@tangle-network/agent-eval 0.203.1 → 0.203.4
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 +15 -0
- package/dist/adapters/http.d.ts +2 -2
- package/dist/analyst/index.d.ts +6 -6
- package/dist/analyst/index.js +1 -1
- package/dist/{benchmark-command-6OBcB0sG.js → benchmark-command-BpgCWzJm.js} +2 -2
- package/dist/{benchmark-command-6OBcB0sG.js.map → benchmark-command-BpgCWzJm.js.map} +1 -1
- package/dist/benchmarks/index.d.ts +4 -4
- package/dist/benchmarks/index.js +2 -2
- package/dist/campaign/index.d.ts +4 -4
- package/dist/campaign/index.js +3 -3
- package/dist/{campaign-BbObY2c6.js → campaign-DYhg4SjN.js} +4 -4
- package/dist/{campaign-BbObY2c6.js.map → campaign-DYhg4SjN.js.map} +1 -1
- package/dist/cli.js +1 -1
- package/dist/contract/index.d.ts +7 -7
- package/dist/contract/index.js +3 -3
- package/dist/{default-registry-CxEBVA6W.d.ts → default-registry-BHEfZHQd.d.ts} +3 -3
- package/dist/{default-registry-CxEBVA6W.d.ts.map → default-registry-BHEfZHQd.d.ts.map} +1 -1
- package/dist/{define-agent-eval-Da8mHyc7.js → define-agent-eval-B25Cc3ki.js} +2 -2
- package/dist/{define-agent-eval-Da8mHyc7.js.map → define-agent-eval-B25Cc3ki.js.map} +1 -1
- package/dist/{define-agent-eval-BbEXdM3H.d.ts → define-agent-eval-D8bPoBM5.d.ts} +5 -5
- package/dist/{define-agent-eval-BbEXdM3H.d.ts.map → define-agent-eval-D8bPoBM5.d.ts.map} +1 -1
- package/dist/{engine-CaSqEVlL.d.ts → engine-F_PvwesR.d.ts} +2 -2
- package/dist/{engine-CaSqEVlL.d.ts.map → engine-F_PvwesR.d.ts.map} +1 -1
- package/dist/evaluation.d.ts +1 -1
- package/dist/experiment/index.d.ts +1 -1
- package/dist/{index-Ce2HACcP.d.ts → index-B8n12TR2.d.ts} +9 -11
- package/dist/index-B8n12TR2.d.ts.map +1 -0
- package/dist/{index-CIkxRsLS.d.ts → index-C9l_-23M.d.ts} +3 -3
- package/dist/{index-CIkxRsLS.d.ts.map → index-C9l_-23M.d.ts.map} +1 -1
- package/dist/{index-CGAwP9l3.d.ts → index-CG7mtu_G.d.ts} +6 -6
- package/dist/{index-CGAwP9l3.d.ts.map → index-CG7mtu_G.d.ts.map} +1 -1
- package/dist/index.d.ts +11 -11
- package/dist/index.js +5 -5
- package/dist/jev.d.ts +1 -1
- package/dist/judge-gate-B1qYnkpi.js +658 -0
- package/dist/judge-gate-B1qYnkpi.js.map +1 -0
- package/dist/judge-gate-BuQd7sSm.d.ts +454 -0
- package/dist/judge-gate-BuQd7sSm.d.ts.map +1 -0
- package/dist/{llm-judge-EUMHOZOQ.js → llm-judge-DyZdrp6I.js} +54 -12
- package/dist/llm-judge-DyZdrp6I.js.map +1 -0
- package/dist/{matrix-BqHiMKIj.d.ts → matrix-W1lbSx72.d.ts} +2 -2
- package/dist/{matrix-BqHiMKIj.d.ts.map → matrix-W1lbSx72.d.ts.map} +1 -1
- package/dist/meta-eval/index.d.ts +2 -132
- package/dist/meta-eval/index.d.ts.map +1 -1
- package/dist/meta-eval/index.js +2 -275
- package/dist/meta-eval/index.js.map +1 -1
- package/dist/multishot/golden/index.d.ts +1 -1
- package/dist/multishot/index.d.ts +2 -2
- package/dist/openapi.json +1 -1
- package/dist/{produced-state-g9BPgu4D.js → produced-state-DLToUXrD.js} +2 -2
- package/dist/{produced-state-g9BPgu4D.js.map → produced-state-DLToUXrD.js.map} +1 -1
- package/dist/rl.d.ts +2 -2
- package/dist/search/index.d.ts +2 -2
- package/dist/{search-kernel-C5jua_UB.d.ts → search-kernel-DNpsHb8P.d.ts} +2 -2
- package/dist/{search-kernel-C5jua_UB.d.ts.map → search-kernel-DNpsHb8P.d.ts.map} +1 -1
- package/dist/{skillopt-optimization-method-DZrdQGZl.js → skillopt-optimization-method-Dz5CGZb8.js} +2 -2
- package/dist/{skillopt-optimization-method-DZrdQGZl.js.map → skillopt-optimization-method-Dz5CGZb8.js.map} +1 -1
- package/dist/{statistical-heldout-CJYAO-4a.d.ts → statistical-heldout-Dq3pnmxd.d.ts} +2 -2
- package/dist/{statistical-heldout-CJYAO-4a.d.ts.map → statistical-heldout-Dq3pnmxd.d.ts.map} +1 -1
- package/dist/{store-tool-spans-fIE31bU8.d.ts → store-tool-spans-CAF3T-BS.d.ts} +3 -3
- package/dist/{store-tool-spans-fIE31bU8.d.ts.map → store-tool-spans-CAF3T-BS.d.ts.map} +1 -1
- package/dist/{tool-groups-DP8b5kjF.d.ts → tool-groups-B7_jUPHp.d.ts} +2 -2
- package/dist/tool-groups-B7_jUPHp.d.ts.map +1 -0
- package/dist/{trace-questions-Cd0Fyhbv.d.ts → trace-questions-Dd3AMf-U.d.ts} +2 -2
- package/dist/{trace-questions-Cd0Fyhbv.d.ts.map → trace-questions-Dd3AMf-U.d.ts.map} +1 -1
- package/dist/trace-repair/index.d.ts +1 -1
- package/dist/traces.d.ts +2 -2
- package/dist/{types-hH9APahd.d.ts → types-DZQpUJse.d.ts} +2 -2
- package/dist/{types-hH9APahd.d.ts.map → types-DZQpUJse.d.ts.map} +1 -1
- package/dist/{types-Duc6xOoW.d.ts → types-ZUy5HGY5.d.ts} +9 -3
- package/dist/{types-Duc6xOoW.d.ts.map → types-ZUy5HGY5.d.ts.map} +1 -1
- package/docs/eval-surface-map.md +12 -2
- package/docs/proofs/campaign-attempt-retention-20261001/README.md +70 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-judge-attribution/attribution_0/attempts/e7949644-3d60-4ab9-817e-f80d0a48f5d8/attempt-1/artifacts/capture.json +3 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-judge-attribution/attribution_0/attempts/e7949644-3d60-4ab9-817e-f80d0a48f5d8/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-judge-attribution/attribution_0/attempts/e7949644-3d60-4ab9-817e-f80d0a48f5d8/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-judge-attribution/attribution_0/attempts/e7949644-3d60-4ab9-817e-f80d0a48f5d8/attempt-1/trace/spans.jsonl +0 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-judge-attribution/attribution_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-judge-attribution/attribution_0/latest-attempt.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-legacy/capture-retention_0/attempts/7c7718f6-c85e-4867-9e3a-13ef9513de80/attempt-1/artifacts/captured-response.json +4 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-legacy/capture-retention_0/attempts/7c7718f6-c85e-4867-9e3a-13ef9513de80/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-legacy/capture-retention_0/attempts/7c7718f6-c85e-4867-9e3a-13ef9513de80/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-legacy/capture-retention_0/attempts/7c7718f6-c85e-4867-9e3a-13ef9513de80/attempt-1/trace/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-legacy/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-legacy/capture-retention_0/captured-response.json +6 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-legacy/capture-retention_0/failure-receipt.json +58 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-legacy/capture-retention_0/latest-attempt.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-legacy/capture-retention_0/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/baseline/capture-retention_0/attempts/3652118b-e1dc-4b17-bc15-6b802dc71eb1/attempt-1/artifacts/captured-response.json +4 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/baseline/capture-retention_0/attempts/3652118b-e1dc-4b17-bc15-6b802dc71eb1/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/baseline/capture-retention_0/attempts/3652118b-e1dc-4b17-bc15-6b802dc71eb1/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/baseline/capture-retention_0/attempts/3652118b-e1dc-4b17-bc15-6b802dc71eb1/attempt-1/trace/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/baseline/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/baseline/capture-retention_0/latest-attempt.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/candidates/42b3fec6b7e55008/capture-retention_0/attempts/40e411fe-ad12-4e6b-a279-2cf07c356a2f/attempt-1/artifacts/captured-response.json +4 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/candidates/42b3fec6b7e55008/capture-retention_0/attempts/40e411fe-ad12-4e6b-a279-2cf07c356a2f/attempt-1/failure-receipt.json +62 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/candidates/42b3fec6b7e55008/capture-retention_0/attempts/40e411fe-ad12-4e6b-a279-2cf07c356a2f/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/candidates/42b3fec6b7e55008/capture-retention_0/attempts/40e411fe-ad12-4e6b-a279-2cf07c356a2f/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/candidates/42b3fec6b7e55008/capture-retention_0/attempts/40e411fe-ad12-4e6b-a279-2cf07c356a2f/attempt-1/trace/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/candidates/42b3fec6b7e55008/capture-retention_0/latest-attempt.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/candidates/87740cc101602a08/capture-retention_0/attempts/9e7da63e-2d19-4ec1-a5ec-ec804e2b0f40/attempt-1/artifacts/captured-response.json +4 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/candidates/87740cc101602a08/capture-retention_0/attempts/9e7da63e-2d19-4ec1-a5ec-ec804e2b0f40/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/candidates/87740cc101602a08/capture-retention_0/attempts/9e7da63e-2d19-4ec1-a5ec-ec804e2b0f40/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/candidates/87740cc101602a08/capture-retention_0/attempts/9e7da63e-2d19-4ec1-a5ec-ec804e2b0f40/attempt-1/trace/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/candidates/87740cc101602a08/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/candidates/87740cc101602a08/capture-retention_0/latest-attempt.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/first-interruption.json +38 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/search/blobs/19147243d263e55b1eaf1c940691947382a7cce17c5f49e4a4da765a473d2a55.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/search/blobs/5dd0d871d559d1c55b2ac04b3fbd26d62f7c9e25d7632d4542502cd292818fc0.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/search/blobs/5de45d632937e75ff90fec9f1f2ec425904d3632da71f6b26f83b1e1eb8cde5a.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/search/blobs/6caba627806e499aa3d9e156aec9d5a731a3ef30e29756d343b85b7217c77c11.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/search/blobs/ce2770851b3b7651f27cedcbad0372b9191e9a9b016665987e3ddd2ad53012b5.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/search/blobs/d8ac4ba8d9844afb4defa5cd277fe3cb944d99e58cf41c07d3ae199b6f467966.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/search/blobs/f59a02035fe96ea2f5582dd715d2e3ad8c184ae959310978962f1ac0780753bd.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/search/ledger.jsonl +23 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-optimization/search/ledger.jsonl.head +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-path-guard/capture-retention_0/attempts/3c8272d6-d35b-477e-9d08-085261b8dc64/attempt-1/artifacts/capture.json +3 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-path-guard/capture-retention_0/attempts/3c8272d6-d35b-477e-9d08-085261b8dc64/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-path-guard/capture-retention_0/attempts/3c8272d6-d35b-477e-9d08-085261b8dc64/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-path-guard/capture-retention_0/attempts/3c8272d6-d35b-477e-9d08-085261b8dc64/attempt-1/trace/spans.jsonl +0 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-path-guard/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-path-guard/capture-retention_0/latest-attempt.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-resume/capture-retention_0/attempts/2b022ccb-adb3-4468-8c21-48ee82d0634a/attempt-1/artifacts/captured-response.json +6 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-resume/capture-retention_0/attempts/2b022ccb-adb3-4468-8c21-48ee82d0634a/attempt-1/failure-receipt.json +81 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-resume/capture-retention_0/attempts/2b022ccb-adb3-4468-8c21-48ee82d0634a/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-resume/capture-retention_0/attempts/2b022ccb-adb3-4468-8c21-48ee82d0634a/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-resume/capture-retention_0/attempts/2b022ccb-adb3-4468-8c21-48ee82d0634a/attempt-1/trace/spans.jsonl +2 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-resume/capture-retention_0/attempts/c4859951-1e7c-4d23-b89a-0ca8fa79e2cf/attempt-1/artifacts/captured-response.json +6 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-resume/capture-retention_0/attempts/c4859951-1e7c-4d23-b89a-0ca8fa79e2cf/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-resume/capture-retention_0/attempts/c4859951-1e7c-4d23-b89a-0ca8fa79e2cf/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-resume/capture-retention_0/attempts/c4859951-1e7c-4d23-b89a-0ca8fa79e2cf/attempt-1/trace/spans.jsonl +2 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-resume/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-resume/capture-retention_0/latest-attempt.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-resume/observation.json +608 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-retry/capture-retention_0/attempts/74c13884-d47e-4b6d-a823-5bd44ee0e6cb/attempt-1/artifacts/captured-response.json +6 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-retry/capture-retention_0/attempts/74c13884-d47e-4b6d-a823-5bd44ee0e6cb/attempt-1/failure-receipt.json +81 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-retry/capture-retention_0/attempts/74c13884-d47e-4b6d-a823-5bd44ee0e6cb/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-retry/capture-retention_0/attempts/74c13884-d47e-4b6d-a823-5bd44ee0e6cb/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-retry/capture-retention_0/attempts/74c13884-d47e-4b6d-a823-5bd44ee0e6cb/attempt-1/trace/spans.jsonl +2 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-retry/capture-retention_0/attempts/74c13884-d47e-4b6d-a823-5bd44ee0e6cb/attempt-2/artifacts/captured-response.json +6 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-retry/capture-retention_0/attempts/74c13884-d47e-4b6d-a823-5bd44ee0e6cb/attempt-2/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-retry/capture-retention_0/attempts/74c13884-d47e-4b6d-a823-5bd44ee0e6cb/attempt-2/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-retry/capture-retention_0/attempts/74c13884-d47e-4b6d-a823-5bd44ee0e6cb/attempt-2/trace/spans.jsonl +2 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-retry/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-retry/capture-retention_0/latest-attempt.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-retry/observation.json +514 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/after-summary.json +1853 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/judge-attribution.json +140 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/qualification.log +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/qualify.mjs +165 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/reproduce.mjs +39 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/reproduction-resume/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/reproduction-resume/capture-retention_0/captured-response.json +5 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/reproduction-resume/capture-retention_0/failure-receipt.json +58 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/reproduction-resume/capture-retention_0/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/reproduction-resume/observer-retained-captures.json +20 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/reproduction-retry/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/reproduction-retry/capture-retention_0/captured-response.json +5 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/reproduction-retry/capture-retention_0/failure-receipt.attempt-1.json +58 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/reproduction-retry/capture-retention_0/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/reproduction-retry/observer-retained-captures.json +20 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/reproduction-summary.json +315 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/reproduction.log +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/source-and-build.json +2409 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/typecheck.log +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/verify-package.log +340 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/attempts/33d57ccb-d4d8-4c23-92e3-c1ddf7db6058/attempt-1/artifacts/captured-response.json +4 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/attempts/33d57ccb-d4d8-4c23-92e3-c1ddf7db6058/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/attempts/33d57ccb-d4d8-4c23-92e3-c1ddf7db6058/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/attempts/33d57ccb-d4d8-4c23-92e3-c1ddf7db6058/attempt-1/trace/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/captured-response.json +6 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/failure-receipt.json +58 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/latest-attempt.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/baseline/capture-retention_0/attempts/6b25a1d2-5324-445b-b59d-bdc7381cd4b4/attempt-1/artifacts/captured-response.json +4 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/baseline/capture-retention_0/attempts/6b25a1d2-5324-445b-b59d-bdc7381cd4b4/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/baseline/capture-retention_0/attempts/6b25a1d2-5324-445b-b59d-bdc7381cd4b4/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/baseline/capture-retention_0/attempts/6b25a1d2-5324-445b-b59d-bdc7381cd4b4/attempt-1/trace/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/baseline/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/baseline/capture-retention_0/latest-attempt.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/first-interruption.json +5 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/search/blobs/5de45d632937e75ff90fec9f1f2ec425904d3632da71f6b26f83b1e1eb8cde5a.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/search/ledger.jsonl +7 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/search/ledger.jsonl.head +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-path-guard/capture-retention_0/attempts/723fa9c6-099e-437c-bddf-a4cbd28a8508/attempt-1/artifacts/capture.json +3 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-path-guard/capture-retention_0/attempts/723fa9c6-099e-437c-bddf-a4cbd28a8508/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-path-guard/capture-retention_0/attempts/723fa9c6-099e-437c-bddf-a4cbd28a8508/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-path-guard/capture-retention_0/attempts/723fa9c6-099e-437c-bddf-a4cbd28a8508/attempt-1/trace/spans.jsonl +0 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-path-guard/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-path-guard/capture-retention_0/latest-attempt.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/attempts/06db3429-c5c6-42c1-8623-08e2da1f563e/attempt-1/artifacts/captured-response.json +6 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/attempts/06db3429-c5c6-42c1-8623-08e2da1f563e/attempt-1/failure-receipt.json +62 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/attempts/06db3429-c5c6-42c1-8623-08e2da1f563e/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/attempts/06db3429-c5c6-42c1-8623-08e2da1f563e/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/attempts/06db3429-c5c6-42c1-8623-08e2da1f563e/attempt-1/trace/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/attempts/d49f7e40-c4d2-4cb6-98d1-5e49f45c6e2c/attempt-1/artifacts/captured-response.json +6 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/attempts/d49f7e40-c4d2-4cb6-98d1-5e49f45c6e2c/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/attempts/d49f7e40-c4d2-4cb6-98d1-5e49f45c6e2c/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/attempts/d49f7e40-c4d2-4cb6-98d1-5e49f45c6e2c/attempt-1/trace/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/latest-attempt.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/observation.json +458 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-retry/capture-retention_0/attempts/cf34f352-67e5-4e8b-a510-0e896bcc96fc/attempt-1/artifacts/captured-response.json +6 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-retry/capture-retention_0/attempts/cf34f352-67e5-4e8b-a510-0e896bcc96fc/attempt-1/failure-receipt.json +62 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-retry/capture-retention_0/attempts/cf34f352-67e5-4e8b-a510-0e896bcc96fc/attempt-1/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-retry/capture-retention_0/attempts/cf34f352-67e5-4e8b-a510-0e896bcc96fc/attempt-1/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-retry/capture-retention_0/attempts/cf34f352-67e5-4e8b-a510-0e896bcc96fc/attempt-1/trace/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-retry/capture-retention_0/attempts/cf34f352-67e5-4e8b-a510-0e896bcc96fc/attempt-2/artifacts/captured-response.json +6 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-retry/capture-retention_0/attempts/cf34f352-67e5-4e8b-a510-0e896bcc96fc/attempt-2/identity.json +10 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-retry/capture-retention_0/attempts/cf34f352-67e5-4e8b-a510-0e896bcc96fc/attempt-2/result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-retry/capture-retention_0/attempts/cf34f352-67e5-4e8b-a510-0e896bcc96fc/attempt-2/trace/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-retry/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-retry/capture-retention_0/latest-attempt.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-retry/observation.json +380 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/before-resume/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/before-resume/capture-retention_0/captured-response.json +6 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/before-resume/capture-retention_0/failure-receipt.json +58 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/before-resume/capture-retention_0/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/before-resume/observation.json +198 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/before-retry/capture-retention_0/cached-result.json +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/before-retry/capture-retention_0/captured-response.json +6 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/before-retry/capture-retention_0/failure-receipt.attempt-1.json +58 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/before-retry/capture-retention_0/spans.jsonl +1 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/before-retry/observation.json +121 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/before-summary.json +321 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/build.attribution.log +491 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/build.final.log +491 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/build.log +495 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/qualification.log +18 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/qualification.wrong-root.log +9 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/qualify.mjs +148 -0
