qBAAqB,eAAe,WAClD,QAAQ,eAAe,UACtB,kBAAkB,QAAQ;;UAQZ,yBAAyB;EACxC;;;;;;;;iBASoB,eACpB,OAAO,WACP,OAAO,eACP,kBAAkB,qBACjB,QAAQ;;UAUM;EACf;EACA;EACA;EACA;EACA;EACA;;;;;;;iBAQc,oBAAoB,mBAAmB,qBAAqB;;UAkB3D;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;iBAIc,kBAAkB,eAAe,kBAAkB;UAwBlD;;EAEf;;EAEA,OAAO;;EAEP;;;;;;;;;iBAkBc,yBACd,eAAe,iBACf,OAAM;;;;;;;;iBCnMQ;;;;;iBAQA,cAAc,gBAAgB;;;;;;;;;;;;;;;;;UCD7B;EACf;;EAEA;;UAGe;EACf;;EAEA;;EAEA;EACA;;EAEA;;;;;;;;;;iBA4Cc,oBACd,SAAS,kBACT;EAAS;IACR;;;;;;;;;;;iBA0Ba,qBACd,SAAS,kBACT,WACA;EAAS;;;;;;;iBC3FK,sBAAsB,WAAW,kBAAkB,UACjE,UAAU,eAAe,WAAW;;;;iBA0CtB,gBAAgB,sBAAsB;UAWrC;;EAEf,YAAY;;;;;EAKZ,WAAW;IAAQ;IAAoB;IAAmB;IAAgB;;;;;iBAK5D,kBAAkB,WAAW,kBAAkB,UAC7D,UAAU,eAAe,WAAW,aACnC;;;;cC/EU,0BAA0B;;cAG1B,mCAAmC;;cAGnC,kCAAkC;;;cCgClC;KAED,mBAAmB;KAEnB;;;UAYK;EACf;EACA;EACA,QAAQ;EACR;;;;UAKe;EACf;EACA;;UAGe;EACf;EACA;;UAGe;EACf;EACA,MAAM;EACN,UAAU;;UAGK;;EAEf;EACA;EACA;EACA;EACA,gBAAgB;;KAGN;UAOK;EACf;EACA,QAAQ;EACR,WAAW;;;EAGX;;UAGe;EACf;EACA,MAAM;;UAGS;EACf;;;EAGA;;UAGe;;EAEf,gBAAgB;;EAEhB,OAAO;;EAEP,YAAY;;KAGF;EAEN;EACA;EACA;EACA;;EAGA;EACA;;KAGM;EAEN;EACA;EACA;;EAGA;;EAEA;EACA;;UAGW;EACf,QAAQ;EACR,MAAM;;UAGS;EACf;EACA;;KAGU;EAEN;EACA;EACA,SAAS;;EAGT;EACA;EACA,SAAS;EACT,SAAS;;EAGT;EACA,SAAS;EACT,OAAO;IAAwB;;;KAGzB;EAEN;EACA;EACA;EACA;EACA;;EAGA;EACA;;;;UAKW;EACf;EACA;EACA;EACA,QAAQ;EACR,UAAU;;UAGF;EACR;EACA;EACA,WAAW;;UAGI,2BAA2B;EAC1C;EACA,MAAM;;UAGS,uCAAuC;EACtD;EACA;EACA;EACA;EACA,SAAS;EACT,UAAU;;UAGK,uCAAuC;EACtD;EACA;EACA;EACA,QAAQ;;UAGO,iCAAiC;EAChD;EACA;EACA;EACA;EACA;IACE;IACA,QAAQ;;EAEV;IACE,OAAO;IACP,OAAO;IACP,WAAW;;EAEb,SAAS;EACT,YAAY;EACZ,iBAAiB;;UAGF,qCAAqC;EACpD;EACA;EACA,eAAe;EACf;IAEM;IACA,OAAO;IACP,QAAQ;;IAGR;IACA,QAAQ;;EAEd;IACM;;IACA;IAAmB,SAAS;;IAC5B;IAAkB,SAAS;;EACjC,YAAY;;UAGG,oCAAoC;EACnD;EACA;EACA;IACM;;IAEA;IACA,QAAQ;;;UAIC,6BAA6B;EAC5C;EACA;IAEM;IACA;;IAGA;IACA,QAAQ;;;KAIJ,oBACR,qBACA,iCACA,iCACA,2BACA,+BACA,8BACA;UAEa;EACf,eAAe;EACf;EACA;EACA,cAAc;EACd,OAAO;EACP,WAAW;;KAGD;EAEN;EACA;EACA;EACA;EACA;;EAGA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGW;EACf;EACA;EACA;EACA;EACA;EACA;EACA;IAAY;IAAgB;IAAgB;;EAC5C;IAAqB;IAAmB;IAAiB;;EACzD;IAAa;IAAkB;IAAkB;;EACjD;IACE;IACA;IACA;IACA;IACA;IACA;;EAEF;EACA;EACA,YAAY;EACZ,UAAU;;UAGK;EACf,SAAS;EACT,MAAM;EACN,YAAY;EACZ,sBAAsB;EACtB,UAAU;EACV,YAAY;EACZ,WAAW;EACX,YAAY;EACZ,OAAO;;UAGQ;EACf,OAAO;;EAEP;EACA,QAAQ;;;;iBA+ZM,0BAA0B,iBAAiB;;;;;;;;;;;;;;;KAsB/C;UAEK;EACf;EACA;EACA,cAAc;;UAGC;WACN;WACA;;WAEA;EACT,OAAO,OAAO,oBAAoB,QAAQ;EAC1C,UAAU,QAAQ;;EAElB,eAAe,QAAQ;;;EAGvB,kBAAkB,QAAQ;;;;;;;EAO1B,oBAAoB,QAAQ;;;;iBAKd,iBAAiB,SAAS,0BAA0B;;cA8BvD,4BAA4B;WAC9B;WACA;WACA;mBACQ;mBACA;EAMjB,YAAY,cAAc,oBAAoB,cAAa;EAerD,UAAU,QAAQ;EAIlB,OAAO,OAAO,oBAAoB,QAAQ;EAU1C,eAAe,QAAQ;EAIvB,kBAAkB,QAAQ;EAI1B,oBAAoB,QAAQ;;;;;;;;;;;;;;;;;;;UC/3BnB;;WAEN;;WAEA;;WAEA;;WAEA;;UAGM;;EAEf;;;;;;;iBAwBc,qBAAqB,MAAM,uBAAuB;;;;iBCpDlD,0BAA0B,2BAA2B,WAAW;;iBAmChE,uBAAuB,2BAA2B,WAAW;;iBA8B7D,iCAAiC,SAAS;;;;iBAa1C,4BAA4B,SAAS;;iBAgBrC,mBAAmB,SAAS;;iBAW5B,YAAY,SAAS;;iBAKrB,kBACd,eAAe,gBACf,iBAAiB;;;;;;;;;;;;;;;;;;;;;UCpGF;;;;;;;WAON;;WAEA,yBAAyB;;;;;;iBAYpB,4BACd,oCACA,OAAM;;;UCnCS;EACf;EACA;EACA;;UAGe,qBAAqB,gBAAgB;EACpD;EACA;EACA;EACA,SACE,QAAQ,SACR,SAAS,4BACR,QAAQ;;UAGI;EACf;EACA;EACA,WAAW;;KAGD,oBAAoB,eAAe,4BAC7C,OAAO,QACP,SAAS,8BACN,qBAAqB,QAAQ;UAEjB;WACN,QAAQ;WACR;;KAGN,sBAAsB,WAAW,KACpC,iBAAiB;EAGjB;;UAGQ,qBAAqB,kBAAkB;EAC/C;EACA;EACA;EACA,aAAa,UAAU;;EAEvB,eAAe;;iBAGD,sBACd,gBAAgB,yBAChB,WACA,kBAAkB,WAAW,UAE7B,WAAW,qBAAqB,UAChC,SAAS,qBAAqB;EAC5B,SAAS;IAAS,UAAU;IAAW,UAAU;MAAc;EAC/D,WAAW,sBAAsB;IAElC,YAAY,WAAW;iBA8CV,qBACd,eAAe,yBACf,WACA,kBAAkB,WAAW,UAE7B,QAAQ,oBAAoB,SAC5B,SAAS,qBAAqB;EAC5B;EACA,SAAS;IAAS,UAAU;IAAW,UAAU;MAAc;;EAEzD;EAAc;;EACd;EAAa,UAAU,sBAAsB;KAEpD,YAAY,WAAW;;;KCxFrB,qBAAqB;KACrB,iBAAiB,SAAS;KAC1B,aAAa,gBAAgB,aAAa,MAAM,mBAAmB;UAcvD;;WAEN;;WAEA;;WAEA;;WAEA;;WAEA;;UAGM;;EAEf,OAAO;IAAQ;IAAiB;MAAkB,QAAQ;;;EAG1D,SAAS,UAAU,UAAU,kBAAkB,QAAQ;;EAEvD,QAAQ,UAAU,WAAW;;;cAIlB,6BAA6B;WAG7B;EAFX,YACE,iBACS;;UAOI;;EAEf;;EAEA;;EAEA;;;;EAIA,MAAM;;UAobS;;EAEf;;EAEA;;EAEA;;;EAGA,YAAY;;;;;iBAoJE,mBAAmB,MAAM,4BAA4B;;;;iBA0FrD,kBACd,SAAS,aACT,uBACC;;;iBAOa,oBAAoB,SAAS,aAAa"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"integrity-fdt8XPAv.js","names":[],"sources":["../src/trace/store.ts","../src/trace/integrity.ts"],"sourcesContent":["/**\n * TraceStore — persistence + query over the TraceSchema v1 corpus.\n *\n * Two implementations ship in the core:\n * - InMemoryTraceStore: dev + tests, fully in-process\n * - FileSystemTraceStore: NDJSON append-only files per entity, suitable\n * for long-running CI jobs; rolled over at 32MB\n *\n * Downstream adapters (DuckDB, Langfuse, R2 parquet) implement this same\n * interface — the rest of the framework is storage-agnostic.\n */\n\nimport type {\n Artifact,\n BudgetLedgerEntry,\n EventKind,\n Run,\n RunStatus,\n Span,\n SpanKind,\n TraceEvent,\n} from './schema'\n\nexport interface RunFilter {\n scenarioId?: string\n variantId?: string\n status?: RunStatus\n since?: number\n until?: number\n tag?: { key: string; value: string }\n parentRunId?: string\n projectId?: string\n chatId?: string\n layer?: import('./schema').RunLayer\n}\n\nexport interface SpanFilter {\n runId?: string\n parentSpanId?: string\n kind?: SpanKind\n name?: string\n toolName?: string\n judgeId?: string\n since?: number\n until?: number\n}\n\nexport interface EventFilter {\n runId?: string\n spanId?: string\n kind?: EventKind\n since?: number\n until?: number\n}\n\nexport interface TraceStore {\n appendRun(run: Run): Promise<void>\n updateRun(runId: string, patch: Partial<Run>): Promise<void>\n appendSpan(span: Span): Promise<void>\n updateSpan(spanId: string, patch: Partial<Span>): Promise<void>\n appendEvent(event: TraceEvent): Promise<void>\n appendArtifact(artifact: Artifact): Promise<void>\n appendBudgetEntry(entry: BudgetLedgerEntry): Promise<void>\n\n getRun(runId: string): Promise<Run | undefined>\n listRuns(filter?: RunFilter): Promise<Run[]>\n spans(filter?: SpanFilter): Promise<Span[]>\n events(filter?: EventFilter): Promise<TraceEvent[]>\n budget(runId: string): Promise<BudgetLedgerEntry[]>\n artifacts(runId: string): Promise<Artifact[]>\n}\n\n// ── In-memory ────────────────────────────────────────────────────────\n\nexport class InMemoryTraceStore implements TraceStore {\n private runs = new Map<string, Run>()\n private allSpans: Span[] = []\n private allEvents: TraceEvent[] = []\n private allArtifacts: Artifact[] = []\n private allBudget: BudgetLedgerEntry[] = []\n\n async appendRun(run: Run): Promise<void> {\n if (this.runs.has(run.runId)) throw new Error(`run ${run.runId} already exists`)\n this.runs.set(run.runId, { ...run })\n }\n\n async updateRun(runId: string, patch: Partial<Run>): Promise<void> {\n const existing = this.runs.get(runId)\n if (!existing) throw new Error(`run ${runId} not found`)\n this.runs.set(runId, { ...existing, ...patch })\n }\n\n async appendSpan(span: Span): Promise<void> {\n this.allSpans.push({ ...span })\n }\n\n async updateSpan(spanId: string, patch: Partial<Span>): Promise<void> {\n const idx = this.allSpans.findIndex((s) => s.spanId === spanId)\n if (idx < 0) throw new Error(`span ${spanId} not found`)\n this.allSpans[idx] = { ...this.allSpans[idx], ...patch } as Span\n }\n\n async appendEvent(event: TraceEvent): Promise<void> {\n this.allEvents.push({ ...event })\n }\n\n async appendArtifact(artifact: Artifact): Promise<void> {\n this.allArtifacts.push({ ...artifact })\n }\n\n async appendBudgetEntry(entry: BudgetLedgerEntry): Promise<void> {\n this.allBudget.push({ ...entry })\n }\n\n async getRun(runId: string): Promise<Run | undefined> {\n const r = this.runs.get(runId)\n return r ? { ...r } : undefined\n }\n\n async listRuns(filter: RunFilter = {}): Promise<Run[]> {\n return [...this.runs.values()].filter((r) => matchesRun(r, filter))\n }\n\n async spans(filter: SpanFilter = {}): Promise<Span[]> {\n return this.allSpans.filter((s) => matchesSpan(s, filter)).map((s) => ({ ...s }))\n }\n\n async events(filter: EventFilter = {}): Promise<TraceEvent[]> {\n return this.allEvents.filter((e) => matchesEvent(e, filter)).map((e) => ({ ...e }))\n }\n\n async budget(runId: string): Promise<BudgetLedgerEntry[]> {\n return this.allBudget.filter((b) => b.runId === runId).map((b) => ({ ...b }))\n }\n\n async artifacts(runId: string): Promise<Artifact[]> {\n return this.allArtifacts.filter((a) => a.runId === runId).map((a) => ({ ...a }))\n }\n}\n\nfunction matchesRun(r: Run, f: RunFilter): boolean {\n if (f.scenarioId && r.scenarioId !== f.scenarioId) return false\n if (f.variantId && r.variantId !== f.variantId) return false\n if (f.status && r.status !== f.status) return false\n if (f.since !== undefined && r.startedAt < f.since) return false\n if (f.until !== undefined && r.startedAt > f.until) return false\n if (f.tag && r.tags?.[f.tag.key] !== f.tag.value) return false\n if (f.parentRunId && r.parentRunId !== f.parentRunId) return false\n if (f.projectId && r.projectId !== f.projectId) return false\n if (f.chatId && r.chatId !== f.chatId) return false\n if (f.layer && r.layer !== f.layer) return false\n return true\n}\n\nfunction matchesSpan(s: Span, f: SpanFilter): boolean {\n if (f.runId && s.runId !== f.runId) return false\n if (f.parentSpanId && s.parentSpanId !== f.parentSpanId) return false\n if (f.kind && s.kind !== f.kind) return false\n if (f.name && s.name !== f.name) return false\n if (f.toolName && (s.kind !== 'tool' || s.toolName !== f.toolName)) return false\n if (f.judgeId && (s.kind !== 'judge' || s.judgeId !== f.judgeId)) return false\n if (f.since !== undefined && s.startedAt < f.since) return false\n if (f.until !== undefined && s.startedAt > f.until) return false\n return true\n}\n\nfunction matchesEvent(e: TraceEvent, f: EventFilter): boolean {\n if (f.runId && e.runId !== f.runId) return false\n if (f.spanId && e.spanId !== f.spanId) return false\n if (f.kind && e.kind !== f.kind) return false\n if (f.since !== undefined && e.timestamp < f.since) return false\n if (f.until !== undefined && e.timestamp > f.until) return false\n return true\n}\n\n// ── Filesystem (NDJSON append-only, one file per entity) ─────────────\n\nexport interface FileSystemTraceStoreOptions {\n dir: string\n /** Roll over NDJSON files when they exceed this size in bytes. Default 32 MB. */\n maxBytes?: number\n}\n\nexport class FileSystemTraceStore implements TraceStore {\n private dir: string\n private maxBytes: number\n /** Lazy in-memory index for queries — populated on first read. */\n private index?: InMemoryTraceStore\n /** Memoized index build — concurrent first reads share one build, and an\n * append racing an in-flight load awaits this so its row isn't lost. */\n private indexPromise?: Promise<InMemoryTraceStore>\n /** Strictly-increasing rollover stamp. Date.now() alone collides when two\n * rollovers land in the same millisecond, overwriting a rolled file. */\n private lastRolloverStamp = 0\n /**\n * Per-file append serialization. stat → conditional-rename → appendFile is a\n * read-modify-write on the active file; without a lock two concurrent appends\n * to the same `name` can both pass the size check (exceeding maxBytes) or one\n * can append to a file the other just renamed away. Each entry chains the\n * next append behind the prior one for that file.\n */\n private appendLocks = new Map<string, Promise<void>>()\n\n constructor(options: FileSystemTraceStoreOptions) {\n this.dir = options.dir\n this.maxBytes = options.maxBytes ?? 32 * 1024 * 1024\n }\n\n private async ensureDir(): Promise<void> {\n const fs = await import('node:fs/promises')\n await fs.mkdir(this.dir, { recursive: true })\n }\n\n private async append(name: string, record: unknown): Promise<void> {\n // If an index load is in flight, wait for it. The row must hit disk AFTER\n // the load finished reading, so it lands in the completed index via the\n // mirror below — never lost in the gap before `this.index` is assigned.\n if (this.indexPromise) await this.indexPromise\n // Chain this append behind any in-flight append for the same file. The tail\n // promise never rejects (errors are isolated to each caller's `result`), so\n // one failed append can't break the lock chain for the next one.\n const prior = this.appendLocks.get(name) ?? Promise.resolve()\n const result = prior.then(() => this.appendLocked(name, record))\n const tail = result.then(\n () => {},\n () => {},\n )\n this.appendLocks.set(name, tail)\n try {\n await result\n } finally {\n // Drop the slot once this append is the current tail, so the map doesn't\n // grow without bound across many appends to the same file.\n if (this.appendLocks.get(name) === tail) this.appendLocks.delete(name)\n }\n }\n\n private async appendLocked(name: string, record: unknown): Promise<void> {\n await this.ensureDir()\n const fs = await import('node:fs/promises')\n const path = await import('node:path')\n const active = path.join(this.dir, `${name}.ndjson`)\n try {\n const stat = await fs.stat(active)\n if (stat.size >= this.maxBytes) {\n // Strictly-increasing, collision-free even within one millisecond — a\n // bare Date.now() would let two same-ms rollovers overwrite each other.\n const stamp = Math.max(Date.now(), this.lastRolloverStamp + 1)\n this.lastRolloverStamp = stamp\n const rolled = path.join(this.dir, `${name}.${stamp}.ndjson`)\n await fs.rename(active, rolled)\n }\n } catch {\n /* file doesn't exist yet */\n }\n await fs.appendFile(active, `${JSON.stringify(record)}\\n`, 'utf8')\n // Mirror genuinely-new rows into the lazy index. Update rows (marked\n // with `_update: true` by updateRun/updateSpan) are applied by those\n // methods directly via the index's update* APIs — re-inserting them\n // here triggers a duplicate-id error once the first read populates\n // the index. Await the mirror: a fire-and-forget would let an index\n // rejection go unhandled and let disk + index diverge silently.\n if (this.index && !(record as { _update?: boolean })?._update) {\n await this.insertInto(name, record)\n }\n }\n\n private async insertInto(name: string, record: unknown): Promise<void> {\n if (!this.index) return\n switch (name) {\n case 'runs':\n await this.index.appendRun(record as Run)\n break\n case 'spans':\n await this.index.appendSpan(record as Span)\n break\n case 'events':\n await this.index.appendEvent(record as TraceEvent)\n break\n case 'artifacts':\n await this.index.appendArtifact(record as Artifact)\n break\n case 'budget':\n await this.index.appendBudgetEntry(record as BudgetLedgerEntry)\n break\n }\n }\n\n private load(): Promise<InMemoryTraceStore> {\n // Memoize: concurrent first reads share one build, and an append racing an\n // in-flight load awaits this same promise (see append()).\n this.indexPromise ??= this.buildIndex()\n return this.indexPromise\n }\n\n private async buildIndex(): Promise<InMemoryTraceStore> {\n const fs = await import('node:fs/promises')\n const path = await import('node:path')\n const store = new InMemoryTraceStore()\n let entries: string[]\n try {\n entries = await fs.readdir(this.dir)\n } catch {\n // No dir yet (first run / empty corpus). Anything past this point — parse\n // errors, missing patch bases — is real corruption and must fail loud.\n this.index = store\n return store\n }\n // Replay files in write order. For one entity, rolled files\n // (`name.<ts>.ndjson`) are older than the active bare file (`name.ndjson`),\n // and readdir order is not chronological — so without this sort a patch in\n // an older rolled file could be applied AFTER (overwrite) a newer patch in\n // the active file, silently restoring a stale field. Sort: per entity,\n // rolled ascending by timestamp, then the bare active file last.\n entries.sort((a, b) => {\n // Group by the same entity base the loop uses (`split('.')[0]`); within an\n // entity, real rollover files (`<base>.<ms>.ndjson`) sort by timestamp\n // ascending and the bare active file (no numeric segment) sorts last.\n // Non-rollover names tie (Infinity) and a stable sort preserves their\n // readdir order.\n const baseA = a.split('.')[0]!\n const baseB = b.split('.')[0]!\n if (baseA !== baseB) return baseA < baseB ? -1 : 1\n const ta = a.match(/^[^.]+\\.(\\d+)\\.ndjson$/)\n const tb = b.match(/^[^.]+\\.(\\d+)\\.ndjson$/)\n const tsA = ta ? Number(ta[1]) : Number.POSITIVE_INFINITY\n const tsB = tb ? Number(tb[1]) : Number.POSITIVE_INFINITY\n return tsA - tsB\n })\n {\n // Two-pass load. Pass 1 indexes all full (base) rows; pass 2 replays the\n // `_update` patch rows that updateRun/updateSpan append. readdir order is\n // not deterministic and rollover splits a stream across files, so a patch\n // can be read before its base. Applying patches in a separate pass — after\n // every base row is indexed — means a patch always finds its target.\n // Doing it in one pass let a patch hit the catch and append a runId-less\n // fragment, corrupting span/run counts cross-instance.\n const runUpdates: Array<Record<string, unknown>> = []\n const spanUpdates: Array<Record<string, unknown>> = []\n for (const file of entries) {\n if (!file.endsWith('.ndjson')) continue\n const full = path.join(this.dir, file)\n const content = await fs.readFile(full, 'utf8')\n const base = file.split('.')[0]\n const lines = content.split('\\n')\n for (let ln = 0; ln < lines.length; ln++) {\n const line = lines[ln]!\n if (!line.trim()) continue\n let record: ReturnType<typeof JSON.parse>\n try {\n record = JSON.parse(line)\n } catch (err) {\n // Fail loud WITH context — a bare SyntaxError loses which file/line\n // is corrupt and which valid rows surround it.\n throw new Error(\n `FileSystemTraceStore: corrupt NDJSON in ${file} line ${ln + 1}: ${\n err instanceof Error ? err.message : String(err)\n } — ${line.slice(0, 120)}`,\n )\n }\n if (base === 'runs') {\n if (record?._update) {\n runUpdates.push(record)\n continue\n }\n // A duplicate full row (same dir loaded twice into one stream) is\n // collapsed last-write-wins rather than thrown.\n try {\n await store.appendRun(record)\n } catch {\n await store.updateRun(record.runId, record)\n }\n } else if (base === 'spans') {\n // `updateSpan` appends an `_update: true` patch row instead of\n // rewriting the original span. Deferred to pass 2 so it collapses\n // onto the original span — otherwise a fresh FileSystemTraceStore\n // reading the same dir reports duplicate spans (one full, one\n // fragment with no runId/kind/name), which breaks any downstream\n // consumer that re-opens the store cross-process (e.g. the OTLP\n // converter).\n if (record?._update) {\n spanUpdates.push(record)\n continue\n }\n await store.appendSpan(record)\n } else if (base === 'events') {\n await store.appendEvent(record)\n } else if (base === 'artifacts') {\n await store.appendArtifact(record)\n } else if (base === 'budget') {\n await store.appendBudgetEntry(record)\n }\n }\n }\n // Pass 2: every base row is now indexed. A patch whose base is genuinely\n // absent (truncated/partial corpus) throws inside update* — we let it\n // propagate rather than silently appending a fragment, so a corrupt\n // corpus fails loud instead of polluting counts.\n for (const record of runUpdates) {\n await store.updateRun(record.runId as string, record as Partial<Run>)\n }\n for (const record of spanUpdates) {\n await store.updateSpan(record.spanId as string, record as Partial<Span>)\n }\n }\n this.index = store\n return store\n }\n\n async appendRun(run: Run): Promise<void> {\n await this.append('runs', run)\n }\n async updateRun(runId: string, patch: Partial<Run>): Promise<void> {\n // NDJSON is append-only; record updates as new rows with the same runId —\n // readers collapse by last-write-wins on load.\n await this.append('runs', { runId, ...patch, _update: true })\n if (this.index) await this.index.updateRun(runId, patch)\n }\n async appendSpan(span: Span): Promise<void> {\n await this.append('spans', span)\n }\n async updateSpan(spanId: string, patch: Partial<Span>): Promise<void> {\n await this.append('spans', { spanId, ...patch, _update: true })\n if (this.index) await this.index.updateSpan(spanId, patch)\n }\n async appendEvent(event: TraceEvent): Promise<void> {\n await this.append('events', event)\n }\n async appendArtifact(artifact: Artifact): Promise<void> {\n await this.append('artifacts', artifact)\n }\n async appendBudgetEntry(entry: BudgetLedgerEntry): Promise<void> {\n await this.append('budget', entry)\n }\n\n async getRun(runId: string): Promise<Run | undefined> {\n return (await this.load()).getRun(runId)\n }\n async listRuns(filter?: RunFilter): Promise<Run[]> {\n return (await this.load()).listRuns(filter)\n }\n async spans(filter?: SpanFilter): Promise<Span[]> {\n return (await this.load()).spans(filter)\n }\n async events(filter?: EventFilter): Promise<TraceEvent[]> {\n return (await this.load()).events(filter)\n }\n async budget(runId: string): Promise<BudgetLedgerEntry[]> {\n return (await this.load()).budget(runId)\n }\n async artifacts(runId: string): Promise<Artifact[]> {\n return (await this.load()).artifacts(runId)\n }\n}\n","/**\n * Run-completion integrity check — at end of run, verify the expected event\n * types were actually captured. The point is the launch-review failure mode:\n * a run *appears* successful but the raw provider events were never written,\n * so a downstream reviewer can't reconstruct what happened.\n *\n * Pattern:\n *\n * const report = await assertRunCaptured(store, runId, {\n * llmSpansMin: 1,\n * judgeSpansMin: 1,\n * rawSink: providerSink, // must have ≥ 1 event for this run\n * requireRawCoverageOfLlmSpans: true, // every llm span has matching raw events\n * })\n * if (!report.ok) throwIfRunIncomplete(report) // or mark run failed and continue\n *\n * The function is read-only on the store and returns a structured report;\n * the caller chooses the failure mode (throw, mark run failed, log warning).\n * `throwIfRunIncomplete` is the convenient strict mode.\n */\n\nimport { CaptureIntegrityError } from '../errors'\nimport type { RawProviderSink } from './raw-provider-sink'\nimport type { TraceStore } from './store'\n\nexport interface RunIntegrityExpectations {\n /** Minimum LLM span count. Default 0 (no requirement). */\n llmSpansMin?: number\n /** Minimum judge span count. Default 0. */\n judgeSpansMin?: number\n /** Minimum tool span count. Default 0. */\n toolSpansMin?: number\n /**\n * Raw provider sink to consult for capture verification. When present,\n * the check requires at least one raw event for the run.\n */\n rawSink?: RawProviderSink\n /** Minimum raw provider event count. Default 0; ignored when `rawSink` absent. */\n rawProviderEventsMin?: number\n /**\n * Every LLM span must have at least one matching raw `request` event\n * (matched by spanId). Catches the common bug where the structured span\n * was emitted but the raw HTTP capture was wired to a different sink.\n */\n requireRawCoverageOfLlmSpans?: boolean\n /** Run outcome must be set (not null/undefined). Default false. */\n requireOutcome?: boolean\n}\n\nexport type RunIntegrityIssueCode =\n | 'no_run'\n | 'missing_llm_spans'\n | 'missing_judge_spans'\n | 'missing_tool_spans'\n | 'missing_raw_events'\n | 'no_raw_sink'\n | 'orphan_llm_span'\n | 'missing_outcome'\n\nexport interface RunIntegrityIssue {\n code: RunIntegrityIssueCode\n message: string\n detail?: Record<string, unknown>\n}\n\nexport interface RunIntegrityReport {\n ok: boolean\n runId: string\n llmSpanCount: number\n judgeSpanCount: number\n toolSpanCount: number\n rawProviderEventCount: number\n /**\n * Coverage of LLM spans by raw provider events keyed on spanId.\n * `total` is the number of LLM spans; `covered` is the count with at\n * least one matching `request` raw event.\n */\n rawSpanCoverage: { covered: number; total: number }\n issues: RunIntegrityIssue[]\n}\n\nexport class RunIntegrityError extends CaptureIntegrityError {\n constructor(public readonly report: RunIntegrityReport) {\n super(\n `Run ${report.runId} failed integrity check: ${report.issues.map((i) => i.code).join(', ')}`,\n )\n }\n}\n\nexport async function assertRunCaptured(\n store: TraceStore,\n runId: string,\n expectations: RunIntegrityExpectations = {},\n): Promise<RunIntegrityReport> {\n const issues: RunIntegrityIssue[] = []\n const run = await store.getRun(runId)\n if (!run) {\n return {\n ok: false,\n runId,\n llmSpanCount: 0,\n judgeSpanCount: 0,\n toolSpanCount: 0,\n rawProviderEventCount: 0,\n rawSpanCoverage: { covered: 0, total: 0 },\n issues: [{ code: 'no_run', message: `Run ${runId} not found in store.` }],\n }\n }\n\n const spans = await store.spans({ runId })\n const llmSpans = spans.filter((s) => s.kind === 'llm')\n const judgeSpans = spans.filter((s) => s.kind === 'judge')\n const toolSpans = spans.filter((s) => s.kind === 'tool')\n\n const llmMin = expectations.llmSpansMin ?? 0\n const judgeMin = expectations.judgeSpansMin ?? 0\n const toolMin = expectations.toolSpansMin ?? 0\n\n if (llmSpans.length < llmMin) {\n issues.push({\n code: 'missing_llm_spans',\n message: `Expected ≥ ${llmMin} LLM spans, found ${llmSpans.length}.`,\n detail: { expected: llmMin, found: llmSpans.length },\n })\n }\n if (judgeSpans.length < judgeMin) {\n issues.push({\n code: 'missing_judge_spans',\n message: `Expected ≥ ${judgeMin} judge spans, found ${judgeSpans.length}.`,\n detail: { expected: judgeMin, found: judgeSpans.length },\n })\n }\n if (toolSpans.length < toolMin) {\n issues.push({\n code: 'missing_tool_spans',\n message: `Expected ≥ ${toolMin} tool spans, found ${toolSpans.length}.`,\n detail: { expected: toolMin, found: toolSpans.length },\n })\n }\n\n let rawEventCount = 0\n let coverage = { covered: 0, total: llmSpans.length }\n\n if (expectations.rawSink) {\n if (!expectations.rawSink.list) {\n issues.push({\n code: 'no_raw_sink',\n message: 'Provided rawSink does not implement list(); cannot verify capture.',\n })\n } else {\n const events = await expectations.rawSink.list({ runId })\n rawEventCount = events.length\n const rawMin = expectations.rawProviderEventsMin ?? 1\n if (rawEventCount < rawMin) {\n issues.push({\n code: 'missing_raw_events',\n message: `Expected ≥ ${rawMin} raw provider events, found ${rawEventCount}.`,\n detail: { expected: rawMin, found: rawEventCount },\n })\n }\n if (expectations.requireRawCoverageOfLlmSpans) {\n const requestEventsBySpan = new Set(\n events.filter((e) => e.direction === 'request' && e.spanId).map((e) => e.spanId!),\n )\n const orphaned = llmSpans.filter((s) => !requestEventsBySpan.has(s.spanId))\n coverage = { covered: llmSpans.length - orphaned.length, total: llmSpans.length }\n if (orphaned.length > 0) {\n issues.push({\n code: 'orphan_llm_span',\n message: `${orphaned.length} LLM span(s) have no matching raw provider request event.`,\n detail: { orphanedSpanIds: orphaned.map((s) => s.spanId) },\n })\n }\n }\n }\n } else if (expectations.requireRawCoverageOfLlmSpans || expectations.rawProviderEventsMin) {\n issues.push({\n code: 'no_raw_sink',\n message: 'Raw coverage required but no rawSink supplied to the integrity check.',\n })\n }\n\n if (expectations.requireOutcome && (run.outcome === undefined || run.outcome === null)) {\n issues.push({\n code: 'missing_outcome',\n message: `Run ${runId} has no outcome recorded.`,\n })\n }\n\n return {\n ok: issues.length === 0,\n runId,\n llmSpanCount: llmSpans.length,\n judgeSpanCount: judgeSpans.length,\n toolSpanCount: toolSpans.length,\n rawProviderEventCount: rawEventCount,\n rawSpanCoverage: coverage,\n issues,\n }\n}\n\n/** Strict mode: throws `RunIntegrityError` when the report isn't ok. */\nexport function throwIfRunIncomplete(report: RunIntegrityReport): void {\n if (!report.ok) throw new RunIntegrityError(report)\n}\n"],"mappings":";;AA0EA,IAAa,qBAAb,MAAsD;CACpD,uBAAe,IAAI,IAAiB;CACpC,WAA2B,CAAC;CAC5B,YAAkC,CAAC;CACnC,eAAmC,CAAC;CACpC,YAAyC,CAAC;CAE1C,MAAM,UAAU,KAAyB;EACvC,IAAI,KAAK,KAAK,IAAI,IAAI,KAAK,GAAG,MAAM,IAAI,MAAM,OAAO,IAAI,MAAM,gBAAgB;EAC/E,KAAK,KAAK,IAAI,IAAI,OAAO,EAAE,GAAG,IAAI,CAAC;CACrC;CAEA,MAAM,UAAU,OAAe,OAAoC;EACjE,MAAM,WAAW,KAAK,KAAK,IAAI,KAAK;EACpC,IAAI,CAAC,UAAU,MAAM,IAAI,MAAM,OAAO,MAAM,WAAW;EACvD,KAAK,KAAK,IAAI,OAAO;GAAE,GAAG;GAAU,GAAG;EAAM,CAAC;CAChD;CAEA,MAAM,WAAW,MAA2B;EAC1C,KAAK,SAAS,KAAK,EAAE,GAAG,KAAK,CAAC;CAChC;CAEA,MAAM,WAAW,QAAgB,OAAqC;EACpE,MAAM,MAAM,KAAK,SAAS,WAAW,MAAM,EAAE,WAAW,MAAM;EAC9D,IAAI,MAAM,GAAG,MAAM,IAAI,MAAM,QAAQ,OAAO,WAAW;EACvD,KAAK,SAAS,OAAO;GAAE,GAAG,KAAK,SAAS;GAAM,GAAG;EAAM;CACzD;CAEA,MAAM,YAAY,OAAkC;EAClD,KAAK,UAAU,KAAK,EAAE,GAAG,MAAM,CAAC;CAClC;CAEA,MAAM,eAAe,UAAmC;EACtD,KAAK,aAAa,KAAK,EAAE,GAAG,SAAS,CAAC;CACxC;CAEA,MAAM,kBAAkB,OAAyC;EAC/D,KAAK,UAAU,KAAK,EAAE,GAAG,MAAM,CAAC;CAClC;CAEA,MAAM,OAAO,OAAyC;EACpD,MAAM,IAAI,KAAK,KAAK,IAAI,KAAK;EAC7B,OAAO,IAAI,EAAE,GAAG,EAAE,IAAI,KAAA;CACxB;CAEA,MAAM,SAAS,SAAoB,CAAC,GAAmB;EACrD,OAAO,CAAC,GAAG,KAAK,KAAK,OAAO,CAAC,CAAC,CAAC,QAAQ,MAAM,WAAW,GAAG,MAAM,CAAC;CACpE;CAEA,MAAM,MAAM,SAAqB,CAAC,GAAoB;EACpD,OAAO,KAAK,SAAS,QAAQ,MAAM,YAAY,GAAG,MAAM,CAAC,CAAC,CAAC,KAAK,OAAO,EAAE,GAAG,EAAE,EAAE;CAClF;CAEA,MAAM,OAAO,SAAsB,CAAC,GAA0B;EAC5D,OAAO,KAAK,UAAU,QAAQ,MAAM,aAAa,GAAG,MAAM,CAAC,CAAC,CAAC,KAAK,OAAO,EAAE,GAAG,EAAE,EAAE;CACpF;CAEA,MAAM,OAAO,OAA6C;EACxD,OAAO,KAAK,UAAU,QAAQ,MAAM,EAAE,UAAU,KAAK,CAAC,CAAC,KAAK,OAAO,EAAE,GAAG,EAAE,EAAE;CAC9E;CAEA,MAAM,UAAU,OAAoC;EAClD,OAAO,KAAK,aAAa,QAAQ,MAAM,EAAE,UAAU,KAAK,CAAC,CAAC,KAAK,OAAO,EAAE,GAAG,EAAE,EAAE;CACjF;AACF;AAEA,SAAS,WAAW,GAAQ,GAAuB;CACjD,IAAI,EAAE,cAAc,EAAE,eAAe,EAAE,YAAY,OAAO;CAC1D,IAAI,EAAE,aAAa,EAAE,cAAc,EAAE,WAAW,OAAO;CACvD,IAAI,EAAE,UAAU,EAAE,WAAW,EAAE,QAAQ,OAAO;CAC9C,IAAI,EAAE,UAAU,KAAA,KAAa,EAAE,YAAY,EAAE,OAAO,OAAO;CAC3D,IAAI,EAAE,UAAU,KAAA,KAAa,EAAE,YAAY,EAAE,OAAO,OAAO;CAC3D,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,IAAI,SAAS,EAAE,IAAI,OAAO,OAAO;CACzD,IAAI,EAAE,eAAe,EAAE,gBAAgB,EAAE,aAAa,OAAO;CAC7D,IAAI,EAAE,aAAa,EAAE,cAAc,EAAE,WAAW,OAAO;CACvD,IAAI,EAAE,UAAU,EAAE,WAAW,EAAE,QAAQ,OAAO;CAC9C,IAAI,EAAE,SAAS,EAAE,UAAU,EAAE,OAAO,OAAO;CAC3C,OAAO;AACT;AAEA,SAAS,YAAY,GAAS,GAAwB;CACpD,IAAI,EAAE,SAAS,EAAE,UAAU,EAAE,OAAO,OAAO;CAC3C,IAAI,EAAE,gBAAgB,EAAE,iBAAiB,EAAE,cAAc,OAAO;CAChE,IAAI,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,OAAO;CACxC,IAAI,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,OAAO;CACxC,IAAI,EAAE,aAAa,EAAE,SAAS,UAAU,EAAE,aAAa,EAAE,WAAW,OAAO;CAC3E,IAAI,EAAE,YAAY,EAAE,SAAS,WAAW,EAAE,YAAY,EAAE,UAAU,OAAO;CACzE,IAAI,EAAE,UAAU,KAAA,KAAa,EAAE,YAAY,EAAE,OAAO,OAAO;CAC3D,IAAI,EAAE,UAAU,KAAA,KAAa,EAAE,YAAY,EAAE,OAAO,OAAO;CAC3D,OAAO;AACT;AAEA,SAAS,aAAa,GAAe,GAAyB;CAC5D,IAAI,EAAE,SAAS,EAAE,UAAU,EAAE,OAAO,OAAO;CAC3C,IAAI,EAAE,UAAU,EAAE,WAAW,EAAE,QAAQ,OAAO;CAC9C,IAAI,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,OAAO;CACxC,IAAI,EAAE,UAAU,KAAA,KAAa,EAAE,YAAY,EAAE,OAAO,OAAO;CAC3D,IAAI,EAAE,UAAU,KAAA,KAAa,EAAE,YAAY,EAAE,OAAO,OAAO;CAC3D,OAAO;AACT;AAUA,IAAa,uBAAb,MAAwD;CACtD;CACA;;CAEA;;;CAGA;;;CAGA,oBAA4B;;;;;;;;CAQ5B,8BAAsB,IAAI,IAA2B;CAErD,YAAY,SAAsC;EAChD,KAAK,MAAM,QAAQ;EACnB,KAAK,WAAW,QAAQ,YAAY,KAAK,OAAO;CAClD;CAEA,MAAc,YAA2B;EAEvC,OAAM,MADW,OAAO,oBAAA,CACf,MAAM,KAAK,KAAK,EAAE,WAAW,KAAK,CAAC;CAC9C;CAEA,MAAc,OAAO,MAAc,QAAgC;EAIjE,IAAI,KAAK,cAAc,MAAM,KAAK;EAKlC,MAAM,UADQ,KAAK,YAAY,IAAI,IAAI,KAAK,QAAQ,QAAQ,EAAA,CACvC,WAAW,KAAK,aAAa,MAAM,MAAM,CAAC;EAC/D,MAAM,OAAO,OAAO,WACZ,CAAC,SACD,CAAC,CACT;EACA,KAAK,YAAY,IAAI,MAAM,IAAI;EAC/B,IAAI;GACF,MAAM;EACR,UAAU;GAGR,IAAI,KAAK,YAAY,IAAI,IAAI,MAAM,MAAM,KAAK,YAAY,OAAO,IAAI;EACvE;CACF;CAEA,MAAc,aAAa,MAAc,QAAgC;EACvE,MAAM,KAAK,UAAU;EACrB,MAAM,KAAK,MAAM,OAAO;EACxB,MAAM,OAAO,MAAM,OAAO;EAC1B,MAAM,SAAS,KAAK,KAAK,KAAK,KAAK,GAAG,KAAK,QAAQ;EACnD,IAAI;GAEF,KAAI,MADe,GAAG,KAAK,MAAM,EAAA,CACxB,QAAQ,KAAK,UAAU;IAG9B,MAAM,QAAQ,KAAK,IAAI,KAAK,IAAI,GAAG,KAAK,oBAAoB,CAAC;IAC7D,KAAK,oBAAoB;IACzB,MAAM,SAAS,KAAK,KAAK,KAAK,KAAK,GAAG,KAAK,GAAG,MAAM,QAAQ;IAC5D,MAAM,GAAG,OAAO,QAAQ,MAAM;GAChC;EACF,QAAQ,CAER;EACA,MAAM,GAAG,WAAW,QAAQ,GAAG,KAAK,UAAU,MAAM,EAAE,KAAK,MAAM;EAOjE,IAAI,KAAK,SAAS,CAAE,QAAkC,SACpD,MAAM,KAAK,WAAW,MAAM,MAAM;CAEtC;CAEA,MAAc,WAAW,MAAc,QAAgC;EACrE,IAAI,CAAC,KAAK,OAAO;EACjB,QAAQ,MAAR;GACE,KAAK;IACH,MAAM,KAAK,MAAM,UAAU,MAAa;IACxC;GACF,KAAK;IACH,MAAM,KAAK,MAAM,WAAW,MAAc;IAC1C;GACF,KAAK;IACH,MAAM,KAAK,MAAM,YAAY,MAAoB;IACjD;GACF,KAAK;IACH,MAAM,KAAK,MAAM,eAAe,MAAkB;IAClD;GACF,KAAK;IACH,MAAM,KAAK,MAAM,kBAAkB,MAA2B;IAC9D;EACJ;CACF;CAEA,OAA4C;EAG1C,KAAK,iBAAiB,KAAK,WAAW;EACtC,OAAO,KAAK;CACd;CAEA,MAAc,aAA0C;EACtD,MAAM,KAAK,MAAM,OAAO;EACxB,MAAM,OAAO,MAAM,OAAO;EAC1B,MAAM,QAAQ,IAAI,mBAAmB;EACrC,IAAI;EACJ,IAAI;GACF,UAAU,MAAM,GAAG,QAAQ,KAAK,GAAG;EACrC,QAAQ;GAGN,KAAK,QAAQ;GACb,OAAO;EACT;EAOA,QAAQ,MAAM,GAAG,MAAM;GAMrB,MAAM,QAAQ,EAAE,MAAM,GAAG,CAAC,CAAC;GAC3B,MAAM,QAAQ,EAAE,MAAM,GAAG,CAAC,CAAC;GAC3B,IAAI,UAAU,OAAO,OAAO,QAAQ,QAAQ,KAAK;GACjD,MAAM,KAAK,EAAE,MAAM,wBAAwB;GAC3C,MAAM,KAAK,EAAE,MAAM,wBAAwB;GAG3C,QAFY,KAAK,OAAO,GAAG,EAAE,IAAI,OAAO,sBAC5B,KAAK,OAAO,GAAG,EAAE,IAAI,OAAO;EAE1C,CAAC;EACD;GAQE,MAAM,aAA6C,CAAC;GACpD,MAAM,cAA8C,CAAC;GACrD,KAAK,MAAM,QAAQ,SAAS;IAC1B,IAAI,CAAC,KAAK,SAAS,SAAS,GAAG;IAC/B,MAAM,OAAO,KAAK,KAAK,KAAK,KAAK,IAAI;IACrC,MAAM,UAAU,MAAM,GAAG,SAAS,MAAM,MAAM;IAC9C,MAAM,OAAO,KAAK,MAAM,GAAG,CAAC,CAAC;IAC7B,MAAM,QAAQ,QAAQ,MAAM,IAAI;IAChC,KAAK,IAAI,KAAK,GAAG,KAAK,MAAM,QAAQ,MAAM;KACxC,MAAM,OAAO,MAAM;KACnB,IAAI,CAAC,KAAK,KAAK,GAAG;KAClB,IAAI;KACJ,IAAI;MACF,SAAS,KAAK,MAAM,IAAI;KAC1B,SAAS,KAAK;MAGZ,MAAM,IAAI,MACR,2CAA2C,KAAK,QAAQ,KAAK,EAAE,IAC7D,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,EAChD,KAAK,KAAK,MAAM,GAAG,GAAG,GACzB;KACF;KACA,IAAI,SAAS,QAAQ;MACnB,IAAI,QAAQ,SAAS;OACnB,WAAW,KAAK,MAAM;OACtB;MACF;MAGA,IAAI;OACF,MAAM,MAAM,UAAU,MAAM;MAC9B,QAAQ;OACN,MAAM,MAAM,UAAU,OAAO,OAAO,MAAM;MAC5C;KACF,OAAO,IAAI,SAAS,SAAS;MAQ3B,IAAI,QAAQ,SAAS;OACnB,YAAY,KAAK,MAAM;OACvB;MACF;MACA,MAAM,MAAM,WAAW,MAAM;KAC/B,OAAO,IAAI,SAAS,UAClB,MAAM,MAAM,YAAY,MAAM;UACzB,IAAI,SAAS,aAClB,MAAM,MAAM,eAAe,MAAM;UAC5B,IAAI,SAAS,UAClB,MAAM,MAAM,kBAAkB,MAAM;IAExC;GACF;GAKA,KAAK,MAAM,UAAU,YACnB,MAAM,MAAM,UAAU,OAAO,OAAiB,MAAsB;GAEtE,KAAK,MAAM,UAAU,aACnB,MAAM,MAAM,WAAW,OAAO,QAAkB,MAAuB;EAE3E;EACA,KAAK,QAAQ;EACb,OAAO;CACT;CAEA,MAAM,UAAU,KAAyB;EACvC,MAAM,KAAK,OAAO,QAAQ,GAAG;CAC/B;CACA,MAAM,UAAU,OAAe,OAAoC;EAGjE,MAAM,KAAK,OAAO,QAAQ;GAAE;GAAO,GAAG;GAAO,SAAS;EAAK,CAAC;EAC5D,IAAI,KAAK,OAAO,MAAM,KAAK,MAAM,UAAU,OAAO,KAAK;CACzD;CACA,MAAM,WAAW,MAA2B;EAC1C,MAAM,KAAK,OAAO,SAAS,IAAI;CACjC;CACA,MAAM,WAAW,QAAgB,OAAqC;EACpE,MAAM,KAAK,OAAO,SAAS;GAAE;GAAQ,GAAG;GAAO,SAAS;EAAK,CAAC;EAC9D,IAAI,KAAK,OAAO,MAAM,KAAK,MAAM,WAAW,QAAQ,KAAK;CAC3D;CACA,MAAM,YAAY,OAAkC;EAClD,MAAM,KAAK,OAAO,UAAU,KAAK;CACnC;CACA,MAAM,eAAe,UAAmC;EACtD,MAAM,KAAK,OAAO,aAAa,QAAQ;CACzC;CACA,MAAM,kBAAkB,OAAyC;EAC/D,MAAM,KAAK,OAAO,UAAU,KAAK;CACnC;CAEA,MAAM,OAAO,OAAyC;EACpD,QAAQ,MAAM,KAAK,KAAK,EAAA,CAAG,OAAO,KAAK;CACzC;CACA,MAAM,SAAS,QAAoC;EACjD,QAAQ,MAAM,KAAK,KAAK,EAAA,CAAG,SAAS,MAAM;CAC5C;CACA,MAAM,MAAM,QAAsC;EAChD,QAAQ,MAAM,KAAK,KAAK,EAAA,CAAG,MAAM,MAAM;CACzC;CACA,MAAM,OAAO,QAA6C;EACxD,QAAQ,MAAM,KAAK,KAAK,EAAA,CAAG,OAAO,MAAM;CAC1C;CACA,MAAM,OAAO,OAA6C;EACxD,QAAQ,MAAM,KAAK,KAAK,EAAA,CAAG,OAAO,KAAK;CACzC;CACA,MAAM,UAAU,OAAoC;EAClD,QAAQ,MAAM,KAAK,KAAK,EAAA,CAAG,UAAU,KAAK;CAC5C;AACF;;;;;;;;;;;;;;;;;;;;;;;ACpXA,IAAa,oBAAb,cAAuC,sBAAsB;CAC/B;CAA5B,YAAY,QAA4C;EACtD,MACE,OAAO,OAAO,MAAM,2BAA2B,OAAO,OAAO,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,KAAK,IAAI,GAC3F;EAH0B,KAAA,SAAA;CAI5B;AACF;AAEA,eAAsB,kBACpB,OACA,OACA,eAAyC,CAAC,GACb;CAC7B,MAAM,SAA8B,CAAC;CACrC,MAAM,MAAM,MAAM,MAAM,OAAO,KAAK;CACpC,IAAI,CAAC,KACH,OAAO;EACL,IAAI;EACJ;EACA,cAAc;EACd,gBAAgB;EAChB,eAAe;EACf,uBAAuB;EACvB,iBAAiB;GAAE,SAAS;GAAG,OAAO;EAAE;EACxC,QAAQ,CAAC;GAAE,MAAM;GAAU,SAAS,OAAO,MAAM;EAAsB,CAAC;CAC1E;CAGF,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,MAAM,CAAC;CACzC,MAAM,WAAW,MAAM,QAAQ,MAAM,EAAE,SAAS,KAAK;CACrD,MAAM,aAAa,MAAM,QAAQ,MAAM,EAAE,SAAS,OAAO;CACzD,MAAM,YAAY,MAAM,QAAQ,MAAM,EAAE,SAAS,MAAM;CAEvD,MAAM,SAAS,aAAa,eAAe;CAC3C,MAAM,WAAW,aAAa,iBAAiB;CAC/C,MAAM,UAAU,aAAa,gBAAgB;CAE7C,IAAI,SAAS,SAAS,QACpB,OAAO,KAAK;EACV,MAAM;EACN,SAAS,cAAc,OAAO,oBAAoB,SAAS,OAAO;EAClE,QAAQ;GAAE,UAAU;GAAQ,OAAO,SAAS;EAAO;CACrD,CAAC;CAEH,IAAI,WAAW,SAAS,UACtB,OAAO,KAAK;EACV,MAAM;EACN,SAAS,cAAc,SAAS,sBAAsB,WAAW,OAAO;EACxE,QAAQ;GAAE,UAAU;GAAU,OAAO,WAAW;EAAO;CACzD,CAAC;CAEH,IAAI,UAAU,SAAS,SACrB,OAAO,KAAK;EACV,MAAM;EACN,SAAS,cAAc,QAAQ,qBAAqB,UAAU,OAAO;EACrE,QAAQ;GAAE,UAAU;GAAS,OAAO,UAAU;EAAO;CACvD,CAAC;CAGH,IAAI,gBAAgB;CACpB,IAAI,WAAW;EAAE,SAAS;EAAG,OAAO,SAAS;CAAO;CAEpD,IAAI,aAAa,SACf,IAAI,CAAC,aAAa,QAAQ,MACxB,OAAO,KAAK;EACV,MAAM;EACN,SAAS;CACX,CAAC;MACI;EACL,MAAM,SAAS,MAAM,aAAa,QAAQ,KAAK,EAAE,MAAM,CAAC;EACxD,gBAAgB,OAAO;EACvB,MAAM,SAAS,aAAa,wBAAwB;EACpD,IAAI,gBAAgB,QAClB,OAAO,KAAK;GACV,MAAM;GACN,SAAS,cAAc,OAAO,8BAA8B,cAAc;GAC1E,QAAQ;IAAE,UAAU;IAAQ,OAAO;GAAc;EACnD,CAAC;EAEH,IAAI,aAAa,8BAA8B;GAC7C,MAAM,sBAAsB,IAAI,IAC9B,OAAO,QAAQ,MAAM,EAAE,cAAc,aAAa,EAAE,MAAM,CAAC,CAAC,KAAK,MAAM,EAAE,MAAO,CAClF;GACA,MAAM,WAAW,SAAS,QAAQ,MAAM,CAAC,oBAAoB,IAAI,EAAE,MAAM,CAAC;GAC1E,WAAW;IAAE,SAAS,SAAS,SAAS,SAAS;IAAQ,OAAO,SAAS;GAAO;GAChF,IAAI,SAAS,SAAS,GACpB,OAAO,KAAK;IACV,MAAM;IACN,SAAS,GAAG,SAAS,OAAO;IAC5B,QAAQ,EAAE,iBAAiB,SAAS,KAAK,MAAM,EAAE,MAAM,EAAE;GAC3D,CAAC;EAEL;CACF;MACK,IAAI,aAAa,gCAAgC,aAAa,sBACnE,OAAO,KAAK;EACV,MAAM;EACN,SAAS;CACX,CAAC;CAGH,IAAI,aAAa,mBAAmB,IAAI,YAAY,KAAA,KAAa,IAAI,YAAY,OAC/E,OAAO,KAAK;EACV,MAAM;EACN,SAAS,OAAO,MAAM;CACxB,CAAC;CAGH,OAAO;EACL,IAAI,OAAO,WAAW;EACtB;EACA,cAAc,SAAS;EACvB,gBAAgB,WAAW;EAC3B,eAAe,UAAU;EACzB,uBAAuB;EACvB,iBAAiB;EACjB;CACF;AACF;;AAGA,SAAgB,qBAAqB,QAAkC;CACrE,IAAI,CAAC,OAAO,IAAI,MAAM,IAAI,kBAAkB,MAAM;AACpD"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"llm-client-Dv5BiKLE.js","names":[],"sources":["../src/judge-families.ts","../src/integrity/served-model.ts","../src/llm-client.ts"],"sourcesContent":["/**\n * Judge model-family classification + cross-family enforcement.\n *\n * A judge ensemble built entirely from one provider family shares that\n * family's blind spots and self-preference — its \"agreement\" is correlated\n * bias, not independent signal. `assertCrossFamily` makes the consumer prove\n * the ensemble spans ≥2 families; `judgeFamily` is the single regex map that\n * replaces the per-consumer copies (tax/legal/creative/gtm each ship one).\n */\n\n/** Provider family a model belongs to. `unknown` when no rule matches. */\nexport type JudgeFamily =\n | 'anthropic'\n | 'openai'\n | 'google'\n | 'meta'\n | 'mistral'\n | 'deepseek'\n | 'xai'\n | 'qwen'\n | 'cohere'\n | 'amazon'\n | 'moonshot'\n | 'zhipu'\n | 'unknown'\n\n/** Explicit `provider/...` prefix → family (models.dev / OpenRouter style). */\nconst PROVIDER_PREFIX: Record<string, JudgeFamily> = {\n anthropic: 'anthropic',\n openai: 'openai',\n 'azure-openai': 'openai',\n google: 'google',\n 'google-vertex': 'google',\n meta: 'meta',\n 'meta-llama': 'meta',\n mistral: 'mistral',\n mistralai: 'mistral',\n deepseek: 'deepseek',\n xai: 'xai',\n qwen: 'qwen',\n alibaba: 'qwen',\n cohere: 'cohere',\n amazon: 'amazon',\n bedrock: 'amazon',\n moonshot: 'moonshot',\n moonshotai: 'moonshot',\n kimi: 'moonshot',\n 'kimi-code': 'moonshot',\n zhipu: 'zhipu',\n zhipuai: 'zhipu',\n zai: 'zhipu',\n 'z-ai': 'zhipu',\n glm: 'zhipu',\n}\n\n/** Fallback model-name patterns when there's no recognised provider prefix. */\nconst NAME_PATTERNS: Array<[RegExp, JudgeFamily]> = [\n [/claude/i, 'anthropic'],\n [/\\b(gpt|davinci|babbage)\\b|^o[134]\\b|[-/]o[134]\\b|gpt-/i, 'openai'],\n [/gemini|palm|gemma|bison/i, 'google'],\n [/llama/i, 'meta'],\n [/mi(s|x)tral|codestral|magistral/i, 'mistral'],\n [/deepseek/i, 'deepseek'],\n [/grok/i, 'xai'],\n [/qwen/i, 'qwen'],\n [/command-?(r|a)?/i, 'cohere'],\n [/\\b(nova|titan)\\b/i, 'amazon'],\n [/\\bkimi\\b|moonshot/i, 'moonshot'],\n [/\\bglm\\b|zhipu|\\bz-?ai\\b/i, 'zhipu'],\n]\n\n/**\n * Classify a model id into its provider family. Strips a `@snapshot` suffix\n * and prefers an explicit `provider/...` prefix; otherwise matches the model\n * name. Returns `unknown` when nothing matches (callers decide whether that's\n * acceptable — `assertCrossFamily` counts it as its own family).\n */\nexport function judgeFamily(modelId: string): JudgeFamily {\n const id = modelId.trim().split('@')[0]!.toLowerCase()\n const slash = id.indexOf('/')\n if (slash > 0) {\n const prefix = id.slice(0, slash)\n const mapped = PROVIDER_PREFIX[prefix]\n if (mapped) return mapped\n }\n for (const [pattern, family] of NAME_PATTERNS) {\n if (pattern.test(id)) return family\n }\n return 'unknown'\n}\n\nexport interface AssertCrossFamilyOptions {\n /** Minimum number of distinct families the ensemble must span. Default 2. */\n minFamilies?: number\n /** When false (default), `unknown`-family models do NOT count toward the\n * family total — an ensemble of all-unclassifiable models is not provably\n * cross-family. Set true to count `unknown` as one shared family. */\n allowUnknown?: boolean\n}\n\nexport class CrossFamilyError extends Error {\n constructor(\n message: string,\n public readonly families: JudgeFamily[],\n public readonly models: string[],\n ) {\n super(message)\n this.name = 'CrossFamilyError'\n }\n}\n\n/**\n * Throw unless the judge models span at least `minFamilies` distinct provider\n * families. Pass the model ids backing your judge ensemble. Fail-loud by\n * design — a correlated single-family ensemble silently inflates agreement.\n *\n * Scope: this reads the ids you REQUEST. It proves the panel was configured\n * across families; it cannot prove the panel RAN across families, because a\n * routing gateway may answer several different ids from one provider. Where\n * the diversity claim is load-bearing (a published leaderboard, a\n * certification, a non-self-judging exclusion), assert on the ids the\n * provider echoed instead: `assertCrossFamilyServed` in\n * ./integrity/served-model.\n */\nexport function assertCrossFamily(\n models: string[],\n opts: AssertCrossFamilyOptions = {},\n): JudgeFamily[] {\n const minFamilies = opts.minFamilies ?? 2\n const families = new Set<JudgeFamily>()\n for (const m of models) {\n const f = judgeFamily(m)\n if (f === 'unknown' && !opts.allowUnknown) continue\n families.add(f)\n }\n const list = [...families].sort()\n if (list.length < minFamilies) {\n throw new CrossFamilyError(\n `judge ensemble spans ${list.length} provider famil${list.length === 1 ? 'y' : 'ies'} ` +\n `(${list.join(', ') || 'none'}) but ${minFamilies} required — a single-family ensemble ` +\n 'is correlated bias, not independent signal',\n list,\n models,\n )\n }\n return list\n}\n","/**\n * Served-model identity: prove the model that ANSWERED is the model that was\n * REQUESTED.\n *\n * A routing gateway can accept `model: \"gpt-4.1-mini\"` and answer from a\n * different model entirely. Every guard that reasons about the requested id —\n * cross-family judge diversity, per-model leaderboard rows, non-self-judging\n * exclusions, cost attribution — then passes while measuring something else.\n * The requested id is an intent; only the id echoed on the response is\n * evidence.\n *\n * `checkServedModel` classifies one requested/served pair; `assertServedModel`\n * and `assertServedModels` are the fail-loud gates; `assertCrossFamilyServed`\n * is the family-diversity rule computed over SERVED ids (the requested-id\n * version lives in ../judge-families and cannot see substitution).\n *\n * Aliasing is tolerated, substitution is not: `openai/gpt-4.1-mini` and\n * `gpt-4.1-mini` are the same request expressed two ways, so a response\n * echoing either satisfies the other. A response echoing `gemini-2.5-flash-lite`\n * does not.