- package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/qualify.wrong-root.mjs +148 -0
- package/docs/proofs/campaign-attempt-retention-20261001/manifest.json +808 -0
- package/package.json +2 -2
- package/dist/index-Ce2HACcP.d.ts.map +0 -1
- package/dist/llm-judge-EUMHOZOQ.js.map +0 -1
- package/dist/series-convergence-CjO2QdRW.js +0 -43
- package/dist/series-convergence-CjO2QdRW.js.map +0 -1
- package/dist/series-convergence-DLy0bQMV.d.ts +0 -126
- package/dist/series-convergence-DLy0bQMV.d.ts.map +0 -1
- package/dist/tool-groups-DP8b5kjF.d.ts.map +0 -1
|
@@ -1,10 +1,10 @@
|
|
|
1
1
|
import { c as CostLedgerHandle, f as CostLedgerSummary, m as CostReceipt } from "./cost-ledger-DbQdN3nO.js";
|
|
2
2
|
import { _ as ProposalFinding } from "./types-6VEehmEg.js";
|
|
3
|
-
import { B as PowerPreflight, ft as EvidenceReceipt, it as CampaignEvidenceContext } from "./statistical-heldout-
|
|
3
|
+
import { B as PowerPreflight, ft as EvidenceReceipt, it as CampaignEvidenceContext } from "./statistical-heldout-Dq3pnmxd.js";
|
|
4
4
|
import { nt as EvaluationClaim } from "./search-ledger-types-p2xKfiDJ.js";
|
|
5
|
-
import { H as SurfaceProposer, R as Scenario, S as JudgeConfig, a as CampaignResult, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, p as Gate, v as GateResult } from "./types-
|
|
6
|
-
import { Oa as FinalEvidencePolicy, Ti as ComparisonCost, _i as OptimizationMethodProvenance, da as openAutoPr, ea as RunImprovementLoopResult, fa as RunEvalOptions, fi as OptimizationMethod, ga as RunCampaignOptions, ha as CampaignCellRetryPolicy, ka as FinalEvidenceUse, na as PremeasuredOptimizationBaseline, ra as RunOptimizationOptions, ur as LoopProvenanceRecord, vi as OptimizationMethodResult, xr as CampaignComparisonUnits } from "./index-
|
|
7
|
-
import { At as SearchHistoryCoverageRow, Dt as SearchHistoryAdmissionOptions, Mt as SearchHistoryReceipt, Yt as CampaignStorage } from "./search-kernel-
|
|
5
|
+
import { H as SurfaceProposer, R as Scenario, S as JudgeConfig, a as CampaignResult, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, p as Gate, v as GateResult } from "./types-ZUy5HGY5.js";
|
|
6
|
+
import { Oa as FinalEvidencePolicy, Ti as ComparisonCost, _i as OptimizationMethodProvenance, da as openAutoPr, ea as RunImprovementLoopResult, fa as RunEvalOptions, fi as OptimizationMethod, ga as RunCampaignOptions, ha as CampaignCellRetryPolicy, ka as FinalEvidenceUse, na as PremeasuredOptimizationBaseline, ra as RunOptimizationOptions, ur as LoopProvenanceRecord, vi as OptimizationMethodResult, xr as CampaignComparisonUnits } from "./index-B8n12TR2.js";
|
|
7
|
+
import { At as SearchHistoryCoverageRow, Dt as SearchHistoryAdmissionOptions, Mt as SearchHistoryReceipt, Yt as CampaignStorage } from "./search-kernel-DNpsHb8P.js";
|
|
8
8
|
import { r as HostedTenant } from "./client-BzL9niqi.js";
|
|
9
9
|
import { d as InsightReport } from "./analyze-runs-Q74o4Hfj.js";
|
|
10
10
|
//#region src/campaign/presets/run-final-comparison.d.ts
|
|
@@ -512,4 +512,4 @@ interface DefinedAgentEval<TScenario extends Scenario, TArtifact> {
|
|
|
512
512
|
declare function defineAgentEval<TScenario extends Scenario, TArtifact>(defaults: DefineAgentEvalOptions<TScenario, TArtifact>): DefinedAgentEval<TScenario, TArtifact>;
|
|
513
513
|
//#endregion
|
|
514
514
|
export { SelfImproveMethodResult as _, DefinedAgentEval as a, SelfImproveMethodOptions as c, SelfImproveProposerOptions as d, SelfImproveProposerResult as f, SelfImproveMethodProvenance as g, selfImprove as h, DefineAgentEvalOptions as i, SelfImproveOptions as l, SelfImproveRunError as m, AgentEvalEvaluateOptions as n, defineAgentEval as o, SelfImproveResult as p, AgentEvalImproveOptions as r, SelfImproveBudget as s, AgentEvalAgent as t, SelfImproveProgressEvent as u };
|
|
515
|
-
//# sourceMappingURL=define-agent-eval-
|
|
515
|
+
//# sourceMappingURL=define-agent-eval-D8bPoBM5.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"define-agent-eval-
|
|
1
|
+
{"version":3,"file":"define-agent-eval-D8bPoBM5.d.ts","names":["_$__0.FinalEvidenceUse","_$__0.GateResult"],"sources":["../src/campaign/presets/run-final-comparison.ts","../src/contract/self-improve-method.ts","../src/contract/self-improve.ts","../src/contract/define-agent-eval.ts"],"mappings":";;;;;;;;;;UAaiB,uBAAuB,kBAAkB,UAAU,mBAC1D,KAAK,mBAAmB,WAAW;EAC3C,iBAAiB;EACjB,eAAe;EACf,sBACE,SAAS,gBACT,UAAU,WACV,KAAK,WAAW,mBAAmB,WAAW,+BAC3C,QAAQ;EACb,MAAM,KAAK,WAAW;EACtB,QAAQ;EACR,gBAAgB;EAChB;EACA;EACA,cAAc,QAAQ,gBAAgB,UAAU,mBAAmB;;;iBAI/C,mBAAmB,kBAAkB,UAAU,WACnE,MAAM,uBAAuB,WAAW,aAAU;;;;;;;;;;;;kBAA5CA;;;;;;;;;;;UCAS;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,oBAAoB,YAAY,0BAA0B;EAC1D,QAAQ;EACR;IACE;IACA;IACA;IACA;IACA;IACA;IACA;IACA,sBAAsB;;EAExB,MAAM,QAAQ,kBAAkB;EAChC;EACA;EACA;EACA;EACA,MAAM;EACN;EACA;;UAGe,wBAAwB,kBAAkB,UAAU,mBAC3D,KACN,0BAA0B,WAAW;EAWvC;IAAa,UAAU;IAAiB,QAAQ;;EAChD,uBAAuB;EACvB;;EAEA,UAAU,0BAA0B,WAAW;EAC/C,QAAQ,KAAK,0BAA0B,WAAW;IAChD;;EAEF,YAAY;EACZ,cAAc,YAAY,0BAA0B,WAAW;;EAE/D,MAAM;;EAEN,YAAY;EACZ,UAAU,0BAA0B,WAAW;EAC/C,KAAK,QAAQ,kBAAkB,mBAAmB,WAAW;IAC3D;IACA,iBAAiB,0BAA0B,WAAW;IACtD,eAAe,0BAA0B,WAAW;IACpD;IACA,QAAQ;IACR,MAAM;IACN,WAAW,kBAAkB;;;;;UCpBhB;;;EAGf;;;;EAIA;;EAEA;;EAEA;;;EAGA;;;EAGA;;;;EAIA;;EAEA,mBAAmB;;;;;;;;;EASnB;;EAEA;;;;;EAKA;;KAGU;EACN;EAA0B;;EAC1B;EAA4B;EAAuB;;EACnD;EAA4B;EAAe;;EAC3C;EAA8B;EAAe;EAAuB;;EAGpE;EAAsB;EAAkB;;EACxC;EAAyB;EAAW;EAAY;EAAa;;UAElD,mBAAmB,kBAAkB,UAAU,mBACtD;;;;;;;;;;;;;;EAcR,QAAQ,SAAS,gBAAgB,UAAU,WAAW,KAAK,oBAAoB,QAAQ;;;;;;;EAQvF;;EAEA;;;EAIA,WAAW;;EAGX,OAAO,YAAY,WAAW;;;EAI9B,iBAAiB;;EAGjB,SAAS;;;;;;;;;;;;;EAcT,sBAAsB,gCAAgC,WAAW;;;;;EAMjE,WAAW,gBAAgB;;;;;;EAO3B,SAAS,mBAAmB,WAAW;;;EAIvC,qBAAqB;;;EAIrB,OAAO,KAAK,WAAW;;;;;;;;EASvB,cAAc,eAAe,gBAAgB,iBAAiB,mBAAmB;;;;EAKjF,UAAU;;;;EAKV;;;EAIA,gBAAgB,QAAQ,uBAAuB;;;;;EAM/C,iBAAiB;IACf,UAAU;IACV;IACA;;;;EAKF;;;;;;;EAQA,YAAY;;;EAIZ,cAAc,OAAO;;;;EAKrB;EACA;EACA;;;;;;;;;;;;;EAcA,eAAe;;;;EAKf,eAAe;;EAGf;;;;;;;;EASA;;;;;;;;;EAUA,oBAAoB,uBAAuB,WAAW;;;EAItD,WAAW;;;;;EAMX,SAAS,uBAAuB,WAAW;;;EAI3C,eAAe,uBAAuB,WAAW;;EAEjD,WAAW;;EAEX,QAAQ;;EAER,gBAAgB;;UAGD,0BAA0B,kBAAkB,UAAU;EACrE;EACA,QAAQ;EACR,gBAAgB;;;;EAIhB;IACE;IACA,aAAa;;;;;EAKf;IACE;IACA,aAAa;IACb,SAAS;;;IAGT;;;;IAIA;;;;;;EAMF;;;EAGA;;;;EAIA,YAAY;;EAEZ;;;EAGA;;EAEA;;EAEA;;EAEA,MAAM;;;EAGN,UAAU;;;EAGV,gBAAgB;;EAEhB;IACE;IACA,MAAM;IACN;IACA,aAAa;;;;;;;;EAQf,SAAS;;;;EAIT,QAAQ;;;;;;EAMR,KAAK,yBAAyB,WAAW;;KAG/B,kBAAkB,kBAAkB,UAAU,aACtD,0BAA0B,WAAW,aACrC,wBAAwB,WAAW;KAE3B,yBAAyB,kBAAkB,UAAU,aAAa,KAC5E,mBAAmB,WAAW;EAG9B,QAAQ,mBAAmB,WAAW;EACtC;EACA,gBAAgB,QAAQ;;KAGd,2BAA2B,kBAAkB,UAAU,aAAa,KAC9E,mBAAmB,WAAW;EAG9B;EACA,gBAAgB,QAAQ;;;cAIb,4BAA4B;WAC9B,MAAM;WACN,UAAU;EAEnB,YAAY,gBAAgB,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA0NtB,YAAY,kBAAkB,UAAU,WACtD,MAAM,yBAAyB,WAAW,aACzC,QAAQ,wBAAwB,WAAW;iBAC9B,YAAY,kBAAkB,UAAU,WACtD,MAAM,2BAA2B,WAAW,aAC3C,QAAQ,0BAA0B,WAAW;iBAChC,YAAY,kBAAkB,UAAU,WACtD,MAAM,mBAAmB,WAAW,aACnC,QAAQ,kBAAkB,WAAW;;;KC9nB5B,eAAe,kBAAkB,UAAU,cACrD,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;KAED,uBAAuB,kBAAkB,UAAU,aAAa,mBAC1E,WACA;UAGe,yBAAyB,kBAAkB,UAAU,mBAC5D,KACN,eAAe,WAAW;;EAI5B,YAAY;;EAEZ,UAAU;;EAEV,QAAQ,eAAe,WAAW;;EAElC,QAAQ,YAAY,WAAW;;EAE/B,SAAS,YAAY,WAAW;;EAEhC;;KAGU,wBAAwB,kBAAkB,UAAU,aAAa,KAC3E,QAAQ,mBAAmB,WAAW;EAGtC,SAAS,QAAQ;EACjB,eAAe,QAAQ;;UAGR,iBAAiB,kBAAkB,UAAU;;WAEnD,oBAAoB;;WAEpB,iBAAiB;;;;;EAK1B,SACE,OAAO,yBAAyB,WAAW,aAC1C,QAAQ,eAAe,WAAW;;;;;;;EAOrC,QACE,OAAO,wBAAwB,WAAW,aACzC,QAAQ,kBAAkB,WAAW;;;;;;;;;iBAU1B,gBAAgB,kBAAkB,UAAU,WAC1D,UAAU,uBAAuB,WAAW,aAC3C,iBAAiB,WAAW"}
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import { c as CostLedgerHandle } from "./cost-ledger-DbQdN3nO.js";
|
|
2
2
|
import { C as BoundedTraceAnalysisStoreOptions, O as TraceAnalysisStore, p as EvidenceRef } from "./types-6VEehmEg.js";
|
|
3
|
-
import "./types-
|
|
3
|
+
import "./types-ZUy5HGY5.js";
|
|
4
4
|
import "./analyze-runs-Q74o4Hfj.js";
|
|
5
5
|
import { z } from "zod";
|
|
6
6
|
//#region src/eval-tools.d.ts
|
|
@@ -168,4 +168,4 @@ interface TraceAnalysisEngine {
|
|
|
168
168
|
declare function resolveTraceAnalystLimits(limits: Partial<TraceAnalystLimits> | undefined): TraceAnalystLimits;
|
|
169
169
|
//#endregion
|
|
170
170
|
export { buildTraceAnalysisToolDescriptors as _, TraceAnalystLimits as a, RawAnalystEvidence as c, evidenceRefsFromRawFinding as d, parseRawFinding as f, UNTRUSTED_TRACE_TEXT as g, TraceAnalysisToolDescriptor as h, TraceAnalysisEngineResult as i, RawAnalystFinding as l, TRACE_ANALYST_TOOL_NAMESPACE as m, TraceAnalysisEngine as n, resolveTraceAnalystLimits as o, BuildTraceAnalysisToolsOptions as p, TraceAnalysisEngineRequest as r, RAW_FINDING_SCHEMA_PROMPT as s, DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS as t, RawAnalystFindingSchema as u, traceAnalystFunctionGroup as v };
|
|
171
|
-
//# sourceMappingURL=engine-
|
|
171
|
+
//# sourceMappingURL=engine-F_PvwesR.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"engine-
|
|
1
|
+
{"version":3,"file":"engine-F_PvwesR.d.ts","names":[],"sources":["../src/eval-tools.ts","../src/trace-analyst/tools.ts","../src/analyst/finding-signature.ts","../src/analyst/engine.ts"],"mappings":";;;;;;;UA4BiB;EACf;EACA;EACA,YAAY;EACZ,UAAU,eAAe;IAAQ,SAAS;QAAkB;;;;cClBjD;;;;;;cAOA;UAGI,uCAAuC;EACtD,OAAO;;UAGQ,oCAAoC;EACnD,kBAAkB;;EAElB;;EAEA;;;iBAMc,kCACd,SAAS,iCACR;iBAyIa,0BAA0B,SAAS;EACjD;EACA;EACA;EACA;EACA,WAAW;;;;cC3KP,0BAAwB,EAAA;;;GAKnB,EAAA,KAAA;KAEC,qBAAqB,EAAE,aAAa;cAiBnC,yBAAuB,EAAA;;;;;;;;;;;;;;;;;GAKzB,EAAA,KAAA;KAEC,oBAAoB,EAAE,aAAa;;;;;;cAOlC;;iBAYG,2BAA2B,SAAS,oBAAoB;iBAQxD,gBACd,cACA,OAAO,aAAa,SAAS,mCAC5B;;;;UCpEc;;EAEf;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;;cAwBW;UAEI;EACf;EACA;EACA;EACA,gBAAgB;;;;;;;;;;;EAWhB,aAAa,SAAS;EACtB,QAAQ;EACR,YAAY;EACZ;EACA,WAAW;EACX,SAAS;EACT,OAAO,iBAAiB,SAAS;;;UAIlB;EACf;EACA,UAAU;;EAEV;;;;;EAKA;;EAEA;EACA,SAAS;;;;;;;;;UAUM;WACN;WACA;WACA;;WAEA;;;;;;;;WAQA,iBAAiB,SAAS;EACnC,QAAQ,SAAS,6BAA6B,QAAQ;;iBAGxC,0BACd,QAAQ,QAAQ,kCACf"}
|
package/dist/evaluation.d.ts
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import { a as CostChannel, c as CostLedgerHandle, g as CostReceiptInput, m as CostReceipt, x as MaximumCharge } from "./cost-ledger-DbQdN3nO.js";
|
|
2
2
|
import { i as AnalystFinding, n as AnalystContext, t as Analyst } from "./types-6VEehmEg.js";
|
|
3
|
-
import { R as Scenario, S as JudgeConfig, w as JudgeScore } from "./types-
|
|
3
|
+
import { R as Scenario, S as JudgeConfig, w as JudgeScore } from "./types-ZUy5HGY5.js";
|
|
4
4
|
//#region src/evaluation.d.ts
|
|
5
5
|
export interface EvaluationResult<T> {
|
|
6
6
|
value: T;
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import { c as ValidationError, r as CaptureIntegrityError } from "../errors-DEE6u6ot.js";
|
|
2
2
|
import { n as bonferroni, r as holm, t as benjaminiHochberg } from "../multiplicity-DIWHvysC.js";
|
|
3
|
-
import { $ as paretoSignificanceGate, A as PairArmsOptions, At as FinalEvidenceReservation, B as PowerPreflight, C as hashJson, Dt as FinalEvidenceMeasurement, E as verifyManifest, Et as FinalEvidenceLedger, Ft as EProcessStep, Gt as pairedRiskDifference, H as powerPreflight, It as eProcess, K as EvidenceVector, Kt as pairedRiskDifferenceExact, L as comparePairedArms, Mt as EProcess, N as PairedArmRow, Nt as EProcessOptions, O as MatchedPair, Ot as FinalEvidenceOutcome, Pt as EProcessState, Q as paretoPolicy, R as pairArms, S as evaluateHypothesis, T as signManifest, Tt as FinalEvidenceError, Ut as mcnemar, V as PowerPreflightOptions, X as PromotionPolicy, Yt as wilson, Z as buildEvidenceVector, _ as sequentialPairedGate, _t as verifyEvidenceReceipt, at as createCampaignEvidenceReceipt, b as SignedManifest, c as pairHoldout, ct as EVIDENCE_RECEIPT_VERSION, d as SequentialDecision, dt as EvidenceBinding, f as SequentialObservation, ft as EvidenceReceipt, g as sequentialDecide, gt as isIndependentEvidence, ht as createEvidenceReceipt, i as PairedHoldout, it as CampaignEvidenceContext, j as PairArmsResult, jt as openFinalEvidenceLedger, kt as FinalEvidenceRecord, lt as EvidenceAuthority, m as SequentialPairedGateOptions, mt as INDEPENDENT_EVIDENCE_AUTHORITY_KINDS, n as HeldoutSignificance, ot as CreateEvidenceReceiptInput, p as SequentialPairedGate, pt as EvidenceReceiptVerification, qt as pairedRiskDifferenceScore, r as HeldoutSignificanceOptions, s as heldoutSignificance, st as EVIDENCE_AUTHORITY_KINDS, ut as EvidenceAuthorityKind, v as HypothesisManifest, w as manifestContentDigest, wt as FinalEvidenceConflictError, x as SignedManifestAlgo, y as HypothesisResult, z as pairRunRecords, zt as ProportionInterval } from "../statistical-heldout-
|
|
3
|
+
import { $ as paretoSignificanceGate, A as PairArmsOptions, At as FinalEvidenceReservation, B as PowerPreflight, C as hashJson, Dt as FinalEvidenceMeasurement, E as verifyManifest, Et as FinalEvidenceLedger, Ft as EProcessStep, Gt as pairedRiskDifference, H as powerPreflight, It as eProcess, K as EvidenceVector, Kt as pairedRiskDifferenceExact, L as comparePairedArms, Mt as EProcess, N as PairedArmRow, Nt as EProcessOptions, O as MatchedPair, Ot as FinalEvidenceOutcome, Pt as EProcessState, Q as paretoPolicy, R as pairArms, S as evaluateHypothesis, T as signManifest, Tt as FinalEvidenceError, Ut as mcnemar, V as PowerPreflightOptions, X as PromotionPolicy, Yt as wilson, Z as buildEvidenceVector, _ as sequentialPairedGate, _t as verifyEvidenceReceipt, at as createCampaignEvidenceReceipt, b as SignedManifest, c as pairHoldout, ct as EVIDENCE_RECEIPT_VERSION, d as SequentialDecision, dt as EvidenceBinding, f as SequentialObservation, ft as EvidenceReceipt, g as sequentialDecide, gt as isIndependentEvidence, ht as createEvidenceReceipt, i as PairedHoldout, it as CampaignEvidenceContext, j as PairArmsResult, jt as openFinalEvidenceLedger, kt as FinalEvidenceRecord, lt as EvidenceAuthority, m as SequentialPairedGateOptions, mt as INDEPENDENT_EVIDENCE_AUTHORITY_KINDS, n as HeldoutSignificance, ot as CreateEvidenceReceiptInput, p as SequentialPairedGate, pt as EvidenceReceiptVerification, qt as pairedRiskDifferenceScore, r as HeldoutSignificanceOptions, s as heldoutSignificance, st as EVIDENCE_AUTHORITY_KINDS, ut as EvidenceAuthorityKind, v as HypothesisManifest, w as manifestContentDigest, wt as FinalEvidenceConflictError, x as SignedManifestAlgo, y as HypothesisResult, z as pairRunRecords, zt as ProportionInterval } from "../statistical-heldout-Dq3pnmxd.js";
|
|
4
4
|
import { c as BOOTSTRAP_GATE_MIN_N, d as PairedBootstrapResult, h as pairedBootstrap, u as PairedBootstrapOptions } from "../paired-promotion-decision-DPKz8ZnF.js";
|
|
5
5
|
import { a as requiredPairedSampleSize, i as pairedMde, n as mcnemarPower, o as requiredSampleSize, r as mcnemarRequiredN, t as mulberry32 } from "../random-DGEyr3zG.js";
|
|
6
6
|
import { at as summarizeEvaluationUnits, it as defineEvaluationClaim, nt as EvaluationClaim, rt as EvaluationUnitSummary } from "../search-ledger-types-p2xKfiDJ.js";
|
|
@@ -3,12 +3,12 @@ import { T as RunPaidCallInput, c as CostLedgerHandle, f as CostLedgerSummary, m
|
|
|
3
3
|
import { a as RunRecord, c as RunSplitTag } from "./run-record-DF-fuu31.js";
|
|
4
4
|
import { O as TraceAnalysisStore, _ as ProposalFinding, f as AnalystUsageReceipt, i as AnalystFinding, l as AnalystRunResult } from "./types-6VEehmEg.js";
|
|
5
5
|
import { M as LlmCallMetadata, ft as ServedModelPolicy, rt as RawProviderSink, x as ChatClient } from "./types-DLxe96rp.js";
|
|
6
|
-
import { Et as FinalEvidenceLedger, P as PairedArmsComparison, ft as EvidenceReceipt, it as CampaignEvidenceContext, kt as FinalEvidenceRecord } from "./statistical-heldout-
|
|
6
|
+
import { Et as FinalEvidenceLedger, P as PairedArmsComparison, ft as EvidenceReceipt, it as CampaignEvidenceContext, kt as FinalEvidenceRecord } from "./statistical-heldout-Dq3pnmxd.js";
|
|
7
7
|
import { a as PairedPromotionDecision, d as PairedBootstrapResult } from "./paired-promotion-decision-DPKz8ZnF.js";
|
|
8
8
|
import { E as SearchExecutionIdentity, G as SearchSplit, V as SearchProposer, a as SearchAttemptAccounting, i as SearchArtifactRef, j as SearchModelIdentity, nt as EvaluationClaim, rt as EvaluationUnitSummary, tt as SearchUnknown, z as SearchOperationRecordedEvent } from "./search-ledger-types-p2xKfiDJ.js";
|
|
9
9
|
import { r as LedgerHash } from "./canonical-CFpojCN5.js";
|
|
10
|
-
import { A as LabeledScenarioWrite, C as JudgeDimension, D as LabeledScenarioSampleArgs, E as LabeledScenarioRecord, H as SurfaceProposer, N as ParetoParent, O as LabeledScenarioSource, R as Scenario, S as JudgeConfig, T as LabelTrust, _ as GateDecision, a as CampaignResult, b as GenerationRecord, c as CampaignTraceWriter, d as DispatchContext, f as DispatchFn, g as GateContribution, i as CampaignCostMeter, j as MutableSurface, k as LabeledScenarioStore, l as CodeSurface, o as CampaignScenarioIdentity, p as Gate, r as CampaignCellResult, u as ComponentSurface, v as GateResult, y as GenerationCandidate } from "./types-
|
|
11
|
-
import { Dt as SearchHistoryAdmissionOptions, L as SearchPolicy, Mt as SearchHistoryReceipt, Yt as CampaignStorage, bt as SearchRunIdentity, gt as SearchLedgerBinding, ht as RegisterSearchNodeInput, i as SearchCellResult, kt as SearchHistoryCoverage, vt as SearchRecorder } from "./search-kernel-
|
|
10
|
+
import { A as LabeledScenarioWrite, C as JudgeDimension, D as LabeledScenarioSampleArgs, E as LabeledScenarioRecord, H as SurfaceProposer, N as ParetoParent, O as LabeledScenarioSource, R as Scenario, S as JudgeConfig, T as LabelTrust, _ as GateDecision, a as CampaignResult, b as GenerationRecord, c as CampaignTraceWriter, d as DispatchContext, f as DispatchFn, g as GateContribution, i as CampaignCostMeter, j as MutableSurface, k as LabeledScenarioStore, l as CodeSurface, o as CampaignScenarioIdentity, p as Gate, r as CampaignCellResult, u as ComponentSurface, v as GateResult, y as GenerationCandidate } from "./types-ZUy5HGY5.js";
|
|
11
|
+
import { Dt as SearchHistoryAdmissionOptions, L as SearchPolicy, Mt as SearchHistoryReceipt, Yt as CampaignStorage, bt as SearchRunIdentity, gt as SearchLedgerBinding, ht as RegisterSearchNodeInput, i as SearchCellResult, kt as SearchHistoryCoverage, vt as SearchRecorder } from "./search-kernel-DNpsHb8P.js";
|
|
12
12
|
import { B as ExternalOptimizerModelBudget, E as AnalystIssueExpectation, F as ExternalOptimizerCallbackLimits, J as ExternalOptimizerWireCounts, K as ExternalOptimizerResumeMode, R as ExternalOptimizerEvaluationObservation, V as ExternalOptimizerModelCall, X as ExternalTextEvaluationRequest, Y as ExternalTextCandidate, m as AnalystBenchmarkLabelState, q as ExternalOptimizerRunnerCommand, t as AgentProfile$1, u as AnalystBenchmarkCase } from "./agent-profile-EpO0HURy.js";
|
|
13
13
|
import { r as DatasetScenario, t as Dataset } from "./dataset-DQqhOCPt.js";
|
|
14
14
|
import { a as CheckerIdentity, n as VerdictCertification, t as DefaultVerdict, x as VerificationStrategySource } from "./verdict-E4eRNf7-.js";
|
|
@@ -236,7 +236,7 @@ interface RunCampaignOptions<TScenario extends Scenario, TArtifact> {
|
|
|
236
236
|
expectUsage?: 'assert' | 'warn' | 'off';
|
|
237
237
|
/** Test seam — override the wall clock for deterministic tests. */
|
|
238
238
|
now?: () => Date;
|
|
239
|
-
/**
|
|
239
|
+
/** Override the trace writer. The supplied directory belongs to this attempt. */
|
|
240
240
|
buildTraceWriter?: (cellId: string, dir: string) => CampaignTraceWriter;
|
|
241
241
|
/** Storage backend for run/cell dirs, the resumability cache, artifacts,
|
|
242
242
|
* and trace spans. Default: the Node filesystem (`fsCampaignStorage`).
|
|
@@ -261,7 +261,7 @@ interface RunCampaignOptions<TScenario extends Scenario, TArtifact> {
|
|
|
261
261
|
generation?: number;
|
|
262
262
|
}) => string | undefined;
|
|
263
263
|
}
|
|
264
|
-
/** Durable `<
|
|
264
|
+
/** Durable `<attempt>/failure-receipt.json` written before a failed cell can
|
|
265
265
|
* trigger campaign-wide cancellation. The cell records dispatch measurements;
|
|
266
266
|
* `cost` covers every settled agent and judge call attributed to this exact run
|
|
267
267
|
* attempt. */
|
|
@@ -285,11 +285,9 @@ interface CampaignCellFailureReceipt<TArtifact = unknown> {
|
|
|
285
285
|
* Bounded in-run retry of failed cells. Every attempt dispatches the same
|
|
286
286
|
* slot and charges the shared cost ledger, so the final cell's `costUsd`,
|
|
287
287
|
* `tokenUsage`, and `costCallIds` cover all attempts. Each retried attempt
|
|
288
|
-
* keeps its
|
|
289
|
-
*
|
|
290
|
-
* final cell records the retry count as `retryAttempts`.
|
|
291
|
-
* spans written by a later attempt replace those of the retried attempt; the
|
|
292
|
-
* per-attempt failure receipts are the durable evidence.
|
|
288
|
+
* keeps its artifacts, trace, result, and failure receipt in a separate
|
|
289
|
+
* execution directory, including attempts resumed in a later invocation.
|
|
290
|
+
* The final cell records the retry count as `retryAttempts`.
|
|
293
291
|
*/
|
|
294
292
|
interface CampaignCellRetryPolicy {
|
|
295
293
|
/** Total attempts per cell, including the first. A positive safe integer. */
|
|
@@ -3392,4 +3390,4 @@ declare function verifyCodeSurface(surface: CodeSurface, worktreeDir?: string):
|
|
|
3392
3390
|
declare function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string;
|
|
3393
3391
|
//#endregion
|
|
3394
3392
|
export { ScoreboardRow as $, RunImprovementLoopOptions as $i, TraceAnalystScenario as $n, GepaOptimizationRecipe as $r, importExternalEvaluations as $t, ScenarioSignal as A, CampaignCoverage as Aa, GepaCandidatePopulationSummary as Ai, CrossSurfaceComponentEvidence as An, DEFAULT_RED_TEAM_CORPUS as Ar, RequirementCheck as At, FinalizedProfileMatrixResult as B, ReferenceEquivalenceJudgeInput as Bi, CrossSurfaceInteractionTask as Bn, CanaryKind as Br, FsLabeledScenarioStoreOptions as Bt, SingleRunLockOptions as C, CacheRead as Ca, compareOptimizationMethods as Ci, CrossSurfaceBootstrapPolicy as Cn, campaignMeanComposite as Cr, extractProducedState as Ct, searchProposalExecution as D, cellCachePath as Da, costFromLedgerSummary as Di, CrossSurfaceCandidateOutcome as Dn, DefaultProductionGateOptions as Dr, LlmCorrectnessCheckerOpts as Dt, searchModelIdentity as E, buildCellSchedule as Ea, combineComparisonCosts as Ei, CrossSurfaceCandidateEvidence as En, DefaultProductionGateCheck as Er, CorrectnessChecker as Et, RetainedDispatchOptions as F, campaignSplitDigest as Fa, ExternalOptimizerObservationSummary as Fi, CrossSurfaceIneligibilityReason as Fn, redTeamDataset as Fr, verifyCompletion as Ft, RunProfileMatrixSegmentOptions as G, runReferenceEquivalenceJudge as Gi, CrossSurfacePairwiseEntry as Gn, SkillOptOptimizationMethodConfig as Gr, ScoredRollout as Gt, ProfileMatrixPlan as H, ReferenceEquivalenceJudgeResult as Hi, CrossSurfacePairCompatibility as Hn, CanaryReport as Hr, RolloutArgumentDiff as Ht, RetainedDispatchOutcome as I, campaignSplitDigestFromIdentities as Ia, ExternalOptimizerSubmittedCandidate as Ii, CrossSurfaceInteractionAwareSelection as In, redTeamReport as Ir, scopedOptimizationMethod as It, runProfileMatrixSegment as J, llmJudge as Ji, CrossSurfaceSelectionPolicy as Jn, skillOptOptimizationMethod as Jr, rolloutArgumentDiff as Jt, createProfileMatrixPlan as K, LlmJudgeDimension as Ki, CrossSurfaceRankedSingle as Kn, SkillOptRunnerCommand as Kr, UngroundedLiteralReport as Kt, createRetainedDispatch as L, readExternalOptimizerObservationArtifact as Li, CrossSurfaceInteractionEffect as Ln, scoreRedTeamOutput as Lr, sequentialOptimizationMethod as Lt, selectDiscriminative as M, assertCampaignSplitIdentity as Ma, readGepaCandidatePopulationArtifact as Mi, CrossSurfaceDistribution as Mn, RedTeamCategory as Mr, TaskGold as Mt, resolveRunDir as N, campaignCoverage as Na, ExternalOptimizerExecutionSummary as Ni, CrossSurfaceEligibility as Nn, RedTeamFinding as Nr, completionVerdict as Nt, searchReceiptAccounting as O, FinalEvidencePolicy as Oa, GepaCandidatePopulationArtifact as Oi, CrossSurfaceCandidateSummary as On, DefaultProductionRewardHackingOptions as Or, ProducedProposal as Ot, tangleTracesRoot as P, campaignScenarioIdentity as Pa, ExternalOptimizerObservationArtifact as Pi, CrossSurfaceEvidenceBreakdown as Pn, RedTeamReport as Pr, createLlmCorrectnessChecker as Pt, ScoreboardRenderOptions as Q, transientDispatchFailure as Qi, TraceAnalystArtifact as Qn, GepaOptimizationMethodConfig as Qr, externalSurface as Qt, CreateProfileMatrixPlanOptions as R, REFERENCE_EQUIVALENCE_INPUT_LIMITS as Ri, CrossSurfaceInteractionPath as Rn, CanaryAlert as Rr, neutralizeText as Rt, SingleRunLock as S, CacheIssueReason as Sa, OptimizationTokenUsage as Si, CrossSurfaceBestSingleSelection as Sn, campaignBreakdown as Sr, ToolCallEventLike as St, campaignCellSearchResult as T, CellScheduleSlot as Ta, ComparisonCost as Ti, CrossSurfaceCandidateComparison as Tn, heldOutGate as Tr, CompletionVerdict as Tt, ProfileMatrixRow as U, ReferenceEquivalenceScenario as Ui, CrossSurfacePairEvidence as Un, runCanaries as Ur, RolloutArgumentDiffOptions as Ut, ProfileMatrixCoverage as V, ReferenceEquivalenceJudgeOptions as Vi, CrossSurfaceNaiveStackSelection as Vn, CanaryOptions as Vr, LabeledScenarioStoreError as Vt, ProfileMatrixSegmentResult as W, createReferenceEquivalenceJudge as Wi, CrossSurfacePairIncompatibilityReason as Wn, composeGate as Wr, RolloutCall as Wt, PlaybackDriver as X, isTransientTransportFailure as Xi, CrossSurfaceTaskRow as Xn, GepaEngineOptions as Xr, GepaEvaluationImport as Xt, PlaybackContext as Y, TransientFailureOptions as Yi, CrossSurfaceSelections as Yn, GepaAdaptiveEngineRun as Yr, ExternalSearchOptimizer as Yt, PlaybackStep as Z, quotaExhaustedUntil as Zi, BuildTraceAnalystSurfaceDispatchOptions as Zn, GepaEngineRun as Zr, GepaPopulationImport as Zt, componentSurfaceIdentityMaterial as _, runCampaign as _a, OptimizationMethodProvenance as _i, analyzeCrossSurfaceInteractions as _n, provenanceRecordPath as _r, assertRealBackend as _t, WorktreeAdapterError as a, runOptimization as aa, ExternalTextOptimizationMethodConfig as ai, EvalFixture as an, LoopProvenanceArgsFromResult as ar, scoreUserStory as at, surfaceDispatchRef as b, PlanCampaignRunOptions as ba, OptimizationMethodScore as bi, CrossSurfaceAdditionRejectionReason as bn, CampaignBreakdown as br, ProposalEventLike as bt, verifyCodeSurface as c, SearchLedgerIntegrityError as ca, ExternalOptimizationExample as ci, EvalFixtureRunPlan as cn, LoopProvenanceEvidence as cr, ProfileDispatchFn as ct, PhoenixEvaluationResultLike as d, openAutoPr as da, CompareOptimizationMethodsOptions as di, LoadEvalFixtureScenariosOptions as dn, buildLoopProvenanceRecord as dr, RunProfileMatrixOptions as dt, RunImprovementLoopResult as ea, GepaRunnerCommand as ei, importGepaPopulation as en, buildTraceAnalystSurfaceDispatch as er, ScoreboardSummary as et, PhoenixEvaluatorLike as f, RunEvalOptions as fa, OptimizationMethod as fi, PlanEvalFixtureRunOptions as fn, campaignMeasurementDigest as fr, RunProfileMatrixResult as ft, codeSurfaceIdentityMaterial as g, RunCampaignOptions as ga, OptimizationMethodPairwise as gi, planEvalFixtureRun as gn, loopProvenanceSpans as gr, BackendIntegrityReport as gt, assertCodeSurfaceIdentity as h, CampaignCellRetryPolicy as ha, OptimizationMethodInput as hi, loadEvalFixtureScenarios as hn, loopProvenanceArgsFromResult as hr, BackendIntegrityError as ht, WorktreeAdapter as i, RunOptimizationResult as ia, externalTextOptimizationMethod as ii, neutralizationGate as in, EmitLoopProvenanceResult as ir, renderScoreboardMarkdown as it, scoreDiscrimination as j, assertCampaignDesign as ja, GepaCandidateSelectionScore as ji, CrossSurfaceCompositionStep as jn, RedTeamCase as jr, SatisfiedBy as jt, DiscriminationScore as k, FinalEvidenceUse as ka, GepaCandidatePopulationCandidate as ki, CrossSurfaceComponent as kn, defaultProductionGate as kr, ProducedState as kt, AutoevalsScoreLike as l, OpenAutoPrOptions as la, ExternalTextEvaluationResponse as li, EvalFixtureScenario as ln, LoopProvenanceOptimizationMethod as lr, ProfileMatrixError as lt, phoenixEvaluatorJudge as m, CampaignCellFailureReceipt as ma, OptimizationMethodComposition as mi, loadEvalFixture as mn, emitLoopProvenance as mr, runProfileMatrix as mt, GitWorktreeAdapterOptions as n, PremeasuredOptimizationBaseline as na, OpenAICompatibleOptimizerModel as ni, recordSkillOptSearch as nn, BuildLoopProvenanceArgs as nr, UserStoryVerdict as nt, gitWorktreeAdapter as o, SearchLedgerConflictError as oa, ExternalTextOptimizerContext as oi, EvalFixtureFile as on, LoopProvenanceBackend as or, scoreboardSummary as ot, autoevalsScorerJudge as p, runEval as pa, OptimizationMethodComparison as pi, discoverEvalFixtures as pn, canonicalDigest as pr, ScenarioRollup as pt, finalizeProfileMatrix as q, LlmJudgeOptions as qi, CrossSurfaceRelativeCost as qn, SkillOptTrainerConfig as qr, classifyUngroundedLiterals as qt, Worktree as r, RunOptimizationOptions as ra, OptimizerModelBudget as ri, NeutralizationGateOptions as rn, EmitLoopProvenanceArgs as rr, makePlaybackDispatch as rt, resolveWorktreePath as s, SearchLedgerError as sa, ExternalTextOptimizerResult as si, EvalFixtureLoadOptions as sn, LoopProvenanceCandidate as sr, userStoryScoreboard as st, CodeSurfaceVerification as t, runImprovementLoop as ta, gepaOptimizationMethod as ti, recordGepaSearch as tn, traceAnalystQualityJudge as tr, UserStory as tt, AutoevalsScorerLike as u, OpenAutoPrResult as ua, decodeExternalTextCandidate as ui, EvalFixtureValidationMode as un, LoopProvenanceRecord as ur, ProfileSummary as ut, renderSurfaceDiff as v, CampaignRunPlan as va, OptimizationMethodResult as vi, AnalyzeCrossSurfaceInteractionsInput as vn, provenanceSpansPath as vr, summarizeBackendIntegrity as vt, acquireSingleRunLock as w, readCachedCell as wa, optimizationTokenUsageFromSummary as wi, CrossSurfaceCandidate as wn, HeldOutGateOptions as wr, CompletionRequirement as wt, surfaceHash as x, planCampaignRun as xa, OptimizationPackageSource as xi, CrossSurfaceAttemptCompleteness as xn, CampaignComparisonUnits as xr, RuntimeEventLike as xt, surfaceContentHash as y, CampaignRunPlanCell as ya, OptimizationMethodRunOptions as yi, CrossSurfaceAdditionDecision as yn, verifyLoopProvenanceRecord as yr, ArtifactEventLike as yt, FinalizeProfileMatrixOptions as z, REFERENCE_EQUIVALENCE_JUDGE_VERSION as zi, CrossSurfaceInteractionReport as zn, CanaryEvaluation as zr, FsLabeledScenarioStore as zt };
|
|
3395
|
-
//# sourceMappingURL=index-
|
|
3393
|
+
//# sourceMappingURL=index-B8n12TR2.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"index-B8n12TR2.d.ts","names":["_$__external_optimizer_process0.ExternalTextEvaluationRequest","AgentProfile","AgentProfile"],"sources":["../src/campaign/coverage.ts","../src/campaign/final-evidence.ts","../src/campaign/cell-schedule.ts","../src/campaign/cell-cache.ts","../src/campaign/plan-campaign-run.ts","../src/campaign/run-campaign.ts","../src/campaign/presets/run-eval.ts","../src/campaign/auto-pr.ts","../src/campaign/search-ledger-errors.ts","../src/campaign/presets/run-optimization.ts","../src/campaign/presets/run-improvement-loop.ts","../src/campaign/transient-failure.ts","../src/llm-judge.ts","../src/reference-equivalence-judge.ts","../src/campaign/external-optimizer-observations.ts","../src/campaign/gepa-candidate-population.ts","../src/campaign/optimization-cost.ts","../src/campaign/presets/compare-optimization-methods.ts","../src/campaign/external-text-evaluation.ts","../src/campaign/external-text-optimization-contract.ts","../src/campaign/external-text-optimization.ts","../src/campaign/optimizer-model.ts","../src/campaign/gepa-optimization-method.ts","../src/campaign/skillopt-optimization-method.ts","../src/campaign/gates/compose.ts","../src/canary.ts","../src/red-team.ts","../src/campaign/gates/default-production-gate.ts","../src/campaign/gates/heldout-gate.ts","../src/campaign/score-utils.ts","../src/campaign/provenance.ts","../src/campaign/analyst-surface.ts","../src/campaign/cross-surface-types.ts","../src/campaign/cross-surface-interaction.ts","../src/campaign/fixtures.ts","../src/campaign/gates/neutralization-gate.ts","../src/campaign/gepa-search-import.ts","../src/campaign/grounded-reflection.ts","../src/campaign/labeled-store/fs-adapter.ts","../src/campaign/neutralize.ts","../src/campaign/optimization-method-composition.ts","../src/artifact-validator.ts","../src/completion-verifier.ts","../src/produced-state.ts","../src/integrity/backend-integrity.ts","../src/campaign/presets/run-profile-matrix.ts","../src/campaign/presets/playback.ts","../src/campaign/presets/segmented-profile-matrix.ts","../src/campaign/retained-dispatch.ts","../src/campaign/run-dir.ts","../src/campaign/scenario-selection.ts","../src/campaign/search-campaign-cell.ts","../src/campaign/single-run-lock.ts","../src/campaign/surface-identity.ts","../src/campaign/upstream-evaluators.ts","../src/campaign/worktree/index.ts"],"mappings":";;;;;;;;;;;;;;;;;;;UAQiB;EACf;EACA;EACA;EACA,iBAAiB;IAAQ;IAAgB;;;;iBAI3B,qBAAqB,kBAAkB,UACrD,oBAAoB,aACpB;;iBA2Bc,yBAAyB,kBAAkB,UACzD,UAAU,YACT,2BAA2B,KAAK;;iBAUnB,kCACd,oBAAoB,4BACpB;;iBAgBc,oBAAoB,kBAAkB,UACpD,oBAAoB,aACpB;;iBAOc,4BACd,oBAAoB,4BACpB,cACA;;iBAQc,iBAAiB,WAAW,kBAAkB,UAC5D,gBAAgB,mBAAmB,cACnC,oBAAoB,aACpB,cACA,6BACC;;;;UC/Ec;EACf,QAAQ;EACR;EACA,iBAAiB;;UAGF;EACf,OAAO;EACP,QAAQ;;;;iBCnBM,kBAAkB,kBAAkB,UAClD,WAAW,aACX,cACA,eACC;EAAQ,UAAU;EAAW;EAAa;EAAgB;;KA4BjD,iBAAiB,kBAAkB,YAAY,kBAClD,kBAAkB;iBAoBX,cAAc,gBAAgB;;;KCqDlC;KASA,UAAU;EAChB;EAAe,MAAM,mBAAmB;;EACxC;EAAgB,QAAQ;;iBAEd,eAAe,WAAW;EACxC,SAAS;EACT;EACA;EACA;IACE,UAAU;;;UCxHG;EACf;EACA;EACA;EACA;EACA;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;;UAGQ,uBAAuB,kBAAkB,UAAU;EAClE,WAAW;EACX,WAAW,WAAW,WAAW;EACjC;EACA,SAAS,YAAY,WAAW;EAChC;EACA;EACA;;EAEA;EACA;;EAEA;EACA,UAAU;;EAEV,aAAa;;EAEb,WAAW,SAAS;;;;;;iBAON,gBAAgB,kBAAkB,UAAU,WAC1D,MAAM,uBAAuB,WAAW,aACvC;;;UCjBc,mBAAmB,kBAAkB,UAAU;EAC9D,WAAW;EACX,UAAU,WAAW,WAAW;;EAEhC,SAAS;;;;;;EAMT;EACA,SAAS,YAAY,WAAW;;EAEhC;;;EAGA;;;EAGA;;;;;;EAMA,cAAc;IAAS,UAAU;IAAW;;;;;;EAK5C;;;;;;EAMA;;;;EAIA,eAAe;EACf;EACA;;EAEA;;;EAGA,aAAa;;EAEb;;EAEA,WAAW,SAAS;;EAEpB;;;;;;;;;;EAUA;;;;;;;;EAQA,YAAY;;;;;;;;;;EAUZ;;;;;EAKA;;;;EAIA;;;EAGA;;;;EAIA;;;;;;;;;;;EAWA;;EAEA,YAAY;;EAEZ,oBAAoB,gBAAgB,gBAAgB;;;;;;EAMpD,UAAU;;;;;;;;;;;;EAYV,iBAAiB;IACf,UAAU;IACV;IACA;;;;;;;UAQa,2BAA2B;EAC1C;EACA;EACA;EACA;IACE;IACA;IACA;MACE;MACA;MACA;;;EAGJ,MAAM,mBAAmB;EACzB,MAAM;;;;;;;;;;UAWS;;EAEf;;;;;EAKA,YAAY,SAAS;;;;;iBAMD,YAAY,kBAAkB,UAAU,WAC5D,MAAM,mBAAmB,WAAW,aACnC,QAAQ,eAAe,WAAW;;;UCrNpB,eAAe,kBAAkB,UAAU,mBAClD,KAAK,mBAAmB,WAAW;EAC3C;;;;;iBAMoB,QAAQ,kBAAkB,UAAU,WACxD,MAAM,eAAe,WAAW,aAC/B,QAAQ,eAAe,WAAW;;;UCHpB,kBAAkB,WAAW,kBAAkB;;EAE9D,QAAQ,eAAe,WAAW;;;EAGlC,MAAM;;;EAGN;;EAEA;EACA;;EAEA;;EAEA;;;EAGA;;EAEA,UAAU;IAAqB;IAAgB;IAAgB;;;UAGhD;EACf;EACA;EACA;EACA;;;;;iBAMc,WAAW,WAAW,kBAAkB,UACtD,SAAS,kBAAkB,WAAW,aACrC;;;;cClDU,0BAA0B;;cAG1B,mCAAmC;;cAGnC,kCAAkC;;;UC8F9B,gCAAgC,WAAW,kBAAkB;;EAE5E;;EAEA,UAAU,eAAe,WAAW;;UAGrB,2BAA2B,kBAAkB,UAAU,mBAC9D,KAAK,mBAAmB,WAAW;;EAE3C,iBAAiB;;;;;;;;EAQjB,sBAAsB,gCAAgC,WAAW;;EAEjE,sBACE,SAAS,gBACT,UAAU,WACV,KAAK,WAAW,mBAAmB,WAAW,+BAC3C,QAAQ;;EAEb,UAAU,gBAAgB;;EAE1B;;EAEA;;;EAGA;;;EAGA;;EAEA,WAAW,cAAc;;;;;;;;EAQzB,qBAAqB;IACnB;IACA;IACA,YAAY;MACV;MACA,UAAU,eAAe,WAAW;MACpC;;IAEF,SAAS;;IAET,aAAa;IACb;QACI,QAAQ,cAAc;;;;;;;;EAQ5B,SAAS;;;;;EAKT,eAAe;;KAGL,uBACV,kBAAkB,UAClB,aACE,2BAA2B,WAAW;UAEzB,sBAAsB,WAAW,kBAAkB;;EAElE,aAAa;IACX,QAAQ;IACR,UAAU;MACR;MACA,SAAS;MACT,UAAU,eAAe,WAAW;;;;EAIxC,iBAAiB;EACjB,eAAe;EACf;;EAEA;;;EAGA;EACA,kBAAkB,eAAe,WAAW;;EAE5C,MAAM;;EAEN,eAAe;;EAEf,gBAAgB;;;iBAII,gBAAgB,kBAAkB,UAAU,WAChE,MAAM,uBAAuB,WAAW,aACvC,QAAQ,sBAAsB,WAAW;;;KCxLhC,0BACV,kBAAkB,UAClB,aACE,uBAAuB,WAAW;;;EAGpC,kBAAkB;EAClB,QAAQ;EACR,gBAAgB;;;;;;;;;EAShB;;;;EAIA,MAAM,KAAK,WAAW;;;;;EAKtB;;EAEA;EACA;;;;;;;;EAQA,cAAc,eAAe,gBAAgB,iBAAiB,mBAAmB;;UAGlE,yBAAyB,WAAW,kBAAkB,kBAC7D,sBAAsB,WAAW;EACzC,mBAAmB,eAAe,WAAW;EAC7C,iBAAiB,eAAe,WAAW;EAC3C,uBAAuB,eAAe,WAAW;EACjD