\n */\n\nimport { AgentEvalError } from '../errors'\nimport { type JudgeFamily, judgeFamily } from '../judge-families'\n\n/** How a served id relates to the id that was requested. */\nexport type ServedModelVerdict =\n /** Byte-identical after normalisation — the requested model answered. */\n | 'exact'\n /** Same model, different spelling (provider prefix, snapshot, tier suffix). */\n | 'alias'\n /** A different model of the SAME provider family answered. */\n | 'substituted-within-family'\n /** A different provider's model answered. */\n | 'substituted-cross-family'\n /** The response carried no model id — identity is unproven either way. */\n | 'unreported'\n\nexport interface ServedModelCheck {\n /** The id the caller asked for. */\n requested: string\n /** The id echoed on the response; `null` when the response omitted it. */\n served: string | null\n requestedFamily: JudgeFamily\n /** `null` when `served` is null. */\n servedFamily: JudgeFamily | null\n verdict: ServedModelVerdict\n /** True for every verdict except `exact` and `alias`. */\n substituted: boolean\n}\n\n/**\n * Reduce a model id to its comparable core: lowercase, no surrounding space,\n * no `provider/` prefix, no `@snapshot` / `:batch` / `:free` tier suffix, no\n * trailing build date, and `.`/`_` folded to `-` so one version is spelled one\n * way.\n *\n * Dropping the build date is what makes snapshot resolution legible as the\n * non-event it is: a router answering `gpt-4o-mini` with\n * `gpt-4o-mini-2024-07-18` pinned a floating alias to a reproducible build —\n * the same model, which is the behaviour we want. Only routing decoration is\n * stripped; version digits are load-bearing, so `deepseek-v3.2` and\n * `deepseek-v4-flash` stay distinct, and comparison is EXACT equality rather\n * than a prefix test (a prefix rule would accept `gpt-5` → `gpt-5-mini`, a\n * silent downgrade wearing the right vendor name).\n */\nexport function normalizeModelId(modelId: string): string {\n let id = modelId.trim().toLowerCase()\n const at = id.indexOf('@')\n if (at > 0) id = id.slice(0, at)\n const colon = id.indexOf(':')\n if (colon > 0) id = id.slice(0, colon)\n const slash = id.lastIndexOf('/')\n if (slash >= 0) id = id.slice(slash + 1)\n return id\n .replace(/-\\d{4}-\\d{2}-\\d{2}$/, '')\n .replace(/-\\d{8}$/, '')\n .replace(/[._]/g, '-')\n .replace(/-+$/, '')\n .trim()\n}\n\n/**\n * Classify one requested/served pair. Pure — no I/O — so it is safe inside\n * response handlers, reducers, and CI gates.\n *\n * `served` is the id echoed by the provider (OpenAI-compatible bodies put it\n * at `model`). `null`/`undefined` means the body omitted it; that is\n * `unreported`, NOT a pass — a provider that does not name what answered has\n * not proven identity, and a transport that drops the field must not read as\n * agreement.\n */\nexport function checkServedModel(\n requested: string,\n served: string | null | undefined,\n): ServedModelCheck {\n const requestedFamily = judgeFamily(requested)\n if (served === null || served === undefined || served.trim() === '') {\n return {\n requested,\n served: null,\n requestedFamily,\n servedFamily: null,\n verdict: 'unreported',\n substituted: true,\n }\n }\n const servedFamily = judgeFamily(served)\n if (requested.trim().toLowerCase() === served.trim().toLowerCase()) {\n return {\n requested,\n served,\n requestedFamily,\n servedFamily,\n verdict: 'exact',\n substituted: false,\n }\n }\n if (normalizeModelId(requested) === normalizeModelId(served)) {\n return {\n requested,\n served,\n requestedFamily,\n servedFamily,\n verdict: 'alias',\n substituted: false,\n }\n }\n return {\n requested,\n served,\n requestedFamily,\n servedFamily,\n verdict:\n requestedFamily === servedFamily ? 'substituted-within-family' : 'substituted-cross-family',\n substituted: true,\n }\n}\n\nexport class ModelSubstitutionError extends AgentEvalError {\n constructor(\n message: string,\n public readonly checks: ReadonlyArray<ServedModelCheck>,\n ) {\n super('model_substitution', message)\n this.name = 'ModelSubstitutionError'\n }\n}\n\nexport interface AssertServedModelOptions {\n /**\n * Accept a different model of the same provider family (e.g. requested\n * `deepseek-v3.2`, served `deepseek-v4-flash`). Default false. Setting this\n * keeps family-level claims valid and forfeits per-model claims.\n */\n allowWithinFamily?: boolean\n /**\n * Accept a response that carried no model id. Default false — an\n * unidentified response cannot support a per-model or per-family claim.\n */\n allowUnreported?: boolean\n /** Prefixed to the thrown message, e.g. the judge or campaign cell name. */\n context?: string\n}\n\n/**\n * The one place the accept/reject policy lives, so a caller that reports\n * substitution (a preflight table, a run record) and a caller that throws on it\n * can never drift apart. A cross-family substitution is never acceptable.\n */\nexport function servedModelAcceptable(\n check: ServedModelCheck,\n opts: AssertServedModelOptions = {},\n): boolean {\n switch (check.verdict) {\n case 'exact':\n case 'alias':\n return true\n case 'unreported':\n return opts.allowUnreported === true\n case 'substituted-within-family':\n return opts.allowWithinFamily === true\n default:\n return false\n }\n}\n\nfunction describe(check: ServedModelCheck): string {\n if (check.verdict === 'unreported')\n return `${check.requested}: response carried no model id (identity unproven)`\n return (\n `${check.requested} (${check.requestedFamily}) → served ${check.served} ` +\n `(${check.servedFamily}) [${check.verdict}]`\n )\n}\n\n/**\n * Throw `ModelSubstitutionError` unless the served id is the requested model.\n * Returns the check on success so callers can record the served id alongside\n * the result.\n */\nexport function assertServedModel(\n requested: string,\n served: string | null | undefined,\n opts: AssertServedModelOptions = {},\n): ServedModelCheck {\n const check = checkServedModel(requested, served)\n if (servedModelAcceptable(check, opts)) return check\n const prefix = opts.context ? `${opts.context}: ` : ''\n throw new ModelSubstitutionError(\n `${prefix}model substitution — ${describe(check)}. The measurement is of the SERVED model, ` +\n 'not the requested one; any per-model or per-family claim from this call is invalid.',\n [check],\n )\n}\n\n/**\n * Batch form: check every pair and throw naming EVERY substitution, so one\n * failure does not hide the rest. Returns all checks on success.\n */\nexport function assertServedModels(\n pairs: ReadonlyArray<{ requested: string; served: string | null | undefined }>,\n opts: AssertServedModelOptions = {},\n): ServedModelCheck[] {\n const checks = pairs.map((p) => checkServedModel(p.requested, p.served))\n const bad = checks.filter((c) => !servedModelAcceptable(c, opts))\n if (bad.length > 0) {\n const prefix = opts.context ? `${opts.context}: ` : ''\n throw new ModelSubstitutionError(\n `${prefix}${bad.length}/${checks.length} call(s) were answered by a different model than ` +\n `requested — ${bad.map(describe).join('; ')}. Per-model and per-family claims from this ` +\n 'run are invalid until the ids are re-measured.',\n checks,\n )\n }\n return checks\n}\n\nexport interface AssertCrossFamilyServedOptions extends AssertServedModelOptions {\n /** Minimum distinct SERVED families required. Default 2. */\n minFamilies?: number\n /** Count `unknown`-family served ids toward the total. Default false. */\n allowUnknown?: boolean\n}\n\nexport class ServedCrossFamilyError extends AgentEvalError {\n constructor(\n message: string,\n public readonly families: JudgeFamily[],\n public readonly checks: ReadonlyArray<ServedModelCheck>,\n ) {\n super('model_substitution', message)\n this.name = 'ServedCrossFamilyError'\n }\n}\n\n/**\n * Family-diversity rule over the models that actually ANSWERED.\n *\n * `assertCrossFamily` (../judge-families) reads the requested ids and so\n * cannot see a gateway that answers three \"different\" requests from one\n * provider. This one asserts no substitution first, then counts families from\n * the served ids — a panel that collapsed to one family under the hood fails\n * here even though its request list looked diverse.\n */\nexport function assertCrossFamilyServed(\n pairs: ReadonlyArray<{ requested: string; served: string | null | undefined }>,\n opts: AssertCrossFamilyServedOptions = {},\n): JudgeFamily[] {\n const checks = assertServedModels(pairs, opts)\n const families = new Set<JudgeFamily>()\n for (const check of checks) {\n const family = check.servedFamily\n if (family === null) continue\n if (family === 'unknown' && !opts.allowUnknown) continue\n families.add(family)\n }\n const list = [...families].sort()\n const minFamilies = opts.minFamilies ?? 2\n if (list.length < minFamilies) {\n const prefix = opts.context ? `${opts.context}: ` : ''\n throw new ServedCrossFamilyError(\n `${prefix}the models that ANSWERED span ${list.length} provider famil` +\n `${list.length === 1 ? 'y' : 'ies'} (${list.join(', ') || 'none'}) but ${minFamilies} ` +\n `required — served ids: ${checks.map((c) => c.served ?? 'unreported').join(', ')}`,\n list,\n checks,\n )\n }\n return list\n}\n","/**\n * LLM client with graceful degrade.\n *\n * OpenAI-compatible `/v1/chat/completions` client with:\n * - Exponential-backoff retry on 429 + 5xx gateway errors (502/503/504).\n * - Retry on transient network errors (fetch failed, AbortError, ECONNRESET).\n * - One retry at temperature 1 when a model explicitly requires it.\n * - Graceful json_schema → json_object degrade on 400 with schema-reject body.\n * - Fenced-JSON stripping (```json ... ```) for models that wrap structured output.\n * - Configurable base URL + api key / bearer, works with LiteLLM proxies, OpenAI\n * directly, cli-bridge subscriptions, and any router that speaks the spec.\n *\n * Usage:\n * const { value, result } = await callLlmJson<MyType>(\n * { model: 'gpt-4o', messages: [...], jsonSchema: { name: 'x', schema: {...} } },\n * { baseUrl: 'https://router.tangle.tools/v1', apiKey: process.env.KEY },\n * )\n *\n * `createChatClient` wraps this implementation for provider-neutral package\n * entry points. Direct callers can use `callLlm` or `callLlmJson`.\n */\n\nimport {\n type CostReceiptInput,\n type CustomTokenPricing,\n costForTokenPricing,\n type MaximumCharge,\n} from './cost-ledger'\nimport { AgentEvalError, CaptureIntegrityError } from './errors'\nimport {\n type AssertServedModelOptions,\n assertServedModel as assertServedModelIdentity,\n checkServedModel,\n} from './integrity/served-model'\nimport {\n defaultProviderRedactor,\n type ProviderRedactor,\n providerFromBaseUrl,\n type RawProviderEvent,\n type RawProviderSink,\n} from './trace/raw-provider-sink'\n\n// ─── Types ──────────────────────────────────────────────────────────────\n\nexport interface LlmMessage {\n role: 'system' | 'user' | 'assistant'\n /**\n * Either a plain text content string OR a multimodal content array\n * (text + image_url parts) for vision-capable models.\n */\n content:\n | string\n | Array<\n | { type: 'text'; text: string }\n | { type: 'image_url'; image_url: { url: string; detail?: 'auto' | 'low' | 'high' } }\n >\n}\n\nexport type LlmThinkingMode = 'enabled' | 'disabled'\n\nexport interface LlmCallRequest {\n model: string\n messages: LlmMessage[]\n /** Optional JSON-mode response format (response_format: json_object). */\n jsonMode?: boolean\n /** Optional structured output via JSON Schema. Falls back to json_object on 400. */\n jsonSchema?: { name: string; schema: Record<string, unknown> }\n temperature?: number\n maxTokens?: number\n /** OpenAI-compatible reasoning mode. Omitted when the provider default should apply. */\n thinking?: LlmThinkingMode\n /** Per-call timeout, default 300s. */\n timeoutMs?: number\n}\n\n/** Conservative priced bound for the exact text request sent to a provider.\n * Returns undefined when output or multimodal input is not bounded, causing a\n * capped CostLedger to reject the call before execution. Pass\n * `customTokenPricing` when package pricing does not cover the model or endpoint. */\nexport function maximumChargeForLlmRequest(\n request: Pick<LlmCallRequest, 'model' | 'messages' | 'jsonSchema' | 'maxTokens' | 'thinking'>,\n options: LlmClientOptions = {},\n): MaximumCharge | undefined {\n if (request.maxTokens === undefined) return undefined\n if (!Number.isInteger(request.maxTokens) || request.maxTokens <= 0) {\n throw new RangeError(`maximumChargeForLlmRequest: maxTokens must be a positive integer`)\n }\n if (\n request.messages.some(\n (message) =>\n Array.isArray(message.content) && message.content.some((part) => part.type === 'image_url'),\n )\n ) {\n return undefined\n }\n\n const attempts = resolveMaximumAttempts(options.maximumAttempts)\n const forceJsonObject = options.jsonSchemaTransport === 'json-object'\n // A byte-level tokenizer cannot emit more input tokens than request bytes.\n // Pricing the complete body also covers role/schema framing omitted from content-only estimates.\n const requestBytes = new TextEncoder().encode(\n JSON.stringify(buildBody(request, forceJsonObject, options.thinking)),\n ).byteLength\n // A rejected response schema can trigger one JSON-mode batch with the same output limit.\n const batches = request.jsonSchema && !forceJsonObject ? 2 : 1\n const usage = {\n inputTokens: requestBytes * attempts * batches,\n outputTokens: request.maxTokens * attempts * batches,\n }\n return options.customTokenPricing\n ? { customTokenPricing: options.customTokenPricing, ...usage }\n : { model: request.model, ...usage }\n}\n\nexport interface LlmUsage {\n promptTokens: number\n completionTokens: number\n totalTokens: number\n /** False when the provider omitted or malformed prompt/completion usage. */\n captured?: boolean\n /** Reasoning-token subset of completionTokens, when reported. */\n reasoningTokens?: number\n /** Proxies populate this when prompt caching is on. */\n cachedPromptTokens?: number\n}\n\nexport interface LlmCallResult {\n /** The text content of the first choice. Empty string if none. */\n content: string\n usage: LlmUsage\n /**\n * Cost in USD. Uses the provider's reported cost when present, otherwise\n * caller-supplied token pricing. `null` when neither is available.\n */\n costUsd: number | null\n /**\n * Model id used for attribution (cost, pricing, log lines). The response's\n * echoed id when the provider sent one, else the requested id.\n *\n * NOT evidence of which model answered — read `servedModel` for that. A\n * provider that omits `model` makes this equal to the request, which is\n * exactly the case an identity check must be able to distinguish.\n */\n model: string\n /**\n * The model id the provider echoed on the response, verbatim; `null` when\n * the body carried none. This is the only field that can witness a gateway\n * substituting a different model for the one requested — compare it with\n * `assertServedModel` / `checkServedModel` (src/integrity/served-model.ts).\n *\n * Optional so hand-built results (mock/custom transports) still typecheck,\n * but omitting it is not a pass: the identity checks read `undefined` as\n * `unreported` and reject it by default. A transport that knows which model\n * answered should say so.\n */\n servedModel?: string | null\n /** Wall-clock duration of the HTTP call (last attempt, if retried). */\n durationMs: number\n /**\n * `finish_reason` echoed from the first choice (`stop`, `length`,\n * `content_filter`, `tool_calls`, ...). `null` when the provider omits it.\n * Exposed so a free-form `callLlm` caller CAN detect a truncated answer\n * (`length`) instead of treating a cut-off completion as complete. Note:\n * `callLlm` does not itself reject on it — acting on this signal is the\n * caller's responsibility (in-repo free-form drivers do not yet enforce it).\n */\n finishReason?: string | null\n /**\n * True when `content.trim()` is empty. An empty completion is a silent zero\n * for free-form `callLlm` callers; this flag is the signal a caller can\n * inspect to fail loud rather than proceed on an empty string. `callLlm`\n * surfaces it but does not throw on it.\n */\n contentEmpty?: boolean\n /** Raw response body. */\n raw: Record<string, unknown>\n}\n\nexport type LlmCallMetadata = Pick<LlmCallResult, 'usage' | 'costUsd' | 'model' | 'durationMs'>\n\n/** Convert a provider result into the canonical paid-call receipt input. */\nexport function costReceiptFromLlm(\n result: LlmCallResult,\n customTokenPricing?: CustomTokenPricing,\n): CostReceiptInput {\n const cachedTokens = result.usage.cachedPromptTokens ?? 0\n const inputTokens = Math.max(0, result.usage.promptTokens - cachedTokens)\n const providerCostUsd = providerReportedCost(result.raw)\n return {\n model: result.model,\n inputTokens,\n outputTokens: result.usage.completionTokens,\n reasoningTokens: result.usage.reasoningTokens,\n cachedTokens: cachedTokens > 0 ? cachedTokens : undefined,\n ...(providerCostUsd === undefined\n ? customTokenPricing && result.usage.captured !== false\n ? { customTokenPricing }\n : result.costUsd === null\n ? {}\n : { estimatedCostUsd: result.costUsd }\n : { actualCostUsd: providerCostUsd }),\n usageUnknown: result.usage.captured === false,\n }\n}\n\nfunction providerReportedCost(raw: Record<string, unknown>): number | undefined {\n const value = raw._response_cost ?? raw.cost_usd\n return typeof value === 'number' && Number.isFinite(value) && value >= 0 ? value : undefined\n}\n\n/** Structured-response failures retain their completed provider receipt. */\nexport function costReceiptFromLlmError(\n error: Error,\n customTokenPricing?: CustomTokenPricing,\n): CostReceiptInput | undefined {\n return error instanceof LlmResponseError\n ? costReceiptFromLlm(error.result, customTokenPricing)\n : undefined\n}\n\nexport class LlmCallError extends AgentEvalError {\n constructor(\n message: string,\n public readonly status: number,\n public readonly body: string,\n public readonly model: string,\n ) {\n super('judge', message)\n }\n}\n\n/** A provider response completed and incurred measurable usage, but its content\n * could not satisfy the caller's response contract. The response envelope is\n * retained so accounting can commit the receipt before the error propagates. */\nexport class LlmResponseError extends AgentEvalError {\n constructor(\n message: string,\n public readonly result: LlmCallResult,\n options?: { cause?: unknown },\n ) {\n super('judge', message, options)\n }\n}\n\nexport interface LlmClientOptions {\n /** Base URL (without trailing slash). Must end at the `/v1` prefix. */\n baseUrl?: string\n /** Bearer token — either `apiKey` or `bearer` populates `Authorization: Bearer ...`. */\n apiKey?: string\n bearer?: string\n /** Override for the `Authorization` header (e.g. `X-Auth: ...`). Takes precedence over apiKey/bearer. */\n authHeader?: { name: string; value: string }\n /** Stable provider idempotency key, reused across retries of this logical call. */\n idempotencyKey?: string\n /** Default timeout in ms. Per-call can override. */\n defaultTimeoutMs?: number\n /**\n * Caller-supplied abort signal — e.g. a campaign-wide cancel. Linked to\n * each attempt's per-attempt timeout controller, so aborting it cancels\n * the in-flight fetch. A caller abort is FATAL: it is not retried even\n * though an AbortError otherwise matches the transient patterns.\n */\n signal?: AbortSignal\n /**\n * Cross-attempt wall-clock budget in ms, measured from the first attempt.\n * Before launching each attempt the loop checks the remaining budget and\n * stops retrying once it is exhausted, rather than waiting the full\n * per-attempt timeout on every retry. Bounds total time independent of\n * total attempts × `timeoutMs`.\n */\n deadlineMs?: number\n /** Total provider attempts. Default 3. */\n maximumAttempts?: number\n /** Token rates used when the provider omits cost or package pricing does not cover the model. */\n customTokenPricing?: CustomTokenPricing\n /**\n * Transport for requests that declare `jsonSchema`. `native` sends\n * `response_format: json_schema`; `json-object` sends the broadly supported\n * JSON mode and relies on the caller to include the schema in model-visible\n * instructions. Default: `native`.\n */\n jsonSchemaTransport?: 'native' | 'json-object'\n /**\n * JSON payload parsing policy. `extract` accepts fenced or prose-prefixed JSON.\n * `exact` requires the complete response content to be one JSON value.\n * Default: `extract`.\n */\n jsonPayloadMode?: 'extract' | 'exact'\n /** Default provider reasoning mode. A per-call request value takes precedence. */\n thinking?: LlmThinkingMode\n /** Fetch implementation — defaults to global `fetch`. Override for custom transport (e.g. tests). */\n fetch?: typeof fetch\n /**\n * Optional raw HTTP capture sink. When provided, every request, response,\n * and error (across all retry attempts) is recorded to the sink, with auth\n * headers and credential-shaped body fields redacted by default. This is\n * the layer-1 forensics primitive: structured `LlmSpan`s record intent,\n * raw events record what actually crossed the wire.\n */\n rawSink?: RawProviderSink\n /**\n * Logical provider id attached to raw events. When omitted, derived from\n * `baseUrl` via `providerFromBaseUrl`.\n */\n provider?: string\n /** Trace context attached to raw events; populated by emitter-aware callers. */\n traceContext?: { runId?: string; spanId?: string }\n /** Override the redaction strategy for this call. Defaults to `defaultProviderRedactor`. */\n redactor?: ProviderRedactor\n /**\n * Reject a response whose echoed model is not the model that was requested.