,qBAAqB;EACrB,YAAY,QAAQ,WAAW,KAAK,WAAW;EAC/C,QAAQ;EACR,gBAAgB;;;;EAIhB;;;;;EAKA;EACA,WAAW,kBAAkB;;;;;iBAMT,mBAAmB,kBAAkB,UAAU,WACnE,MAAM,0BAA0B,WAAW,aAC1C,QAAQ,yBAAyB,WAAW;;;UCrE9B;;;;;;;WAON;;WAEA,yBAAyB;;;;;WAKzB;;;;;;;;;;;;;;;;;;;;;;;;iBAiFK,oBAAoB,qCAAqC;;;;;iBAiCzD,4BACd,oCACA,OAAM;;;;;;;;;;;;iBAyBQ,yBACd,OAAM,2BACJ,SAAS;;;;;KCzID,6BAA6B;UAExB,gBAAgB,WAAW,kBAAkB,WAAW;;;;EAIvE,MAAM;;;EAGN,aAAa;;EAEb;;EAEA;EACA;EACA;;;EAGA,UAAU;;;;;;;;;;;;;;;;EAgBV;IAAY;;IAAwB;IAAuB;;;;;;EAK3D;;EAEA,aAAa,UAAU;;;EAGvB,cAAc;IAAS,UAAU;IAAW,UAAU;;;EAEtD,aAAa;EACb;IAAmB;IAAc,QAAQ,EAAE;;;;;;;;;;;;;iBAoB7B,SAAS,qBAAqB,kBAAkB,WAAW,UACzE,cACA,gBACA,MAAM,gBAAgB,WAAW,aAChC,YAAY,WAAW;;;cC9Gb;cAEA;WACX;WACA;WACA;;UAGe,qCAAqC;EACpD;EACA;;UAGe;EACf;EACA;EACA;;UAGe;;EAEf,MAAM;;EAEN;;EAEA,SAAS;;EAET,aAAa;;UAGE,wCAAwC;EACvD;EACA;EACA;EACA;;;iBA0Bc,gCACd,SAAS,mCACR,oBAAoB;;iBA2CD,6BACpB,OAAO,gCACP,SAAS,mCACR,QAAQ;;;UCtGM;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;;WAEN,WAAW;;WAEX;WACA;WACA;;WAEA;aACE;aACA;;;UAII;WACN,SAAS;WACT,uBAAuB;;WAEvB,qBAAqB;;;;;;;;;;iBAWhB,yCAAyC;EACvD,SAAS;EACT,UAAU;IACR;;;UCrDa;WACN;WACA;WACA;WACA;WACA;WACA;WACA;WACA;WACA;WACA;WACA;;UAGM;WACN;WACA;;UAGM;;WAEN;WACA,WAAW;WACX;WACA;;WAEA;;WAEA;WACA,0BAA0B;WAC1B;;UAGM;WACN,SAAS;WACT;WACA;WACA,qBAAqB;;;;;;;;;;iBAWhB,oCAAoC;EAClD,SAAS;EACT,UAAU;IACR;;;;UC1Da;;EAEf;;EAEA,gBAAgB;EAChB;EACA;;;iBAuDc,sBAAsB,SAAS,oBAAoB;;iBAYnD,uBACd,SAAS;EAAgB;EAAe,MAAM;KAC7C;;;;KCDS,6BAA6B,kBAAkB,UAAU,aAAa,KAChF,mBAAmB,WAAW;UAIf;EACf;;EAEA;EACA;EACA;EACA;EACA;;EAEA;;UAGe;EACf;EACA;;UAGe;EACf;EACA;;UAGe;;EAEf;;EAEA;;EAEA;EACA;;EAEA;EACA;EACA;;UAGe;;EAEf,QAAQ;;EAER,SAAS;;EAET,UAAU;;EAEV,SAAS;;EAET;;EAEA;EACA;;EAEA;EACA;;EAEA;;EAEA;;;;;;EAMA;EACA;EACA,aAAa;;EAEb,eAAe;;EAEf,0BAA0B;;EAE1B,kBAAkB;;EAElB,oBAAoB;;;UAIL,wBAAwB,kBAAkB,UAAU;;WAE1D,iBAAiB;;WAEjB,yBAAyB;;WAEzB,6BAA6B;;WAE7B,sBACP,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;;WAEJ,iBAAiB,YAAY,WAAW;;WAExC;WACA;;WAEA,YAAY,SAAS,6BAA6B,WAAW;;WAE7D,YAAY;;EAErB,cACE,QAAQ,mBAAmB,WAAW,YACtC,OAAO,wBAAwB,WAAW,aACzC,QAAQ;;WAEF,iBAAiB,SAAS,mBAAmB;;;UAIvC;EACf;EACA;EACA,QAAQ;IACN;IACA;IACA,QAAQ;;;UAIK;;EAEf,eAAe;EACf,cAAc;;EAEd,MAAM;;EAEN;;EAEA,aAAa;;EAEb,gBAAgB;;;UAID,mBAAmB,kBAAkB,WAAW,UAAU;;EAEzE;EACA,WACE,OAAO,wBAAwB,WAAW,eACvC,QAAQ;;UAGE;EACf;IAAa,UAAU;IAAiB,QAAQ;;EAChD;;EAEA;;EAEA;;EAEA;;EAEA;IAAU;IAAa;;EACvB,UAAU;;EAEV,kBAAkB;;EAElB;;EAEA,aAAa;;EAEb,gBAAgB;IACd;IACA;IACA;IACA;;;EAGF,YAAY;IACV;IACA;IACA;IACA;IACA;;EAEF,eAAe;EACf,cAAc;;EAEd;;UAGe;;EAEf;EACA;;EAEA;EACA;EACA;;EAEA;EACA,UAAU;;UAGK;EACf,QAAQ;EACR,gBAAgB;;EAEhB,QAAQ;EACR,MAAM;;EAEN,UAAU;EACV;EACA,OAAO;EACP;;EAEA;;EAEA,kBAAkB;;EAElB,UAAU;;EAEV,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,eAAe;;UAGA,kCAAkC,kBAAkB,UAAU,mBACrE,KAAK,mBAAmB,WAAW,kDACzC;;EAEF,WAAW;EACX,QAAQ;EACR,gBAAgB;EAChB,SAAS,mBAAmB,WAAW;EACvC,iBAAiB;;EAEjB,gBAAgB;;EAEhB,oBAAoB;;EAEpB,eAAe;;EAEf,sBACE,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;EACb,QAAQ,YAAY,WAAW;;;EAG/B;;EAEA,yBAAyB,6BAA6B,WAAW;;EAEjE;;;EAGA;;EAEA;;;;;iBAMoB,2BAA2B,kBAAkB,UAAU,WAC3E,MAAM,kCAAkC,WAAW,aAClD,QAAQ;;iBA4qBK,kCACd,SAAS,mBACT,mBAAmB,gBAClB;;;UChgCc;EACf;EACA;;UAGe;EACf;EACA;IACE;IACA,YAAY;IACZ;IACA;;;IAGA;IACA;MACE;MACA;;;;iBAwHU,4BAA4B,WAAW,wBAAwB;;;UC3I9D;WACN;WACA;WACA;WACA;WACA;WACA,eAAe;WACf,mBAAmB;WACnB,uBAAuB;WACvB;WACA;;WAEA;WACA;WACA;WACA,QAAQ;;WAER,MAAM;WACN,WACP,SAFaA,kCAGV,QAAQ;;UAGE;EACf,eAAe;EACf;EACA;IACM;;IACA;;IACA;IAAkB;;;;;;;;;UAqBT,qCACf,kBAAkB,UAClB;EAEA;EACA,QAAQ,KAAK;EACb;EACA;EACA;EACA;;EAEA;;EAEA;;EAEA,SAAS;EACT;EACA;EACA,oBAAoB,UAAU;EAC9B,oBAAoB,UAAU,WAAW,UAAU;EACnD,MAAM,SAAS,iCAAiC,QAAQ;;;;;;;;;;iBCN1C,+BAA+B,kBAAkB,UAAU,WACzE,QAAQ,qCAAqC,WAAW,aACvD,mBAAmB,WAAW;;;KCvErB,uBAAuB;;UAGlB;EACf;EACA,QAAQ;;EAER,MAAM;;EAEN;;;;;;;EAOA,oBAAoB;;;;;;;EAOpB;;;;;UC0Ce;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;;;;EAKA,eAAe;;;UAIA,sBAAsB;;EAErC;;;KAIU,wBAAwB;;;;;;;KAQxB;EAEN;EACA,KAAK;;EAGL;EACA,eAAe;;EAGf;EACA,eAAe;;EAEf;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;;EAGA;EACA,eAAe;EACf;;EAGA;EACA,eAAe;EACf;;EAGA;EACA,kBAAkB;EAClB,cAAc;EACd;;;KAIM,oBAAoB;UAEf,6BAA6B,kBAAkB,UAAU;;EAExE;;EAEA,QAAQ;;EAER;;EAEA;;EAEA;;;;;EAKA;;EAEA;;EAEA;;EAEA,2BAA2B,QAAQ;;EAEnC;;;;;;EAMA,YAAY;;;;;;EAMZ,oBAAoB,UAAU;;EAE9B,oBAAoB,UAAU,WAAW,UAAU;;EAEnD,SAAS;;;;;EAKT;EACA,SAAS;;;;;;;;;;;EAWT;IAAiB,UAAU;IAAmB;;;iBA6ChC,uBAAuB,kBAAkB,UAAU,WACjE,QAAQ,6BAA6B,WAAW,aAC/C,mBAAmB,WAAW;;;UC3LhB;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;;;;EAKA,YAAY;;KAGF,wBAAwB;UAEnB,iCAAiC,kBAAkB,UAAU;EAC5E;;EAEA;EACA;;EAEA;EACA,SAAS;;;;;EAKT,WAAW;;EAEX;;EAEA;EACA;;EAEA;;EAEA,2BAA2B,QAAQ;EACnC;EACA,oBAAoB,UAAU;EAC9B,oBAAoB,UAAU,WAAW,UAAU;EACnD,SAAS;EACT,SAAS;;;;;;;;;;;;EAYT;IAAiB,UAAU;IAAmB;;;;iBAIhC,2BAA2B,kBAAkB,UAAU,WACrE,QAAQ,iCAAiC,WAAW,aACnD,mBAAmB,WAAW;;;;;;iBC1HjB,YAAY,qBAAqB,kBAAkB,WAAW,aACzE,OAAO,MAAM,KAAK,WAAW,cAC/B,KAAK,WAAW;;;KCkBP;KAEA;UAEK;EACf,MAAM;EACN,UAAU;EACV;;;EAGA,UAAU;;UAGK;EACf,QAAQ;;EAER,QAAQ,OAAO;;EAEf,aAAa;;UAGE;EACf,MAAM;EACN;EACA;EACA;;UAGe;;;;;;;;;;EAUf;IACE;IACA;;IAEA;;;;;;;;;;;;;EAcF;IACE;IACA;IACA;IACA;;;;;;;;;;EAWF;IACE,WAAW,KAAK;IAChB;IACA;IACA;IACA;;;;;;;;iBASY,YAAY,MAAM,aAAa,OAAM,gBAAqB;;;KClG9D;UAUK;EACf,UAAU;;EAEV;;;;;EAKA;;EAEA;;EAEA;;UAGe,oBAAoB;EACnC,SAAS;;UAGM;EACf;EACA,UAAU;EACV;EACA;EACA;;UAGe;EACf,UAAU;EACV,oBAAoB,OAAO;EAC3B;;;cA0DW,yBAAyB;iBA6FtB,eAAe,aAAY,gBAAqB;;;;;iBAkBhD,mBACd,gBACA,qBACA,QAAQ,cACP;;iBA0Fa,cAAc,UAAU,mBAAmB;;;KCvS/C;KAOA,wCAAwC,KAClD;EAGA,SAAS,YAAY;;UAGN;;;EAGf,kBAAkB;;;;;;EAMlB,8BAA8B;;;;EAI9B;;EAEA;;EAEA;;EAEA;;;;EAIA;;;;EAIA;;;;EAIA;;;;EAIA;;;EAGA;;;;EAIA,iBAAiB;;;EAGjB,aAAa;;EAEb,gBAAgB;;EAEhB,SAAS;;;;;EAKT,iBAAiB;;;;;iBAMH,sBAAsB,WAAW,kBAAkB,UACjE,SAAS,+BACR,KAAK,WAAW;;;UCnFF,mBAAmB,kBAAkB,WAAW;EAC/D,WAAW;;;EAGX;;EAEA;;;;EAIA;;EAEA;;EAEA;;;;;;;;iBASc,YAAY,WAAW,kBAAkB,UACvD,SAAS,mBAAmB,aAC3B,KAAK,WAAW;;;UCjCF;EACf;EACA;;EAEA;;EAEA,8BAA8B;;;;;iBAoEhB,sBAAsB,WAAW,kBAAkB,UACjE,UAAU,eAAe,WAAW;UAoBrB;;EAEf,YAAY;;;;;EAKZ,WAAW;IAAQ;IAAoB;IAAmB;IAAgB;;;;;iBAK5D,kBAAkB,WAAW,kBAAkB,UAC7D,UAAU,eAAe,WAAW,aACnC;;;UC3Dc;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA;;;;EAIA,cAAc,SAAS;;EAEvB;;EAEA,UAAU,YAAY;;EAEtB;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;EACA;EACA;EACA;;UAGe;EACf;IACE;IACA;;EAEF;IACE;IACA;IACA;IACA,eAAe;IACf;MACE;MACA;MACA;MACA;;;EAGJ;;UAGe;EACf;EACA,MAAM;EACN;EACA,aAAa;;;;;;UAOE;EACf;;EAEA;EACA;EACA;EACA;;EAEA;EACA;;EAEA;EACA;;EAEA;;EAEA,YAAY;;EAEZ,qBAAqB;;EAErB,UAAU;;EAEV;EACA,QAAQ;;EAER;IACE,UAAU;IACV;IACA;IACA,mBAAmB;;;;;;EAMrB;;EAEA;;EAEA;;;EAGA;;EAEA,SAAS;EACT;EACA;;UAGe,wBAAwB,WAAW,kBAAkB;EACpE;EACA;EACA;EACA,iBAAiB;EACjB,eAAe;EACf;EACA;;EAEA,wBAAwB,eAAe,WAAW;;EAElD,aAAa;IACX;IACA,YAAY;IACZ;;;IAGA,UAAU;MACR;MACA,SAAS;MACT,UAAU,eAAe,WAAW;;;EAGxC,MAAM;;;;EAIN;EACA,mBAAmB,eAAe,WAAW;EAC7C,iBAAiB,eAAe,WAAW;EAC3C,QAAQ;;EAER,8BAA8B;EAC9B,qBAAqB;EACrB,uBAAuB,eAAe,WAAW;;EAEjD,cAAc,cAAc;EAC5B;EACA;EACA,qBAAqB;;UAGN,6BAA6B,WAAW,kBAAkB;EACzE;EACA;EACA;EACA,iBAAiB;EACjB,QAAQ,yBAAyB,WAAW;EAC5C,cAAc,cAAc;EAC5B;EACA;EACA,8BAA8B;;;iBAIhB,6BAA6B,WAAW,kBAAkB,UACxE,OAAO,6BAA6B,WAAW,aAC9C,wBAAwB,WAAW;;iBA0CtB,0BAA0B,WAAW,kBAAkB,UACrE,MAAM,wBAAwB,WAAW,aACxC;;iBA+Ra,0BAA0B,WAAW,kBAAkB,UACrE,UAAU,eAAe,WAAW;;iBAqCtB,2BAA2B,QAAQ,uBAAuB;;;iBAc1D,gBAAgB;;;;;;;;;;;;iBAuHhB,oBACd,QAAQ,sBACR;EAAQ;IACP;;iBA4Ia,qBAAqB;;;;iBAMrB,oBAAoB;UAInB;EACf,QAAQ;EACR,OAAO;;EAEP;EACA;;UAGe,uBAAuB,WAAW,kBAAkB,kBAC3D,wBAAwB,WAAW;;EAE3C,SAAS;;;EAGT,eAAe;;;;;;;;;;;;iBAaK,mBAAmB,WAAW,kBAAkB,UACpE,MAAM,uBAAuB,WAAW,aACvC,QAAQ;;;UCh5BM,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;;;;KCtDU;;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;;;;KCZP;UAUK;;EAEf,SAAS;;EAET;EACA;;EAEA;;EAEA;;;;;;;iBAQoB,qBAAqB;EACzC,UAAU;EACV,YAAY;EACZ,UAAU;;EAEV,QAAQ,WAAW,qCAAqC;;EAExD,QACE,QAAQ,kCACR,OAAO,qCACJ,oBAAoB;IACvB,QAAQ;UA0EK;EACf;;EAEA;;;;;;;;iBASoB,0BAA0B;EAC9C,UAAU;EACV,cAAc;EACd,UAAU;EACV,UAAU;;EAEV,UAAU,sBAAsB;;EAEhC,YAAY;IACV,QAAQ;;iBAwGI,gBAAgB,WAAW,wBAAwB;;UAKzD;EACR;EACA;EACA;EACA,UAAU;EACV,SAAS;EACT;EACA,iBAAiB;EACjB,yBAAyB;EACzB,6BAA6B;EAC7B;EACA,cAAc;EACd,sBAAsB;;;;;;;;;;;;;;iBAeF,iBACpB,OAAO;EAEC,YAAY;;EACZ;EAAkB,UAAU;KAEnC,QAAQ;;;;;;;;iBAsBW,qBACpB,OAAO;EAA8B,UAAU;IAC9C,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC7TM;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;;;UChFc;;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;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBCvNM,eAAe;;;;iBCpBf,yBAAyB,UAAU,UAAU,GAAG;EAC9D;EACA,QAAQ,mBAAmB,GAAG;EAC9B,UAAU,SAAS,mBAAmB;EACtC,QAAQ,UAAU,gBAAgB,UAAU,mBAAmB;IAC7D,mBAAmB,GAAG;;iBAuDV,6BAA6B,UAAU,UAAU,GAAG;EAClE;EACA,kBAAkB,mBAAmB,GAAG;IACtC,mBAAmB,GAAG;;;;;;;;;;;;;;;UC9DT;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER,WAAW;;;;;KCsBD;UAEK;;EAEf;;EAEA;;EAEA;;EAEA,cAAc;;UAGC;EACf;EACA,cAAc;;UAGC;EACf;EACA;EACA;;EAEA;;;UAIe;;EAEf,WAAW;;EAEX,WAAW;;EAEX;;UAGe;EACf;EACA;;EAEA;;;;;;EAMA;;EAEA;;;;;;;;;EASA;;EAEA;;EAEA;;;;;UAMe,0BAA0B;EACzC;EACA,cAAc;;EAEd;;EAEA;;EAEA;;;;;;;;;iBAUc,kBAAkB;EAChC;EACA,cAAc;;;EAGd,gBAAgB;IACd;;;;;;UAqCa;EACf,UAAU;EACV,SAAS;EACT;;;;;;;;;;;;UAae;GAEb,aAAa,uBACb,kBACC;IAAU;IAAkB;;EAC/B,cAAc;;;;;;;;;iBAsLM,iBACpB,MAAM,UACN,OAAO,eACP,kBAAkB,qBACjB,QAAQ;UAyGM;EACf;;EAEA,aAAa;EACb;EACA,WAAW;EACX,SAAS;;EAET;;;;;;EAMA;;;;;;EAMA,UAAU;;;;;;;;iBA2CI,4BACd,MAAM,YACN,OAAM,4BACL;;;;UCnhBc;EACf;EACA;;;;;;;UAQe;EACf;EACA;EACA;EACA;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;EAIA;;;;;;;;KASU,mBACR,oBACA,oBACA;EACE;;;;;;;;;;;;;;;iBAsBU,qBAAqB,iBAAiB,qBAAqB;;;UCvD1D;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;cAQW,8BAA8B;WAGvB,QAAQ;EAF1B,YACE,iBACgB,QAAQ;;;;;;;iBAWZ,0BACd,SAAS,cAAc,aACtB;;;;;;;;iBAuHa,kBACd,SAAS,cAAc,YACvB;EAAQ;IACP;;;;;;cClHU,2BAA2B;EACtC,YAAY;;;;;KAQF,kBAAkB,kBAAkB,UAAU,cACxD,SAASC,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;UAEI,wBAAwB,kBAAkB,UAAU;;EAEnE,UAAUA;;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;;;;;;;EAMvB,aAAa;IAAS,SAASA;IAAc,UAAU;IAAW;;;EAElE;;EAEA;;UAGe;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;;;;;iBAoOhC,iBAAiB,kBAAkB,UAAU,WACjE,MAAM,wBAAwB,WAAW,aACxC,QAAQ,uBAAuB,WAAW;;;;;UC9Z5B;;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;;;UCpKS;EACf;EACA;EACA;EACA;EACA;;UAGe,kBAAkB,kBAAkB,UAAU;WACpD;WACA;WACA;WACA;WACA,mBAAmBC;WACnB,oBAAoB;WACpB,iBAAiB,YAAY,WAAW;WACxC;WACA;WACA,UAAU,YAAY,wBAAwB,WAAW;WACzD;;WAEA;WACA,WAAW,YAAY,wBAAwB,WAAW;WAC1D;WACA;WACA,aAAa,UAAU;WACvB,cACP,UAAU,WACV,UAAU;IACL;IAAgB;;WACd,eAAe;;UAGT,+BAA+B,kBAAkB,UAAU,mBAClE,KACN,wBAAwB,WAAW;;EAgBrC;;UAGe;EACf;EACA;EACA;EACA;EACA;;UAGe,+BAA+B,kBAAkB,UAAU;EAC1E,MAAM,kBAAkB,WAAW;;EAEnC;;EAEA,yBAAyB;EACzB,UAAU,kBAAkB,WAAW;EACvC;EACA,UAAU;EACV;EACA;EACA;EACA,eAAe,wBAAwB,WAAW;EAClD,gBAAgB,wBAAwB,WAAW;EACnD,YAAY;;UAGG,2BAA2B,WAAW,kBAAkB;EACvE;EACA;EACA,QAAQ,uBAAuB,WAAW;EAC1C,UAAU;;UAGK,6BAA6B,kBAAkB,UAAU;EACxE,MAAM,kBAAkB,WAAW;EACnC;EACA,UAAU;EACV;EACA;EACA,YAAY;;UAGG,6BAA6B,WAAW,kBAAkB,kBACjE,uBAAuB,WAAW;EAC1C,UAAU;;iBAGI,wBAAwB,kBAAkB,UAAU,WAClE,MAAM,+BAA+B,WAAW,aAC/C,kBAAkB,WAAW;iBA0HV,wBAAwB,kBAAkB,UAAU,WACxE,MAAM,+BAA+B,WAAW,aAC/C,QAAQ,2BAA2B,WAAW;iBAkD3B,sBAAsB,kBAAkB,UAAU,WACtE,MAAM,6BAA6B,WAAW,aAC7C,QAAQ,6BAA6B,WAAW;;;KClRvC,wBAAwB;EAE9B;EACA,OAAO;EACP,YAAY;EACZ,eAAe;EACf;;EAEA;EAAkB;EAA2B,YAAY;EAAY;;EACrE;EAAkB;EAA+B,YAAY;;;UAGlD,wBAAwB,qBAAqB;EAC5D;EACA;;EAEA,QAAQ,SAAS,OAAO;;EAExB,QAAQ,mBAAmB;EAC3B,UAAU;;EAEV;;;;;;;;;;iBAWc,uBAAuB,qBAAqB,GAC1D,SAAS,wBAAwB,MAAM;;EAoOrC,IAAG;IACD,MAAM;IACN;IACA,gBAAgB,QAAQ;IACxB,SAAS;MACP,QAAQ,wBAAwB;;EAsFpC,aAAa,YAAY;;;;;;;;;iBCnYb;;;;;iBAQA,cAAc,gBAAgB;;;;;;;;;;;;;;;;;UCD7B;EACf;;EAEA;;UAGe;EACf;;EAEA;;EAEA;EACA;;EAEA;;;;;;;;;;iBA4Cc,oBACd,SAAS,kBACT;EAAS;IACR;;;;;;;;;;;iBA0Ba,qBACd,SAAS,kBACT,WACA;EAAS;;;;;;;;;;;iBC1EK,yBAAyB,WACvC,MAAM,mBAAmB,YACzB;EAAS,WAAW;EAAyB;IAC5C;;iBAkBa,oBAAoB,eAAe,mBAAmB;;;iBAyDtD,wBACd,UAAU,cAAc,eACvB;;;iBA+Ba,wBACd,UAAU,cAAc,cACxB,kBACA,QAAQ,sDACP;;;;;;;;;;;;;;;;;;UC5Hc;;WAEN;;WAEA;;WAEA;;WAEA;;UAGM;;EAEf;;;;;;;iBAwBc,qBAAqB,MAAM,uBAAuB;;;;iBCnDlD,0BAA0B,2BAA2B,WAAW;;;;;;;;;;iBAyEhE,iCAAiC,SAAS;;;;iBAoC1C,4BAA4B,SAAS;;iBAgBrC,mBAAmB,SAAS;;iBAW5B,YAAY,SAAS;;iBA4BrB,kBACd,eAAe,gBACf,iBAAiB;;iBAoCH,mBAAmB,SAAS,gBAAgB;;;UC3M3C;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;;;;iBAgGrD,kBACd,SAAS,aACT,uBACC;;;iBAOa,oBAAoB,SAAS,aAAa"}
|
|
@@ -4,8 +4,8 @@ import { R as LlmThinkingMode, x as ChatClient } from "./types-DLxe96rp.js";
|
|
|
4
4
|
import { F as ExternalOptimizerCallbackLimits, V as ExternalOptimizerModelCall, W as ExternalOptimizerModelExecutionObservation, q as ExternalOptimizerRunnerCommand } from "./agent-profile-EpO0HURy.js";
|
|
5
5
|
import { o as Severity } from "./multi-layer-verifier-BUaQ4C17.js";
|
|
6
6
|
import { l as ExactCapableAnalyst } from "./exact-types-BBMMYk3g.js";
|
|
7
|
-
import { i as TraceAnalystDefinition } from "./default-registry-
|
|
8
|
-
import { n as TraceAnalysisEngine } from "./engine-
|
|
7
|
+
import { i as TraceAnalystDefinition } from "./default-registry-BHEfZHQd.js";
|
|
8
|
+
import { n as TraceAnalysisEngine } from "./engine-F_PvwesR.js";
|
|
9
9
|
import { C as SupervisorRunSources, w as SupervisorRunTree } from "./types-v6nRcIqg.js";
|
|
10
10
|
//#region src/semantic-concept-judge.d.ts
|
|
11
11
|
/**
|
|
@@ -476,4 +476,4 @@ declare const KNOWLEDGE_POISONING_KIND_SPEC: TraceAnalystDefinition;
|
|
|
476
476
|
declare const DEFAULT_TRACE_ANALYST_KINDS: readonly TraceAnalystDefinition[];
|
|
477
477
|
//#endregion
|
|
478
478
|
export { SemanticConceptJudgeInput as A, DspyRlmTraceEngineOptions as C, ConceptFinding as D, createChatTraceEngine as E, SemanticConceptJudgeResult as M, runSemanticConceptJudge as N, ConceptSpec as O, renderFindingSubject as S, ChatTraceEngineOptions as T, FindingSubject as _, FAILURE_MODE_KIND_SPEC as a, findingSubjectGrammarPromptFor as b, emitControlIntegrityFindings as c, FindingsStore as d, PersistedFinding as f, FINDING_SUBJECT_SYNTAX as g, FINDING_SUBJECT_KINDS as h, IMPROVEMENT_KIND_SPEC as i, SemanticConceptJudgeOptions as j, SEMANTIC_CONCEPT_JUDGE_VERSION as k, DiffPolicy as l, diffFindings as m, KNOWLEDGE_POISONING_KIND_SPEC as n, CONTROL_INTEGRITY_ANALYST as o, defaultIsMaterial as p, KNOWLEDGE_GAP_KIND_SPEC as r, ControlIntegrityAnalyst as s, DEFAULT_TRACE_ANALYST_KINDS as t, FindingsDiff as u, FindingSubjectKind as v, createDspyRlmTraceEngine as w, parseFindingSubject as x, KIND_EXPECTED_SUBJECTS as y };
|
|
479
|
-
//# sourceMappingURL=index-
|
|
479
|
+
//# sourceMappingURL=index-C9l_-23M.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index-
|
|
1
|
+
{"version":3,"file":"index-C9l_-23M.d.ts","names":[],"sources":["../src/semantic-concept-judge.ts","../src/analyst/chat-trace-engine.ts","../src/analyst/trace-tool-callback.ts","../src/analyst/dspy-rlm-engine.ts","../src/analyst/finding-subject.ts","../src/analyst/findings-store.ts","../src/analyst/kinds/control-integrity.ts","../src/analyst/kinds/failure-mode.ts","../src/analyst/kinds/improvement.ts","../src/analyst/kinds/knowledge-gap.ts","../src/analyst/kinds/knowledge-poisoning.ts","../src/analyst/kinds/index.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAuDY;UAEK;EACf;;EAEA;;EAEA;;EAEA,aAAa;;UAGE;EACf;EACA;;EAEA;EACA;EACA,UAAU;;UAGK;;EAEf;;EAEA;;EAEA,aAAa;IAAQ;IAAc;;;EAEnC,kBAAkB;;EAElB;EACA;;UAGe;EACf;EACA;;EAEA;EACA;EACA;EACA,UAAU;EACV;EACA;EACA;;EAEA;EACA;;;;;;;;KASU;UAQK;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,MAAM;;EAEN,UAAU;EACV,aAAa;EACb;EACA,WAAW;EACX,SAAS;;;;;;EAMT,iBAAiB;;EAEjB,oBAAoB,QAAQ,OAAO;;cAKxB;;;;;;iBAkGS,wBACpB,OAAO,2BACP,SAAS,8BACR,QAAQ;;;UCjMM;;;;;EAKf,MAAM;;EAEN;;EAEA,UAAU;;EAEV;;EAEA;;EAEA,WAAW;;EAEX;;;;;;;;;;EAUA;;;;;;iBAkCc,sBAAsB,SAAS,yBAAyB;;;KCrG5D,0BAA0B;;;UCYrB;;EAEf,MAAM;;EAEN;;EAEA,kBAAkB,aAAa;EAC/B;;EAEA,UAAU;;EAEV;;EAEA;;;;;;EAMA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,kBAAkB,QAAQ;;EAE1B;;;;;;;;;;EAUA;;EAEA,SAAS;;EAET;;;iBAIc,yBAAyB,SAAS,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KCtClE;EAEN;EAAwB;EAAc;;EACtC;EAAyB;;EACzB;EAAuB;;EACvB;EAAyB;;EAGzB;EAAuB;;EACvB;EAAe;;EACf;EAAkB;EAAc;;EAChC;EAAkB;;EAClB;EAAa;EAAgB;;EAC7B;EAAc;;EACd;EAAkB;;EAClB;EAAkB;;EAClB;EAAwB;;EACxB;EAAuB;;EACvB;EAAc;;EACd;EAAa;EAAgB;;EAC7B;EAAgB;;EAChB;EAAqB;;EACrB;EAAuB;;EAEvB;EAA4B;;EAC5B;EAA2B;;EAE3B;EAAiB;;KAEX,qBAAqB;cAEpB,uBAAuB,cAAc;;;;;;;;;;;;;;;;;iBAgGlC,oBAAoB,iCAAiC;;;;;;;iBAoHrD,qBAAqB,GAAG;;;;;;;;;;cA8D3B,wBAAwB,SAAS,OAAO;;;;;;;;;cAoExC,wBAAwB,eAAe,cAAc;;iBAqElD,+BAA+B;;;;;;;;UCxc9B,yBAAyB;EACxC;;cAGW;WAGiB;mBAFX;EAEjB,YAA4B;EAItB,OAAO,eAAe,UAAU,mBAAmB;;EAQzD,WAAW;;EAkBX,QAAQ,gBAAgB;;UAOT;;EAEf,UAAU;;EAEV,aAAa;;EAEb,WAAW;;;;;EAKX,SAAS;IAAQ,UAAU;IAAkB,SAAS;;;UAGvC;;;;;;;;EAQf,cAAc,UAAU,gBAAgB,SAAS;;;;;;iBAOnC,kBAAkB,GAAG,gBAAgB,GAAG;;;;;iBAWxC,aACd,UAAU,oBACV,SAAS,oBACT,SAAQ,aACP;;;;iBC7Fa,6BACd,OAAO,uBAAuB,mBAC9B,qBACC;;cA4BU,mCACA,oBAAoB,uBAAuB;WAE7C;WACA;WAEA;WACA;IAAS;IAAgC;;WACzC;WACA;aACP;aACA;;EAGI,QACJ,OAAO,uBAAuB,mBAC9B,KAAK,iBACJ,QAAQ;;cAUA,2BAAyB;;;cCHzB,wBAAwB;;;cC3BxB,uBAAuB;;;cCGvB,yBAAyB;;;cCPzB,+BAA+B;;;;;;;;;;;;cCP/B,sCAAsC"}
|
|
@@ -1,13 +1,13 @@
|
|
|
1
1
|
import { c as CostLedgerHandle } from "./cost-ledger-DbQdN3nO.js";
|
|
2
2
|
import { a as RunRecord, c as RunSplitTag, n as RunCostProvenance } from "./run-record-DF-fuu31.js";
|
|
3
3
|
import "./types-6VEehmEg.js";
|
|
4
|
-
import "./statistical-heldout-
|
|
5
|
-
import "./types-
|
|
6
|
-
import "./index-
|
|
4
|
+
import "./statistical-heldout-Dq3pnmxd.js";
|
|
5
|
+
import "./types-ZUy5HGY5.js";
|
|
6
|
+
import "./index-B8n12TR2.js";
|
|
7
7
|
import { p as TraceSpanEvent } from "./client-BzL9niqi.js";
|
|
8
8
|
import { d as InsightReport, t as AnalyzeRunsOptions } from "./analyze-runs-Q74o4Hfj.js";
|
|
9
|
-
import "./define-agent-eval-
|
|
10
|
-
import "./default-registry-
|
|
9
|
+
import "./define-agent-eval-D8bPoBM5.js";
|
|
10
|
+
import "./default-registry-BHEfZHQd.js";
|
|
11
11
|
import { AgentCandidateBenchmarkCellRef, AgentCandidateBenchmarkSuiteInputs, AgentCandidateBenchmarkTask, AgentCandidateBenchmarkTaskMaterial, AgentCandidateBundle, AgentCandidateEvaluationPolicy, AgentCandidateExperiment, AgentCandidateExperimentMaterial, AgentCandidateExperimentMeasurement, AgentImprovementCost, AgentImprovementMeasuredComparison, AgentProfileImprovementExperiment, AgentProfileImprovementExperimentMaterial, AgentProfileImprovementMeasuredComparison, AgentProfileImprovementMeasurement, AgentProfileImprovementRunCell, AgentProfileImprovementRunReceipt, AgentProfileImprovementSuiteInputs, AgentProfileImprovementTask, AgentProfileImprovementTaskMaterial, CandidateExecutionEvidence, Sha256Digest } from "@tangle-network/agent-interface";
|
|
12
12
|
//#region src/contract/measured-comparison.d.ts
|
|
13
13
|
interface SealCandidateBenchmarkSuiteOptions {
|
|
@@ -687,4 +687,4 @@ interface FromOtelSpansOptions {
|
|
|
687
687
|
declare function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[];
|
|
688
688
|
//#endregion
|
|
689
689
|
export { FromRunRecordDirResult as $, CodeAgentSessionActionStatus as A, AgentTraceFile as B, fromOpenCodeSession as C, measuredComparisonFromCandidateExperiment as Ct, streamCodeAgentJsonlFile as D, sealCandidateExperiment as Dt, parseCodeAgentJsonlFile as E, sealCandidateBenchmarkTask as Et, CodeAgentSessionTerminalStatus as F, PartitionByAuthoringModelResult as G, AgentTraceRange as H, observeCodeAgentSession as I, EvalReportingSuiteInput as J, parseAgentTrace as K, AgentTraceContributor as L, CodeAgentSessionExecutionReceipt as M, CodeAgentSessionObservation as N, CodeAgentSessionAction as O, verifyCandidateExperiment as Ot, CodeAgentSessionSource as P, FromRunRecordDirOptions as Q, AgentTraceContributorType as R, fromKimiCodeSession as S, evaluatePairedMeasurements as St, parseCodeAgentJsonl as T, sealCandidateBenchmarkSuite as Tt, AgentTraceRecord as U, AgentTraceIndex as V, AuthoringProvenance as W, EvalReportingSuiteResult as X, EvalReportingSuiteOptions as Y, evalReportingSuite as Z, CodeAgentSessionIntakeResult as _, PairedMeasurement as _t, FromFeedbackTableOptions as a, RunAgentProfileImprovementExperimentOptions as at, fromClaudeCodeSession as b, RunCandidateExperimentOptions as bt, CodeAgentStoreObservation as c, runAgentProfileImprovementExperiment as ct, CodeAgentStoreScope as d, sealAgentProfileImprovementTask as dt, RunRecordRejection as et, CodeAgentStoreSession as f, verifyAgentProfileImprovementExperimentComparison as ft, CodeAgentSessionIntakeOptions as g, EvaluatePairedMeasurementsOptions as gt, CodeAgentSessionDiagnostic as h, CompareCandidateExperimentOptions as ht, FeedbackTableRow as i, CompareAgentProfileImprovementExperimentOptions as it, CodeAgentSessionActionSurface as j, CodeAgentSessionActionKind as k, verifyCandidateExperimentComparison as kt, CodeAgentStoreRejection as l, sealAgentProfileImprovementExperiment as lt, CodeAgentJsonlLine as m, CandidateExperimentRun as mt, fromOtelSpans as n, AgentProfileImprovementExperimentExecutionInput as nt, FromFeedbackTableResult as o, SealAgentProfileImprovementSuiteOptions as ot, observeCodeAgentStore as p, CandidateExperimentExecutionInput as pt, partitionRunsByAuthoringModel as q, FeedbackTableMeta as r, AgentProfileImprovementExperimentRun as rt, fromFeedbackTable as s, measuredComparisonFromAgentProfileImprovementExperiment as st, FromOtelSpansOptions as t, fromRunRecordDir as tt, CodeAgentStoreRejections as u, sealAgentProfileImprovementSuite as ut, CodeAgentSessionMetrics as v, PairedMeasurementAdapter as vt, fromPiSession as w, runCandidateExperiment as wt, fromCodexSession as x, SealCandidateBenchmarkSuiteOptions as xt, ParsedCodeAgentJsonl as y, PairedMeasurementEvaluation as yt, AgentTraceConversation as z };
|
|
690
|
-
//# sourceMappingURL=index-
|
|
690
|
+
//# sourceMappingURL=index-CG7mtu_G.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index-
|
|
1
|
+
{"version":3,"file":"index-CG7mtu_G.d.ts","names":[],"sources":["../src/contract/measured-comparison.ts","../src/contract/profile-measured-comparison.ts","../src/contract/intake/run-record-dir.ts","../src/contract/eval-reporting-suite.ts","../src/contract/intake/agent-trace.ts","../src/contract/intake/code-agent-observation.ts","../src/contract/intake/code-agent-session.ts","../src/contract/intake/code-agent-store.ts","../src/contract/intake/feedback-table.ts","../src/contract/intake/otel-spans.ts"],"mappings":";;;;;;;;;;;;UAiCiB;EACf,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,QAAQ;EACR,MAAM;EACN,eAAe;EACf;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QAAQ,OAAO,oCAAoC,QAAQ;;EAE3D;;EAEA,aAAa;EACb,SAAS;;UAGM;EACf,cAAc;EACd;IACE;IACA,MAAM;;;UAIO;EACf,YAAY;EACZ,cAAc;EACd;IACE;IACA,MAAM;;EAER,aAAa;EACb;EACA,YAAY;EACZ;EACA,WAAW;;;UAII,kBAAkB;EACjC;EACA,UAAU;EACV,WAAW;;;UAII,yBAAyB;EACxC,MAAM,KAAK;EACX,WAAW,KAAK;IAAkB;IAAc;;EAChD,QAAQ,KAAK;EACb,eAAe,KAAK,OAAO;EAC3B,UAAU,KAAK;EACf,UAAU,KAAK;EACf,OAAO,KAAK;;UAGG,kCAAkC;EACjD,uBAAuB,kBAAkB;EACzC,QAAQ;EACR,SAAS,yBAAyB;;EAElC;;EAEA,kBAAkB;;EAElB,kBAAkB;;;KAIR,8BAA8B,KACxC;EAGA,iBAAiB;EACjB,WAAW;EACX;;;iBAIc,2BACd,UAAU,sCACT;;iBAQa,4BACd,SAAS,qCACR;;iBAiBa,wBACd,UAAU,mCACT;iBAQa,0BAA0B,iBAAiB;;iBAarC,uBACpB,SAAS,gCACR,QAAQ;;;;;;;;iBAkEK,2BAA2B,MACzC,SAAS,kCAAkC,QAC1C;;iBA8Ya,0CACd,SAAS,oCACR;;iBAuEa,oCACd,iBACC;;;UCjrBc;EACf,aAAa;EACb,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,aAAa;EACb,MAAM;EACN,SAAS;EACT;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QACE,OAAO,kDACN,QAAQ;;EAEX;;EAEA,aAAa;EACb,SAAS;;UAGM;EACf,cAAc;EACd;IACE;IACA,MAAM;;;UAIO;EACf,YAAY;EACZ,cAAc;EACd;IACE;IACA,MAAM;;EAER,aAAa;EACb;EACA,YAAY;EACZ;EACA,WAAW;;;iBAIG,gCACd,UAAU,sCACT;;iBAQa,iCACd,SAAS,0CACR;;iBAqBa,sCACd,UAAU,4CACT;;;;;;iBAsBmB,qCACpB,SAAS,8CACR,QAAQ;;iBA0CK,wDACd,SAAS,kDACR;;iBAuEa,kDACd,iBACC;;;;UC7Oc;;EAEf;;EAEA;;EAEA;;UAGe;;;;;;EAMf;;;;;;;EAOA,WAAW;;;;;;EAMX;;UAGe;;EAEf,MAAM;;EAEN,UAAU;;EAEV;;;;;;;;;;iBAkBoB,iBACpB,cACA,UAAS,0BACR,QAAQ;;;;;KC7CC,0BAA0B;UAErB;;;;;EAKf,UAAU,KAAK;;EAEf,OAAO;;;;;;;;;;EAUP;;;;UAKe;;;EAGf,QAAQ;;EAER;;IAEE;;IAEA;;;IAGA;;IAEA;;;IAGA,UAAU;;;EAGZ;;;;;;;iBAUoB,mBACpB,OAAO,yBACP,UAAS,4BACR,QAAQ;;;KChEC;UAEK;EACf,MAAM;;EAEN;;UAGe;EACf;EACA;EACA;;;EAGA,cAAc;;UAGC;EACf;EACA,cAAc;EACd,QAAQ;;UAGO;EACf;EACA,eAAe;;UAGA;EACf;EACA;EACA;EACA;IAAQ;IAAc;;EACtB;IAAS;IAAe;;EACxB,OAAO;;;;UAOQ;EACf;;EAEA;;EAEA;EACA;EACA;;EAEA;;EAEA;;KAGU,kBAAkB,YAAY;;;;;;iBAW1B,gBAAgB,SAAS,qBAAqB;UAmE7C;;;;EAIf,SAAS,YAAY;;;EAGrB,cAAc;;;;;;;;iBASA,8BACd,MAAM,aACN,OAAO,kBACN;;;KCjLS;KAEA;KAEA;KAEA;KASA;UAEK;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA,MAAM;EACN,SAAS;EACT;EACA,QAAQ;EACR;EACA;EACA,UAAU;;UAGK;EACf,QAAQ;EACR;EACA;EACA;IACE,QAAQ;IACR;;EAEF,SAAS;;UAGM;EACf,QAAQ;EACR;EACA;EACA,YAAY;;;;;;;iBAeE,wBACd,SAAS,iCACR;;;UCjCc;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf,QAAQ;EACR;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,WAAW;EACX;;UAGe;EACf,MAAM;EACN,aAAa;EACb,SAAS;EACT,cAAc;;UAGC;EACf;EACA;EACA;EACA;EACA;EACA;EACA,WAAW;EACX;EACA;EACA;EACA;EACA;EACA;;;EAGA,iBAAiB;;;EAGjB,YAAY;;;;KAKF;EACN;EAAe;EAAoB;;EACnC;EAAmB;;iBAcT,oBAAoB,gBAAgB;;;;;;;;;;iBAuB7B,yBAAyB,eAAe,eAAe;;;;;iBA6BxD,wBAAwB,eAAe,QAAQ;iBAUrD,iBACd,SAAS,gCACR;iBAIa,sBACd,SAAS,gCACR;iBAIa,oBACd,SAAS,gCACR;iBAIa,oBACd,SAAS,gCACR;iBAIa,cACd,SAAS,gCACR;;;;UC/Lc;;;;;WAKN;;WAEA,QAAQ;;;;;WAKR;;;;;WAKA;aACE;;aAEA;;;;KAKD;;UAUK;WACN;WACA;WACA;WACA;WACA,aAAa;;;UAIP;WACN;WACA,UAAU,SAAS,QAAQ,OAAO;;WAElC,QAAQ;aACN;aACA,QAAQ;;;;;;;;;KAUT;WAEG;WACA,OAAO;WACP;WACA,mBAAmB;;WAEnB;;WAEA;;WAEA;WACA,WAAW,SAAS,QAAQ,OAAO;WACnC,YAAY;;WAGZ;WACA,OAAO;WACP;;WAEA;WACA,YAAY;;;;;;;;;iBAaL,sBACpB,OAAO,sBACN,QAAQ;;;UCtHM;;;EAGf;;EAEA;;;EAGA;;;EAGA,WAAW;;UAGI;EACf;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,WAAW;;;EAGX,SAAS;;UAGM;;EAEf,SAAS;;;EAGT,OAAO;;;;EAIP;IAAU;IAAa;;;;;EAIvB;;UAGe;EACf,MAAM;;;EAGN,aAAa;IAAQ;IAAe;IAAe;;;iBAGrC,kBAAkB,MAAM,2BAA2B;;;UC3BlD;EACf,OAAO;;EAEP,eAAe;;EAEf;;;;;;EAMA,eAAe,eAAe,gBAAgB;;iBAGhC,cAAc,MAAM,uBAAuB"}
|
package/dist/index.d.ts
CHANGED
|
@@ -6,18 +6,18 @@ import { _ as roundTripRunRecord, a as RunRecord, b as searchCellRunId, c as Run
|
|
|
6
6
|
import { A as DEFAULT_TRACE_ANALYST_BUDGETS, B as TraceAnalystSpan, F as SearchTraceResult, G as ViewTraceOversized, H as TraceAnalystSpanStatus, I as SpanMatchRecord, K as ViewTraceResult, L as TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, M as ErrorCluster, N as QueryTracesPage, O as TraceAnalysisStore, P as SearchSpanResult, R as TraceAnalystByteBudgets, S as createBoundedTraceAnalysisStore, U as TraceAnalystTraceSummary, V as TraceAnalystSpanKind, W as ViewSpansResult, _ as ProposalFinding, b as makeFinding, c as AnalystRunInputs, d as AnalystSeverity, f as AnalystUsageReceipt, i as AnalystFinding, j as DatasetOverview, l as AnalystRunResult, n as AnalystContext, p as EvidenceRef, s as AnalystRunEvent, t as Analyst, u as AnalystRunSummary, x as makeProposalFinding, y as computeFindingId, z as TraceAnalystFilters } from "./types-6VEehmEg.js";
|
|
7
7
|
import { $ as NoopRawProviderSink, A as createChatClient, B as LlmToolCall, C as ChatResponse, Ct as judgeFamily, D as MockTransportOpts, E as CustomTransportOpts, F as LlmChargeBounds, G as costReceiptFromLlmError, H as LlmToolDefinition, I as LlmMessage, J as stripFencedJson, K as isTransientLlmError, L as LlmResponseError, M as LlmCallMetadata, N as LlmCallRequest, O as OpenAiCompatibleTransportOpts, P as LlmCallResult, S as ChatRequest, St as assertCrossFamily, T as CreateChatClientOpts, U as LlmUsage, V as LlmToolChoice, W as costReceiptFromLlm, Y as FileSystemRawProviderSink, Z as InMemoryRawProviderSink, _ as ScenarioResult, _t as checkServedModel, a as CompletionCriterion, b as ChatCallOpts, bt as CrossFamilyError, c as JudgeInput, ct as AssertServedModelOptions, d as PersonaConfig, dt as ServedModelCheck, f as ProductClientConfig, ft as ServedModelPolicy, g as ScenarioFile, gt as assertServedModels, h as Scenario, ht as assertServedModel, i as CollectedArtifacts, j as LlmCallError, k as SandboxSdkTransportOpts, l as JudgeRubric, lt as ModelSubstitutionError, m as RubricDimension, mt as assertCrossFamilyServed, n as ArtifactResult, nt as RawProviderEvent, o as DriverState, p as RouteMap, pt as ServedModelVerdict, q as maximumChargeForLlmRequest, r as CheckResult, rt as RawProviderSink, s as JudgeFn, st as AssertCrossFamilyServedOptions, t as ArtifactCheck, u as JudgeScore, ut as ServedCrossFamilyError, v as Turn, vt as servedModelAcceptable, w as ChatTransport, x as ChatClient, xt as JudgeFamily, y as TurnResult, yt as AssertCrossFamilyOptions, z as LlmTokenLogprob } from "./types-DLxe96rp.js";
|
|
8
8
|
import { a as ContinuousCalibrationResult, d as calibrateJudgeContinuous, f as continuousAgreement, h as verbosityBias, i as ContinuousAgreementOptions, l as VerbosityBiasResult, n as CandidateScore, o as GoldenItem, r as ContinuousAgreement, t as CalibrationResult, u as calibrateJudge } from "./judge-calibration-DFtEMlde.js";
|
|
9
|
-
import {
|
|
9
|
+
import { G as CorpusAgreementPerDimension, H as SeriesConvergenceResult, J as corpusInterRaterAgreement, K as CorpusAgreementReport, U as analyzeSeries, V as SeriesConvergenceOptions, W as CorpusAgreementOptions, X as interRaterReliability, Y as corpusInterRaterAgreementFromJudgeScores, _ as assertJudgeMayGate, a as JudgeCalibrationSet, b as registerCalibrationSet, c as JudgeGateDecision, d as JudgeGateRefusedError, f as JudgeGateStatus, g as addCalibrationVerdicts, h as OwnerVerdict, i as JudgeAgreement, l as JudgeGateInput, m as JudgeRunMeta, n as CalibrationMerge, o as JudgeConfusion, p as JudgeRun, q as CorpusScoreRecord, r as DEFAULT_JUDGE_GATE_POLICY, s as JudgeExampleAgreement, t as CalibrationExample, u as JudgeGatePolicy, v as judgeGateDecision, x as snapshotFromJudgeAgreement, y as measureJudgeAgreement } from "./judge-gate-BuQd7sSm.js";
|
|
10
10
|
import { a as partialCredit, c as spearmanR, d as weightedMean, i as confidenceInterval, l as summarizeNumberSeries, n as WeightedCompositeInput, o as pearsonR, r as WeightedCompositeResult, s as ranks, t as SeriesDistribution, u as weightedComposite } from "./descriptive-B2iPaT9J.js";
|
|
11
11
|
import { n as bonferroni, r as holm, t as benjaminiHochberg } from "./multiplicity-DIWHvysC.js";
|
|
12
|
-
import { A as PairArmsOptions, Bt as RiskDifferenceResult, C as hashJson, Ct as verifyAttestation, D as ComparePairedArmsOptions, F as PairedCorrectness, Ft as EProcessStep, Gt as pairedRiskDifference, Ht as isBinaryOutcomeVector, I as PairedMetricDelta, It as eProcess, Jt as passAtK, Kt as pairedRiskDifferenceExact, L as comparePairedArms, Lt as ExactRiskDifferenceResult, M as PairRunRecordsResult, Mt as EProcess, N as PairedArmRow, Nt as EProcessOptions, O as MatchedPair, P as PairedArmsComparison, Pt as EProcessState, R as pairArms, Rt as McNemarResult, St as attest, Ut as mcnemar, Vt as ScoreRiskDifferenceResult, Wt as pairedBinaryScale, Yt as wilson, bt as AttestationVerification, et as Objective, j as PairArmsResult, k as MatchedRunRecordPair, nt as dominates, qt as pairedRiskDifferenceScore, rt as paretoFrontier, tt as ParetoResult, vt as ATTESTATION_ALGORITHM, w as manifestContentDigest, xt as AttestedReport, yt as AttestationProvenance, z as pairRunRecords, zt as ProportionInterval } from "./statistical-heldout-
|
|
12
|
+
import { A as PairArmsOptions, Bt as RiskDifferenceResult, C as hashJson, Ct as verifyAttestation, D as ComparePairedArmsOptions, F as PairedCorrectness, Ft as EProcessStep, Gt as pairedRiskDifference, Ht as isBinaryOutcomeVector, I as PairedMetricDelta, It as eProcess, Jt as passAtK, Kt as pairedRiskDifferenceExact, L as comparePairedArms, Lt as ExactRiskDifferenceResult, M as PairRunRecordsResult, Mt as EProcess, N as PairedArmRow, Nt as EProcessOptions, O as MatchedPair, P as PairedArmsComparison, Pt as EProcessState, R as pairArms, Rt as McNemarResult, St as attest, Ut as mcnemar, Vt as ScoreRiskDifferenceResult, Wt as pairedBinaryScale, Yt as wilson, bt as AttestationVerification, et as Objective, j as PairArmsResult, k as MatchedRunRecordPair, nt as dominates, qt as pairedRiskDifferenceScore, rt as paretoFrontier, tt as ParetoResult, vt as ATTESTATION_ALGORITHM, w as manifestContentDigest, xt as AttestedReport, yt as AttestationProvenance, z as pairRunRecords, zt as ProportionInterval } from "./statistical-heldout-Dq3pnmxd.js";
|
|
13
13
|
import { _ as pairedSignTest, a as PairedPromotionDecision, c as BOOTSTRAP_GATE_MIN_N, d as PairedBootstrapResult, f as PairedSignTestResult, g as pairedDeltaTieFraction, h as pairedBootstrap, i as PairedMcNemarEvidence, l as DECISION_PAIRED_DELTA_STATISTIC, m as SignTestAlternative, n as PairedDecisionShape, o as PairedPromotionDecisionOptions, p as PairedTTestResult, r as PairedDecisionStatistic, s as decidePairedPromotion, t as PairedDecisionMethod, u as PairedBootstrapOptions, v as pairedTTest } from "./paired-promotion-decision-DPKz8ZnF.js";
|
|
14
14
|
import { a as requiredPairedSampleSize, i as pairedMde, n as mcnemarPower, o as requiredSampleSize, r as mcnemarRequiredN, t as mulberry32 } from "./random-DGEyr3zG.js";