\n * A routing gateway can accept one id and answer from another, which\n * silently invalidates every per-model and per-family claim downstream.\n * `true` uses the strict default (aliases pass, substitutions and\n * unidentified responses throw `ModelSubstitutionError`); pass an options\n * object to relax a specific case. Off by default — turning it on for a\n * measurement run is the point.\n */\n assertServedModel?: boolean | AssertServedModelOptions\n}\n\n// ─── Internals ──────────────────────────────────────────────────────────\n\nconst DEFAULT_BASE_URL = 'https://router.tangle.tools/v1'\n// Flagship / reasoning models routinely take several minutes on large prompts (a\n// reflection over many failures, a long tool transcript). A tight cap aborts a\n// legitimately-slow but healthy call — and because every retry attempt re-uses\n// the same window, such a model aborts on ALL attempts and the loop throws. The\n// default is generous enough to let those complete, bounded enough that a truly\n// hung call still fails over after retries, and tunable per deployment via\n// TANGLE_LLM_TIMEOUT_MS. Per-call `req.timeoutMs` / `opts.defaultTimeoutMs`\n// still win for callers that know their model's latency.\nconst DEFAULT_TIMEOUT_MS = Number(process.env.TANGLE_LLM_TIMEOUT_MS) || 300_000\nconst DEFAULT_MAXIMUM_ATTEMPTS =\n process.env.TANGLE_LLM_MAXIMUM_ATTEMPTS === undefined\n ? 3\n : Number(process.env.TANGLE_LLM_MAXIMUM_ATTEMPTS)\n\nfunction resolveMaximumAttempts(configured: number | undefined): number {\n const attempts = configured ?? DEFAULT_MAXIMUM_ATTEMPTS\n if (!Number.isInteger(attempts) || attempts <= 0) {\n throw new RangeError('LLM maximum attempts must be a positive integer')\n }\n return attempts\n}\n\nfunction providerTokenCount(value: unknown): number | undefined {\n return typeof value === 'number' && Number.isSafeInteger(value) && value >= 0 ? value : undefined\n}\n\nconst RETRYABLE_STATUS = new Set([429, 502, 503, 504])\n\n/**\n * Transient transport/network error signatures, matched against an error's\n * name, message, and `code`. Covers fetch/undici network failures, aborts\n * and timeouts, and — critically — HTTP/2 transport faults a keep-alive\n * connection raises mid-response: `terminated`, `NGHTTP2_INTERNAL_ERROR`,\n * `UND_ERR_*`, `other side closed`. Those last ones carry no clean HTTP\n * status; unrecognised, they escape the retry loop and surface as an\n * uncaught rejection.\n */\nconst TRANSIENT_ERROR_PATTERNS: readonly RegExp[] = [\n /AbortError/i,\n /TimeoutError/i,\n /this operation was aborted/i,\n /fetch failed/i,\n /ECONNRESET/i,\n /ETIMEDOUT/i,\n /EAI_AGAIN/i,\n /socket hang up/i,\n /stream.*ended.*unexpectedly/i,\n /terminated/i,\n /other side closed/i,\n /NGHTTP2/i,\n /UND_ERR/i,\n]\n\n/**\n * True when an error is a transient transport/network fault worth retrying,\n * as opposed to a deterministic failure (4xx schema reject, JSON parse) that\n * a retry cannot fix. Inspects `LlmCallError.status`, then the error's\n * name/message/code, then recurses into `error.cause` — undici nests the\n * real socket fault one or more levels under `.cause`.\n *\n * This is the retry classifier for the package: `callLlm` and\n * `withJudgeRetry` both route through it, so connection failures are treated\n * consistently across transports.\n */\nexport function isTransientLlmError(err: unknown): boolean {\n return classifyTransient(err, 0)\n}\n\nfunction classifyTransient(err: unknown, depth: number): boolean {\n if (err instanceof LlmCallError) return RETRYABLE_STATUS.has(err.status)\n if (!(err instanceof Error)) return false\n // Foreign transport errors can carry a numeric HTTP status without being an\n // LlmCallError. A retryable status is decisive.\n const status = (err as { status?: unknown }).status\n if (typeof status === 'number' && RETRYABLE_STATUS.has(status)) return true\n const code = (err as { code?: unknown }).code\n const haystack = `${err.name}\\n${err.message}\\n${typeof code === 'string' ? code : ''}`\n if (TRANSIENT_ERROR_PATTERNS.some((p) => p.test(haystack))) return true\n const cause = (err as { cause?: unknown }).cause\n if (depth < 4 && cause instanceof Error && cause !== err) {\n return classifyTransient(cause, depth + 1)\n }\n return false\n}\n\nfunction parseRetryAfter(headers: Headers): number | null {\n const h = headers.get('retry-after')\n if (!h) return null\n const asNumber = Number(h)\n if (Number.isFinite(asNumber) && asNumber > 0) return asNumber * 1000\n const asDate = Date.parse(h)\n if (Number.isFinite(asDate)) return Math.max(0, asDate - Date.now())\n return null\n}\n\n/** Exponential backoff: 500ms, 1s, 2s, 4s, ... capped at 16s. Attempt is 0-indexed. */\nexport function backoffMs(attempt: number): number {\n return Math.min(500 * 2 ** attempt, 16_000)\n}\n\nfunction buildHeaders(opts: LlmClientOptions): Record<string, string> {\n const headers: Record<string, string> = {\n 'Content-Type': 'application/json',\n Accept: 'application/json',\n }\n if (opts.authHeader) {\n headers[opts.authHeader.name] = opts.authHeader.value\n } else if (opts.bearer || opts.apiKey) {\n headers.Authorization = `Bearer ${opts.bearer ?? opts.apiKey}`\n }\n if (opts.idempotencyKey) headers['Idempotency-Key'] = opts.idempotencyKey\n return headers\n}\n\nfunction isSchemaRejection(status: number, body: string): boolean {\n if (status !== 400) return false\n const lower = body.toLowerCase()\n return (\n lower.includes('response_format') ||\n lower.includes('json_schema') ||\n lower.includes('is unavailable') ||\n lower.includes('not supported')\n )\n}\n\nfunction isTemperatureOneRejection(status: number, body: string): boolean {\n if (status !== 400 || !/temperature/i.test(body)) return false\n return (\n /temperature[^.\\n]{0,120}\\b(?:only|must|should|required|requires?)\\b[^.\\n]{0,40}\\b1(?:\\.0+)?\\b/i.test(\n body,\n ) || /\\bonly\\s+1(?:\\.0+)?\\s+is\\s+allowed\\b[^.\\n]{0,120}\\btemperature\\b/i.test(body)\n )\n}\n\nfunction buildBody(\n req: LlmCallRequest,\n forceJsonObject: boolean,\n defaultThinking?: LlmThinkingMode,\n): Record<string, unknown> {\n const body: Record<string, unknown> = {\n model: req.model,\n messages: req.messages,\n temperature: req.temperature ?? 0,\n }\n if (req.maxTokens != null) {\n if (usesMaxCompletionTokens(req.model)) body.max_completion_tokens = req.maxTokens\n else body.max_tokens = req.maxTokens\n }\n const thinking = req.thinking ?? defaultThinking\n if (thinking !== undefined) {\n body.thinking = { type: thinking }\n }\n\n if (req.jsonSchema && !forceJsonObject) {\n body.response_format = {\n type: 'json_schema',\n json_schema: { name: req.jsonSchema.name, schema: req.jsonSchema.schema, strict: true },\n }\n } else if (req.jsonMode || req.jsonSchema) {\n body.response_format = { type: 'json_object' }\n }\n\n return body\n}\n\nfunction usesMaxCompletionTokens(model: string): boolean {\n return /^gpt-5(?:[.-]|$)/i.test(model)\n}\n\nasync function sleep(ms: number): Promise<void> {\n return new Promise((resolve) => setTimeout(resolve, ms))\n}\n\n/**\n * Combine the per-attempt timeout signal with an optional caller signal into\n * one signal the fetch listens on. Prefers the native `AbortSignal.any`; falls\n * back to manual wiring on runtimes that predate it. The caller signal is also\n * propagated to the timeout controller so aborting it cancels the in-flight\n * fetch immediately.\n */\nfunction linkSignals(timeoutController: AbortController, caller?: AbortSignal): AbortSignal {\n if (!caller) return timeoutController.signal\n if (typeof (AbortSignal as { any?: unknown }).any === 'function') {\n return AbortSignal.any([timeoutController.signal, caller])\n }\n if (caller.aborted) {\n timeoutController.abort()\n } else {\n caller.addEventListener('abort', () => timeoutController.abort(), { once: true })\n }\n return timeoutController.signal\n}\n\n/** True once the cross-attempt wall-clock budget (if any) is exhausted. */\nfunction deadlineExceeded(start: number, deadlineMs: number | undefined): boolean {\n return deadlineMs != null && Date.now() - start >= deadlineMs\n}\n\n// ─── Public API ─────────────────────────────────────────────────────────\n\n/**\n * Strip a ```json / ``` code fence if the model emitted one.\n * Idempotent for naked JSON. Some models (claude-code via router, certain\n * deepseek models) wrap output even under json_object.\n */\nexport function stripFencedJson(raw: string): string {\n const trimmed = raw.trim()\n const m = trimmed.match(/^```(?:json)?\\s*\\n?([\\s\\S]*?)\\n?```\\s*$/)\n return m ? m[1]!.trim() : trimmed\n}\n\nexport function extractJsonPayload(raw: string): string {\n const stripped = stripFencedJson(raw)\n try {\n JSON.parse(stripped)\n return stripped\n } catch {\n // A response that declares a JSON root must parse as that complete root.\n // Scanning onward could turn a truncated object into one of its valid nested\n // arrays or objects and silently change the response schema.\n if (stripped.startsWith('{') || stripped.startsWith('[')) return stripped\n }\n\n // Only prose-leading responses may contain a recoverable JSON payload.\n const starts = [...stripped.matchAll(/[[{]/g)]\n .map((match) => match.index)\n .filter((index) => index != null)\n for (const start of starts) {\n const candidate = extractBalancedJson(stripped, start)\n if (!candidate) continue\n try {\n JSON.parse(candidate)\n return candidate\n } catch {\n // Keep scanning; earlier braces may belong to prose.\n }\n }\n\n return stripped\n}\n\nfunction extractBalancedJson(input: string, start: number): string | null {\n const opener = input[start]\n const closer = opener === '{' ? '}' : opener === '[' ? ']' : null\n if (!closer) return null\n\n const stack: string[] = [closer]\n let isInString = false\n let isEscaped = false\n\n for (let i = start + 1; i < input.length; i++) {\n const char = input[i]!\n if (isEscaped) {\n isEscaped = false\n continue\n }\n if (char === '\\\\') {\n isEscaped = isInString\n continue\n }\n if (char === '\"') {\n isInString = !isInString\n continue\n }\n if (isInString) continue\n\n if (char === '{') stack.push('}')\n else if (char === '[') stack.push(']')\n else if (char === stack[stack.length - 1]) {\n stack.pop()\n if (stack.length === 0) return input.slice(start, i + 1)\n }\n }\n\n return null\n}\n\n/**\n * Low-level call. Returns raw content + usage + cost. Retries on transient\n * failures; does NOT degrade schema here — callers that want graceful\n * degrade use `callLlmJson`.\n */\nexport async function callLlm(\n req: LlmCallRequest,\n opts: LlmClientOptions = {},\n): Promise<LlmCallResult> {\n const baseUrl = (opts.baseUrl ?? DEFAULT_BASE_URL).replace(/\\/+$/, '')\n const url = `${baseUrl}/chat/completions`\n const endpoint = '/chat/completions'\n const timeoutMs = req.timeoutMs ?? opts.defaultTimeoutMs ?? DEFAULT_TIMEOUT_MS\n const maximumAttempts = resolveMaximumAttempts(opts.maximumAttempts)\n const fetchFn = opts.fetch ?? globalThis.fetch\n const headers = buildHeaders(opts)\n const provider = opts.provider ?? providerFromBaseUrl(baseUrl)\n const sink = opts.rawSink\n const redactor = opts.redactor ?? defaultProviderRedactor\n const traceContext = opts.traceContext\n const callerSignal = opts.signal\n const deadlineMs = opts.deadlineMs\n const deadlineStart = Date.now()\n if (opts.customTokenPricing) {\n costForTokenPricing(opts.customTokenPricing, { inputTokens: 0, outputTokens: 0 })\n }\n\n let lastErr: unknown\n let effectiveRequest = req\n for (let attempt = 0; attempt < maximumAttempts; attempt++) {\n // A caller cancel is fatal — never retried. Checking before each attempt\n // means an already-aborted signal short-circuits without firing fetch.\n if (callerSignal?.aborted) {\n throw new DOMException('callLlm aborted by caller signal', 'AbortError')\n }\n // Stop retrying once the cross-attempt budget is spent rather than burning\n // a full per-attempt timeout on each remaining retry.\n if (attempt > 0 && deadlineExceeded(deadlineStart, deadlineMs)) {\n throw lastErr instanceof Error ? lastErr : new Error(String(lastErr))\n }\n const controller = new AbortController()\n const attemptSignal = linkSignals(controller, callerSignal)\n const timeoutHandle = setTimeout(() => controller.abort(), timeoutMs)\n const started = Date.now()\n const requestBody = buildBody(\n effectiveRequest,\n opts.jsonSchemaTransport === 'json-object',\n opts.thinking,\n )\n let attemptErrorRecorded = false\n if (sink) {\n await recordRaw(sink, redactor, {\n eventId: cryptoEventId(),\n runId: traceContext?.runId,\n spanId: traceContext?.spanId,\n provider,\n model: req.model,\n endpoint,\n baseUrl,\n attemptIndex: attempt,\n direction: 'request',\n timestamp: started,\n requestHeaders: headers,\n requestBody,\n redactedFields: [],\n })\n }\n\n try {\n const res = await fetchFn(url, {\n method: 'POST',\n headers,\n body: JSON.stringify(requestBody),\n signal: attemptSignal,\n })\n clearTimeout(timeoutHandle)\n const responseHeaders = sink ? headersToObject(res.headers) : undefined\n\n if (!res.ok) {\n const body = await res.text()\n if (sink) {\n await recordRaw(sink, redactor, {\n eventId: cryptoEventId(),\n runId: traceContext?.runId,\n spanId: traceContext?.spanId,\n provider,\n model: req.model,\n endpoint,\n baseUrl,\n attemptIndex: attempt,\n direction: 'error',\n timestamp: Date.now(),\n durationMs: Date.now() - started,\n statusCode: res.status,\n responseHeaders,\n responseBody: body,\n errorMessage: `HTTP ${res.status}`,\n redactedFields: [],\n })\n attemptErrorRecorded = true\n }\n const err = new LlmCallError(\n `LLM call failed with HTTP ${res.status}`,\n res.status,\n body,\n req.model,\n )\n if (\n isTemperatureOneRejection(res.status, body) &&\n effectiveRequest.temperature !== 1 &&\n attempt < maximumAttempts - 1 &&\n !deadlineExceeded(deadlineStart, deadlineMs)\n ) {\n lastErr = err\n effectiveRequest = { ...effectiveRequest, temperature: 1 }\n continue\n }\n if (\n RETRYABLE_STATUS.has(res.status) &&\n attempt < maximumAttempts - 1 &&\n !deadlineExceeded(deadlineStart, deadlineMs)\n ) {\n lastErr = err\n const retryAfter = parseRetryAfter(res.headers)\n await sleep(retryAfter ?? backoffMs(attempt))\n continue\n }\n throw err\n }\n\n const text = await res.text()\n let json: Record<string, unknown>\n try {\n json = JSON.parse(text) as Record<string, unknown>\n } catch (parseErr) {\n if (sink) {\n await recordRaw(sink, redactor, {\n eventId: cryptoEventId(),\n runId: traceContext?.runId,\n spanId: traceContext?.spanId,\n provider,\n model: req.model,\n endpoint,\n baseUrl,\n attemptIndex: attempt,\n direction: 'error',\n timestamp: Date.now(),\n durationMs: Date.now() - started,\n statusCode: res.status,\n responseHeaders,\n responseBody: text,\n errorMessage: `non-JSON response: ${parseErr instanceof Error ? parseErr.message : String(parseErr)}`,\n redactedFields: [],\n })\n attemptErrorRecorded = true\n }\n throw parseErr\n }\n if (sink) {\n await recordRaw(sink, redactor, {\n eventId: cryptoEventId(),\n runId: traceContext?.runId,\n spanId: traceContext?.spanId,\n provider,\n model: req.model,\n endpoint,\n baseUrl,\n attemptIndex: attempt,\n direction: 'response',\n timestamp: Date.now(),\n durationMs: Date.now() - started,\n statusCode: res.status,\n responseHeaders,\n responseBody: json,\n redactedFields: [],\n })\n }\n const choice = (\n json.choices as\n | Array<{ message?: { content?: string }; finish_reason?: string | null }>\n | undefined\n )?.[0]\n const usageRaw =\n json.usage && typeof json.usage === 'object' && !Array.isArray(json.usage)\n ? (json.usage as Record<string, unknown>)\n : undefined\n const promptTokens = providerTokenCount(usageRaw?.prompt_tokens)\n const completionTokens = providerTokenCount(usageRaw?.completion_tokens)\n const totalTokens = providerTokenCount(usageRaw?.total_tokens)\n const completionDetails =\n usageRaw?.completion_tokens_details &&\n typeof usageRaw.completion_tokens_details === 'object' &&\n !Array.isArray(usageRaw.completion_tokens_details)\n ? (usageRaw.completion_tokens_details as Record<string, unknown>)\n : undefined\n const reasoningRaw = completionDetails?.reasoning_tokens\n const reasoningTokens =\n reasoningRaw === undefined ? undefined : providerTokenCount(reasoningRaw)\n const cachedRaw =\n usageRaw?.prompt_tokens_details &&\n typeof usageRaw.prompt_tokens_details === 'object' &&\n !Array.isArray(usageRaw.prompt_tokens_details)\n ? (usageRaw.prompt_tokens_details as Record<string, unknown>).cached_tokens\n : undefined\n const cachedPromptTokens = cachedRaw === undefined ? undefined : providerTokenCount(cachedRaw)\n const usageCaptured =\n promptTokens !== undefined &&\n completionTokens !== undefined &&\n (reasoningRaw === undefined ||\n (reasoningTokens !== undefined && reasoningTokens <= completionTokens)) &&\n (cachedRaw === undefined ||\n (cachedPromptTokens !== undefined && cachedPromptTokens <= promptTokens)) &&\n (totalTokens === undefined || totalTokens === promptTokens + completionTokens)\n const costFromProxy = (json._response_cost ?? json.cost_usd) as number | undefined\n const content = choice?.message?.content ?? ''\n\n const configuredCost =\n typeof costFromProxy !== 'number' && usageCaptured && opts.customTokenPricing\n ? costForTokenPricing(opts.customTokenPricing, {\n inputTokens: promptTokens! - (cachedPromptTokens ?? 0),\n ...(cachedPromptTokens ? { cachedTokens: cachedPromptTokens } : {}),\n outputTokens: completionTokens!,\n })\n : undefined\n\n // The echoed id, kept separate from the attribution id: a provider that\n // omits it must read as \"unproven\", never as \"the model I asked for\".\n const servedModel =\n typeof json.model === 'string' && json.model.trim() !== '' ? json.model : null\n if (opts.assertServedModel) {\n assertServedModelIdentity(\n req.model,\n servedModel,\n opts.assertServedModel === true ? {} : opts.assertServedModel,\n )\n }\n\n return {\n content,\n finishReason: choice?.finish_reason ?? null,\n contentEmpty: content.trim().length === 0,\n usage: {\n promptTokens: promptTokens ?? 0,\n completionTokens: completionTokens ?? 0,\n totalTokens: totalTokens ?? (promptTokens ?? 0) + (completionTokens ?? 0),\n captured: usageCaptured,\n reasoningTokens,\n cachedPromptTokens,\n },\n costUsd: typeof costFromProxy === 'number' ? costFromProxy : (configuredCost ?? null),\n model: servedModel ?? req.model,\n servedModel,\n durationMs: Date.now() - started,\n raw: json,\n }\n } catch (err) {\n clearTimeout(timeoutHandle)\n lastErr = err\n // A caller cancel is fatal even though an AbortError matches the\n // transient patterns — a cancelled call must surface immediately, not\n // be retried against the same dead intent.\n if (callerSignal?.aborted) {\n if (sink && !attemptErrorRecorded) {\n await recordRaw(sink, redactor, {\n eventId: cryptoEventId(),\n runId: traceContext?.runId,\n spanId: traceContext?.spanId,\n provider,\n model: req.model,\n endpoint,\n baseUrl,\n attemptIndex: attempt,\n direction: 'error',\n timestamp: Date.now(),\n durationMs: Date.now() - started,\n errorMessage: err instanceof Error ? err.message : String(err),\n redactedFields: [],\n })\n }\n throw err\n }\n if (sink && !attemptErrorRecorded) {\n // Record only if neither the !res.ok branch nor the JSON.parse catch\n // already produced an error event for this attempt. Covers network\n // failures, timeouts, and aborts.\n await recordRaw(sink, redactor, {\n eventId: cryptoEventId(),\n runId: traceContext?.runId,\n spanId: traceContext?.spanId,\n provider,\n model: req.model,\n endpoint,\n baseUrl,\n attemptIndex: attempt,\n direction: 'error',\n timestamp: Date.now(),\n durationMs: Date.now() - started,\n errorMessage: err instanceof Error ? err.message : String(err),\n redactedFields: [],\n })\n }\n if (\n attempt < maximumAttempts - 1 &&\n isTransientLlmError(err) &&\n !deadlineExceeded(deadlineStart, deadlineMs)\n ) {\n await sleep(backoffMs(attempt))\n continue\n }\n throw err\n }\n }\n throw lastErr instanceof Error ? lastErr : new Error(String(lastErr))\n}\n\nasync function recordRaw(\n sink: RawProviderSink,\n redactor: ProviderRedactor,\n event: RawProviderEvent,\n): Promise<void> {\n // Errors from sinks must not crash the LLM call. Forensic capture is\n // best-effort; the structured trace is the system of record.\n try {\n await sink.record(redactor(event))\n } catch {\n // Intentionally swallowed.\n }\n}\n\nfunction headersToObject(h: Headers): Record<string, string> {\n const out: Record<string, string> = {}\n h.forEach((value, key) => {\n out[key] = value\n })\n return out\n}\n\nfunction cryptoEventId(): string {\n if (typeof globalThis.crypto?.randomUUID === 'function') return globalThis.crypto.randomUUID()\n return `${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 10)}`\n}\n\n/**\n * Structured-output call. Returns parsed JSON plus the raw result envelope.\n * Degrades `jsonSchema` → `jsonMode` on a 400 that names the schema param —\n * critical for deepseek-v3/v4, kimi-k2.6, and other models that don't accept\n * the `response_format.json_schema` shape but DO accept `json_object`.\n */\nexport async function callLlmJson<T = unknown>(\n req: LlmCallRequest,\n opts: LlmClientOptions = {},\n): Promise<{ value: T; result: LlmCallResult }> {\n const result = await callLlmStructured(req, opts)\n const value = parseJsonResult<T>(result, opts.jsonPayloadMode ?? 'extract')\n return { value, result }\n}\n\n/** Shared schema-to-JSON-mode fallback that preserves the raw result. */\nasync function callLlmStructured(\n req: LlmCallRequest,\n opts: LlmClientOptions = {},\n): Promise<LlmCallResult> {\n try {\n return await callLlm({ ...req, jsonMode: req.jsonMode ?? !req.jsonSchema }, opts)\n } catch (err) {\n if (\n opts.jsonSchemaTransport !