|
|
15
15
|
import { C as RankTestMethod, D as WilcoxonSignedRankResult, E as WILCOXON_EXACT_MAX_N, O as mannWhitneyU, S as MannWhitneyResult, T as RankTestOptions, _ as bootstrapCi, a as ReleaseConfidenceIssue, b as MANN_WHITNEY_EXACT_MAX_STATES, f as evaluateReleaseConfidence, g as Verdict, i as ReleaseConfidenceInput, k as wilcoxonSignedRank, m as BootstrapResult, o as ReleaseConfidenceMetrics, p as BootstrapOptions, s as ReleaseConfidenceScorecard, t as ActionableSideInfo, u as ReleaseTraceEvidence, w as RankTestMethodRequest, x as MANN_WHITNEY_EXACT_MAX_WORK, y as DEFAULT_PERMUTATIONS } from "./release-confidence-CkoXFx9I.js";
|
|
16
16
|
import { a as RunFilter, i as InMemoryTraceStore, n as FileSystemTraceStore, o as SpanFilter, s as TraceStore, t as EventFilter } from "./store-C4avFjIi.js";
|
|
17
17
|
import { a as MintedRolloutLine } from "./schema-C-ziyJD7.js";
|
|
18
18
|
import { a as mintRolloutRows, c as ScorePreference, i as RolloutSearchLineage, n as MintRolloutResult, t as MintRolloutOptions } from "./mint-C1W8w5wo.js";
|
|
19
|
-
import { R as Scenario$1, S as JudgeConfig, _ as GateDecision, a as CampaignResult, w as JudgeScore$1 } from "./types-
|
|
20
|
-
import { Ar as DEFAULT_RED_TEAM_CORPUS, At as RequirementCheck, Br as CanaryKind, Ct as extractProducedState, Dt as LlmCorrectnessCheckerOpts, Et as CorrectnessChecker, Fr as redTeamDataset, Ft as verifyCompletion, Hr as CanaryReport, Ir as redTeamReport, Ji as llmJudge, Lr as scoreRedTeamOutput, Mr as RedTeamCategory, Mt as TaskGold, Nr as RedTeamFinding, Nt as completionVerdict, Ot as ProducedProposal, Pr as RedTeamReport, Pt as createLlmCorrectnessChecker, Qi as transientDispatchFailure, Rr as CanaryAlert, St as ToolCallEventLike, Tt as CompletionVerdict, Ur as runCanaries, Vr as CanaryOptions, Zi as quotaExhaustedUntil, _a as runCampaign, _t as assertRealBackend, bt as ProposalEventLike, ga as RunCampaignOptions, gt as BackendIntegrityReport, ha as CampaignCellRetryPolicy, ht as BackendIntegrityError, jr as RedTeamCase, jt as SatisfiedBy, kt as ProducedState, ma as CampaignCellFailureReceipt, qi as LlmJudgeOptions, vt as summarizeBackendIntegrity, wt as CompletionRequirement, xt as RuntimeEventLike, yt as ArtifactEventLike, zr as CanaryEvaluation } from "./index-
|
|
19
|
+
import { R as Scenario$1, S as JudgeConfig, _ as GateDecision, a as CampaignResult, w as JudgeScore$1 } from "./types-ZUy5HGY5.js";
|
|
20
|
+
import { Ar as DEFAULT_RED_TEAM_CORPUS, At as RequirementCheck, Br as CanaryKind, Ct as extractProducedState, Dt as LlmCorrectnessCheckerOpts, Et as CorrectnessChecker, Fr as redTeamDataset, Ft as verifyCompletion, Hr as CanaryReport, Ir as redTeamReport, Ji as llmJudge, Lr as scoreRedTeamOutput, Mr as RedTeamCategory, Mt as TaskGold, Nr as RedTeamFinding, Nt as completionVerdict, Ot as ProducedProposal, Pr as RedTeamReport, Pt as createLlmCorrectnessChecker, Qi as transientDispatchFailure, Rr as CanaryAlert, St as ToolCallEventLike, Tt as CompletionVerdict, Ur as runCanaries, Vr as CanaryOptions, Zi as quotaExhaustedUntil, _a as runCampaign, _t as assertRealBackend, bt as ProposalEventLike, ga as RunCampaignOptions, gt as BackendIntegrityReport, ha as CampaignCellRetryPolicy, ht as BackendIntegrityError, jr as RedTeamCase, jt as SatisfiedBy, kt as ProducedState, ma as CampaignCellFailureReceipt, qi as LlmJudgeOptions, vt as summarizeBackendIntegrity, wt as CompletionRequirement, xt as RuntimeEventLike, yt as ArtifactEventLike, zr as CanaryEvaluation } from "./index-B8n12TR2.js";
|
|
21
21
|
import { a as ProfileAxisSpec, c as expandProfileAxes, i as HarnessType, l as harnessAxisOf, n as CODING_HARNESSES, o as agentProfileHash, r as HARNESS_NATIVE_MODEL, s as agentProfileId, t as AgentProfile } from "./agent-profile-EpO0HURy.js";
|
|
22
22
|
import { i as DatasetSplit, n as DatasetManifest, r as DatasetScenario, t as Dataset } from "./dataset-DQqhOCPt.js";
|
|
23
23
|
import { C as defineEquivalenceCheck, S as buildEquivalenceRecord, _ as StrategyChecker, a as CheckerIdentity, b as VerificationStrategyProfile, c as EquivalenceCheckDefinition, d as EquivalenceCheckerInput, f as EquivalenceCheckerResult, g as EquivalenceRecord, h as EquivalenceProtocolError, i as equivalenceVerdict, l as EquivalenceCheckSpec, m as EquivalenceObligationStatus, n as VerdictCertification, o as CheckerOutcome, p as EquivalenceObligation, r as certificationEvidenceDigest, s as EquivalenceArm, t as DefaultVerdict, u as EquivalenceChecker, v as VERIFICATION_STRATEGIES, w as runEquivalenceCheck, x as VerificationStrategySource, y as VERIFICATION_STRATEGY_SOURCES } from "./verdict-E4eRNf7-.js";
|
|
@@ -27,11 +27,11 @@ import { a as FAILURE_BLAME, c as FailureContext, d as INFRA_FAILURE_BLAMES, f a
|
|
|
27
27
|
import { d as InsightReport, i as analyzeRuns, t as AnalyzeRunsOptions } from "./analyze-runs-Q74o4Hfj.js";
|
|
28
28
|
import { i as ExactAnalystRunEvent, l as ExactCapableAnalyst, o as ExactAnalystRunResult } from "./exact-types-BBMMYk3g.js";
|
|
29
29
|
import { c as RegistryRunOpts, i as BudgetPolicy, n as AnalystRegistry, r as AnalystRegistryOptions, s as ExactRegistryRunOpts } from "./registry-CT-hMSkR.js";
|
|
30
|
-
import { _ as SelfImproveMethodResult, a as DefinedAgentEval, c as SelfImproveMethodOptions, d as SelfImproveProposerOptions, f as SelfImproveProposerResult, g as SelfImproveMethodProvenance, h as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveOptions, n as AgentEvalEvaluateOptions, o as defineAgentEval, p as SelfImproveResult, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveProgressEvent } from "./define-agent-eval-
|
|
31
|
-
import { a as createTraceAnalyst, i as TraceAnalystDefinition, n as buildDefaultAnalystRegistry, r as CreateTraceAnalystOptions, t as DefaultAnalystRegistryOptions } from "./default-registry-
|
|
32
|
-
import { i as TraceAnalysisEngineResult, l as RawAnalystFinding, n as TraceAnalysisEngine, t as DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS } from "./engine-
|
|
33
|
-
import { A as SemanticConceptJudgeInput, C as DspyRlmTraceEngineOptions, D as ConceptFinding, E as createChatTraceEngine, M as SemanticConceptJudgeResult, N as runSemanticConceptJudge, O as ConceptSpec, T as ChatTraceEngineOptions, _ as FindingSubject, a as FAILURE_MODE_KIND_SPEC, d as FindingsStore, f as PersistedFinding, j as SemanticConceptJudgeOptions, k as SEMANTIC_CONCEPT_JUDGE_VERSION, l as DiffPolicy, m as diffFindings, t as DEFAULT_TRACE_ANALYST_KINDS, u as FindingsDiff, v as FindingSubjectKind, w as createDspyRlmTraceEngine } from "./index-
|
|
34
|
-
import { i as BenchmarkEvaluation } from "./types-
|
|
30
|
+
import { _ as SelfImproveMethodResult, a as DefinedAgentEval, c as SelfImproveMethodOptions, d as SelfImproveProposerOptions, f as SelfImproveProposerResult, g as SelfImproveMethodProvenance, h as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveOptions, n as AgentEvalEvaluateOptions, o as defineAgentEval, p as SelfImproveResult, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveProgressEvent } from "./define-agent-eval-D8bPoBM5.js";
|
|
31
|
+
import { a as createTraceAnalyst, i as TraceAnalystDefinition, n as buildDefaultAnalystRegistry, r as CreateTraceAnalystOptions, t as DefaultAnalystRegistryOptions } from "./default-registry-BHEfZHQd.js";
|
|
32
|
+
import { i as TraceAnalysisEngineResult, l as RawAnalystFinding, n as TraceAnalysisEngine, t as DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS } from "./engine-F_PvwesR.js";
|
|
33
|
+
import { A as SemanticConceptJudgeInput, C as DspyRlmTraceEngineOptions, D as ConceptFinding, E as createChatTraceEngine, M as SemanticConceptJudgeResult, N as runSemanticConceptJudge, O as ConceptSpec, T as ChatTraceEngineOptions, _ as FindingSubject, a as FAILURE_MODE_KIND_SPEC, d as FindingsStore, f as PersistedFinding, j as SemanticConceptJudgeOptions, k as SEMANTIC_CONCEPT_JUDGE_VERSION, l as DiffPolicy, m as diffFindings, t as DEFAULT_TRACE_ANALYST_KINDS, u as FindingsDiff, v as FindingSubjectKind, w as createDspyRlmTraceEngine } from "./index-C9l_-23M.js";
|
|
34
|
+
import { i as BenchmarkEvaluation } from "./types-DZQpUJse.js";
|
|
35
35
|
import { a as SpanHandle, n as CaptureWrite, o as TraceEmitter, s as TraceEmitterOptions, t as CaptureErrorHandler } from "./emitter-HcCmFQl_.js";
|
|
36
36
|
import { n as SandboxDriver } from "./sandbox-harness-BlSOu4LX.js";
|
|
37
37
|
import { S as shareAllowed, b as redactForShare, d as ShareSafetyVerdict, f as assessShareSafety, i as RedactOptions, o as RedactionProfile, r as REDACTION_VERSION, s as RedactionReport, u as ShareSafetyStatus, x as redactText, y as redact } from "./redact-Cveoq6fJ.js";
|
|
@@ -40,7 +40,7 @@ import { n as TrajectoryStep, r as buildTrajectory, t as Trajectory } from "./tr
|
|
|
40
40
|
import { a as runCounterfactual, i as CounterfactualRunner, n as CounterfactualMutation, r as CounterfactualResult, t as CounterfactualContext } from "./counterfactual-CPaWN99-.js";
|
|
41
41
|
import { A as AnalystFindingDigest, B as ControlRunResult, C as createFeedbackTrajectory, D as renderPreferenceMemoryMarkdown, E as feedbackTrajectoryToOptimizerRow, F as ControlActionOutcome, G as StopDecision, H as ControlRuntimeError, I as ControlBudget, J as subjectiveEval, K as objectiveEval, L as ControlContext, M as AnalystRunDigest, N as analystFindingDigest, O as summarizePreferenceMemory, P as analystRunDigest, R as ControlDecision, S as controlRunToFeedbackTrajectory, T as feedbackTrajectoriesToOptimizerRows, U as ControlSeverity, V as ControlRuntimeConfig, W as ControlStep, _ as PreferenceMemoryEntry, a as FeedbackLabel, b as analystRunToReviewRequests, c as FeedbackOptimizerRow, d as FeedbackTask, f as FeedbackTrajectory, g as InMemoryFeedbackTrajectoryStore, h as FileSystemFeedbackTrajectoryStore, i as FeedbackAttempt, j as AnalystReviewDecision, k as withAssignedFeedbackSplit, l as FeedbackOutcome, m as FeedbackTrajectoryStore, n as AnalystReviewRequest, o as FeedbackLabelKind, p as FeedbackTrajectoryFilter, q as runAgentControlLoop, r as FeedbackArtifactType, s as FeedbackLabelSource, t as AnalystFeedbackTrajectoryOptions, u as FeedbackSplitPolicy, v as ProposedSideEffect, w as feedbackTrajectoriesToDatasetScenarios, x as assignFeedbackSplit, y as analystRunToFeedbackTrajectory, z as ControlEvalResult } from "./feedback-trajectory-C6NJxXEa.js";
|
|
42
42
|
import { y as OUTPUT_VALUE } from "./attribute-vocabulary-CcXqHODT.js";
|
|
43
|
-
import { C as inferDomainKeywords, H as captureFetchToRawSink, I as OtlpExport, J as extractUsageFromSse, K as extractUsage, M as AnalyzeTracesResult, N as analyzeTraces, O as scoreTraceInsightReadiness, P as OtlpFlatLine, R as OtlpSpan, S as domainEvidencePattern, W as ExtractedUsage, g as TraceInsightSuite, h as TraceInsightReadiness, k as tokenizeDomainWords, n as toolSpansToTraceAnalysisStore, o as otlpTextToTraceAnalysisStore, r as OtlpFileTraceStore, v as buildTraceInsightContext, x as describeTraceInsightScope, y as buildTraceInsightPrompt, z as exportRunAsOtlp } from "./store-tool-spans-
|
|
43
|
+
import { C as inferDomainKeywords, H as captureFetchToRawSink, I as OtlpExport, J as extractUsageFromSse, K as extractUsage, M as AnalyzeTracesResult, N as analyzeTraces, O as scoreTraceInsightReadiness, P as OtlpFlatLine, R as OtlpSpan, S as domainEvidencePattern, W as ExtractedUsage, g as TraceInsightSuite, h as TraceInsightReadiness, k as tokenizeDomainWords, n as toolSpansToTraceAnalysisStore, o as otlpTextToTraceAnalysisStore, r as OtlpFileTraceStore, v as buildTraceInsightContext, x as describeTraceInsightScope, y as buildTraceInsightPrompt, z as exportRunAsOtlp } from "./store-tool-spans-CAF3T-BS.js";
|
|
44
44
|
import { a as RunIntegrityReport, o as assertRunCaptured, t as RunIntegrityError } from "./integrity-BWO6RPxv.js";
|
|
45
45
|
import { c as judgeSpans, f as runsForScenario, i as argHash } from "./query-DBUQhehH.js";
|
|
46
46
|
import { a as SteeringChange, c as CampaignRunContext, d as CampaignVariant, f as EvalCampaignOptions, h as runEvalCampaign, i as Researcher, l as CampaignRunOutcome, m as FailedRun, n as ExperimentResult, o as CampaignFactoryParams, p as EvalCampaignResult, r as FailureMode, s as CampaignIntegrityPolicy, t as ExperimentPlan, u as CampaignScenario } from "./researcher-zTpxNJBy.js";
|
|
@@ -3695,5 +3695,5 @@ export declare class PairwiseSteeringOptimizer {
|
|
|
3695
3695
|
optimize(rows: SteeringOptimizationRow[], config?: SteeringOptimizerConfig): SteeringOptimizationResult;
|
|
3696
3696
|
}
|
|
3697
3697
|
//#endregion
|
|
3698
|
-
export { AGENT_PROFILE_KINDS, ATTESTATION_ALGORITHM, type ActionExecutionPolicy, type ActionPolicyDecision, type ActionableSideInfo, type ActiveLearningOptions, type AgentEvalAgent, AgentEvalError, type AgentEvalErrorCode, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, type AgentProfile, type AgentProfileCell, type AgentProfileCellInput, type AgentProfileDimensionValue, type AgentProfileJson, type AgentProfileSourceInput, type Analyst, type AnalystContext, type AnalystFeedbackTrajectoryOptions, type AnalystFinding, type AnalystFindingDigest, AnalystRegistry, type AnalystRegistryOptions, type AnalystReviewDecision, type AnalystReviewRequest, type AnalystRunDigest, type AnalystRunEvent, type AnalystRunInputs, type AnalystRunResult, type AnalystRunSummary, type AnalystSeverity, type AnalystUsageReceipt, type AnalyzeRunsOptions, type AnalyzeTracesResult, type ArgumentCheck, type Artifact, type ArtifactCheck, type ArtifactEventLike, type ArtifactResult, type AssertCapabilityHeadroomOptions, type AssertCrossFamilyOptions, type AssertCrossFamilyServedOptions, type AssertServedModelOptions, type AssertSingleBackendOptions, type AttestationProvenance, type AttestationVerification, type AttestedReport, BOOTSTRAP_GATE_MIN_N, type BackendDescriptor, BackendIntegrityError, type BackendIntegrityReport, type BenchmarkEvaluation, type BlendWeights, type BootstrapOptions, type BootstrapResult, type BoundedProcessInput, type BoundedProcessResult, type BudgetLedgerEntry, type BudgetPolicy, type BudgetSpec, CODING_HARNESSES, type CalibrationResult, type CampaignCellFailureReceipt, type CampaignCellRetryPolicy, type CampaignFactoryParams, type CampaignIntegrityPolicy, type CampaignResult, type CampaignRunContext, type CampaignRunOutcome, type CampaignScenario, type CampaignVariant, type CanaryAlert, type CanaryEvaluation, type CanaryKind, type CanaryLeak, type CanaryOptions, type CanaryReport, type CandidateScore, type CapabilityHeadroomOptions, type CapabilityHeadroomResult, type CaptureErrorHandler, type CaptureWrite, type CellVerdict, type ChannelRollup, type ChatCallOpts, type ChatClient, type ChatRequest, type ChatResponse, type ChatTraceEngineOptions, type ChatTransport, type CheckResult, type CheckerIdentity, type CheckerOutcome, type CliffsMagnitude, type CollectedArtifacts, type CommandRunner, type ComparePairedArmsOptions, type CompletionCriterion, type CompletionRequirement, type CompletionVerdict, type ConceptFinding, type ConceptSpec, ConfigError, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type ContractCheckResult, type ContractLintFinding, type ContractRule, type ContractRuleExecution, type ContractRun, type ContractSpan, type ContractStatus, type ContractVerdict, type ControlActionOutcome, type ControlBudget, type ControlContext, type ControlDecision, type ControlEvalResult, type ControlRunResult, type ControlRuntimeConfig, type ControlRuntimeError, type ControlStep, type CorpusAgreementOptions, type CorpusAgreementPerDimension, type CorpusAgreementReport, type CorpusScoreRecord, type CorrectnessChecker, CostAccountingIncompleteError, CostCallConflictError, CostCeilingReachedError, type CostChannel, type CostEntry, CostLedger, type CostLedgerFilter, type CostLedgerHandle, type CostLedgerOptions, type CostLedgerPersistence, CostLedgerPersistenceError, type CostLedgerSummary, type CostProvenance, type CostReceipt, CostReceiptCaptureError, type CostReceiptInput, CostReservationExceededError, type CostResult, type CostSummary, type CostUsage, type CounterfactualContext, type CounterfactualMutation, type CounterfactualResult, type CounterfactualRunner, type CreateChatClientOpts, type CreateTraceAnalystOptions, CrossFamilyError, type CustomTokenPricing, type CustomTransportOpts, DECISION_PAIRED_DELTA_STATISTIC, DEFAULT_FAILURE_REASON_RULES, DEFAULT_PERMUTATIONS, DEFAULT_RED_TEAM_CORPUS, DEFAULT_TRACE_ANALYST_BUDGETS, DEFAULT_TRACE_ANALYST_KINDS, DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS, type DataAcquisitionPlan, type DatasetManifest, type DatasetOverview, type DatasetScenario, type DatasetSplit, type DefaultAnalystRegistryOptions, type DefaultVerdict, type DefineAgentEvalOptions, type DefinedAgentEval, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DetectorEvent, type DetectorSignal, type DiffPolicy, type DiffScorecardOptions, type JudgeScore as DimensionJudgeScore, type DirEntry, type DiscoverPersonasOptions, type DiscoveredPersona, type DriverState, type DspyRlmTraceEngineOptions, type EProcess, type EProcessOptions, type EProcessState, type EProcessStep, type EnsembleAggregate, type EnsembleJudgeOptions, type EquivalenceArm, type EquivalenceCheckDefinition, type EquivalenceCheckSpec, type EquivalenceChecker, type EquivalenceCheckerInput, type EquivalenceCheckerResult, type EquivalenceObligation, type EquivalenceObligationStatus, EquivalenceProtocolError, type EquivalenceRecord, type ErrorCluster, type ErrorCountPattern, type ErrorStreakOptions, type EvalCampaignOptions, type EvalCampaignResult, type EventFilter, type EvidenceRef, type ExactAnalystRunEvent, type ExactAnalystRunResult, type ExactCapableAnalyst, type ExactRegistryRunOpts, type ExactRiskDifferenceResult, type ExperimentPlan, type ExperimentRep, type ExperimentResult, type ExperimentStats, type ExtractOptions, type ExtractResult, type ExtractedUsage, FAILURE_BLAME, FAILURE_CLASSES, FAILURE_MODE_KIND_SPEC, type FailedRun, type FailureBlame, type FailureClass, type FailureClassification, type FailureContext, type FailureMode, type FailureReasonRule, type FailureRule, type FeedbackArtifactType, type FeedbackAttempt, type FeedbackLabel, type FeedbackLabelKind, type FeedbackLabelSource, type FeedbackOptimizerRow, type FeedbackOutcome, type FeedbackSplitPolicy, type FeedbackTask, type FeedbackTrajectory, type FeedbackTrajectoryFilter, type FeedbackTrajectoryStore, type FieldDestination, FileSystemFeedbackTrajectoryStore, FileSystemRawProviderSink, FileSystemTraceStore, type Finding, type FindingSubject, type FindingSubjectKind, type FindingToPolicyEditOptions, type FindingsDiff, FindingsStore, type GainDistributionBin, type GainDistributionFigureSpec, type GainDistributionOptions, type GateDecision, type GateEvidence, type GenericSpan, type GoldenItem, HARNESS_NATIVE_MODEL, type HarnessType, type HeadroomInput, HeldOutGate, type HeldOutGateConfig, type GateDecision$1 as HeldOutGateDecision, type HeldOutGateRejectionCode, type HeldOutPartition, type HiddenCriteriaGrader, type HiddenGradeResult, type HiddenLeak, INFRA_FAILURE_BLAMES, type ImprovementThresholds, type ImprovementVerdictResult, InMemoryFeedbackTrajectoryStore, InMemoryRawProviderSink, InMemoryTraceStore, type InsightReport, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InterimReleaseConfidence, type InterimReleaseConfidenceInput, JudgeError, type JudgeFamily, type JudgeFn, type JudgeInput, type JudgeRetryOutcome, type JudgeRetryPolicy, type JudgeRubric, type JudgeScore$1 as JudgeScore, type JudgeScoreInput, type JudgeScoresRecord, type JudgeSpan, type JudgeVerdict, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type KnowledgeAcquisitionMode, type KnowledgeBundle, type KnowledgeFreshness, type KnowledgeImportance, type KnowledgeReadinessReport, type KnowledgeRequirement, type KnowledgeRequirementCategory, type KnowledgeSensitivity, type Layer, type LayerResult, type LayerStatus, type LeaderboardOptions, type LeaderboardRow, LlmCallError, type LlmCallMetadata, type LlmCallRequest, type LlmCallResult, type LlmChargeBounds, type LlmCorrectnessCheckerOpts, type LlmJudgeOptions, type LlmMessage, LlmResponseError, type LlmReviewerConfig, type LlmSpan, type LlmTokenLogprob, type LlmToolCall, type LlmToolChoice, type LlmToolDefinition, type LlmUsage, MANN_WHITNEY_EXACT_MAX_STATES, MANN_WHITNEY_EXACT_MAX_WORK, type MannWhitneyResult, type MatchedPair, type MatchedRunRecordPair, type MaximumCharge, type McNemarResult, type MintRolloutOptions, type MintRolloutResult, type MintedRolloutLine, type MockTransportOpts, type ModelEndpointCheck, type ModelEndpointRequest, type ModelSeats, ModelSubstitutionError, MultiLayerVerifier, type NoLeakOptions, NoopRawProviderSink, NotFoundError, OUTPUT_VALUE, type Objective, type OpenAiCompatibleTransportOpts, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OrderMode, type OtlpExport, OtlpFileTraceStore, type OtlpFlatLine, type OtlpSpan, type PaidCallResult, type PairArmsOptions, type PairArmsResult, type PairRunRecordsResult, type PairedArmRow, type PairedArmsComparison, type PairedBootstrapOptions, type PairedBootstrapResult, type PairedCorrectness, type PairedDecisionMethod, type PairedDecisionShape, type PairedDecisionStatistic, type PairedDeltaTestOptions, type PairedDeltaTestResult, type PairedEvalueOptions, type PairedEvalueSequence, type PairedEvalueStep, type PairedMcNemarEvidence, type PairedMetricDelta, type PairedPromotionAlternative, type PairedPromotionAlternativeCell, type PairedPromotionCallRates, type PairedPromotionDecision, type PairedPromotionDecisionOptions, type PairedPromotionPowerCall, type PairedPromotionPowerOptions, type PairedPromotionPowerResult, type PairedSignTestResult, type PairedTTestResult, type PairedValueLaw, type ParetoPoint, type ParetoResult, type PartitionHeldOutOptions, type PendingCostCall, type PendingCostCallView, type PersistedFinding, type PersonaConfig, type PolicyEdit, type PolicyEditAdmission, type PolicyEditAdmissionOptions, type PolicyEditAxis, type PolicyEditCandidateRecord, type PolicyEditChange, type PolicyEditExpectedGain, type PolicyEditGainDirection, type PolicyEditGainUnit, type PolicyEditInit, type PolicyEditRisk, type PolicyEditSchemaVersion, type PolicyEditSource, type PolicyEditTarget, type PolicyEditTargetSurface, type PreferenceMemoryEntry, type ProducedProposal, type ProducedState, type ProductBenchmarkExportOptions, type ProductBenchmarkExportResult, type ProductBenchmarkManifest, type ProductBenchmarkSingleRunExportOptions, type ProductBenchmarkSplit, type ProductBenchmarkValidationReport, type ProductClientConfig, type Scenario as ProductScenario, type ProfileAxisSpec, type ProjectRuntimeTrajectoryEvidenceOptions, type PromptHandle, type ProportionInterval, type ProposalEventLike, type ProposalFinding, type ProposeFn, type ProposeInput, type ProposeOutput, type ProposeReviewConfig, type ProposeReviewControlAction, type ProposeReviewControlConfig, type ProposeReviewControlResult, type ProposeReviewControlState, type ProposeReviewReport, type ProposedSideEffect, type QueryTracesPage, REDACTION_VERSION, type RankTestMethod, type RankTestMethodRequest, type RankTestOptions, type RawAnalystFinding, type RawProviderEvent, type RawProviderSink, type RecordRunsOptions, type RedTeamCase, type RedTeamCategory, type RedTeamFinding, type RedTeamReport, type RedactOptions, type RedactionProfile, type RedactionReport, type ReferenceReplayCaseRun, type ReferenceReplayRun, type ReferenceReplaySplit, type ReflectionContext, type ReflectionProposal, type RegistryRunOpts, type ReleaseConfidenceInput, type ReleaseConfidenceIssue, type ReleaseConfidenceMetrics, type ReleaseConfidenceScorecard, type ReleaseTraceEvidence, type RepeatedActionOptions, type RequiredPairsForPairedPromotionOptions, type RequiredPairsForPairedPromotionResult, type RequirementCheck, type ResearchReport, type ResearchReportOptions, type Researcher, type RetryWrite, type Review, type ReviewFn, type ReviewInput, type ReviewMemoryEntry, type ReviewMemoryStore, type RiskDifferenceResult, type RolloutSearchLineage, type RouteMap, type RoutedField, type RubricDimension, type Run, type RunCampaignOptions, type RunCaptureReport, type RunCommandInput, type RunCommandResult, type RunCostProvenance, type RunFilter, RunIntegrityError, type RunIntegrityReport, type RunJudgeMetadata, type RunOutcome, type RunPaidCallInput, type RunRecord, type RunRecordBackend, RunRecordValidationError, type RunScore, type RunScoreWeights, type RunSearchCoordinates, type RunSplitTag, type RunStatus, type RunTaskFailure, type RunTerminalOutcome, type RunTokenUsage, type RunTraceRef, type RuntimeEventLike, type RuntimeTrajectoryEvidenceProjection, type RuntimeTrajectoryEvidenceSummary, type RuntimeTrajectoryHookEvent, type RuntimeTrajectoryRecord, type RuntimeTrajectoryRunRecord, SEMANTIC_CONCEPT_JUDGE_VERSION, type SandboxDriver, type SandboxSdkTransportOpts, type SatisfiedBy, type Scenario$1 as Scenario, type ScenarioCost, type ScenarioFile, type ScenarioResult, type ScorePreference, type ScoreRiskDifferenceResult, type Scorecard, type ScorecardCell, type ScorecardCellDiff, type ScorecardDiff, type ScorecardEntry, type ScorecardLogLine, type SearchSpanResult, type SearchTraceResult, type SelfImproveBudget, type SelfImproveMethodOptions, type SelfImproveMethodProvenance, type SelfImproveMethodResult, type SelfImproveOptions, type SelfImproveProgressEvent, type SelfImproveProposerOptions, type SelfImproveProposerResult, type SelfImproveResult, type SemanticConceptJudgeInput, type SemanticConceptJudgeOptions, type SemanticConceptJudgeResult, type SequentialDecision, type SeriesConvergenceOptions, type SeriesConvergenceResult, type SeriesDistribution, ServedCrossFamilyError, type ServedModelCheck, type ServedModelPolicy, type ServedModelVerdict, type Severity, type ShareSafetyStatus, type ShareSafetyVerdict, type SignTestAlternative, type SingleBackendDivergence, type SingleBackendReport, type Span, type SpanFilter, type SpanHandle, type SpanMatchRecord, type SpanPredicate, type SplitCoverage, type SteeringBundle, type SteeringChange, type SteeringOptimizationResult, type SteeringOptimizationRow, type StopDecision, type StrategyChecker, type StreamingDetector, type SummaryTable, type SummaryTableOptions, type SummaryTableRow, type SynthesisTarget, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, type TaskGold, type TaskHeadroom, type ToolCallEventLike, type ToolSpan, type ToolStats, type ToolUseMetrics, type ToolUseOptions, type TraceAnalysisEngine, type TraceAnalysisEngineResult, type TraceAnalysisStore, type TraceAnalystByteBudgets, type TraceAnalystDefinition, type TraceAnalystFilters, type TraceAnalystSpan, type TraceAnalystSpanKind, type TraceAnalystSpanStatus, type TraceAnalystTraceSummary, type TraceContract, type TraceContractSpec, TraceEmitter, type TraceEmitterOptions, type TraceEvent, type TraceInsightReadiness, type TraceInsightSuite, type TraceStore, type TrainSelectionTestOptions, type TrainSelectionTestPartition, type Trajectory, type TrajectoryStep, type TreatmentGate, type TreatmentGateInput, type TreatmentGateOptions, type TrialTrace, type Turn, type TurnResult, UNKNOWN_MODEL, type UserQuestion, VERIFICATION_STRATEGIES, VERIFICATION_STRATEGY_SOURCES, ValidationError, type VerbosityBiasResult, type Verdict, type VerdictCacheStore, type VerdictCertification, type Verification, type VerificationReport, type VerificationStrategyProfile, type VerificationStrategySource, type VerifyFn, type VerifyOptions, type ViewSpansResult, type ViewTraceOversized, type ViewTraceResult, type ViteDeployRunnerInput, WILCOXON_EXACT_MAX_N, type WeightedCompositeInput, type WeightedCompositeResult, type WilcoxonSignedRankResult, type WranglerDeployRunnerInput, agentProfileCellHashMaterial, agentProfileCellKey, agentProfileHash, agentProfileId, analystFindingDigest, analystRunDigest, analystRunToFeedbackTrajectory, analystRunToReviewRequests, analyzeRuns, analyzeSeries, analyzeTraces, argHash, assertCrossFamily, assertCrossFamilyServed, assertRealBackend, assertRunCaptured, assertServedModel, assertServedModels, assessShareSafety, assignFeedbackSplit, attest, benjaminiHochberg, bonferroni, bootstrapCi, budgetBreachView, buildAgentProfileCell, buildDefaultAnalystRegistry, buildEquivalenceRecord, buildTraceInsightContext, buildTraceInsightPrompt, buildTrajectory, calibrateJudge, calibrateJudgeContinuous, captureFetchToRawSink, certificationEvidenceDigest, checkServedModel, classifyFailure, classifyFailureReason, comparePairedArms, completionVerdict, computeFindingId, computeToolUseMetrics, confidenceInterval, continuousAgreement, controlRunToFeedbackTrajectory, corpusInterRaterAgreement, corpusInterRaterAgreementFromJudgeScores, costForTokenPricing, costForUsage, costReceiptFromLlm, costReceiptFromLlmError, createBoundedTraceAnalysisStore, createChatClient, createChatTraceEngine, createDspyRlmTraceEngine, createFeedbackTrajectory, createLlmCorrectnessChecker, createTraceAnalyst, decidePairedPromotion, defineAgentEval, defineEquivalenceCheck, describeTraceInsightScope, diffFindings, domainEvidencePattern, dominates, eProcess, equivalenceVerdict, evaluateInterimReleaseConfidence, evaluateReleaseConfidence, expandProfileAxes, exportRunAsOtlp, extractProducedState, extractUsage, extractUsageFromSse, failureBlame, failureClusterView, feedbackTrajectoriesToDatasetScenarios, feedbackTrajectoriesToOptimizerRows, feedbackTrajectoryToOptimizerRow, gainHistogram, gradeSemanticStatus, groupRunsByAgentProfileCell, harnessAxisOf, hashJson, holm, inferDomainKeywords, interRaterReliability, iqr, isBinaryOutcomeVector, isJudgeSpan, isLlmSpan, isRunRecord, isToolSpan, isTransientLlmError, judgeAgreementView, judgeFamily, judgeSpans, llmJudge, makeFinding, makeProposalFinding, manifestContentDigest, mannWhitneyU, maximumChargeForLlmRequest, mcnemar, mcnemarPower, mcnemarRequiredN, mintRolloutRows, modelHasSnapshot, modelPriceKey, mulberry32, objectiveEval, otlpTextToTraceAnalysisStore, pairArms, pairRunRecords, pairedBinaryScale, pairedBootstrap, pairedDeltaTieFraction, pairedEvalueSequence, pairedMde, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, pairedSignTest, pairedTTest, paretoChart, paretoFrontier, parseRunRecordSafe, partialCredit, passAtK, pearsonR, index_d_exports as profile, quotaExhaustedUntil, ranks, redTeamDataset, redTeamReport, redact, redactForShare, redactText, renderPreferenceMemoryMarkdown, requiredPairedSampleSize, requiredSampleSize, roundTripRunRecord, runAgentControlLoop, runCampaign, runCanaries, runCostFloorUsd, runCounterfactual, runEquivalenceCheck, runEvalCampaign, runSemanticConceptJudge, runTaskScore, runsForScenario, scoreRedTeamOutput, scoreTraceInsightReadiness, searchCellRunId, selfImprove, servedModelAcceptable, shareAllowed, spearmanR, stripFencedJson, subjectiveEval, summarizeBackendIntegrity, summarizeNumberSeries, summarizePreferenceMemory, summaryTable, toAgentProfileJson, tokenizeDomainWords, toolSpansToTraceAnalysisStore, toolWasteView, transientDispatchFailure, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyAttestation, verifyCompletion, weightedComposite, weightedMean, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit };
|
|
3698
|
+