== 'json-object' &&\n err instanceof LlmCallError &&\n isSchemaRejection(err.status, err.body) &&\n req.jsonSchema\n ) {\n const degradedReq: LlmCallRequest = { ...req, jsonMode: true, jsonSchema: undefined }\n return await callLlm(degradedReq, opts)\n }\n throw err\n }\n}\n\nfunction parseJsonResult<T>(\n result: LlmCallResult,\n jsonPayloadMode: NonNullable<LlmClientOptions['jsonPayloadMode']>,\n): T {\n try {\n if (result.finishReason === 'length') {\n throw new Error(\n `LLM returned truncated JSON content (model=${result.model}, finishReason=length)`,\n )\n }\n return parseJsonSafely<T>(result.content, result.model, jsonPayloadMode)\n } catch (error) {\n if (error instanceof LlmResponseError) throw error\n const cause = error instanceof Error ? error : new Error(String(error))\n throw new LlmResponseError(cause.message, result, { cause })\n }\n}\n\nfunction parseJsonSafely<T>(\n content: string,\n model: string,\n jsonPayloadMode: NonNullable<LlmClientOptions['jsonPayloadMode']>,\n): T {\n const payload = jsonPayloadMode === 'exact' ? content : extractJsonPayload(content)\n try {\n return JSON.parse(payload) as T\n } catch {\n throw new Error(`LLM returned non-JSON content (model=${model})`)\n }\n}\n\n// ─── Route assertion ────────────────────────────────────────────────────\n\nexport type LlmRouteAssertionReason =\n | 'no_explicit_base_url'\n | 'base_url_blocked'\n | 'base_url_not_allowed'\n | 'no_auth'\n | 'wrong_provider'\n\nexport class LlmRouteAssertionError extends CaptureIntegrityError {\n constructor(\n message: string,\n public readonly reason: LlmRouteAssertionReason,\n public readonly baseUrl: string,\n ) {\n super(message)\n }\n}\n\nexport interface LlmRouteRequirements {\n /**\n * Throw if `opts.baseUrl` is undefined, i.e. the call would fall back to\n * `DEFAULT_BASE_URL`. Set this for evaluation runs where silently using\n * the public/free-tier router is a defect — the launch reviewer needs to\n * know exactly which provider answered.\n */\n requireExplicitBaseUrl?: boolean\n /**\n * Allowlist of acceptable base URLs. Strings match by prefix\n * (case-insensitive); RegExps test against the full base URL.\n */\n allowedBaseUrls?: Array<string | RegExp>\n /** Blocklist that takes precedence over `allowedBaseUrls`. */\n blockedBaseUrls?: Array<string | RegExp>\n /** Throw if no auth header / api key is configured. */\n requireAuth?: boolean\n /**\n * Logical provider id the configured `baseUrl` is expected to match (via\n * `providerFromBaseUrl`). Mainly useful when paired with `requireExplicitBaseUrl`.\n */\n expectedProvider?: string\n}\n\n/**\n * Fail-loud assertion that the configured LLM client points at the route\n * the caller intends. Designed for the matrix-runner preflight: invoke\n * once before any LLM call to catch misconfiguration before a sweep burns\n * dollars on the wrong provider.\n *\n * Throws `LlmRouteAssertionError`. Pure — no I/O — so it's safe to call\n * from constructors and CI gates.\n */\nexport function assertLlmRoute(opts: LlmClientOptions, req: LlmRouteRequirements = {}): void {\n const baseUrlExplicit = opts.baseUrl !== undefined\n const baseUrl = (opts.baseUrl ?? DEFAULT_BASE_URL).replace(/\\/+$/, '')\n\n if (req.requireExplicitBaseUrl && !baseUrlExplicit) {\n throw new LlmRouteAssertionError(\n `assertLlmRoute: requireExplicitBaseUrl set but opts.baseUrl is undefined; would fall back to ${DEFAULT_BASE_URL}.`,\n 'no_explicit_base_url',\n baseUrl,\n )\n }\n\n if (req.blockedBaseUrls?.some((p) => matchUrl(baseUrl, p))) {\n throw new LlmRouteAssertionError(\n `assertLlmRoute: baseUrl ${baseUrl} matches a blocked pattern.`,\n 'base_url_blocked',\n baseUrl,\n )\n }\n\n if (req.allowedBaseUrls && req.allowedBaseUrls.length > 0) {\n const ok = req.allowedBaseUrls.some((p) => matchUrl(baseUrl, p))\n if (!ok) {\n throw new LlmRouteAssertionError(\n `assertLlmRoute: baseUrl ${baseUrl} is not in the allowed list (${req.allowedBaseUrls.map(describePattern).join(', ')}).`,\n 'base_url_not_allowed',\n baseUrl,\n )\n }\n }\n\n if (req.requireAuth && !opts.apiKey && !opts.bearer && !opts.authHeader) {\n throw new LlmRouteAssertionError(\n `assertLlmRoute: requireAuth set but no apiKey, bearer, or authHeader was supplied.`,\n 'no_auth',\n baseUrl,\n )\n }\n\n if (req.expectedProvider) {\n const actual = opts.provider ?? providerFromBaseUrl(baseUrl)\n if (actual !== req.expectedProvider) {\n throw new LlmRouteAssertionError(\n `assertLlmRoute: expected provider ${req.expectedProvider} but baseUrl ${baseUrl} resolves to ${actual}.`,\n 'wrong_provider',\n baseUrl,\n )\n }\n }\n}\n\nfunction matchUrl(url: string, pattern: string | RegExp): boolean {\n if (pattern instanceof RegExp) return pattern.test(url)\n return url.toLowerCase().startsWith(pattern.toLowerCase())\n}\n\nfunction describePattern(p: string | RegExp): string {\n return p instanceof RegExp ? p.source : p\n}\n\n/**\n * Probe whether a model is reachable. Returns latency + null error on\n * success; `ok=false` + error message on any failure (HTTP, timeout,\n * network, parse). Designed for sweep preflights — fail loud at the\n * boundary before burning a 30-leaf run on a misconfigured router.\n *\n * Sends a tiny `ping` message with `maxTokens=64`. Reasoning models\n * (glm-5.1, deepseek-v4) can burn the entire budget on internal reasoning\n * for short prompts, so don't tighten this further. We don't validate\n * content.\n *\n * Reachability and identity are separate answers: `ok` means the route\n * answered, `servedModel` / `substituted` say WHICH model answered. A gateway\n * that serves another provider's model returns `ok: true` with\n * `substituted: true` — inspect both before treating the id as measured.\n */\nexport async function probeLlm(\n model: string,\n opts: LlmClientOptions & { timeoutMs?: number } = {},\n): Promise<{\n ok: boolean\n latencyMs: number\n error: string | null\n /** Id echoed by the provider; `null` when it sent none or the probe failed. */\n servedModel: string | null\n /** True when the echoed id is a different model than `model` (or absent). */\n substituted: boolean\n}> {\n const start = Date.now()\n try {\n const result = await callLlm(\n {\n model,\n messages: [{ role: 'user', content: 'ping' }],\n maxTokens: 64,\n timeoutMs: opts.timeoutMs ?? 30_000,\n },\n opts,\n )\n return {\n ok: true,\n latencyMs: Date.now() - start,\n error: null,\n servedModel: result.servedModel ?? null,\n substituted: checkServedModel(model, result.servedModel).substituted,\n }\n } catch (err) {\n return {\n ok: false,\n latencyMs: Date.now() - start,\n error: err instanceof Error ? err.message : String(err),\n servedModel: null,\n substituted: false,\n }\n }\n}\n\n/**\n * Stateful client — construct once with defaults, call many times.\n * Thin wrapper around the free functions; exists for callers that want\n * to inject a single configured instance into multiple primitives.\n */\nexport class LlmClient {\n readonly maximumAttempts: number\n private readonly opts: LlmClientOptions\n\n constructor(opts: LlmClientOptions = {}) {\n this.opts = opts\n this.maximumAttempts = resolveMaximumAttempts(opts.maximumAttempts)\n }\n\n call(req: LlmCallRequest, per?: LlmClientOptions): Promise<LlmCallResult> {\n const options = { ...this.opts, ...per }\n return req.jsonSchema ? callLlmStructured(req, options) : callLlm(req, options)\n }\n\n callJson<T = unknown>(\n req: LlmCallRequest,\n per?: LlmClientOptions,\n ): Promise<{ value: T; result: LlmCallResult }> {\n return callLlmJson<T>(req, { ...this.opts, ...per })\n }\n}\n"],"mappings":";;;;;AA2BA,MAAM,kBAA+C;CACnD,WAAW;CACX,QAAQ;CACR,gBAAgB;CAChB,QAAQ;CACR,iBAAiB;CACjB,MAAM;CACN,cAAc;CACd,SAAS;CACT,WAAW;CACX,UAAU;CACV,KAAK;CACL,MAAM;CACN,SAAS;CACT,QAAQ;CACR,QAAQ;CACR,SAAS;CACT,UAAU;CACV,YAAY;CACZ,MAAM;CACN,aAAa;CACb,OAAO;CACP,SAAS;CACT,KAAK;CACL,QAAQ;CACR,KAAK;AACP;;AAGA,MAAM,gBAA8C;CAClD,CAAC,WAAW,WAAW;CACvB,CAAC,0DAA0D,QAAQ;CACnE,CAAC,4BAA4B,QAAQ;CACrC,CAAC,UAAU,MAAM;CACjB,CAAC,oCAAoC,SAAS;CAC9C,CAAC,aAAa,UAAU;CACxB,CAAC,SAAS,KAAK;CACf,CAAC,SAAS,MAAM;CAChB,CAAC,oBAAoB,QAAQ;CAC7B,CAAC,qBAAqB,QAAQ;CAC9B,CAAC,sBAAsB,UAAU;CACjC,CAAC,4BAA4B,OAAO;AACtC;;;;;;;AAQA,SAAgB,YAAY,SAA8B;CACxD,MAAM,KAAK,QAAQ,KAAK,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,EAAE,CAAE,YAAY;CACrD,MAAM,QAAQ,GAAG,QAAQ,GAAG;CAC5B,IAAI,QAAQ,GAAG;EACb,MAAM,SAAS,GAAG,MAAM,GAAG,KAAK;EAChC,MAAM,SAAS,gBAAgB;EAC/B,IAAI,QAAQ,OAAO;CACrB;CACA,KAAK,MAAM,CAAC,SAAS,WAAW,eAC9B,IAAI,QAAQ,KAAK,EAAE,GAAG,OAAO;CAE/B,OAAO;AACT;AAWA,IAAa,mBAAb,cAAsC,MAAM;CAGxB;CACA;CAHlB,YACE,SACA,UACA,QACA;EACA,MAAM,OAAO;EAHG,KAAA,WAAA;EACA,KAAA,SAAA;EAGhB,KAAK,OAAO;CACd;AACF;;;;;;;;;;;;;;AAeA,SAAgB,kBACd,QACA,OAAiC,CAAC,GACnB;CACf,MAAM,cAAc,KAAK,eAAe;CACxC,MAAM,2BAAW,IAAI,IAAiB;CACtC,KAAK,MAAM,KAAK,QAAQ;EACtB,MAAM,IAAI,YAAY,CAAC;EACvB,IAAI,MAAM,aAAa,CAAC,KAAK,cAAc;EAC3C,SAAS,IAAI,CAAC;CAChB;CACA,MAAM,OAAO,CAAC,GAAG,QAAQ,CAAC,CAAC,KAAK;CAChC,IAAI,KAAK,SAAS,aAChB,MAAM,IAAI,iBACR,wBAAwB,KAAK,OAAO,iBAAiB,KAAK,WAAW,IAAI,MAAM,MAAM,IAC/E,KAAK,KAAK,IAAI,KAAK,OAAO,QAAQ,YAAY,kFAEpD,MACA,MACF;CAEF,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AChFA,SAAgB,iBAAiB,SAAyB;CACxD,IAAI,KAAK,QAAQ,KAAK,CAAC,CAAC,YAAY;CACpC,MAAM,KAAK,GAAG,QAAQ,GAAG;CACzB,IAAI,KAAK,GAAG,KAAK,GAAG,MAAM,GAAG,EAAE;CAC/B,MAAM,QAAQ,GAAG,QAAQ,GAAG;CAC5B,IAAI,QAAQ,GAAG,KAAK,GAAG,MAAM,GAAG,KAAK;CACrC,MAAM,QAAQ,GAAG,YAAY,GAAG;CAChC,IAAI,SAAS,GAAG,KAAK,GAAG,MAAM,QAAQ,CAAC;CACvC,OAAO,GACJ,QAAQ,uBAAuB,EAAE,CAAC,CAClC,QAAQ,WAAW,EAAE,CAAC,CACtB,QAAQ,SAAS,GAAG,CAAC,CACrB,QAAQ,OAAO,EAAE,CAAC,CAClB,KAAK;AACV;;;;;;;;;;;AAYA,SAAgB,iBACd,WACA,QACkB;CAClB,MAAM,kBAAkB,YAAY,SAAS;CAC7C,IAAI,WAAW,QAAQ,WAAW,KAAA,KAAa,OAAO,KAAK,MAAM,IAC/D,OAAO;EACL;EACA,QAAQ;EACR;EACA,cAAc;EACd,SAAS;EACT,aAAa;CACf;CAEF,MAAM,eAAe,YAAY,MAAM;CACvC,IAAI,UAAU,KAAK,CAAC,CAAC,YAAY,MAAM,OAAO,KAAK,CAAC,CAAC,YAAY,GAC/D,OAAO;EACL;EACA;EACA;EACA;EACA,SAAS;EACT,aAAa;CACf;CAEF,IAAI,iBAAiB,SAAS,MAAM,iBAAiB,MAAM,GACzD,OAAO;EACL;EACA;EACA;EACA;EACA,SAAS;EACT,aAAa;CACf;CAEF,OAAO;EACL;EACA;EACA;EACA;EACA,SACE,oBAAoB,eAAe,8BAA8B;EACnE,aAAa;CACf;AACF;AAEA,IAAa,yBAAb,cAA4C,eAAe;CAGvC;CAFlB,YACE,SACA,QACA;EACA,MAAM,sBAAsB,OAAO;EAFnB,KAAA,SAAA;EAGhB,KAAK,OAAO;CACd;AACF;;;;;;AAuBA,SAAgB,sBACd,OACA,OAAiC,CAAC,GACzB;CACT,QAAQ,MAAM,SAAd;EACE,KAAK;EACL,KAAK,SACH,OAAO;EACT,KAAK,cACH,OAAO,KAAK,oBAAoB;EAClC,KAAK,6BACH,OAAO,KAAK,sBAAsB;EACpC,SACE,OAAO;CACX;AACF;AAEA,SAAS,SAAS,OAAiC;CACjD,IAAI,MAAM,YAAY,cACpB,OAAO,GAAG,MAAM,UAAU;CAC5B,OACE,GAAG,MAAM,UAAU,IAAI,MAAM,gBAAgB,aAAa,MAAM,OAAO,IACnE,MAAM,aAAa,KAAK,MAAM,QAAQ;AAE9C;;;;;;AAOA,SAAgB,kBACd,WACA,QACA,OAAiC,CAAC,GAChB;CAClB,MAAM,QAAQ,iBAAiB,WAAW,MAAM;CAChD,IAAI,sBAAsB,OAAO,IAAI,GAAG,OAAO;CAE/C,MAAM,IAAI,uBACR,GAFa,KAAK,UAAU,GAAG,KAAK,QAAQ,MAAM,GAExC,uBAAuB,SAAS,KAAK,EAAE,gIAEjD,CAAC,KAAK,CACR;AACF;;;;;AAMA,SAAgB,mBACd,OACA,OAAiC,CAAC,GACd;CACpB,MAAM,SAAS,MAAM,KAAK,MAAM,iBAAiB,EAAE,WAAW,EAAE,MAAM,CAAC;CACvE,MAAM,MAAM,OAAO,QAAQ,MAAM,CAAC,sBAAsB,GAAG,IAAI,CAAC;CAChE,IAAI,IAAI,SAAS,GAEf,MAAM,IAAI,uBACR,GAFa,KAAK,UAAU,GAAG,KAAK,QAAQ,MAAM,KAEtC,IAAI,OAAO,GAAG,OAAO,OAAO,+DACvB,IAAI,IAAI,QAAQ,CAAC,CAAC,KAAK,IAAI,EAAE,6FAE9C,MACF;CAEF,OAAO;AACT;AASA,IAAa,yBAAb,cAA4C,eAAe;CAGvC;CACA;CAHlB,YACE,SACA,UACA,QACA;EACA,MAAM,sBAAsB,OAAO;EAHnB,KAAA,WAAA;EACA,KAAA,SAAA;EAGhB,KAAK,OAAO;CACd;AACF;;;;;;;;;;AAWA,SAAgB,wBACd,OACA,OAAuC,CAAC,GACzB;CACf,MAAM,SAAS,mBAAmB,OAAO,IAAI;CAC7C,MAAM,2BAAW,IAAI,IAAiB;CACtC,KAAK,MAAM,SAAS,QAAQ;EAC1B,MAAM,SAAS,MAAM;EACrB,IAAI,WAAW,MAAM;EACrB,IAAI,WAAW,aAAa,CAAC,KAAK,cAAc;EAChD,SAAS,IAAI,MAAM;CACrB;CACA,MAAM,OAAO,CAAC,GAAG,QAAQ,CAAC,CAAC,KAAK;CAChC,MAAM,cAAc,KAAK,eAAe;CACxC,IAAI,KAAK,SAAS,aAEhB,MAAM,IAAI,uBACR,GAFa,KAAK,UAAU,GAAG,KAAK,QAAQ,MAAM,GAExC,gCAAgC,KAAK,OAAO,iBACjD,KAAK,WAAW,IAAI,MAAM,MAAM,IAAI,KAAK,KAAK,IAAI,KAAK,OAAO,QAAQ,YAAY,0BAC3D,OAAO,KAAK,MAAM,EAAE,UAAU,YAAY,CAAC,CAAC,KAAK,IAAI,KACjF,MACA,MACF;CAEF,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACnNA,SAAgB,2BACd,SACA,UAA4B,CAAC,GACF;CAC3B,IAAI,QAAQ,cAAc,KAAA,GAAW,OAAO,KAAA;CAC5C,IAAI,CAAC,OAAO,UAAU,QAAQ,SAAS,KAAK,QAAQ,aAAa,GAC/D,MAAM,IAAI,WAAW,kEAAkE;CAEzF,IACE,QAAQ,SAAS,MACd,YACC,MAAM,QAAQ,QAAQ,OAAO,KAAK,QAAQ,QAAQ,MAAM,SAAS,KAAK,SAAS,WAAW,CAC9F,GAEA;CAGF,MAAM,WAAW,uBAAuB,QAAQ,eAAe;CAC/D,MAAM,kBAAkB,QAAQ,wBAAwB;CAGxD,MAAM,eAAe,IAAI,YAAY,CAAC,CAAC,OACrC,KAAK,UAAU,UAAU,SAAS,iBAAiB,QAAQ,QAAQ,CAAC,CACtE,CAAC,CAAC;CAEF,MAAM,UAAU,QAAQ,cAAc,CAAC,kBAAkB,IAAI;CAC7D,MAAM,QAAQ;EACZ,aAAa,eAAe,WAAW;EACvC,cAAc,QAAQ,YAAY,WAAW;CAC/C;CACA,OAAO,QAAQ,qBACX;EAAE,oBAAoB,QAAQ;EAAoB,GAAG;CAAM,IAC3D;EAAE,OAAO,QAAQ;EAAO,GAAG;CAAM;AACvC;;AAqEA,SAAgB,mBACd,QACA,oBACkB;CAClB,MAAM,eAAe,OAAO,MAAM,sBAAsB;CACxD,MAAM,cAAc,KAAK,IAAI,GAAG,OAAO,MAAM,eAAe,YAAY;CACxE,MAAM,kBAAkB,qBAAqB,OAAO,GAAG;CACvD,OAAO;EACL,OAAO,OAAO;EACd;EACA,cAAc,OAAO,MAAM;EAC3B,iBAAiB,OAAO,MAAM;EAC9B,cAAc,eAAe,IAAI,eAAe,KAAA;EAChD,GAAI,oBAAoB,KAAA,IACpB,sBAAsB,OAAO,MAAM,aAAa,QAC9C,EAAE,mBAAmB,IACrB,OAAO,YAAY,OACjB,CAAC,IACD,EAAE,kBAAkB,OAAO,QAAQ,IACvC,EAAE,eAAe,gBAAgB;EACrC,cAAc,OAAO,MAAM,aAAa;CAC1C;AACF;AAEA,SAAS,qBAAqB,KAAkD;CAC9E,MAAM,QAAQ,IAAI,kBAAkB,IAAI;CACxC,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,KAAK,SAAS,IAAI,QAAQ,KAAA;AACrF;;AAGA,SAAgB,wBACd,OACA,oBAC8B;CAC9B,OAAO,iBAAiB,mBACpB,mBAAmB,MAAM,QAAQ,kBAAkB,IACnD,KAAA;AACN;AAEA,IAAa,eAAb,cAAkC,eAAe;CAG7B;CACA;CACA;CAJlB,YACE,SACA,QACA,MACA,OACA;EACA,MAAM,SAAS,OAAO;EAJN,KAAA,SAAA;EACA,KAAA,OAAA;EACA,KAAA,QAAA;CAGlB;AACF;;;;AAKA,IAAa,mBAAb,cAAsC,eAAe;CAGjC;CAFlB,YACE,SACA,QACA,SACA;EACA,MAAM,SAAS,SAAS,OAAO;EAHf,KAAA,SAAA;CAIlB;AACF;AAiFA,MAAM,mBAAmB;AASzB,MAAM,qBAAqB,OAAO,QAAQ,IAAI,qBAAqB,KAAK;AACxE,MAAM,2BACJ,QAAQ,IAAI,gCAAgC,KAAA,IACxC,IACA,OAAO,QAAQ,IAAI,2BAA2B;AAEpD,SAAS,uBAAuB,YAAwC;CACtE,MAAM,WAAW,cAAc;CAC/B,IAAI,CAAC,OAAO,UAAU,QAAQ,KAAK,YAAY,GAC7C,MAAM,IAAI,WAAW,iDAAiD;CAExE,OAAO;AACT;AAEA,SAAS,mBAAmB,OAAoC;CAC9D,OAAO,OAAO,UAAU,YAAY,OAAO,cAAc,KAAK,KAAK,SAAS,IAAI,QAAQ,KAAA;AAC1F;AAEA,MAAM,mCAAmB,IAAI,IAAI;CAAC;CAAK;CAAK;CAAK;AAAG,CAAC;;;;;;;;;;AAWrD,MAAM,2BAA8C;CAClD;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;;;;;;;;;;;;AAaA,SAAgB,oBAAoB,KAAuB;CACzD,OAAO,kBAAkB,KAAK,CAAC;AACjC;AAEA,SAAS,kBAAkB,KAAc,OAAwB;CAC/D,IAAI,eAAe,cAAc,OAAO,iBAAiB,IAAI,IAAI,MAAM;CACvE,IAAI,EAAE,eAAe,QAAQ,OAAO;CAGpC,MAAM,SAAU,IAA6B;CAC7C,IAAI,OAAO,WAAW,YAAY,iBAAiB,IAAI,MAAM,GAAG,OAAO;CACvE,MAAM,OAAQ,IAA2B;CACzC,MAAM,WAAW,GAAG,IAAI,KAAK,IAAI,IAAI,QAAQ,IAAI,OAAO,SAAS,WAAW,OAAO;CACnF,IAAI,yBAAyB,MAAM,MAAM,EAAE,KAAK,QAAQ,CAAC,GAAG,OAAO;CACnE,MAAM,QAAS,IAA4B;CAC3C,IAAI,QAAQ,KAAK,iBAAiB,SAAS,UAAU,KACnD,OAAO,kBAAkB,OAAO,QAAQ,CAAC;CAE3C,OAAO;AACT;AAEA,SAAS,gBAAgB,SAAiC;CACxD,MAAM,IAAI,QAAQ,IAAI,aAAa;CACnC,IAAI,CAAC,GAAG,OAAO;CACf,MAAM,WAAW,OAAO,CAAC;CACzB,IAAI,OAAO,SAAS,QAAQ,KAAK,WAAW,GAAG,OAAO,WAAW;CACjE,MAAM,SAAS,KAAK,MAAM,CAAC;CAC3B,IAAI,OAAO,SAAS,MAAM,GAAG,OAAO,KAAK,IAAI,GAAG,SAAS,KAAK,IAAI,CAAC;CACnE,OAAO;AACT;;AAGA,SAAgB,UAAU,SAAyB;CACjD,OAAO,KAAK,IAAI,MAAM,KAAK,SAAS,IAAM;AAC5C;AAEA,SAAS,aAAa,MAAgD;CACpE,MAAM,UAAkC;EACtC,gBAAgB;EAChB,QAAQ;CACV;CACA,IAAI,KAAK,YACP,QAAQ,KAAK,WAAW,QAAQ,KAAK,WAAW;MAC3C,IAAI,KAAK,UAAU,KAAK,QAC7B,QAAQ,gBAAgB,UAAU,KAAK,UAAU,KAAK;CAExD,IAAI,KAAK,gBAAgB,QAAQ,qBAAqB,KAAK;CAC3D,OAAO;AACT;AAEA,SAAS,kBAAkB,QAAgB,MAAuB;CAChE,IAAI,WAAW,KAAK,OAAO;CAC3B,MAAM,QAAQ,KAAK,YAAY;CAC/B,OACE,MAAM,SAAS,iBAAiB,KAChC,MAAM,SAAS,aAAa,KAC5B,MAAM,SAAS,gBAAgB,KAC/B,MAAM,SAAS,eAAe;AAElC;AAEA,SAAS,0BAA0B,QAAgB,MAAuB;CACxE,IAAI,WAAW,OAAO,CAAC,eAAe,KAAK,IAAI,GAAG,OAAO;CACzD,OACE,iGAAiG,KAC/F,IACF,KAAK,oEAAoE,KAAK,IAAI;AAEtF;AAEA,SAAS,UACP,KACA,iBACA,iBACyB;CACzB,MAAM,OAAgC;EACpC,OAAO,IAAI;EACX,UAAU,IAAI;EACd,aAAa,IAAI,eAAe;CAClC;CACA,IAAI,IAAI,aAAa,MACnB,IAAI,wBAAwB,IAAI,KAAK,GAAG,KAAK,wBAAwB,IAAI;MACpE,KAAK,aAAa,IAAI;CAE7B,MAAM,WAAW,IAAI,YAAY;CACjC,IAAI,aAAa,KAAA,GACf,KAAK,WAAW,EAAE,MAAM,SAAS;CAGnC,IAAI,IAAI,cAAc,CAAC,iBACrB,KAAK,kBAAkB;EACrB,MAAM;EACN,aAAa;GAAE,MAAM,IAAI,WAAW;GAAM,QAAQ,IAAI,WAAW;GAAQ,QAAQ;EAAK;CACxF;MACK,IAAI,IAAI,YAAY,IAAI,YAC7B,KAAK,kBAAkB,EAAE,MAAM,cAAc;CAG/C,OAAO;AACT;AAEA,SAAS,wBAAwB,OAAwB;CACvD,OAAO,oBAAoB,KAAK,KAAK;AACvC;AAEA,eAAe,MAAM,IAA2B;CAC9C,OAAO,IAAI,SAAS,YAAY,WAAW,SAAS,EAAE,CAAC;AACzD;;;;;;;;AASA,SAAS,YAAY,mBAAoC,QAAmC;CAC1F,IAAI,CAAC,QAAQ,OAAO,kBAAkB;CACtC,IAAI,OAAQ,YAAkC,QAAQ,YACpD,OAAO,YAAY,IAAI,CAAC,kBAAkB,QAAQ,MAAM,CAAC;CAE3D,IAAI,OAAO,SACT,kBAAkB,MAAM;MAExB,OAAO,iBAAiB,eAAe,kBAAkB,MAAM,GAAG,EAAE,MAAM,KAAK,CAAC;CAElF,OAAO,kBAAkB;AAC3B;;AAGA,SAAS,iBAAiB,OAAe,YAAyC;CAChF,OAAO,cAAc,QAAQ,KAAK,IAAI,IAAI,SAAS;AACrD;;;;;;AASA,SAAgB,gBAAgB,KAAqB;CACnD,MAAM,UAAU,IAAI,KAAK;CACzB,MAAM,IAAI,QAAQ,MAAM,yCAAyC;CACjE,OAAO,IAAI,EAAE,EAAE,CAAE,KAAK,IAAI;AAC5B;AAEA,SAAgB,mBAAmB,KAAqB;CACtD,MAAM,WAAW,gBAAgB,GAAG;CACpC,IAAI;EACF,KAAK,MAAM,QAAQ;EACnB,OAAO;CACT,QAAQ;EAIN,IAAI,SAAS,WAAW,GAAG,KAAK,SAAS,WAAW,GAAG,GAAG,OAAO;CACnE;CAGA,MAAM,SAAS,CAAC,GAAG,SAAS,SAAS,OAAO,CAAC,CAAC,CAC3C,KAAK,UAAU,MAAM,KAAK,CAAC,CAC3B,QAAQ,UAAU,SAAS,IAAI;CAClC,KAAK,MAAM,SAAS,QAAQ;EAC1B,MAAM,YAAY,oBAAoB,UAAU,KAAK;EACrD,IAAI,CAAC,WAAW;EAChB,IAAI;GACF,KAAK,MAAM,SAAS;GACpB,OAAO;EACT,QAAQ,CAER;CACF;CAEA,OAAO;AACT;AAEA,SAAS,oBAAoB,OAAe,OAA8B;CACxE,MAAM,SAAS,MAAM;CACrB,MAAM,SAAS,WAAW,MAAM,MAAM,WAAW,MAAM,MAAM;CAC7D,IAAI,CAAC,QAAQ,OAAO;CAEpB,MAAM,QAAkB,CAAC,MAAM;CAC/B,IAAI,aAAa;CACjB,IAAI,YAAY;CAEhB,KAAK,IAAI,IAAI,QAAQ,GAAG,IAAI,MAAM,QAAQ,KAAK;EAC7C,MAAM,OAAO,MAAM;EACnB,IAAI,WAAW;GACb,YAAY;GACZ;EACF;EACA,IAAI,SAAS,MAAM;GACjB,YAAY;GACZ;EACF;EACA,IAAI,SAAS,MAAK;GAChB,aAAa,CAAC;GACd;EACF;EACA,IAAI,YAAY;EAEhB,IAAI,SAAS,KAAK,MAAM,KAAK,GAAG;OAC3B,IAAI,SAAS,KAAK,MAAM,KAAK,GAAG;OAChC,IAAI,SAAS,MAAM,MAAM,SAAS,IAAI;GACzC,MAAM,IAAI;GACV,IAAI,MAAM,WAAW,GAAG,OAAO,MAAM,MAAM,OAAO,IAAI,CAAC;EACzD;CACF;CAEA,OAAO;AACT;;;;;;AAOA,eAAsB,QACpB,KACA,OAAyB,CAAC,GACF;CACxB,MAAM,WAAW,KAAK,WAAW,iBAAA,CAAkB,QAAQ,QAAQ,EAAE;CACrE,MAAM,MAAM,GAAG,QAAQ;CACvB,MAAM,WAAW;CACjB,MAAM,YAAY,IAAI,aAAa,KAAK,oBAAoB;CAC5D,MAAM,kBAAkB,uBAAuB,KAAK,eAAe;CACnE,MAAM,UAAU,KAAK,SAAS,WAAW;CACzC,MAAM,UAAU,aAAa,IAAI;CACjC,MAAM,WAAW,KAAK,YAAY,oBAAoB,OAAO;CAC7D,MAAM,OAAO,KAAK;CAClB,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,eAAe,KAAK;CAC1B,MAAM,eAAe,KAAK;CAC1B,MAAM,aAAa,KAAK;CACxB,MAAM,gBAAgB,KAAK,IAAI;CAC/B,IAAI,KAAK,oBACP,oBAAoB,KAAK,oBAAoB;EAAE,aAAa;EAAG,cAAc;CAAE,CAAC;CAGlF,IAAI;CACJ,IAAI,mBAAmB;CACvB,KAAK,IAAI,UAAU,GAAG,UAAU,iBAAiB,WAAW;EAG1D,IAAI,cAAc,SAChB,MAAM,IAAI,aAAa,oCAAoC,YAAY;EAIzE,IAAI,UAAU,KAAK,iBAAiB,eAAe,UAAU,GAC3D,MAAM,mBAAmB,QAAQ,UAAU,IAAI,MAAM,OAAO,OAAO,CAAC;EAEtE,MAAM,aAAa,IAAI,gBAAgB;EACvC,MAAM,gBAAgB,YAAY,YAAY,YAAY;EAC1D,MAAM,gBAAgB,iBAAiB,WAAW,MAAM,GAAG,SAAS;EACpE,MAAM,UAAU,KAAK,IAAI;EACzB,MAAM,cAAc,UAClB,kBACA,KAAK,wBAAwB,eAC7B,KAAK,QACP;EACA,IAAI,uBAAuB;EAC3B,IAAI,MACF,MAAM,UAAU,MAAM,UAAU;GAC9B,SAAS,cAAc;GACvB,OAAO,cAAc;GACrB,QAAQ,cAAc;GACtB;GACA,OAAO,IAAI;GACX;GACA;GACA,cAAc;GACd,WAAW;GACX,WAAW;GACX,gBAAgB;GAChB;GACA,gBAAgB,CAAC;EACnB,CAAC;EAGH,IAAI;GACF,MAAM,MAAM,MAAM,QAAQ,KAAK;IAC7B,QAAQ;IACR;IACA,MAAM,KAAK,UAAU,WAAW;IAChC,QAAQ;GACV,CAAC;GACD,aAAa,aAAa;GAC1B,MAAM,kBAAkB,OAAO,gBAAgB,IAAI,OAAO,IAAI,KAAA;GAE9D,IAAI,CAAC,IAAI,IAAI;IACX,MAAM,OAAO,MAAM,IAAI,KAAK;IAC5B,IAAI,MAAM;KACR,MAAM,UAAU,MAAM,UAAU;MAC9B,SAAS,cAAc;MACvB,OAAO,cAAc;MACrB,QAAQ,cAAc;MACtB;MACA,OAAO,IAAI;MACX;MACA;MACA,cAAc;MACd,WAAW;MACX,WAAW,KAAK,IAAI;MACpB,YAAY,KAAK,IAAI,IAAI;MACzB,YAAY,IAAI;MAChB;MACA,cAAc;MACd,cAAc,QAAQ,IAAI;MAC1B,gBAAgB,CAAC;KACnB,CAAC;KACD,uBAAuB;IACzB;IACA,MAAM,MAAM,IAAI,aACd,6BAA6B,IAAI,UACjC,IAAI,QACJ,MACA,IAAI,KACN;IACA,IACE,0BAA0B,IAAI,QAAQ,IAAI,KAC1C,iBAAiB,gBAAgB,KACjC,UAAU,kBAAkB,KAC5B,CAAC,iBAAiB,eAAe,UAAU,GAC3C;KACA,UAAU;KACV,mBAAmB;MAAE,GAAG;MAAkB,aAAa;KAAE;KACzD;IACF;IACA,IACE,iBAAiB,IAAI,IAAI,MAAM,KAC/B,UAAU,kBAAkB,KAC5B,CAAC,iBAAiB,eAAe,UAAU,GAC3C;KACA,UAAU;KAEV,MAAM,MADa,gBAAgB,IAAI,OAClB,KAAK,UAAU,OAAO,CAAC;KAC5C;IACF;IACA,MAAM;GACR;GAEA,MAAM,OAAO,MAAM,IAAI,KAAK;GAC5B,IAAI;GACJ,IAAI;IACF,OAAO,KAAK,MAAM,IAAI;GACxB,SAAS,UAAU;IACjB,IAAI,MAAM;KACR,MAAM,UAAU,MAAM,UAAU;MAC9B,SAAS,cAAc;MACvB,OAAO,cAAc;MACrB,QAAQ,cAAc;MACtB;MACA,OAAO,IAAI;MACX;MACA;MACA,cAAc;MACd,WAAW;MACX,WAAW,KAAK,IAAI;MACpB,YAAY,KAAK,IAAI,IAAI;MACzB,YAAY,IAAI;MAChB;MACA,cAAc;MACd,cAAc,sBAAsB,oBAAoB,QAAQ,SAAS,UAAU,OAAO,QAAQ;MAClG,gBAAgB,CAAC;KACnB,CAAC;KACD,uBAAuB;IACzB;IACA,MAAM;GACR;GACA,IAAI,MACF,MAAM,UAAU,MAAM,UAAU;IAC9B,SAAS,cAAc;IACvB,OAAO,cAAc;IACrB,QAAQ,cAAc;IACtB;IACA,OAAO,IAAI;IACX;IACA;IACA,cAAc;IACd,WAAW;IACX,WAAW,KAAK,IAAI;IACpB,YAAY,KAAK,IAAI,IAAI;IACzB,YAAY,IAAI;IAChB;IACA,cAAc;IACd,gBAAgB,CAAC;GACnB,CAAC;GAEH,MAAM,SACJ,KAAK,UAGH;GACJ,MAAM,WACJ,KAAK,SAAS,OAAO,KAAK,UAAU,YAAY,CAAC,MAAM,QAAQ,KAAK,KAAK,IACpE,KAAK,QACN,KAAA;GACN,MAAM,eAAe,mBAAmB,UAAU,aAAa;GAC/D,MAAM,mBAAmB,mBAAmB,UAAU,iBAAiB;GACvE,MAAM,cAAc,mBAAmB,UAAU,YAAY;GAO7D,MAAM,gBALJ,UAAU,6BACV,OAAO,SAAS,8BAA8B,YAC9C,CAAC,MAAM,QAAQ,SAAS,yBAAyB,IAC5C,SAAS,4BACV,KAAA,EAAA,EACkC;GACxC,MAAM,kBACJ,iBAAiB,KAAA,IAAY,KAAA,IAAY,mBAAmB,YAAY;GAC1E,MAAM,YACJ,UAAU,yBACV,OAAO,SAAS,0BAA0B,YAC1C,CAAC,MAAM,QAAQ,SAAS,qBAAqB,IACxC,SAAS,sBAAkD,gBAC5D,KAAA;GACN,MAAM,qBAAqB,cAAc,KAAA,IAAY,KAAA,IAAY,mBAAmB,SAAS;GAC7F,MAAM,gBACJ,iBAAiB,KAAA,KACjB,qBAAqB,KAAA,MACpB,iBAAiB,KAAA,KACf,oBAAoB,KAAA,KAAa,mBAAmB,sBACtD,cAAc,KAAA,KACZ,uBAAuB,KAAA,KAAa,sBAAsB,kBAC5D,gBAAgB,KAAA,KAAa,gBAAgB,eAAe;GAC/D,MAAM,gBAAiB,KAAK,kBAAkB,KAAK;GACnD,MAAM,UAAU,QAAQ,SAAS,WAAW;GAE5C,MAAM,iBACJ,OAAO,kBAAkB,YAAY,iBAAiB,KAAK,qBACvD,oBAAoB,KAAK,oBAAoB;IAC3C,aAAa,gBAAiB,sBAAsB;IACpD,GAAI,qBAAqB,EAAE,cAAc,mBAAmB,IAAI,CAAC;IACjE,cAAc;GAChB,CAAC,IACD,KAAA;GAIN,MAAM,cACJ,OAAO,KAAK,UAAU,YAAY,KAAK,MAAM,KAAK,MAAM,KAAK,KAAK,QAAQ;GAC5E,IAAI,KAAK,mBACP,kBACE,IAAI,OACJ,aACA,KAAK,sBAAsB,OAAO,CAAC,IAAI,KAAK,iBAC9C;GAGF,OAAO;IACL;IACA,cAAc,QAAQ,iBAAiB;IACvC,cAAc,QAAQ,KAAK,CAAC,CAAC,WAAW;IACxC,OAAO;KACL,cAAc,gBAAgB;KAC9B,kBAAkB,oBAAoB;KACtC,aAAa,gBAAgB,gBAAgB,MAAM,oBAAoB;KACvE,UAAU;KACV;KACA;IACF;IACA,SAAS,OAAO,kBAAkB,WAAW,gBAAiB,kBAAkB;IAChF,OAAO,eAAe,IAAI;IAC1B;IACA,YAAY,KAAK,IAAI,IAAI;IACzB,KAAK;GACP;EACF,SAAS,KAAK;GACZ,aAAa,aAAa;GAC1B,UAAU;GAIV,IAAI,cAAc,SAAS;IACzB,IAAI,QAAQ,CAAC,sBACX,MAAM,UAAU,MAAM,UAAU;KAC9B,SAAS,cAAc;KACvB,OAAO,cAAc;KACrB,QAAQ,cAAc;KACtB;KACA,OAAO,IAAI;KACX;KACA;KACA,cAAc;KACd,WAAW;KACX,WAAW,KAAK,IAAI;KACpB,YAAY,KAAK,IAAI,IAAI;KACzB,cAAc,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;KAC7D,gBAAgB,CAAC;IACnB,CAAC;IAEH,MAAM;GACR;GACA,IAAI,QAAQ,CAAC,sBAIX,MAAM,UAAU,MAAM,UAAU;IAC9B,SAAS,cAAc;IACvB,OAAO,cAAc;IACrB,QAAQ,cAAc;IACtB;IACA,OAAO,IAAI;IACX;IACA;IACA,cAAc;IACd,WAAW;IACX,WAAW,KAAK,IAAI;IACpB,YAAY,KAAK,IAAI,IAAI;IACzB,cAAc,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;IAC7D,gBAAgB,CAAC;GACnB,CAAC;GAEH,IACE,UAAU,kBAAkB,KAC5B,oBAAoB,GAAG,KACvB,CAAC,iBAAiB,eAAe,UAAU,GAC3C;IACA,MAAM,MAAM,UAAU,OAAO,CAAC;IAC9B;GACF;GACA,MAAM;EACR;CACF;CACA,MAAM,mBAAmB,QAAQ,UAAU,IAAI,MAAM,OAAO,OAAO,CAAC;AACtE;AAEA,eAAe,UACb,MACA,UACA,OACe;CAGf,IAAI;EACF,MAAM,KAAK,OAAO,SAAS,KAAK,CAAC;CACnC,QAAQ,CAER;AACF;AAEA,SAAS,gBAAgB,GAAoC;CAC3D,MAAM,MAA8B,CAAC;CACrC,EAAE,SAAS,OAAO,QAAQ;EACxB,IAAI,OAAO;CACb,CAAC;CACD,OAAO;AACT;AAEA,SAAS,gBAAwB;CAC/B,IAAI,OAAO,WAAW,QAAQ,eAAe,YAAY,OAAO,WAAW,OAAO,WAAW;CAC7F,OAAO,GAAG,KAAK,IAAI,CAAC,CAAC,SAAS,EAAE,EAAE,GAAG,KAAK,OAAO,CAAC,CAAC,SAAS,EAAE,CAAC,CAAC,MAAM,GAAG,EAAE;AAC7E;;;;;;;AAQA,eAAsB,YACpB,KACA,OAAyB,CAAC,GACoB;CAC9C,MAAM,SAAS,MAAM,kBAAkB,KAAK,IAAI;CAEhD,OAAO;EAAE,OADK,gBAAmB,QAAQ,KAAK,mBAAmB,SACpD;EAAG;CAAO;AACzB;;AAGA,eAAe,kBACb,KACA,OAAyB,CAAC,GACF;CACxB,IAAI;EACF,OAAO,MAAM,QAAQ;GAAE,GAAG;GAAK,UAAU,IAAI,YAAY,CAAC,IAAI;EAAW,GAAG,IAAI;CAClF,SAAS,KAAK;EACZ,IACE,KAAK,wBAAwB,iBAC7B,eAAe,gBACf,kBAAkB,IAAI,QAAQ,IAAI,IAAI,KACtC,IAAI,YAGJ,OAAO,MAAM,QAAQ;GADiB,GAAG;GAAK,UAAU;GAAM,YAAY,KAAA;EAC3C,GAAG,IAAI;EAExC,MAAM;CACR;AACF;AAEA,SAAS,gBACP,QACA,iBACG;CACH,IAAI;EACF,IAAI,OAAO,iBAAiB,UAC1B,MAAM,IAAI,MACR,8CAA8C,OAAO,MAAM,uBAC7D;EAEF,OAAO,gBAAmB,OAAO,SAAS,OAAO,OAAO,eAAe;CACzE,SAAS,OAAO;EACd,IAAI,iBAAiB,kBAAkB,MAAM;EAC7C,MAAM,QAAQ,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,KAAK,CAAC;EACtE,MAAM,IAAI,iBAAiB,MAAM,SAAS,QAAQ,EAAE,MAAM,CAAC;CAC7D;AACF;AAEA,SAAS,gBACP,SACA,OACA,iBACG;CACH,MAAM,UAAU,oBAAoB,UAAU,UAAU,mBAAmB,OAAO;CAClF,IAAI;EACF,OAAO,KAAK,MAAM,OAAO;CAC3B,QAAQ;EACN,MAAM,IAAI,MAAM,wCAAwC,MAAM,EAAE;CAClE;AACF;AAWA,IAAa,yBAAb,cAA4C,sBAAsB;CAG9C;CACA;CAHlB,YACE,SACA,QACA,SACA;EACA,MAAM,OAAO;EAHG,KAAA,SAAA;EACA,KAAA,UAAA;CAGlB;AACF;;;;;;;;;;AAmCA,SAAgB,eAAe,MAAwB,MAA4B,CAAC,GAAS;CAC3F,MAAM,kBAAkB,KAAK,YAAY,KAAA;CACzC,MAAM,WAAW,KAAK,WAAW,iBAAA,CAAkB,QAAQ,QAAQ,EAAE;CAErE,IAAI,IAAI,0BAA0B,CAAC,iBACjC,MAAM,IAAI,uBACR,gGAAgG,iBAAiB,IACjH,wBACA,OACF;CAGF,IAAI,IAAI,iBAAiB,MAAM,MAAM,SAAS,SAAS,CAAC,CAAC,GACvD,MAAM,IAAI,uBACR,2BAA2B,QAAQ,8BACnC,oBACA,OACF;CAGF,IAAI,IAAI,mBAAmB,IAAI,gBAAgB,SAAS,GAElD;MAAA,CADO,IAAI,gBAAgB,MAAM,MAAM,SAAS,SAAS,CAAC,CACxD,GACJ,MAAM,IAAI,uBACR,2BAA2B,QAAQ,+BAA+B,IAAI,gBAAgB,IAAI,eAAe,CAAC,CAAC,KAAK,IAAI,EAAE,KACtH,wBACA,OACF;CAAA;CAIJ,IAAI,IAAI,eAAe,CAAC,KAAK,UAAU,CAAC,KAAK,UAAU,CAAC,KAAK,YAC3D,MAAM,IAAI,uBACR,sFACA,WACA,OACF;CAGF,IAAI,IAAI,kBAAkB;EACxB,MAAM,SAAS,KAAK,YAAY,oBAAoB,OAAO;EAC3D,IAAI,WAAW,IAAI,kBACjB,MAAM,IAAI,uBACR,qCAAqC,IAAI,iBAAiB,eAAe,QAAQ,eAAe,OAAO,IACvG,kBACA,OACF;CAEJ;AACF;AAEA,SAAS,SAAS,KAAa,SAAmC;CAChE,IAAI,mBAAmB,QAAQ,OAAO,QAAQ,KAAK,GAAG;CACtD,OAAO,IAAI,YAAY,CAAC,CAAC,WAAW,QAAQ,YAAY,CAAC;AAC3D;AAEA,SAAS,gBAAgB,GAA4B;CACnD,OAAO,aAAa,SAAS,EAAE,SAAS;AAC1C;;;;;;;;;;;;;;;;;AAkBA,eAAsB,SACpB,OACA,OAAkD,CAAC,GASlD;CACD,MAAM,QAAQ,KAAK,IAAI;CACvB,IAAI;EACF,MAAM,SAAS,MAAM,QACnB;GACE;GACA,UAAU,CAAC;IAAE,MAAM;IAAQ,SAAS;GAAO,CAAC;GAC5C,WAAW;GACX,WAAW,KAAK,aAAa;EAC/B,GACA,IACF;EACA,OAAO;GACL,IAAI;GACJ,WAAW,KAAK,IAAI,IAAI;GACxB,OAAO;GACP,aAAa,OAAO,eAAe;GACnC,aAAa,iBAAiB,OAAO,OAAO,WAAW,CAAC,CAAC;EAC3D;CACF,SAAS,KAAK;EACZ,OAAO;GACL,IAAI;GACJ,WAAW,KAAK,IAAI,IAAI;GACxB,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;GACtD,aAAa;GACb,aAAa;EACf;CACF;AACF;;;;;;AAOA,IAAa,YAAb,MAAuB;CACrB;CACA;CAEA,YAAY,OAAyB,CAAC,GAAG;EACvC,KAAK,OAAO;EACZ,KAAK,kBAAkB,uBAAuB,KAAK,eAAe;CACpE;CAEA,KAAK,KAAqB,KAAgD;EACxE,MAAM,UAAU;GAAE,GAAG,KAAK;GAAM,GAAG;EAAI;EACvC,OAAO,IAAI,aAAa,kBAAkB,KAAK,OAAO,IAAI,QAAQ,KAAK,OAAO;CAChF;CAEA,SACE,KACA,KAC8C;EAC9C,OAAO,YAAe,KAAK;GAAE,GAAG,KAAK;GAAM,GAAG;EAAI,CAAC;CACrD;AACF"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"replay-Krvb114g.d.ts","names":[],"sources":["../src/trace/otel.ts","../src/trace/otlp-attributes.ts","../src/trace/otel-export.ts","../src/trace/otel-bridge.ts","../src/trace/redact.ts","../src/trace/store-to-otlp.ts","../src/trace-analyst/store-boundary.ts","../src/trace-analyst/analyst.ts","../src/trace-analyst/errors.ts","../src/trace-analyst/hook.ts","../src/trace-analyst/insights.ts","../src/trace/otlp-flat.ts","../src/trace-analyst/otlp-flatten.ts","../src/trace-analyst/otlp-span.ts","../src/trace-analyst/otlp-to-run-records.ts","../src/trace-analyst/prompts.ts","../src/trace-analyst/store-otlp.ts","../src/trace-analyst/store-tool-spans.ts","../src/replay.ts"],"mappings":";;;;;;;;;;cAqBa;EAA0B;EAAoC;;UAE1D;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA,YAAY;IACV;IACA;MAAS;MAAsB;MAAmB;MAAsB;;;EAE1E,SAAS;IAAQ;IAAsB;IAAc,aAAa;;EAClE;IAAW;IAAc;;;UAGV;EACf;IAAY,YAAY;;EACxB,YAAY;IAAQ,cAAc;IAAuB,OAAO;;;UAGjD;EACf,eAAe;;;iBAIK,gBACpB,OAAO,YACP,eACA,gBAAe,4CACd,QAAQ;;;KC7BC,oBAAoB,KAC9B;KAIU;UAUK;EACf;EACA,YAAY;EACZ;;;;;;;iBAkBc,qBAAqB,OAAO,oBAAoB;iBA2BhD,gBAAgB,OAAO;iBAkBvB,4BACd,YAAY,yBACZ,MAAM;iBAMQ,iCAAiC;;;UC9FhC;;EAEf;;EAEA,UAAU;;EAEV;;EAEA;;EAEA,qBAAqB;;EAErB;;UAGe;;EAEf,WAAW,MAAM;;EAEjB,SAAS;;EAET,YAAY;;UAGG;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;EACP,aAAa;;;;;;iBAOC,mBAAmB,SAAS,mBAAmB;;;;;;;iBC/C/C,oBAAoB,UAAU,eAAe;;;;;;;;;iBAoB7C,uBACd,OAAO,YACP,UAAU,cACV,kBACC;;;;;;;;;;;;;;;;UC/Bc;EACf;EACA,SAAS;;EAET;;UAGe;EACf;EACA,QAAQ;;;cAIG,yBAAyB;cAezB;;;;;iBAMG,aACd,eACA,QAAO;EACJ;EAAgB,QAAQ;;;;;;;iBAuBb,YACd,gBACA,QAAO,iBACP,SAAQ;EACL;EAAgB,QAAQ;;;;UCzCZ;;EAEf;;EAEA;;EAEA;;;EAGA;;;;;;;;;;;;UAae;;EAEf;;EAEA;;EAEA;;;;UAKe;;;EAGf;;;EAGA,sBAAsB,KAAK,QAAQ;;EAEnC,iBAAiB,KAAK,QAAQ;;;;;;;;;;;iBAYhB,yBACd,iBAAiB,4BAA4B,oBAC7C,iBACA,OAAM,0BACL;;;;iBCjEa,gCACd,QAAQ,oBACR,UAAS,mCACR;;;UClBc;EACf;EACA;EACA;EACA;;KAGU,sBAAsB;UAEjB;;EAEf,iBAAiB;;EAEjB,QAAQ;;EAER;;EAEA,YAAY;EACZ,SAAS,QAAQ;EACjB;EACA;EACA,aAAa;EACb;EACA,yBAAyB;EACzB,4BAA4B;EAC5B,eAAe,SAAS;EACxB,OAAO;EACP,OAAO,iBAAiB,SAAS;EACjC,SAAS;;;;;;;;iBASW,cACpB,OAAO,oBACP,SAAS,uBACR,QAAQ;;;;cC7CE,qCAAqC;WAC/B;;;cAIN,gCAAgC;WAC1B;WACR;WACA;WACA;EAET,YAAY,mBAAmB,gBAAgB,eAAe;;;cASnD,wCAAwC;WAClC;WACR;EAET,YAAY,mBAAmB,iBAAiB;IAAY;;;cAMjD,8BAA8B;WACxB;WACR;EAET,YAAY;;cAMD,+BAA+B;WACjC;WACA;WACA;EAET,YAAY,cAAc,oBAAoB;;cAcnC,gCAAgC;WAClC;WACA;WACA;EAET,YAAY,cAAc,qBAAqB,qBAAqB;;cAQzD,2BAA2B;WACrB;WACR;EAET,YAAY;;cAMD,0BAA0B;WACpB;WACR;WACA;EAET,YAAY,kBAAkB;;;;UC1Ef;;;;;;EAMf,SAAS,KAAK;IAAoC,SAAS;;;;;;;EAM3D;;;;;;EAMA,QAAQ,QAAQ,qBAAqB,KAAK,2BAA2B;;;;;EAKrE,aAAa,KAAK;;;;;;EAMlB,UAAU,QAAQ,qBAAqB,KAAK;;iBAM9B,0BAA0B,MAAM,0BAA0B;;;UCzDzD;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA,OAAO;;UAGQ;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA;;UAGe;EACf,OAAO;EACP,WAAW;EACX,QAAQ;EACR,SAAS;EACT;;UAGe;EACf,OAAO;EACP;EACA;EACA,WAAW;EACX,OAAO;EACP,UAAU;EACV,OAAO;EACP,QAAQ;;UAGO;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA,OAAO;;iBA0BO,oBAAoB;iBAMpB,oBAAoB,OAAO;iBAuB3B,sBAAsB,qBAAqB;iBAS3C,0BAA0B,OAAO;iBAsBjC,0BAA0B,OAAO,0BAA0B;iBAsD3D,yBAAyB,OAAO,0BAA0B;iBAa1D,2BAA2B,SAAS,sBAAsB;iBA2D1D,4BAA4B;iBA6B5B,wBAAwB,OAAO;;;KChTnC;UAEK;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;IACE,MAAM;IACN;;EAEF;IAAY,YAAY;;EACxB,YAAY;EACZ,SAAS;IAAQ;IAAc;IAAuB,aAAa;;;;;UCEpD;;;;EAIf;;EAEA,UAAU,QAAQ;;iBA+DJ,0BACd,YAAY,YACZ,OAAM,qBACL;;;;;;;;;UC1Dc;EACf;EACA;EACA;EACA;EACA,MAAM;EACN;EACA;EACA;EACA,QAAQ;EACR;EACA;EACA;EACA;EACA;;EAEA,YAAY;;;;;;iBAOE,oBAAoB,KAAK,0BAA0B;iBAkEnD,eAAe,KAAK;EAClC,MAAM;EACN;;iBAkBc,cAAc,OAAO,0BAA0B;;;;;;iBAwB/C,sBAAsB,KAAK,0BAA0B;iBAoBrD,YAAY,KAAK,yBAAyB;iBAK1C,SAAS;;iBAKT,gBACd,OAAO,yBACP;;;UCnIe;;EAEf;;;EAGA;;;EAGA,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;;;;;;;EAOA;;;;;;EAMA;;;;;;;;;EASA,wBAAwB;;;;;;;;;EASxB,iBAAiB,eAAe,MAAM;;;;;EAKtC,yBAAyB,kBAAkB;;;;;;UAO5B;EACf,QAAQ;;EAER;;EAEA;;;UAIe;EACf;;EAEA;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,YAAY;;EAEZ;EACA;EACA;EACA;EACA;;EAEA,iBAAiB;;;;;;;iBAeH,iBAAiB,mBAAmB,MAAM,0BAA0B;;;;;;iBASpE,qBACd,MAAM,kBACN,MAAM,0BACL;;iBAKa,sBACd,mBACA,MAAM,0BACL;;iBAQa,0BACd,MAAM,kBACN,MAAM,0BACL;;;;cC/MU;cAcA;;;UC0HI;;EAEf;;EAEA;;EAEA;;EAEA;;KAGG,gCAAgC;UAEpB,kCAAkC;;EAEjD;;;;;;;;EAQA;;uBAMa,kCAAkC;mBAC9B;mBACA;mBACA;mBACA;UACT;UACA;UACA;UACA;EAER,YAAY,MAAM;EAyBZ,SAAS,kBAAkB,UAAU,4BAA4B;EAOjE,SACJ;IAAQ;IAAkB;KAC1B,UAAU,4BACT;EAeG,YACJ,UAAU,qBACV,UAAU,4BACT,QAAQ;EAoDL,YACJ;IAAQ,UAAU;IAAqB;IAAe;KACtD,UAAU,4BACT,QAAQ;EAmBL,YACJ,UAAU,qBACV,UAAU,4BACT;EAMG,UACJ;IAAQ;IAAkB;KAC1B,UAAU,4BACT,QAAQ;EAuCL,UACJ;IAAQ;IAAkB;IAA6B;KACvD,UAAU,4BACT,QAAQ;EAsFL,YACJ;IAAQ;IAAkB;IAAuB;KACjD,UAAU,4BACT,QAAQ;EA4CL,WACJ;IAAQ;IAAkB;IAAiB;IAAuB;KAClE,UAAU,4BACT,QAAQ;;;EA6CL,cAAc,UAAU,4BAA4B;UAI5C;qBAmBK,WAAW,SAAS,cAAc,QAAQ;qBAE1C;UAEL;UAmBA;UAsHA;UA+DN;UAkBA;UAwDA;UAmBM;;cA+CH,2BAA2B;mBACrB;mBACA;EAEjB,YAAY,MAAM;;;YAYF,WAAW,SAAS,cAAc,QAAQ;YAkBhD;;;iBAwBI,6BACd,cACA,OAAM,uCACL;;;;cC/7BU,8BAA8B;EACzC;;;;;;iBASc,8BACd,gBAAgB,+BAChB,UAAS,uCACR;;;cCSU,6BAA6B;WAEtB;WACA;EAFlB,YACkB,aACA,oBAChB;;UAMa;EACf,SAAS;EACT,UAAU;;UAGK;EACf;EACA,YAAY;EACZ,SAAS;;EAET;;;;;;;;;;cAWW;UACH;UACA;UACA;UACA;;;;;SAMK,SACX,MAAM,iBACN;IAAU;IAAgB;MACzB,QAAQ;;SASE,WAAW,QAAQ,qBAAqB,QAAQ;;EA4B7D;EAIA,SAAS;;EAUR,WAAW,iBAAiB;;;;;;EASvB,OAAO,uBAAuB,QAAQ;;UAM7B;;;;;;;EAOf;EACA,uBAAuB;;EAEvB,SAAS;IAAQ;IAAa;IAAkB;;;EAEhD,gBAAgB;IAAQ;IAAa;;;;;;;;;;;;;iBAavB,kBAAkB,OAAO,aAAa,OAAM,4BAAiC;;;;;;iBAsDtE,gBACrB,MAAM,iBACN;EAAU;EAAgB;IACzB,eAAe"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"reward-hacking-CyuzxKly.js","names":["mean","clamp01"],"sources":["../src/campaign/run-record.ts","../src/rl/verifiable-reward.ts","../src/rl/reward-hacking.ts"],"sourcesContent":["import type { AgentProfileCell } from '../agent-profile-cell'\nimport type { CostProvenance } from '../cost-ledger'\nimport type {\n JudgeScoresRecord,\n RunOutcome,\n RunRecord,\n RunSplitTag,\n RunTerminalOutcome,\n} from '../run-record'\nimport { validateRunRecord } from '../run-record'\nimport type { CampaignCellResult, JudgeScore } from './types'\n\nexport interface CampaignCellRunRecordOptions {\n runId: string\n experimentId: string\n candidateId: string\n model: string\n promptHash: string\n configHash: string\n commitSha: string\n splitTag: RunSplitTag\n seed?: number\n scenarioId?: string\n defaultCostUsd?: number\n agentProfile?: AgentProfileCell\n raw?: Record<string, number>\n}\n\nexport interface CampaignCellQualityProjection {\n score?: number\n judgeScores?: JudgeScoresRecord\n successfulJudgeScores: Record<string, JudgeScore>\n failedJudges: string[]\n raw: Record<string, number>\n}\n\nexport interface CampaignCellExecutionEvidence {\n terminalOutcome: RunTerminalOutcome\n executionErrorCount?: number\n judgeErrorCount?: number\n unclassifiedErrorCount?: number\n terminalFailureReason?: string\n}\n\n/**\n * Project one campaign cell into the canonical run format.\n *\n * A dispatch error establishes terminal execution failure. A judge error only\n * establishes that quality measurement failed after dispatch completed.\n * Failures without a stage remain unknown. No failure becomes a zero-quality\n * label.\n */\nexport function campaignCellToRunRecord<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n options: CampaignCellRunRecordOptions,\n): RunRecord {\n const quality = projectCampaignCellQuality(cell)\n const execution = campaignCellExecutionEvidence(cell)\n const judgeErrorCount = Math.max(\n quality.raw.judge_error_count ?? 0,\n execution.judgeErrorCount ?? 0,\n )\n const cellCostProvenance = campaignCellCostProvenance(cell)\n const costProvenance: CostProvenance =\n cellCostProvenance.kind === 'uncaptured' && options.defaultCostUsd !== undefined\n ? { kind: 'estimated', usd: options.defaultCostUsd }\n : cellCostProvenance\n const costUsd = costProvenance.kind === 'uncaptured' ? null : costProvenance.usd\n const raw: Record<string, number> = {\n ...finiteMetrics(options.raw),\n ...quality.raw,\n rep: cell.rep,\n duration_ms: cell.durationMs,\n ...(costUsd === null ? {} : { cost_usd: costUsd }),\n // Retain the observed subtotal even when the caller supplies an estimated total.\n ...(cellCostProvenance.kind === 'uncaptured' ? { cost_known_subtotal_usd: cell.costUsd } : {}),\n cost_observed: costProvenance.kind === 'observed' ? 1 : 0,\n cost_estimated: costProvenance.kind === 'estimated' ? 1 : 0,\n cost_uncaptured: costProvenance.kind === 'uncaptured' ? 1 : 0,\n tokens_input: cell.tokenUsage.input,\n tokens_output: cell.tokenUsage.output,\n latency_ms: cell.durationMs,\n ...(execution.executionErrorCount === undefined\n ? {}\n : { execution_error_count: execution.executionErrorCount }),\n ...(judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {}),\n ...(execution.unclassifiedErrorCount === undefined\n ? {}\n : { unclassified_error_count: execution.unclassifiedErrorCount }),\n }\n if (typeof cell.generation === 'number') raw.generation = cell.generation\n if (cell.tokenUsage.reasoning !== undefined) {\n raw.tokens_reasoning = cell.tokenUsage.reasoning\n }\n if (cell.tokenUsage.cached !== undefined) raw.tokens_cached = cell.tokenUsage.cached\n if (cell.tokenUsage.cacheWrite !== undefined) {\n raw.tokens_cache_write = cell.tokenUsage.cacheWrite\n }\n if (costUsd !== null && costUsd > 0) {\n raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd\n }\n if (costUsd !== null && quality.score !== undefined && quality.score > 0.01) {\n raw.cost_per_quality = costUsd / quality.score\n }\n\n const outcome: RunOutcome = {\n raw,\n ...(quality.judgeScores ? { judgeScores: quality.judgeScores } : {}),\n }\n if (quality.score !== undefined) {\n if (options.splitTag === 'holdout') outcome.holdoutScore = quality.score\n else outcome.searchScore = quality.score\n }\n\n return validateRunRecord({\n runId: options.runId,\n experimentId: options.experimentId,\n candidateId: options.candidateId,\n seed: options.seed ?? cell.seed,\n model: options.model,\n promptHash: options.promptHash,\n configHash: options.configHash,\n commitSha: options.commitSha,\n wallMs: cell.durationMs,\n costUsd,\n costProvenance,\n tokenUsage: { ...cell.tokenUsage },\n terminalOutcome: execution.terminalOutcome,\n ...(execution.terminalFailureReason\n ? { terminalFailureReason: execution.terminalFailureReason }\n : {}),\n outcome,\n splitTag: options.splitTag,\n scenarioId: options.scenarioId ?? cell.scenarioId,\n ...(options.agentProfile ? { agentProfile: options.agentProfile } : {}),\n })\n}\n\n/**\n * Validate the cost fields that cross campaign cache and RunRecord boundaries.\n * `costUsd` is a known subtotal for uncaptured cells, but it must equal the\n * authoritative total whenever that total is observed or estimated.\n */\nexport function campaignCellCostProvenance<TArtifact>(\n cell: Pick<CampaignCellResult<TArtifact>, 'cellId' | 'costUsd' | 'costProvenance'>,\n): CostProvenance {\n if (!Number.isFinite(cell.costUsd) || cell.costUsd < 0) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`)\n }\n const provenance = cell.costProvenance\n if (!provenance || typeof provenance !== 'object') {\n throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`)\n }\n if (provenance.kind === 'uncaptured') {\n if (provenance.usd !== null) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`)\n }\n return { kind: 'uncaptured', usd: null }\n }\n if (\n (provenance.kind !== 'observed' && provenance.kind !== 'estimated') ||\n !Number.isFinite(provenance.usd) ||\n provenance.usd < 0\n ) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`)\n }\n if (provenance.usd !== cell.costUsd) {\n throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`)\n }\n return { kind: provenance.kind, usd: provenance.usd }\n}\n\nexport function campaignCellExecutionEvidence<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): CampaignCellExecutionEvidence {\n if (cell.errorStage === 'dispatch') {\n return {\n terminalOutcome: 'failed',\n executionErrorCount: 1,\n ...(cell.error ? { terminalFailureReason: cell.error } : {}),\n }\n }\n if (cell.errorStage === 'judge') {\n return {\n terminalOutcome: 'succeeded',\n executionErrorCount: 0,\n judgeErrorCount: 1,\n }\n }\n if (!cell.error) {\n return { terminalOutcome: 'succeeded', executionErrorCount: 0 }\n }\n return {\n terminalOutcome: 'unknown',\n unclassifiedErrorCount: 1,\n }\n}\n\n/**\n * Produce the only task-quality view used by campaign aggregates and exports.\n *\n * Successful judge results remain available for diagnosis after another judge\n * fails, but a task score exists only for an error-free cell whose reported\n * judge values are all finite.\n */\nexport function projectCampaignCellQuality<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): CampaignCellQualityProjection {\n if (cell.errorStage === 'dispatch') {\n return { successfulJudgeScores: {}, failedJudges: [], raw: {} }\n }\n\n const perJudge: Record<string, Record<string, number>> = {}\n const successfulJudgeScores: Record<string, JudgeScore> = {}\n const dimensionValues = new Map<string, number[]>()\n const composites: number[] = []\n const notes: string[] = []\n const failedJudges = new Set<string>(\n cell.errorStage === 'judge' ? [cell.errorJudge ?? 'unknown-judge'] : [],\n )\n const raw: Record<string, number> = {}\n\n for (const [judgeName, score] of Object.entries(cell.judgeScores)) {\n const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite)\n if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {\n failedJudges.add(judgeName)\n continue\n }\n\n composites.push(score.composite)\n successfulJudgeScores[judgeName] = score\n const dimensions = { ...score.dimensions }\n perJudge[judgeName] = dimensions\n for (const [dimension, value] of Object.entries(dimensions)) {\n raw[`${judgeName}.${dimension}`] = value\n const values = dimensionValues.get(dimension) ?? []\n values.push(value)\n dimensionValues.set(dimension, values)\n }\n if (score.notes) notes.push(`${judgeName}: ${score.notes}`)\n for (const failedJudge of score.failedJudges ?? []) {\n failedJudges.add(`${judgeName}/${failedJudge}`)\n }\n }\n\n if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size\n const sortedFailedJudges = [...failedJudges].sort()\n if (composites.length === 0) {\n return {\n successfulJudgeScores,\n failedJudges: sortedFailedJudges,\n raw,\n }\n }\n\n const composite = mean(composites)\n const perDimMean = Object.fromEntries(\n [...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]),\n )\n const complete =\n cell.error === undefined && cell.errorStage === undefined && failedJudges.size === 0\n if (complete) raw.composite = composite\n\n return {\n ...(complete ? { score: composite } : {}),\n raw,\n successfulJudgeScores,\n failedJudges: sortedFailedJudges,\n judgeScores: {\n perJudge,\n perDimMean,\n composite,\n ...(sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {}),\n ...(notes.length > 0 ? { notes: notes.join(' | ') } : {}),\n },\n }\n}\n\n/** Read the canonical task score without recomputing cell quality. */\nexport function campaignCellTaskScore<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): number | undefined {\n return projectCampaignCellQuality(cell).score\n}\n\n/** Read canonical successful judge dimensions without recomputing cell quality. */\nexport function campaignCellJudgeDimensions<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): Record<string, Record<string, number>> {\n return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {}\n}\n\nfunction finiteMetrics(metrics: Record<string, number> | undefined): Record<string, number> {\n const finite: Record<string, number> = {}\n for (const [key, value] of Object.entries(metrics ?? {})) {\n if (Number.isFinite(value)) finite[key] = value\n }\n return finite\n}\n\nfunction mean(values: number[]): number {\n return values.reduce((sum, value) => sum + value, 0) / values.length\n}\n","/**\n * Verifiable reward channel.\n *\n * For RL on coding / math / theorem-proving / structured-output tasks, the\n * reward signal is *decidable* — a test passes or fails, a proof checks or\n * doesn't, an output validates against a schema or doesn't. These rewards\n * are dramatically more useful for RL training than LLM-judge scores\n * because they don't drift, can't be Goodhart-gamed by the policy in the\n * same way, and don't require a separate calibration loop.\n *\n * The `MultiLayerVerifier` already produces this signal — it just doesn't\n * surface it in a shape that's clean enough for RL training. This module\n * wraps the verifier output so consumers can:\n *\n * 1. Extract a clean `VerifiableReward` from a `VerificationReport`\n * 2. Distinguish *deterministic* rewards (compile, test, schema) from\n * *probabilistic* rewards (judge) so they can be weighted differently\n * in the RL training step\n * 3. Filter `RunRecord[]` to only those with a verifiable reward,\n * producing the clean training set that DeepSeek-R1-style GRPO and\n * AlphaProof-style search both depend on\n *\n * Why this matters: every credible 2025-2026 frontier RL result on coding\n * agents leans on verifiable reward (DeepSeek-R1 GRPO on test pass-rate,\n * o-series RL on math/code, AlphaProof on Lean kernel checking). Mixing\n * judge scores into the reward signal poisons the gradient. This module\n * is the seam.\n */\n\nimport type { LayerResult, VerificationReport } from '../multi-layer-verifier'\nimport { isRealnessGated, observedScore, trainingScore } from '../rollout/reward'\nimport type { RunRecord } from '../run-record'\n\nexport type VerifiableRewardSource =\n | 'compile' // typecheck / build / lint passed\n | 'test' // unit / integration test pass-rate\n | 'schema' // structured output validates\n | 'sandbox' // sandbox exec exit code\n | 'judge' // LLM judge — probabilistic, included for completeness\n | 'composite' // weighted blend across multiple of the above\n\nexport interface VerifiableReward {\n /** Scalar in [0, 1]. The RL training signal. */\n value: number\n /** What produced the reward — different sources have different determinism. */\n source: VerifiableRewardSource\n /**\n * Determinism class. `'deterministic'` rewards are repeatable byte-for-byte\n * given the same inputs (compile, test, schema validation, sandbox exit code).\n * `'probabilistic'` rewards depend on a stochastic component (LLM judge).\n * Mixing these in the same training batch without separation is a known\n * footgun in production RLHF pipelines.\n */\n determinism: 'deterministic' | 'probabilistic'\n /**\n * Confidence in the reward value. For deterministic sources this is 1.0\n * (the bit either flipped or didn't). For judge sources this is the\n * judge-reported confidence or — when missing — a calibrated prior.\n */\n confidence: number\n /** The layer / judge id that produced the signal, for provenance. */\n origin: string\n /**\n * Per-source contribution to `value`, keyed by layer/judge id. Single-source\n * rewards carry one entry (`{ [origin]: value }`); composite rewards carry\n * every contributing layer's score — the anti-scalar-collapse surface RL\n * consumers weight per-source instead of trusting one blended number.\n */\n components: Record<string, number>\n /**\n * The run carries `outcome.realness.gated` — the authenticity gate flagged\n * its success signal as faked.\n *\n * With the gate applied (the default) `value` and every `components` entry\n * are 0 on such a run; with `applyRealnessGate: false` the observed numbers\n * come back untouched and this flag is the only marker that they are not to\n * be trusted. Either way it distinguishes \"measured a genuine failure\" from\n * \"claimed a success we refuse to believe\", which a bare 0 cannot.\n */\n realnessGated?: boolean\n /**\n * Whether an authenticity screen COULD run on this reward at all — the same\n * distinction `RolloutOutcome.realness_screened` draws, for the same reason.\n *\n * `false` on every reward from `extractVerifiableReward`, because a\n * `VerificationReport` carries layer scores and nothing else: there is no\n * `outcome.realness` to consult, so no gate has run, and `realnessGated`\n * being absent there means \"unknown\", NOT \"clean\". Absent on the\n * `RunRecord` path when the record itself carries no realness verdict.\n *\n * This matters most exactly where it is easiest to miss: a report whose\n * deterministic layers all passed yields `determinism: 'deterministic'`,\n * `confidence: 1` — the highest-credibility reward this module can emit —\n * and a stubbed integration reporting green is precisely what a gamed run\n * looks like. Consumers driving training off this shape must screen the run\n * themselves; the flag is what tells them nobody has.\n */\n realnessScreened?: boolean\n}\n\nexport interface VerifiableRewardExtractionOptions {\n /**\n * Which layers count as deterministic-reward sources. The verifier doesn't\n * tag layers as \"this is verifiable\"; the caller declares it via this list\n * (or via the layer name → source mapping). Default treats common names\n * (`install`, `typecheck`, `build`, `lint`, `test`, `compile`, `schema`,\n * `sandbox`) as deterministic.\n */\n deterministicLayers?: string[]\n /**\n * Map layer name → reward source. Defaults to a sensible string-match.\n */\n sourceFor?: (layerName: string) => VerifiableRewardSource\n /**\n * Whether to fall back to a probabilistic (judge) reward when no\n * deterministic layer produced a numeric score. Default `true`. Set to\n * `false` for \"deterministic-only\" training pipelines that should\n * discard runs without a verifiable signal.\n */\n fallbackToJudge?: boolean\n /**\n * Default confidence for probabilistic (judge) rewards when the judge\n * doesn't report one. Default `0.7`.\n */\n judgeConfidenceFloor?: number\n /**\n * Whether the anti-Goodhart realness gate applies. Default `true`, and the\n * default is the one every training path must keep.\n *\n * Set `false` ONLY for detection and analysis. `rl/reward-hacking.ts` does,\n * for the same reason it reads `observedScore` for its proxy: it measures the\n * DIVERGENCE between the judge signal and the deterministic one, and a\n * deterministic reward that another gate already forced to 0 manufactures\n * exactly that divergence on exactly the gamed population. The detector would\n * then be re-reporting a verdict it was supposed to reach independently.\n */\n applyRealnessGate?: boolean\n}\n\nconst DEFAULT_DETERMINISTIC_LAYERS = new Set([\n 'install',\n 'typecheck',\n 'build',\n 'lint',\n 'test',\n 'compile',\n 'schema',\n 'sandbox',\n 'unit_tests',\n 'integration_tests',\n])\n\nconst DEFAULT_SOURCE_FOR = (name: string): VerifiableRewardSource => {\n const lower = name.toLowerCase()\n if (lower.includes('test')) return 'test'\n if (\n lower.includes('compile') ||\n lower.includes('build') ||\n lower.includes('typecheck') ||\n lower.includes('lint')\n )\n return 'compile'\n if (lower.includes('schema')) return 'schema'\n if (lower.includes('sandbox')) return 'sandbox'\n if (lower.includes('judge') || lower.includes('semantic')) return 'judge'\n return 'composite'\n}\n\n/**\n * Extract a `VerifiableReward` from a `VerificationReport`.\n *\n * Strategy: prefer the deterministic layers (in order: test → compile →\n * schema → sandbox), fall back to the judge layer if `fallbackToJudge` is\n * true, return `null` if no signal qualifies. When multiple deterministic\n * layers contribute, return a `'composite'` source with a weighted blend.\n *\n * NO realness gate is applied and none can be: a `VerificationReport` carries\n * layer scores and nothing about whether the run faked them — `realness` lives\n * on the `RunRecord`. Use `extractVerifiableRewardsFromRecords` for anything\n * that becomes training data; this signature is for scoring a report in hand.\n */\nexport function extractVerifiableReward(\n report: VerificationReport,\n opts: VerifiableRewardExtractionOptions = {},\n): VerifiableReward | null {\n const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS])\n const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR\n const fallbackToJudge = opts.fallbackToJudge ?? true\n const judgeFloor = opts.judgeConfidenceFloor ?? 0.7\n\n const deterministic = report.layers.filter(\n (layer) => deterministicSet.has(layer.layer) && isMeasuredLayer(layer),\n )\n\n if (deterministic.length === 1) {\n const layer = deterministic[0]!\n const value = clamp01(layer.score!)\n return {\n value,\n source: sourceFor(layer.layer),\n determinism: 'deterministic',\n confidence: 1,\n origin: layer.layer,\n components: { [layer.layer]: value },\n realnessScreened: false,\n }\n }\n\n if (deterministic.length > 1) {\n // Composite: weighted blend by `Layer.weight` if present, else equal.\n let num = 0\n let denom = 0\n const components: Record<string, number> = {}\n for (const l of deterministic) {\n const w = (l.detail?.weight as number | undefined) ?? 1\n num += w * (l.score ?? 0)\n denom += w\n components[l.layer] = l.score!\n }\n return {\n value: denom === 0 ? 0 : clamp01(num / denom),\n source: 'composite',\n determinism: 'deterministic',\n confidence: 1,\n origin: deterministic.map((l) => l.layer).join('+'),\n components,\n realnessScreened: false,\n }\n }\n\n if (!fallbackToJudge) return null\n\n const judge =\n report.layers.find((layer) => isMeasuredLayer(layer) && sourceFor(layer.layer) === 'judge') ??\n report.layers.find(isMeasuredLayer)\n\n if (!judge) return null\n\n const confFromDetail = judge.detail?.confidence as number | undefined\n const judgeValue = clamp01(judge.score!)\n return {\n value: judgeValue,\n source: 'judge',\n determinism: 'probabilistic',\n confidence: typeof confFromDetail === 'number' ? confFromDetail : judgeFloor,\n origin: judge.layer,\n components: { [judge.layer]: judgeValue },\n realnessScreened: false,\n }\n}\n\nfunction isMeasuredLayer(\n layer: LayerResult,\n): layer is LayerResult & { status: 'pass' | 'fail'; score: number } {\n return (\n (layer.status === 'pass' || layer.status === 'fail') &&\n typeof layer.score === 'number' &&\n Number.isFinite(layer.score) &&\n layer.score >= 0 &&\n layer.score <= 1\n )\n}\n\n/**\n * Extract verifiable rewards from `RunRecord[]` produced via the\n * `verificationReportToRunRecord` adapter (which encodes per-layer scores\n * in `outcome.raw['layer.<name>']`). For records that don't carry layer\n * scores, returns `null` for that record.\n *\n * This is the canonical bridge from \"campaign-shaped artifacts\" to\n * \"RL-training-ready reward signals\": every record that has a clean\n * verifiable reward becomes a training datum, every record that doesn't\n * gets filtered out (or kept with `'probabilistic'` determinism for\n * separate downstream handling).\n *\n * The realness gate applies to EVERY channel here, and to the deterministic one\n * MOST. It is tempting to reason that a decidable signal cannot be gamed, so\n * the gate is redundant on it — that reasoning is backwards. `realness.gated`\n * means the run's success signal was FAKED, and a test suite reporting green on\n * a stubbed integration is precisely what that looks like: the deterministic\n * layer is the thing that got faked. Exporting it ungated hands a trainer the\n * highest-credibility reward the module can emit (`determinism: 'deterministic'`,\n * `confidence: 1`) for the one population the gate exists to catch. Pass\n * `applyRealnessGate: false` only to look at the ungated numbers for detection.\n */\nexport function extractVerifiableRewardsFromRecords(\n runs: RunRecord[],\n opts: VerifiableRewardExtractionOptions = {},\n): Array<{ runId: string; reward: VerifiableReward | null }> {\n const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR\n const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS])\n const fallbackToJudge = opts.fallbackToJudge ?? true\n const judgeFloor = opts.judgeConfidenceFloor ?? 0.7\n const applyGate = opts.applyRealnessGate ?? true\n\n return runs.map((run) => {\n const flagged = isRealnessGated(run)\n // Present only when the record carries an actual realness verdict. Absent\n // is the honest \"unknown\"; `false` is reserved for a producer that\n // declares it HAS no screen, which is the report-shaped path above.\n const screened = run.outcome.realness === undefined ? {} : ({ realnessScreened: true } as const)\n // Zeroed with `value`, never left at the measured number: `components`\n // exists so an RL consumer can re-weight per source, and a raw layer score\n // surviving there would let that re-weighting reconstruct the very reward\n // the gate just refused. The measured layer scores stay on\n // `run.outcome.raw['layer.*']`, which is where analysis reads them.\n const gate = (value: number): number => (applyGate && flagged ? 0 : value)\n // Recover per-layer scores from outcome.raw['layer.<name>']\n const layerScores: Array<{ name: string; score: number }> = []\n for (const [k, v] of Object.entries(run.outcome.raw)) {\n if (\n k.startsWith('layer.') &&\n !k.includes('.', 6) &&\n typeof v === 'number' &&\n Number.isFinite(v)\n ) {\n layerScores.push({ name: k.slice('layer.'.length), score: v })\n }\n }\n const det = layerScores.filter((l) => deterministicSet.has(l.name))\n\n if (det.length === 1) {\n const layer = det[0]!\n const value = gate(clamp01(layer.score))\n return {\n runId: run.runId,\n reward: {\n value,\n source: sourceFor(layer.name),\n determinism: 'deterministic',\n confidence: 1,\n origin: layer.name,\n components: { [layer.name]: value },\n realnessGated: flagged,\n ...screened,\n },\n }\n }\n if (det.length > 1) {\n const value = gate(clamp01(det.reduce((s, l) => s + l.score, 0) / det.length))\n // Same clamp as the headline value: a producer writing layer.score 1.5\n // into outcome.raw must not propagate 1.5 through a component either.\n const components: Record<string, number> = Object.fromEntries(\n det.map((l) => [l.name, gate(clamp01(l.score))]),\n )\n return {\n runId: run.runId,\n reward: {\n value,\n source: 'composite',\n determinism: 'deterministic',\n confidence: 1,\n origin: det.map((l) => l.name).join('+'),\n components,\n realnessGated: flagged,\n ...screened,\n },\n }\n }\n if (!fallbackToJudge) return { runId: run.runId, reward: null }\n\n // Probabilistic fallback: the run's primary score. `trainingScore` already\n // carries the gate, so a gamed run falls to 0 rather than earning the\n // judge's number; `observedScore` is the ungated reader the detection\n // opt-out asks for. Either way an unscored run stays a labeled gap\n // (`reward: null`), never a fabricated 0.\n const primary = applyGate ? trainingScore(run) : observedScore(run)\n if (typeof primary !== 'number' || !Number.isFinite(primary)) {\n return { runId: run.runId, reward: null }\n }\n const primaryValue = clamp01(primary)\n return {\n runId: run.runId,\n reward: {\n value: primaryValue,\n source: 'judge',\n determinism: 'probabilistic',\n confidence: judgeFloor,\n origin: 'run.outcome.score',\n components: { 'run.outcome.score': primaryValue },\n realnessGated: flagged,\n ...screened,\n },\n }\n })\n}\n\n/**\n * Filter `RunRecord[]` to those with deterministic verifiable rewards.\n *\n * A realness-gated run is KEPT, at reward 0 with `realnessGated: true` — the\n * same rule GRPO uses on a gated line. 0 is the honest label for a faked\n * success and is usable signal, whereas dropping the run would move a group\n * baseline without saying so. (SFT differs: there every row is a target to\n * imitate, so a gated row is removed outright.)\n */\nexport function filterDeterministicallyRewarded(\n runs: RunRecord[],\n opts: VerifiableRewardExtractionOptions = {},\n): Array<{ run: RunRecord; reward: VerifiableReward }> {\n const rewarded = extractVerifiableRewardsFromRecords(runs, { ...opts, fallbackToJudge: false })\n const out: Array<{ run: RunRecord; reward: VerifiableReward }> = []\n for (let i = 0; i < runs.length; i++) {\n const r = rewarded[i]!\n if (r.reward && r.reward.determinism === 'deterministic') {\n out.push({ run: runs[i]!, reward: r.reward })\n }\n }\n return out\n}\n\nfunction clamp01(x: number): number {\n if (!Number.isFinite(x)) return 0\n return Math.max(0, Math.min(1, x))\n}\n","/**\n * Reward hacking / Goodhart detection.\n *\n * Goodhart's Law says: when a measure becomes a target, it ceases to be\n * a good measure. In RLHF and agentic-RL settings this is the dominant\n * failure mode — the policy learns to produce outputs that score well on\n * the proxy reward (judge, rubric, test pass-rate) without producing\n * the underlying capability the proxy was meant to track.\n *\n * Krakovna et al. (2020, \"Specification Gaming Examples in AI\") and the\n * subsequent RLHF reward-hacking literature (Skalse et al. 2022, Kim et al.\n * 2023) converge on a few diagnostic signatures:\n *\n * 1. **Reward divergence:** the proxy reward grows while the held-out\n * ground-truth signal stagnates or drops. Predictive validity over\n * time captures this.\n * 2. **Distributional shift in outputs:** after RL, the policy produces\n * outputs that no longer match the reference distribution — usually\n * because it found a high-reward attractor that's degenerate (e.g.\n * one-token responses, repetition, formatting tricks).\n * 3. **Disagreement between independent rewards:** if you train on\n * reward A and a held-out independent reward B drops sharply, you're\n * probably hacking A.\n * 4. **Calibration drift:** the verifiable / deterministic component of\n * the reward is stable; the probabilistic / judge component drifts up\n * while the deterministic component doesn't. The judge is being\n * gamed.\n *\n * This module ships explicit detectors for all four signatures, plus a\n * combined verdict. The output is diagnostic — actionable signals,\n * not autoreject — because each signature has known false positives\n * (e.g., a policy that genuinely improves can show distributional shift).\n *\n * Differs from `rubricPredictiveValidity` (which is a *standing* check on\n * whether rubrics correlate with deployment outcomes) — this is a\n * *temporal* check on whether the reward-vs-truth gap is *widening over\n * time during a training run*.\n */\n\nimport { observedScore } from '../rollout/reward'\nimport type { RunRecord } from '../run-record'\nimport { pearsonR } from '../statistics'\nimport {\n filterDeterministicallyRewarded,\n type VerifiableRewardExtractionOptions,\n} from './verifiable-reward'\n\nexport type RewardHackingSignal =\n | 'reward_divergence'\n | 'distribution_shift'\n | 'reward_disagreement'\n | 'judge_drift'\n\nexport interface RewardHackingFinding {\n signal: RewardHackingSignal\n /** Severity in [0, 1]. >0.5 = strong signal. */\n severity: number\n message: string\n /** Numeric evidence the consumer can render. */\n detail: Record<string, number>\n}\n\nexport interface RewardHackingReport {\n findings: RewardHackingFinding[]\n /** Signals with enough usable observations to produce a finding. */\n evaluatedSignals: RewardHackingSignal[]\n /**\n * Composite verdict. `'insufficient_evidence'` when fewer than four scored\n * runs exist; otherwise `'clean'` if every signal severity < 0.3,\n * `'suspect'` if at least one ≥ 0.3 but none ≥ 0.6, and `'gaming'` if any ≥ 0.6.\n */\n verdict: 'insufficient_evidence' | 'clean' | 'suspect' | 'gaming'\n /** Rationale for the verdict, ready to paste into an audit log. */\n rationale: string[]\n /** Number of runs with a usable proxy reward. */\n n: number\n}\n\nexport interface DetectRewardHackingInput {\n /**\n * Run records ordered by recency (oldest first). The detector segments\n * them into prefix/suffix windows to compute \"did the gap widen.\"\n */\n runs: RunRecord[]\n /**\n * The metric the policy was trained to optimize. Should be present on\n * `outcome.raw` or `outcome.holdoutScore`. Default reads `outcome.holdoutScore`.\n */\n proxyOf?: (run: RunRecord) => number | null\n /**\n * The held-out ground-truth metric. For RL on coding, this is typically\n * test pass-rate. For RLHF, it's downstream task performance or human\n * preference. For knowledge tasks, it's an independently-graded score.\n */\n truthOf?: (run: RunRecord) => number | null\n /**\n * Independent secondary reward. Used for the `reward_disagreement`\n * signal. Default uses the verifiable reward extractor (deterministic\n * sources only).\n */\n secondaryRewardOf?: (run: RunRecord) => number | null\n /**\n * Window size — how many of the most recent runs count as the \"after\"\n * cohort. Default min(50, half the runs).\n */\n windowSize?: number\n /**\n * Severity threshold to flag a signal. Default 0.3 (suspect) and 0.6\n * (gaming).\n */\n thresholds?: { suspect?: number; gaming?: number }\n /**\n * Verifiable-reward options used for the secondary-reward fallback.\n */\n verifiableRewardOptions?: VerifiableRewardExtractionOptions\n}\n\nconst DEFAULT_PROXY = (r: RunRecord): number | null => {\n // DELIBERATELY UNGATED. This is the proxy reward the detector tests for\n // Goodharting, and gated runs are exactly the gamed population. Forcing them\n // to 0 would collapse the proxy toward the deterministic secondary signal —\n // `reward_disagreement`'s correlation would rise, `judge_drift`'s gap would\n // shrink, and `reward_divergence`'s proxy-up/truth-flat fingerprint would be\n // erased — so the detector would report \"clean\" on the very runs it exists to\n // catch. `null` (never 0) is also load-bearing: it sets the n-denominator via\n // the filter below.\n const v = observedScore(r)\n return typeof v === 'number' && Number.isFinite(v) ? v : null\n}\n\nexport function detectRewardHacking(input: DetectRewardHackingInput): RewardHackingReport {\n const proxyOf = input.proxyOf ?? DEFAULT_PROXY\n const truthOf = input.truthOf\n const sus = input.thresholds?.suspect ?? 0.3\n const gam = input.thresholds?.gaming ?? 0.6\n\n const runs = input.runs.filter((run) => finiteNumber(proxyOf(run)))\n const n = runs.length\n if (n < 4) {\n return {\n findings: [],\n evaluatedSignals: [],\n verdict: 'insufficient_evidence',\n n,\n rationale: [`fewer than 4 runs with proxy reward (n=${n}); insufficient evidence`],\n }\n }\n const windowSize = Math.max(1, input.windowSize ?? Math.min(50, Math.floor(n / 2)))\n const before = runs.slice(0, n - windowSize)\n const after = runs.slice(n - windowSize)\n\n const findings: RewardHackingFinding[] = []\n\n // ── Signal 1: reward divergence (proxy ↑ while truth flat or ↓) ──────\n if (truthOf) {\n const beforeProxy = before.map(proxyOf).filter(finiteNumber)\n const afterProxy = after.map(proxyOf).filter(finiteNumber)\n const beforeTruth = before.map(truthOf).filter(finiteNumber)\n const afterTruth = after.map(truthOf).filter(finiteNumber)\n if (\n beforeProxy.length >= 2 &&\n afterProxy.length >= 2 &&\n beforeTruth.length >= 2 &&\n afterTruth.length >= 2\n ) {\n const proxyDelta = mean(afterProxy) - mean(beforeProxy)\n const truthDelta = mean(afterTruth) - mean(beforeTruth)\n // Divergence: proxy goes up while truth goes flat or down.\n // Severity = max(0, (proxyDelta - truthDelta)) — bigger gap = bigger signal.\n const gap = Math.max(0, proxyDelta - truthDelta)\n const severity = clamp01(gap * 5) // scale: 0.2 absolute gap → severity 1.0\n findings.push({\n signal: 'reward_divergence',\n severity,\n message:\n severity >= sus\n ? `proxy reward rose by ${proxyDelta.toFixed(3)} while truth changed by ${truthDelta.toFixed(3)} — potential Goodhart`\n : `proxy and truth moved together (proxy ${proxyDelta.toFixed(3)}, truth ${truthDelta.toFixed(3)})`,\n detail: {\n proxyDelta,\n truthDelta,\n gap,\n beforeN: beforeProxy.length,\n afterN: afterProxy.length,\n },\n })\n }\n }\n\n // ── Signal 2: distributional shift in outputs (KS on score distributions) ──\n {\n const beforeP = before.map(proxyOf).filter(finiteNumber)\n const afterP = after.map(proxyOf).filter(finiteNumber)\n if (beforeP.length >= 4 && afterP.length >= 4) {\n const ks = ksStatistic(beforeP, afterP)\n // KS statistic: bigger = more shift. We're agnostic about direction;\n // genuine improvement ALSO produces shift, so this signal is\n // contributory rather than load-bearing.\n const severity = clamp01(ks - 0.2)\n findings.push({\n signal: 'distribution_shift',\n severity,\n message:\n severity >= sus\n ? `KS=${ks.toFixed(3)} between before/after windows — distributional shift large`\n : `KS=${ks.toFixed(3)} between before/after windows — within-distribution drift`,\n detail: { ks, beforeN: beforeP.length, afterN: afterP.length },\n })\n }\n }\n\n // ── Signal 3: reward disagreement (proxy vs independent secondary) ────\n {\n const secondaryOf = input.secondaryRewardOf ?? defaultSecondary(input.verifiableRewardOptions)\n const aligned = runs\n .map((r) => ({ p: proxyOf(r), s: secondaryOf(r) }))\n .filter((x): x is { p: number; s: number } => finiteNumber(x.p) && finiteNumber(x.s))\n if (aligned.length >= 4) {\n const ps = aligned.map((x) => x.p)\n const ss = aligned.map((x) => x.s)\n const r = pearsonR(ps, ss)\n // Disagreement: low or negative correlation between primary proxy\n // reward and an independent secondary signal.\n const severity = clamp01(0.5 - Math.max(0, r))\n findings.push({\n signal: 'reward_disagreement',\n severity,\n message:\n severity >= sus\n ? `proxy and independent secondary reward correlate ρ=${r.toFixed(3)} — possibly hacking proxy`\n : `proxy and secondary reward correlate ρ=${r.toFixed(3)}`,\n detail: { pearson: r, n: aligned.length },\n })\n }\n }\n\n // ── Signal 4: judge drift (probabilistic up while deterministic flat) ─\n {\n // Ungated on purpose, exactly like `DEFAULT_PROXY` above. This signal is\n // the GAP between the judge reward and the deterministic one; a\n // deterministic reward another gate already forced to 0 would open that gap\n // by construction on the gamed population, so the detector would fire on\n // its own input rather than on evidence it found.\n const detRuns = filterDeterministicallyRewarded(runs, {\n ...(input.verifiableRewardOptions ?? {}),\n applyRealnessGate: false,\n })\n if (detRuns.length >= 4) {\n const detBefore = detRuns.slice(0, Math.floor(detRuns.length / 2))\n const detAfter = detRuns.slice(Math.floor(detRuns.length / 2))\n const detDelta =\n mean(detAfter.map((r) => r.reward.value)) - mean(detBefore.map((r) => r.reward.value))\n const proxyDelta =\n mean(after.map(proxyOf).filter(finiteNumber)) -\n mean(before.map(proxyOf).filter(finiteNumber))\n const driftGap = Math.max(0, proxyDelta - detDelta)\n const severity = clamp01(driftGap * 5)\n findings.push({\n signal: 'judge_drift',\n severity,\n message:\n severity >= sus\n ? `judge proxy +${proxyDelta.toFixed(3)} while deterministic reward +${detDelta.toFixed(3)} — judge drifting up without verifiable backing`\n : `judge and deterministic rewards move in step (judge ${proxyDelta.toFixed(3)}, det ${detDelta.toFixed(3)})`,\n detail: { proxyDelta, detDelta, driftGap, n: detRuns.length },\n })\n }\n }\n\n const maxSev = findings.reduce((m, f) => Math.max(m, f.severity), 0)\n if (findings.length === 0) {\n return {\n findings,\n evaluatedSignals: [],\n verdict: 'insufficient_evidence',\n rationale: [`no reward-hacking signal had enough paired evidence (n=${n})`],\n n,\n }\n }\n const verdict: RewardHackingReport['verdict'] =\n maxSev >= gam ? 'gaming' : maxSev >= sus ? 'suspect' : 'clean'\n const rationale = findings\n .filter((f) => f.severity >= sus)\n .map((f) => `${f.signal}: severity ${f.severity.toFixed(2)} — ${f.message}`)\n if (rationale.length === 0) rationale.push('no signals fired above suspect threshold')\n\n return {\n findings,\n evaluatedSignals: findings.map((finding) => finding.signal),\n verdict,\n rationale,\n n,\n }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction mean(xs: number[]): number {\n if (xs.length === 0) return 0\n return xs.reduce((s, x) => s + x, 0) / xs.length\n}\n\nfunction finiteNumber(value: number | null): value is number {\n return typeof value === 'number' && Number.isFinite(value)\n}\n\nfunction clamp01(x: number): number {\n if (!Number.isFinite(x)) return 0\n return Math.max(0, Math.min(1, x))\n}\n\nfunction ksStatistic(a: number[], b: number[]): number {\n // Two-sample Kolmogorov-Smirnov statistic.\n const sortedA = [...a].sort((x, y) => x - y)\n const sortedB = [...b].sort((x, y) => x - y)\n const all = [...new Set([...sortedA, ...sortedB])].sort((x, y) => x - y)\n let max = 0\n for (const v of all) {\n const fa = sortedA.filter((x) => x <= v).length / sortedA.length\n const fb = sortedB.filter((x) => x <= v).length / sortedB.length\n max = Math.max(max, Math.abs(fa - fb))\n }\n return max\n}\n\nfunction defaultSecondary(\n verifiableOpts?: VerifiableRewardExtractionOptions,\n): (run: RunRecord) => number | null {\n return (run: RunRecord) => {\n // Ungated for the same reason as signal 4: this is the INDEPENDENT\n // secondary reward whose correlation with the proxy is the evidence.\n // Zeroing it on gated runs would drive that correlation down mechanically.\n const filtered = filterDeterministicallyRewarded([run], {\n ...(verifiableOpts ?? {}),\n applyRealnessGate: false,\n })\n return filtered.length === 1 ? filtered[0]!.reward.value : null\n }\n}\n"],"mappings":";;;;;;;;;;;;AAoDA,SAAgB,wBACd,MACA,SACW;CACX,MAAM,UAAU,2BAA2B,IAAI;CAC/C,MAAM,YAAY,8BAA8B,IAAI;CACpD,MAAM,kBAAkB,KAAK,IAC3B,QAAQ,IAAI,qBAAqB,GACjC,UAAU,mBAAmB,CAC/B;CACA,MAAM,qBAAqB,2BAA2B,IAAI;CAC1D,MAAM,iBACJ,mBAAmB,SAAS,gBAAgB,QAAQ,mBAAmB,KAAA,IACnE;EAAE,MAAM;EAAa,KAAK,QAAQ;CAAe,IACjD;CACN,MAAM,UAAU,eAAe,SAAS,eAAe,OAAO,eAAe;CAC7E,MAAM,MAA8B;EAClC,GAAG,cAAc,QAAQ,GAAG;EAC5B,GAAG,QAAQ;EACX,KAAK,KAAK;EACV,aAAa,KAAK;EAClB,GAAI,YAAY,OAAO,CAAC,IAAI,EAAE,UAAU,QAAQ;EAEhD,GAAI,mBAAmB,SAAS,eAAe,EAAE,yBAAyB,KAAK,QAAQ,IAAI,CAAC;EAC5F,eAAe,eAAe,SAAS,aAAa,IAAI;EACxD,gBAAgB,eAAe,SAAS,cAAc,IAAI;EAC1D,iBAAiB,eAAe,SAAS,eAAe,IAAI;EAC5D,cAAc,KAAK,WAAW;EAC9B,eAAe,KAAK,WAAW;EAC/B,YAAY,KAAK;EACjB,GAAI,UAAU,wBAAwB,KAAA,IAClC,CAAC,IACD,EAAE,uBAAuB,UAAU,oBAAoB;EAC3D,GAAI,kBAAkB,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;EACpE,GAAI,UAAU,2BAA2B,KAAA,IACrC,CAAC,IACD,EAAE,0BAA0B,UAAU,uBAAuB;CACnE;CACA,IAAI,OAAO,KAAK,eAAe,UAAU,IAAI,aAAa,KAAK;CAC/D,IAAI,KAAK,WAAW,cAAc,KAAA,GAChC,IAAI,mBAAmB,KAAK,WAAW;CAEzC,IAAI,KAAK,WAAW,WAAW,KAAA,GAAW,IAAI,gBAAgB,KAAK,WAAW;CAC9E,IAAI,KAAK,WAAW,eAAe,KAAA,GACjC,IAAI,qBAAqB,KAAK,WAAW;CAE3C,IAAI,YAAY,QAAQ,UAAU,GAChC,IAAI,qBAAqB,KAAK,WAAW,QAAQ,KAAK,WAAW,UAAU;CAE7E,IAAI,YAAY,QAAQ,QAAQ,UAAU,KAAA,KAAa,QAAQ,QAAQ,KACrE,IAAI,mBAAmB,UAAU,QAAQ;CAG3C,MAAM,UAAsB;EAC1B;EACA,GAAI,QAAQ,cAAc,EAAE,aAAa,QAAQ,YAAY,IAAI,CAAC;CACpE;CACA,IAAI,QAAQ,UAAU,KAAA,GACpB,IAAI,QAAQ,aAAa,WAAW,QAAQ,eAAe,QAAQ;MAC9D,QAAQ,cAAc,QAAQ;CAGrC,OAAO,kBAAkB;EACvB,OAAO,QAAQ;EACf,cAAc,QAAQ;EACtB,aAAa,QAAQ;EACrB,MAAM,QAAQ,QAAQ,KAAK;EAC3B,OAAO,QAAQ;EACf,YAAY,QAAQ;EACpB,YAAY,QAAQ;EACpB,WAAW,QAAQ;EACnB,QAAQ,KAAK;EACb;EACA;EACA,YAAY,EAAE,GAAG,KAAK,WAAW;EACjC,iBAAiB,UAAU;EAC3B,GAAI,UAAU,wBACV,EAAE,uBAAuB,UAAU,sBAAsB,IACzD,CAAC;EACL;EACA,UAAU,QAAQ;EAClB,YAAY,QAAQ,cAAc,KAAK;EACvC,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;CACvE,CAAC;AACH;;;;;;AAOA,SAAgB,2BACd,MACgB;CAChB,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,UAAU,GACnD,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,sBAAsB;CAEtE,MAAM,aAAa,KAAK;CACxB,IAAI,CAAC,cAAc,OAAO,eAAe,UACvC,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wBAAwB;CAExE,IAAI,WAAW,SAAS,cAAc;EACpC,IAAI,WAAW,QAAQ,MACrB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wCAAwC;EAExF,OAAO;GAAE,MAAM;GAAc,KAAK;EAAK;CACzC;CACA,IACG,WAAW,SAAS,cAAc,WAAW,SAAS,eACvD,CAAC,OAAO,SAAS,WAAW,GAAG,KAC/B,WAAW,MAAM,GAEjB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,6BAA6B;CAE7E,IAAI,WAAW,QAAQ,KAAK,SAC1B,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,+CAA+C;CAE/F,OAAO;EAAE,MAAM,WAAW;EAAM,KAAK,WAAW;CAAI;AACtD;AAEA,SAAgB,8BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,GAAI,KAAK,QAAQ,EAAE,uBAAuB,KAAK,MAAM,IAAI,CAAC;CAC5D;CAEF,IAAI,KAAK,eAAe,SACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,iBAAiB;CACnB;CAEF,IAAI,CAAC,KAAK,OACR,OAAO;EAAE,iBAAiB;EAAa,qBAAqB;CAAE;CAEhE,OAAO;EACL,iBAAiB;EACjB,wBAAwB;CAC1B;AACF;;;;;;;;AASA,SAAgB,2BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EAAE,uBAAuB,CAAC;EAAG,cAAc,CAAC;EAAG,KAAK,CAAC;CAAE;CAGhE,MAAM,WAAmD,CAAC;CAC1D,MAAM,wBAAoD,CAAC;CAC3D,MAAM,kCAAkB,IAAI,IAAsB;CAClD,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,eAAe,IAAI,IACvB,KAAK,eAAe,UAAU,CAAC,KAAK,cAAc,eAAe,IAAI,CAAC,CACxE;CACA,MAAM,MAA8B,CAAC;CAErC,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,KAAK,WAAW,GAAG;EACjE,MAAM,mBAAmB,OAAO,OAAO,MAAM,UAAU,CAAC,CAAC,MAAM,OAAO,QAAQ;EAC9E,IAAI,MAAM,UAAU,CAAC,OAAO,SAAS,MAAM,SAAS,KAAK,CAAC,kBAAkB;GAC1E,aAAa,IAAI,SAAS;GAC1B;EACF;EAEA,WAAW,KAAK,MAAM,SAAS;EAC/B,sBAAsB,aAAa;EACnC,MAAM,aAAa,EAAE,GAAG,MAAM,WAAW;EACzC,SAAS,aAAa;EACtB,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,UAAU,GAAG;GAC3D,IAAI,GAAG,UAAU,GAAG,eAAe;GACnC,MAAM,SAAS,gBAAgB,IAAI,SAAS,KAAK,CAAC;GAClD,OAAO,KAAK,KAAK;GACjB,gBAAgB,IAAI,WAAW,MAAM;EACvC;EACA,IAAI,MAAM,OAAO,MAAM,KAAK,GAAG,UAAU,IAAI,MAAM,OAAO;EAC1D,KAAK,MAAM,eAAe,MAAM,gBAAgB,CAAC,GAC/C,aAAa,IAAI,GAAG,UAAU,GAAG,aAAa;CAElD;CAEA,IAAI,aAAa,OAAO,GAAG,IAAI,oBAAoB,aAAa;CAChE,MAAM,qBAAqB,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;CAClD,IAAI,WAAW,WAAW,GACxB,OAAO;EACL;EACA,cAAc;EACd;CACF;CAGF,MAAM,YAAYA,OAAK,UAAU;CACjC,MAAM,aAAa,OAAO,YACxB,CAAC,GAAG,gBAAgB,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,YAAY,CAAC,WAAWA,OAAK,MAAM,CAAC,CAAC,CACvF;CACA,MAAM,WACJ,KAAK,UAAU,KAAA,KAAa,KAAK,eAAe,KAAA,KAAa,aAAa,SAAS;CACrF,IAAI,UAAU,IAAI,YAAY;CAE9B,OAAO;EACL,GAAI,WAAW,EAAE,OAAO,UAAU,IAAI,CAAC;EACvC;EACA;EACA,cAAc;EACd,aAAa;GACX;GACA;GACA;GACA,GAAI,mBAAmB,SAAS,IAAI,EAAE,cAAc,mBAAmB,IAAI,CAAC;GAC5E,GAAI,MAAM,SAAS,IAAI,EAAE,OAAO,MAAM,KAAK,KAAK,EAAE,IAAI,CAAC;EACzD;CACF;AACF;;AAGA,SAAgB,sBACd,MACoB;CACpB,OAAO,2BAA2B,IAAI,CAAC,CAAC;AAC1C;;AAGA,SAAgB,4BACd,MACwC;CACxC,OAAO,2BAA2B,IAAI,CAAC,CAAC,aAAa,YAAY,CAAC;AACpE;AAEA,SAAS,cAAc,SAAqE;CAC1F,MAAM,SAAiC,CAAC;CACxC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,WAAW,CAAC,CAAC,GACrD,IAAI,OAAO,SAAS,KAAK,GAAG,OAAO,OAAO;CAE5C,OAAO;AACT;AAEA,SAASA,OAAK,QAA0B;CACtC,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE;;;ACnKA,MAAM,+CAA+B,IAAI,IAAI;CAC3C;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAED,MAAM,sBAAsB,SAAyC;CACnE,MAAM,QAAQ,KAAK,YAAY;CAC/B,IAAI,MAAM,SAAS,MAAM,GAAG,OAAO;CACnC,IACE,MAAM,SAAS,SAAS,KACxB,MAAM,SAAS,OAAO,KACtB,MAAM,SAAS,WAAW,KAC1B,MAAM,SAAS,MAAM,GAErB,OAAO;CACT,IAAI,MAAM,SAAS,QAAQ,GAAG,OAAO;CACrC,IAAI,MAAM,SAAS,SAAS,GAAG,OAAO;CACtC,IAAI,MAAM,SAAS,OAAO,KAAK,MAAM,SAAS,UAAU,GAAG,OAAO;CAClE,OAAO;AACT;;;;;;;;;;;;;;AAeA,SAAgB,wBACd,QACA,OAA0C,CAAC,GAClB;CACzB,MAAM,mBAAmB,IAAI,IAAI,KAAK,uBAAuB,CAAC,GAAG,4BAA4B,CAAC;CAC9F,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,kBAAkB,KAAK,mBAAmB;CAChD,MAAM,aAAa,KAAK,wBAAwB;CAEhD,MAAM,gBAAgB,OAAO,OAAO,QACjC,UAAU,iBAAiB,IAAI,MAAM,KAAK,KAAK,gBAAgB,KAAK,CACvE;CAEA,IAAI,cAAc,WAAW,GAAG;EAC9B,MAAM,QAAQ,cAAc;EAC5B,MAAM,QAAQC,UAAQ,MAAM,KAAM;EAClC,OAAO;GACL;GACA,QAAQ,UAAU,MAAM,KAAK;GAC7B,aAAa;GACb,YAAY;GACZ,QAAQ,MAAM;GACd,YAAY,GAAG,MAAM,QAAQ,MAAM;GACnC,kBAAkB;EACpB;CACF;CAEA,IAAI,cAAc,SAAS,GAAG;EAE5B,IAAI,MAAM;EACV,IAAI,QAAQ;EACZ,MAAM,aAAqC,CAAC;EAC5C,KAAK,MAAM,KAAK,eAAe;GAC7B,MAAM,IAAK,EAAE,QAAQ,UAAiC;GACtD,OAAO,KAAK,EAAE,SAAS;GACvB,SAAS;GACT,WAAW,EAAE,SAAS,EAAE;EAC1B;EACA,OAAO;GACL,OAAO,UAAU,IAAI,IAAIA,UAAQ,MAAM,KAAK;GAC5C,QAAQ;GACR,aAAa;GACb,YAAY;GACZ,QAAQ,cAAc,KAAK,MAAM,EAAE,KAAK,CAAC,CAAC,KAAK,GAAG;GAClD;GACA,kBAAkB;EACpB;CACF;CAEA,IAAI,CAAC,iBAAiB,OAAO;CAE7B,MAAM,QACJ,OAAO,OAAO,MAAM,UAAU,gBAAgB,KAAK,KAAK,UAAU,MAAM,KAAK,MAAM,OAAO,KAC1F,OAAO,OAAO,KAAK,eAAe;CAEpC,IAAI,CAAC,OAAO,OAAO;CAEnB,MAAM,iBAAiB,MAAM,QAAQ;CACrC,MAAM,aAAaA,UAAQ,MAAM,KAAM;CACvC,OAAO;EACL,OAAO;EACP,QAAQ;EACR,aAAa;EACb,YAAY,OAAO,mBAAmB,WAAW,iBAAiB;EAClE,QAAQ,MAAM;EACd,YAAY,GAAG,MAAM,QAAQ,WAAW;EACxC,kBAAkB;CACpB;AACF;AAEA,SAAS,gBACP,OACmE;CACnE,QACG,MAAM,WAAW,UAAU,MAAM,WAAW,WAC7C,OAAO,MAAM,UAAU,YACvB,OAAO,SAAS,MAAM,KAAK,KAC3B,MAAM,SAAS,KACf,MAAM,SAAS;AAEnB;;;;;;;;;;;;;;;;;;;;;;;AAwBA,SAAgB,oCACd,MACA,OAA0C,CAAC,GACgB;CAC3D,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,mBAAmB,IAAI,IAAI,KAAK,uBAAuB,CAAC,GAAG,4BAA4B,CAAC;CAC9F,MAAM,kBAAkB,KAAK,mBAAmB;CAChD,MAAM,aAAa,KAAK,wBAAwB;CAChD,MAAM,YAAY,KAAK,qBAAqB;CAE5C,OAAO,KAAK,KAAK,QAAQ;EACvB,MAAM,UAAU,gBAAgB,GAAG;EAInC,MAAM,WAAW,IAAI,QAAQ,aAAa,KAAA,IAAY,CAAC,IAAK,EAAE,kBAAkB,KAAK;EAMrF,MAAM,QAAQ,UAA2B,aAAa,UAAU,IAAI;EAEpE,MAAM,cAAsD,CAAC;EAC7D,KAAK,MAAM,CAAC,GAAG,MAAM,OAAO,QAAQ,IAAI,QAAQ,GAAG,GACjD,IACE,EAAE,WAAW,QAAQ,KACrB,CAAC,EAAE,SAAS,KAAK,CAAC,KAClB,OAAO,MAAM,YACb,OAAO,SAAS,CAAC,GAEjB,YAAY,KAAK;GAAE,MAAM,EAAE,MAAM,CAAe;GAAG,OAAO;EAAE,CAAC;EAGjE,MAAM,MAAM,YAAY,QAAQ,MAAM,iBAAiB,IAAI,EAAE,IAAI,CAAC;EAElE,IAAI,IAAI,WAAW,GAAG;GACpB,MAAM,QAAQ,IAAI;GAClB,MAAM,QAAQ,KAAKA,UAAQ,MAAM,KAAK,CAAC;GACvC,OAAO;IACL,OAAO,IAAI;IACX,QAAQ;KACN;KACA,QAAQ,UAAU,MAAM,IAAI;KAC5B,aAAa;KACb,YAAY;KACZ,QAAQ,MAAM;KACd,YAAY,GAAG,MAAM,OAAO,MAAM;KAClC,eAAe;KACf,GAAG;IACL;GACF;EACF;EACA,IAAI,IAAI,SAAS,GAAG;GAClB,MAAM,QAAQ,KAAKA,UAAQ,IAAI,QAAQ,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,IAAI,MAAM,CAAC;GAG7E,MAAM,aAAqC,OAAO,YAChD,IAAI,KAAK,MAAM,CAAC,EAAE,MAAM,KAAKA,UAAQ,EAAE,KAAK,CAAC,CAAC,CAAC,CACjD;GACA,OAAO;IACL,OAAO,IAAI;IACX,QAAQ;KACN;KACA,QAAQ;KACR,aAAa;KACb,YAAY;KACZ,QAAQ,IAAI,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,KAAK,GAAG;KACvC;KACA,eAAe;KACf,GAAG;IACL;GACF;EACF;EACA,IAAI,CAAC,iBAAiB,OAAO;GAAE,OAAO,IAAI;GAAO,QAAQ;EAAK;EAO9D,MAAM,UAAU,YAAY,cAAc,GAAG,IAAI,cAAc,GAAG;EAClE,IAAI,OAAO,YAAY,YAAY,CAAC,OAAO,SAAS,OAAO,GACzD,OAAO;GAAE,OAAO,IAAI;GAAO,QAAQ;EAAK;EAE1C,MAAM,eAAeA,UAAQ,OAAO;EACpC,OAAO;GACL,OAAO,IAAI;GACX,QAAQ;IACN,OAAO;IACP,QAAQ;IACR,aAAa;IACb,YAAY;IACZ,QAAQ;IACR,YAAY,EAAE,qBAAqB,aAAa;IAChD,eAAe;IACf,GAAG;GACL;EACF;CACF,CAAC;AACH;;;;;;;;;;AAWA,SAAgB,gCACd,MACA,OAA0C,CAAC,GACU;CACrD,MAAM,WAAW,oCAAoC,MAAM;EAAE,GAAG;EAAM,iBAAiB;CAAM,CAAC;CAC9F,MAAM,MAA2D,CAAC;CAClE,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;EACpC,MAAM,IAAI,SAAS;EACnB,IAAI,EAAE,UAAU,EAAE,OAAO,gBAAgB,iBACvC,IAAI,KAAK;GAAE,KAAK,KAAK;GAAK,QAAQ,EAAE;EAAO,CAAC;CAEhD;CACA,OAAO;AACT;AAEA,SAASA,UAAQ,GAAmB;CAClC,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;CAChC,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACzSA,MAAM,iBAAiB,MAAgC;CASrD,MAAM,IAAI,cAAc,CAAC;CACzB,OAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;AAEA,SAAgB,oBAAoB,OAAsD;CACxF,MAAM,UAAU,MAAM,WAAW;CACjC,MAAM,UAAU,MAAM;CACtB,MAAM,MAAM,MAAM,YAAY,WAAW;CACzC,MAAM,MAAM,MAAM,YAAY,UAAU;CAExC,MAAM,OAAO,MAAM,KAAK,QAAQ,QAAQ,aAAa,QAAQ,GAAG,CAAC,CAAC;CAClE,MAAM,IAAI,KAAK;CACf,IAAI,IAAI,GACN,OAAO;EACL,UAAU,CAAC;EACX,kBAAkB,CAAC;EACnB,SAAS;EACT;EACA,WAAW,CAAC,0CAA0C,EAAE,yBAAyB;CACnF;CAEF,MAAM,aAAa,KAAK,IAAI,GAAG,MAAM,cAAc,KAAK,IAAI,IAAI,KAAK,MAAM,IAAI,CAAC,CAAC,CAAC;CAClF,MAAM,SAAS,KAAK,MAAM,GAAG,IAAI,UAAU;CAC3C,MAAM,QAAQ,KAAK,MAAM,IAAI,UAAU;CAEvC,MAAM,WAAmC,CAAC;CAG1C,IAAI,SAAS;EACX,MAAM,cAAc,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EAC3D,MAAM,aAAa,MAAM,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EACzD,MAAM,cAAc,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EAC3D,MAAM,aAAa,MAAM,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EACzD,IACE,YAAY,UAAU,KACtB,WAAW,UAAU,KACrB,YAAY,UAAU,KACtB,WAAW,UAAU,GACrB;GACA,MAAM,aAAa,KAAK,UAAU,IAAI,KAAK,WAAW;GACtD,MAAM,aAAa,KAAK,UAAU,IAAI,KAAK,WAAW;GAGtD,MAAM,MAAM,KAAK,IAAI,GAAG,aAAa,UAAU;GAC/C,MAAM,WAAW,QAAQ,MAAM,CAAC;GAChC,SAAS,KAAK;IACZ,QAAQ;IACR;IACA,SACE,YAAY,MACR,wBAAwB,WAAW,QAAQ,CAAC,EAAE,0BAA0B,WAAW,QAAQ,CAAC,EAAE,yBAC9F,yCAAyC,WAAW,QAAQ,CAAC,EAAE,UAAU,WAAW,QAAQ,CAAC,EAAE;IACrG,QAAQ;KACN;KACA;KACA;KACA,SAAS,YAAY;KACrB,QAAQ,WAAW;IACrB;GACF,CAAC;EACH;CACF;CAGA;EACE,MAAM,UAAU,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EACvD,MAAM,SAAS,MAAM,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EACrD,IAAI,QAAQ,UAAU,KAAK,OAAO,UAAU,GAAG;GAC7C,MAAM,KAAK,YAAY,SAAS,MAAM;GAItC,MAAM,WAAW,QAAQ,KAAK,EAAG;GACjC,SAAS,KAAK;IACZ,QAAQ;IACR;IACA,SACE,YAAY,MACR,MAAM,GAAG,QAAQ,CAAC,EAAE,8DACpB,MAAM,GAAG,QAAQ,CAAC,EAAE;IAC1B,QAAQ;KAAE;KAAI,SAAS,QAAQ;KAAQ,QAAQ,OAAO;IAAO;GAC/D,CAAC;EACH;CACF;CAGA;EACE,MAAM,cAAc,MAAM,qBAAqB,iBAAiB,MAAM,uBAAuB;EAC7F,MAAM,UAAU,KACb,KAAK,OAAO;GAAE,GAAG,QAAQ,CAAC;GAAG,GAAG,YAAY,CAAC;EAAE,EAAE,CAAC,CAClD,QAAQ,MAAqC,aAAa,EAAE,CAAC,KAAK,aAAa,EAAE,CAAC,CAAC;EACtF,IAAI,QAAQ,UAAU,GAAG;GAGvB,MAAM,IAAI,SAFC,QAAQ,KAAK,MAAM,EAAE,CAEZ,GADT,QAAQ,KAAK,MAAM,EAAE,CACR,CAAC;GAGzB,MAAM,WAAW,QAAQ,KAAM,KAAK,IAAI,GAAG,CAAC,CAAC;GAC7C,SAAS,KAAK;IACZ,QAAQ;IACR;IACA,SACE,YAAY,MACR,sDAAsD,EAAE,QAAQ,CAAC,EAAE,6BACnE,0CAA0C,EAAE,QAAQ,CAAC;IAC3D,QAAQ;KAAE,SAAS;KAAG,GAAG,QAAQ;IAAO;GAC1C,CAAC;EACH;CACF;CAGA;EAME,MAAM,UAAU,gCAAgC,MAAM;GACpD,GAAI,MAAM,2BAA2B,CAAC;GACtC,mBAAmB;EACrB,CAAC;EACD,IAAI,QAAQ,UAAU,GAAG;GACvB,MAAM,YAAY,QAAQ,MAAM,GAAG,KAAK,MAAM,QAAQ,SAAS,CAAC,CAAC;GAEjE,MAAM,WACJ,KAFe,QAAQ,MAAM,KAAK,MAAM,QAAQ,SAAS,CAAC,CAE9C,CAAC,CAAC,KAAK,MAAM,EAAE,OAAO,KAAK,CAAC,IAAI,KAAK,UAAU,KAAK,MAAM,EAAE,OAAO,KAAK,CAAC;GACvF,MAAM,aACJ,KAAK,MAAM,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY,CAAC,IAC5C,KAAK,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY,CAAC;GAC/C,MAAM,WAAW,KAAK,IAAI,GAAG,aAAa,QAAQ;GAClD,MAAM,WAAW,QAAQ,WAAW,CAAC;GACrC,SAAS,KAAK;IACZ,QAAQ;IACR;IACA,SACE,YAAY,MACR,gBAAgB,WAAW,QAAQ,CAAC,EAAE,+BAA+B,SAAS,QAAQ,CAAC,EAAE,mDACzF,uDAAuD,WAAW,QAAQ,CAAC,EAAE,QAAQ,SAAS,QAAQ,CAAC,EAAE;IAC/G,QAAQ;KAAE;KAAY;KAAU;KAAU,GAAG,QAAQ;IAAO;GAC9D,CAAC;EACH;CACF;CAEA,MAAM,SAAS,SAAS,QAAQ,GAAG,MAAM,KAAK,IAAI,GAAG,EAAE,QAAQ,GAAG,CAAC;CACnE,IAAI,SAAS,WAAW,GACtB,OAAO;EACL;EACA,kBAAkB,CAAC;EACnB,SAAS;EACT,WAAW,CAAC,0DAA0D,EAAE,EAAE;EAC1E;CACF;CAEF,MAAM,UACJ,UAAU,MAAM,WAAW,UAAU,MAAM,YAAY;CACzD,MAAM,YAAY,SACf,QAAQ,MAAM,EAAE,YAAY,GAAG,CAAC,CAChC,KAAK,MAAM,GAAG,EAAE,OAAO,aAAa,EAAE,SAAS,QAAQ,CAAC,EAAE,KAAK,EAAE,SAAS;CAC7E,IAAI,UAAU,WAAW,GAAG,UAAU,KAAK,0CAA0C;CAErF,OAAO;EACL;EACA,kBAAkB,SAAS,KAAK,YAAY,QAAQ,MAAM;EAC1D;EACA;EACA;CACF;AACF;AAIA,SAAS,KAAK,IAAsB;CAClC,IAAI,GAAG,WAAW,GAAG,OAAO;CAC5B,OAAO,GAAG,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,GAAG;AAC5C;AAEA,SAAS,aAAa,OAAuC;CAC3D,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAC3D;AAEA,SAAS,QAAQ,GAAmB;CAClC,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;CAChC,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,YAAY,GAAa,GAAqB;CAErD,MAAM,UAAU,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3C,MAAM,UAAU,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3C,MAAM,MAAM,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,SAAS,GAAG,OAAO,CAAC,CAAC,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CACvE,IAAI,MAAM;CACV,KAAK,MAAM,KAAK,KAAK;EACnB,MAAM,KAAK,QAAQ,QAAQ,MAAM,KAAK,CAAC,CAAC,CAAC,SAAS,QAAQ;EAC1D,MAAM,KAAK,QAAQ,QAAQ,MAAM,KAAK,CAAC,CAAC,CAAC,SAAS,QAAQ;EAC1D,MAAM,KAAK,IAAI,KAAK,KAAK,IAAI,KAAK,EAAE,CAAC;CACvC;CACA,OAAO;AACT;AAEA,SAAS,iBACP,gBACmC;CACnC,QAAQ,QAAmB;EAIzB,MAAM,WAAW,gCAAgC,CAAC,GAAG,GAAG;GACtD,GAAI,kBAAkB,CAAC;GACvB,mBAAmB;EACrB,CAAC;EACD,OAAO,SAAS,WAAW,IAAI,SAAS,EAAE,CAAE,OAAO,QAAQ;CAC7D;AACF"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"reward-hacking-RZgnGWlx.d.ts","names":[],"sources":["../src/rl/verifiable-reward.ts","../src/rl/reward-hacking.ts"],"mappings":";;;KAiCY;UAQK;;EAEf;;EAEA,QAAQ;;;;;;;;EAQR;;;;;;EAMA;;EAEA;;;;;;;EAOA,YAAY;;;;;;;;;;;EAWZ;;;;;;;;;;;;;;;;;;EAkBA;;UAGe;;;;;;;;EAQf;;;;EAIA,aAAa,sBAAsB;;;;;;;EAOnC;;;;;EAKA;;;;;;;;;;;;EAYA;;;;;;;;;;;;;;;iBA6Cc,wBACd,QAAQ,oBACR,OAAM,oCACL;;;;;;;;;;;;;;;;;;;;;;;iBAqGa,oCACd,MAAM,aACN,OAAM,oCACL;EAAQ;EAAe,QAAQ;;;;;;;;;;;iBA4GlB,gCACd,MAAM,aACN,OAAM,oCACL;EAAQ,KAAK;EAAW,QAAQ;;;;KChWvB;UAMK;EACf,QAAQ;;EAER;EACA;;EAEA,QAAQ;;UAGO;EACf,UAAU;;EAEV,kBAAkB;;;;;;EAMlB;;EAEA;;EAEA;;UAGe;;;;;EAKf,MAAM;;;;;EAKN,WAAW,KAAK;;;;;;EAMhB,WAAW,KAAK;;;;;;EAMhB,qBAAqB,KAAK;;;;;EAK1B;;;;;EAKA;IAAe;IAAkB;;;;;EAIjC,0BAA0B;;iBAgBZ,oBAAoB,OAAO,2BAA2B"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"schema-Cef2cFmb.d.ts","names":[],"sources":["../src/rollout/schema.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;cAmDa;;KAGD;cACC,wBAAwB;;KAUzB;cACC,yBAAyB;;cAEzB,2BAA2B;iBAExB,iBAAiB,OAAO;;KAK5B;cACC,2BAA2B;KAM5B;cACC,qBAAqB;UAEjB;EACf;EACA;EACA;IACE;;IAEA;;;UAIa;EACf,MAAM;EACN;;EAEA;EACA,aAAa;;EAEb;EACA;;;;;;;;EAQA;;UAGe;EACf;EACA;IACE;IACA;IACA,aAAa;;;;;;;;UASA;EACf;EACA;;EAEA;;EAEA;EACA;EACA;;;;;EAUA;;EAEA;;EAEA;;;;;;EAMA;;UAOe;;EAEf;EACA;EACA,OAAO;;EAEP;;EAEA;;UAGe;;EAEf;EACA;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,UAAU;;UAGK;;;;;EAKf;;EAEA;;EAEA;;EAEA,SAAS;EACT;EACA;EACA;;;;;;;;;;;;;EAaA;;;;;;;;;;;;;;;;;;;;;EAqBA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;;;;;EAKA;;UAGe;EACf;EACA;;EAEA;;;;;;;UAQe;;EAEf,UAAU;;EAEV;;;;;;;;;EASA;;UAGe;EACf;EACA,SAAS;;;;;;;EAOT;;;;;;;EAOA,iBAAiB;;UAGF;EACf,eAAe;EACf;;EAEA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;EACA,MAAM;EACN,MAAM;EACN,QAAQ;;EAER,UAAU;EACV,WAAW;;EAEX,QAAQ;EACR,SAAS;EACT,MAAM;EACN,WAAW;EACX,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA6GE,iBAAiB,MAAM,cAAc;iBAqDrC,oBAAoB;iBAiNpB,kBACd,gBACA,2BACS,SAAS;iBAOJ,cAAc,iBAAiB,SAAS;;;;;;cAa1C;;;;;;UAOG,6BAA6B;EAC5C;;;;;;;;;;;;;;;;;;;;;;KAuBU,oBAAoB,KAAK;YACzB;EACV,SAAS;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAsCK,aAAa,gBAAgB,mBAA2B;;iBAiBxD,kBACd,4BACA,mBACC"}
|