export { AGENT_PROFILE_KINDS, ATTESTATION_ALGORITHM, type ActionExecutionPolicy, type ActionPolicyDecision, type ActionableSideInfo, type ActiveLearningOptions, type AgentEvalAgent, AgentEvalError, type AgentEvalErrorCode, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, type AgentProfile, type AgentProfileCell, type AgentProfileCellInput, type AgentProfileDimensionValue, type AgentProfileJson, type AgentProfileSourceInput, type Analyst, type AnalystContext, type AnalystFeedbackTrajectoryOptions, type AnalystFinding, type AnalystFindingDigest, AnalystRegistry, type AnalystRegistryOptions, type AnalystReviewDecision, type AnalystReviewRequest, type AnalystRunDigest, type AnalystRunEvent, type AnalystRunInputs, type AnalystRunResult, type AnalystRunSummary, type AnalystSeverity, type AnalystUsageReceipt, type AnalyzeRunsOptions, type AnalyzeTracesResult, type ArgumentCheck, type Artifact, type ArtifactCheck, type ArtifactEventLike, type ArtifactResult, type AssertCapabilityHeadroomOptions, type AssertCrossFamilyOptions, type AssertCrossFamilyServedOptions, type AssertServedModelOptions, type AssertSingleBackendOptions, type AttestationProvenance, type AttestationVerification, type AttestedReport, BOOTSTRAP_GATE_MIN_N, type BackendDescriptor, BackendIntegrityError, type BackendIntegrityReport, type BenchmarkEvaluation, type BlendWeights, type BootstrapOptions, type BootstrapResult, type BoundedProcessInput, type BoundedProcessResult, type BudgetLedgerEntry, type BudgetPolicy, type BudgetSpec, CODING_HARNESSES, type CalibrationExample, type CalibrationMerge, type CalibrationResult, type CampaignCellFailureReceipt, type CampaignCellRetryPolicy, type CampaignFactoryParams, type CampaignIntegrityPolicy, type CampaignResult, type CampaignRunContext, type CampaignRunOutcome, type CampaignScenario, type CampaignVariant, type CanaryAlert, type CanaryEvaluation, type CanaryKind, type CanaryLeak, type CanaryOptions, type CanaryReport, type CandidateScore, type CapabilityHeadroomOptions, type CapabilityHeadroomResult, type CaptureErrorHandler, type CaptureWrite, type CellVerdict, type ChannelRollup, type ChatCallOpts, type ChatClient, type ChatRequest, type ChatResponse, type ChatTraceEngineOptions, type ChatTransport, type CheckResult, type CheckerIdentity, type CheckerOutcome, type CliffsMagnitude, type CollectedArtifacts, type CommandRunner, type ComparePairedArmsOptions, type CompletionCriterion, type CompletionRequirement, type CompletionVerdict, type ConceptFinding, type ConceptSpec, ConfigError, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type ContractCheckResult, type ContractLintFinding, type ContractRule, type ContractRuleExecution, type ContractRun, type ContractSpan, type ContractStatus, type ContractVerdict, type ControlActionOutcome, type ControlBudget, type ControlContext, type ControlDecision, type ControlEvalResult, type ControlRunResult, type ControlRuntimeConfig, type ControlRuntimeError, type ControlStep, type CorpusAgreementOptions, type CorpusAgreementPerDimension, type CorpusAgreementReport, type CorpusScoreRecord, type CorrectnessChecker, CostAccountingIncompleteError, CostCallConflictError, CostCeilingReachedError, type CostChannel, type CostEntry, CostLedger, type CostLedgerFilter, type CostLedgerHandle, type CostLedgerOptions, type CostLedgerPersistence, CostLedgerPersistenceError, type CostLedgerSummary, type CostProvenance, type CostReceipt, CostReceiptCaptureError, type CostReceiptInput, CostReservationExceededError, type CostResult, type CostSummary, type CostUsage, type CounterfactualContext, type CounterfactualMutation, type CounterfactualResult, type CounterfactualRunner, type CreateChatClientOpts, type CreateTraceAnalystOptions, CrossFamilyError, type CustomTokenPricing, type CustomTransportOpts, DECISION_PAIRED_DELTA_STATISTIC, DEFAULT_FAILURE_REASON_RULES, DEFAULT_JUDGE_GATE_POLICY, DEFAULT_PERMUTATIONS, DEFAULT_RED_TEAM_CORPUS, DEFAULT_TRACE_ANALYST_BUDGETS, DEFAULT_TRACE_ANALYST_KINDS, DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS, type DataAcquisitionPlan, type DatasetManifest, type DatasetOverview, type DatasetScenario, type DatasetSplit, type DefaultAnalystRegistryOptions, type DefaultVerdict, type DefineAgentEvalOptions, type DefinedAgentEval, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DetectorEvent, type DetectorSignal, type DiffPolicy, type DiffScorecardOptions, type JudgeScore as DimensionJudgeScore, type DirEntry, type DiscoverPersonasOptions, type DiscoveredPersona, type DriverState, type DspyRlmTraceEngineOptions, type EProcess, type EProcessOptions, type EProcessState, type EProcessStep, type EnsembleAggregate, type EnsembleJudgeOptions, type EquivalenceArm, type EquivalenceCheckDefinition, type EquivalenceCheckSpec, type EquivalenceChecker, type EquivalenceCheckerInput, type EquivalenceCheckerResult, type EquivalenceObligation, type EquivalenceObligationStatus, EquivalenceProtocolError, type EquivalenceRecord, type ErrorCluster, type ErrorCountPattern, type ErrorStreakOptions, type EvalCampaignOptions, type EvalCampaignResult, type EventFilter, type EvidenceRef, type ExactAnalystRunEvent, type ExactAnalystRunResult, type ExactCapableAnalyst, type ExactRegistryRunOpts, type ExactRiskDifferenceResult, type ExperimentPlan, type ExperimentRep, type ExperimentResult, type ExperimentStats, type ExtractOptions, type ExtractResult, type ExtractedUsage, FAILURE_BLAME, FAILURE_CLASSES, FAILURE_MODE_KIND_SPEC, type FailedRun, type FailureBlame, type FailureClass, type FailureClassification, type FailureContext, type FailureMode, type FailureReasonRule, type FailureRule, type FeedbackArtifactType, type FeedbackAttempt, type FeedbackLabel, type FeedbackLabelKind, type FeedbackLabelSource, type FeedbackOptimizerRow, type FeedbackOutcome, type FeedbackSplitPolicy, type FeedbackTask, type FeedbackTrajectory, type FeedbackTrajectoryFilter, type FeedbackTrajectoryStore, type FieldDestination, FileSystemFeedbackTrajectoryStore, FileSystemRawProviderSink, FileSystemTraceStore, type Finding, type FindingSubject, type FindingSubjectKind, type FindingToPolicyEditOptions, type FindingsDiff, FindingsStore, type GainDistributionBin, type GainDistributionFigureSpec, type GainDistributionOptions, type GateDecision, type GateEvidence, type GenericSpan, type GoldenItem, HARNESS_NATIVE_MODEL, type HarnessType, type HeadroomInput, HeldOutGate, type HeldOutGateConfig, type GateDecision$1 as HeldOutGateDecision, type HeldOutGateRejectionCode, type HeldOutPartition, type HiddenCriteriaGrader, type HiddenGradeResult, type HiddenLeak, INFRA_FAILURE_BLAMES, type ImprovementThresholds, type ImprovementVerdictResult, InMemoryFeedbackTrajectoryStore, InMemoryRawProviderSink, InMemoryTraceStore, type InsightReport, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InterimReleaseConfidence, type InterimReleaseConfidenceInput, type JudgeAgreement, type JudgeCalibrationSet, type JudgeConfusion, JudgeError, type JudgeExampleAgreement, type JudgeFamily, type JudgeFn, type JudgeGateDecision, type JudgeGateInput, type JudgeGatePolicy, JudgeGateRefusedError, type JudgeGateStatus, type JudgeInput, type JudgeRetryOutcome, type JudgeRetryPolicy, type JudgeRubric, type JudgeRun, type JudgeRunMeta, type JudgeScore$1 as JudgeScore, type JudgeScoreInput, type JudgeScoresRecord, type JudgeSpan, type JudgeVerdict, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type KnowledgeAcquisitionMode, type KnowledgeBundle, type KnowledgeFreshness, type KnowledgeImportance, type KnowledgeReadinessReport, type KnowledgeRequirement, type KnowledgeRequirementCategory, type KnowledgeSensitivity, type Layer, type LayerResult, type LayerStatus, type LeaderboardOptions, type LeaderboardRow, LlmCallError, type LlmCallMetadata, type LlmCallRequest, type LlmCallResult, type LlmChargeBounds, type LlmCorrectnessCheckerOpts, type LlmJudgeOptions, type LlmMessage, LlmResponseError, type LlmReviewerConfig, type LlmSpan, type LlmTokenLogprob, type LlmToolCall, type LlmToolChoice, type LlmToolDefinition, type LlmUsage, MANN_WHITNEY_EXACT_MAX_STATES, MANN_WHITNEY_EXACT_MAX_WORK, type MannWhitneyResult, type MatchedPair, type MatchedRunRecordPair, type MaximumCharge, type McNemarResult, type MintRolloutOptions, type MintRolloutResult, type MintedRolloutLine, type MockTransportOpts, type ModelEndpointCheck, type ModelEndpointRequest, type ModelSeats, ModelSubstitutionError, MultiLayerVerifier, type NoLeakOptions, NoopRawProviderSink, NotFoundError, OUTPUT_VALUE, type Objective, type OpenAiCompatibleTransportOpts, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OrderMode, type OtlpExport, OtlpFileTraceStore, type OtlpFlatLine, type OtlpSpan, type OwnerVerdict, type PaidCallResult, type PairArmsOptions, type PairArmsResult, type PairRunRecordsResult, type PairedArmRow, type PairedArmsComparison, type PairedBootstrapOptions, type PairedBootstrapResult, type PairedCorrectness, type PairedDecisionMethod, type PairedDecisionShape, type PairedDecisionStatistic, type PairedDeltaTestOptions, type PairedDeltaTestResult, type PairedEvalueOptions, type PairedEvalueSequence, type PairedEvalueStep, type PairedMcNemarEvidence, type PairedMetricDelta, type PairedPromotionAlternative, type PairedPromotionAlternativeCell, type PairedPromotionCallRates, type PairedPromotionDecision, type PairedPromotionDecisionOptions, type PairedPromotionPowerCall, type PairedPromotionPowerOptions, type PairedPromotionPowerResult, type PairedSignTestResult, type PairedTTestResult, type PairedValueLaw, type ParetoPoint, type ParetoResult, type PartitionHeldOutOptions, type PendingCostCall, type PendingCostCallView, type PersistedFinding, type PersonaConfig, type PolicyEdit, type PolicyEditAdmission, type PolicyEditAdmissionOptions, type PolicyEditAxis, type PolicyEditCandidateRecord, type PolicyEditChange, type PolicyEditExpectedGain, type PolicyEditGainDirection, type PolicyEditGainUnit, type PolicyEditInit, type PolicyEditRisk, type PolicyEditSchemaVersion, type PolicyEditSource, type PolicyEditTarget, type PolicyEditTargetSurface, type PreferenceMemoryEntry, type ProducedProposal, type ProducedState, type ProductBenchmarkExportOptions, type ProductBenchmarkExportResult, type ProductBenchmarkManifest, type ProductBenchmarkSingleRunExportOptions, type ProductBenchmarkSplit, type ProductBenchmarkValidationReport, type ProductClientConfig, type Scenario as ProductScenario, type ProfileAxisSpec, type ProjectRuntimeTrajectoryEvidenceOptions, type PromptHandle, type ProportionInterval, type ProposalEventLike, type ProposalFinding, type ProposeFn, type ProposeInput, type ProposeOutput, type ProposeReviewConfig, type ProposeReviewControlAction, type ProposeReviewControlConfig, type ProposeReviewControlResult, type ProposeReviewControlState, type ProposeReviewReport, type ProposedSideEffect, type QueryTracesPage, REDACTION_VERSION, type RankTestMethod, type RankTestMethodRequest, type RankTestOptions, type RawAnalystFinding, type RawProviderEvent, type RawProviderSink, type RecordRunsOptions, type RedTeamCase, type RedTeamCategory, type RedTeamFinding, type RedTeamReport, type RedactOptions, type RedactionProfile, type RedactionReport, type ReferenceReplayCaseRun, type ReferenceReplayRun, type ReferenceReplaySplit, type ReflectionContext, type ReflectionProposal, type RegistryRunOpts, type ReleaseConfidenceInput, type ReleaseConfidenceIssue, type ReleaseConfidenceMetrics, type ReleaseConfidenceScorecard, type ReleaseTraceEvidence, type RepeatedActionOptions, type RequiredPairsForPairedPromotionOptions, type RequiredPairsForPairedPromotionResult, type RequirementCheck, type ResearchReport, type ResearchReportOptions, type Researcher, type RetryWrite, type Review, type ReviewFn, type ReviewInput, type ReviewMemoryEntry, type ReviewMemoryStore, type RiskDifferenceResult, type RolloutSearchLineage, type RouteMap, type RoutedField, type RubricDimension, type Run, type RunCampaignOptions, type RunCaptureReport, type RunCommandInput, type RunCommandResult, type RunCostProvenance, type RunFilter, RunIntegrityError, type RunIntegrityReport, type RunJudgeMetadata, type RunOutcome, type RunPaidCallInput, type RunRecord, type RunRecordBackend, RunRecordValidationError, type RunScore, type RunScoreWeights, type RunSearchCoordinates, type RunSplitTag, type RunStatus, type RunTaskFailure, type RunTerminalOutcome, type RunTokenUsage, type RunTraceRef, type RuntimeEventLike, type RuntimeTrajectoryEvidenceProjection, type RuntimeTrajectoryEvidenceSummary, type RuntimeTrajectoryHookEvent, type RuntimeTrajectoryRecord, type RuntimeTrajectoryRunRecord, SEMANTIC_CONCEPT_JUDGE_VERSION, type SandboxDriver, type SandboxSdkTransportOpts, type SatisfiedBy, type Scenario$1 as Scenario, type ScenarioCost, type ScenarioFile, type ScenarioResult, type ScorePreference, type ScoreRiskDifferenceResult, type Scorecard, type ScorecardCell, type ScorecardCellDiff, type ScorecardDiff, type ScorecardEntry, type ScorecardLogLine, type SearchSpanResult, type SearchTraceResult, type SelfImproveBudget, type SelfImproveMethodOptions, type SelfImproveMethodProvenance, type SelfImproveMethodResult, type SelfImproveOptions, type SelfImproveProgressEvent, type SelfImproveProposerOptions, type SelfImproveProposerResult, type SelfImproveResult, type SemanticConceptJudgeInput, type SemanticConceptJudgeOptions, type SemanticConceptJudgeResult, type SequentialDecision, type SeriesConvergenceOptions, type SeriesConvergenceResult, type SeriesDistribution, ServedCrossFamilyError, type ServedModelCheck, type ServedModelPolicy, type ServedModelVerdict, type Severity, type ShareSafetyStatus, type ShareSafetyVerdict, type SignTestAlternative, type SingleBackendDivergence, type SingleBackendReport, type Span, type SpanFilter, type SpanHandle, type SpanMatchRecord, type SpanPredicate, type SplitCoverage, type SteeringBundle, type SteeringChange, type SteeringOptimizationResult, type SteeringOptimizationRow, type StopDecision, type StrategyChecker, type StreamingDetector, type SummaryTable, type SummaryTableOptions, type SummaryTableRow, type SynthesisTarget, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, type TaskGold, type TaskHeadroom, type ToolCallEventLike, type ToolSpan, type ToolStats, type ToolUseMetrics, type ToolUseOptions, type TraceAnalysisEngine, type TraceAnalysisEngineResult, type TraceAnalysisStore, type TraceAnalystByteBudgets, type TraceAnalystDefinition, type TraceAnalystFilters, type TraceAnalystSpan, type TraceAnalystSpanKind, type TraceAnalystSpanStatus, type TraceAnalystTraceSummary, type TraceContract, type TraceContractSpec, TraceEmitter, type TraceEmitterOptions, type TraceEvent, type TraceInsightReadiness, type TraceInsightSuite, type TraceStore, type TrainSelectionTestOptions, type TrainSelectionTestPartition, type Trajectory, type TrajectoryStep, type TreatmentGate, type TreatmentGateInput, type TreatmentGateOptions, type TrialTrace, type Turn, type TurnResult, UNKNOWN_MODEL, type UserQuestion, VERIFICATION_STRATEGIES, VERIFICATION_STRATEGY_SOURCES, ValidationError, type VerbosityBiasResult, type Verdict, type VerdictCacheStore, type VerdictCertification, type Verification, type VerificationReport, type VerificationStrategyProfile, type VerificationStrategySource, type VerifyFn, type VerifyOptions, type ViewSpansResult, type ViewTraceOversized, type ViewTraceResult, type ViteDeployRunnerInput, WILCOXON_EXACT_MAX_N, type WeightedCompositeInput, type WeightedCompositeResult, type WilcoxonSignedRankResult, type WranglerDeployRunnerInput, addCalibrationVerdicts, agentProfileCellHashMaterial, agentProfileCellKey, agentProfileHash, agentProfileId, analystFindingDigest, analystRunDigest, analystRunToFeedbackTrajectory, analystRunToReviewRequests, analyzeRuns, analyzeSeries, analyzeTraces, argHash, assertCrossFamily, assertCrossFamilyServed, assertJudgeMayGate, assertRealBackend, assertRunCaptured, assertServedModel, assertServedModels, assessShareSafety, assignFeedbackSplit, attest, benjaminiHochberg, bonferroni, bootstrapCi, budgetBreachView, buildAgentProfileCell, buildDefaultAnalystRegistry, buildEquivalenceRecord, buildTraceInsightContext, buildTraceInsightPrompt, buildTrajectory, calibrateJudge, calibrateJudgeContinuous, captureFetchToRawSink, certificationEvidenceDigest, checkServedModel, classifyFailure, classifyFailureReason, comparePairedArms, completionVerdict, computeFindingId, computeToolUseMetrics, confidenceInterval, continuousAgreement, controlRunToFeedbackTrajectory, corpusInterRaterAgreement, corpusInterRaterAgreementFromJudgeScores, costForTokenPricing, costForUsage, costReceiptFromLlm, costReceiptFromLlmError, createBoundedTraceAnalysisStore, createChatClient, createChatTraceEngine, createDspyRlmTraceEngine, createFeedbackTrajectory, createLlmCorrectnessChecker, createTraceAnalyst, decidePairedPromotion, defineAgentEval, defineEquivalenceCheck, describeTraceInsightScope, diffFindings, domainEvidencePattern, dominates, eProcess, equivalenceVerdict, evaluateInterimReleaseConfidence, evaluateReleaseConfidence, expandProfileAxes, exportRunAsOtlp, extractProducedState, extractUsage, extractUsageFromSse, failureBlame, failureClusterView, feedbackTrajectoriesToDatasetScenarios, feedbackTrajectoriesToOptimizerRows, feedbackTrajectoryToOptimizerRow, gainHistogram, gradeSemanticStatus, groupRunsByAgentProfileCell, harnessAxisOf, hashJson, holm, inferDomainKeywords, interRaterReliability, iqr, isBinaryOutcomeVector, isJudgeSpan, isLlmSpan, isRunRecord, isToolSpan, isTransientLlmError, judgeAgreementView, judgeFamily, judgeGateDecision, judgeSpans, llmJudge, makeFinding, makeProposalFinding, manifestContentDigest, mannWhitneyU, maximumChargeForLlmRequest, mcnemar, mcnemarPower, mcnemarRequiredN, measureJudgeAgreement, mintRolloutRows, modelHasSnapshot, modelPriceKey, mulberry32, objectiveEval, otlpTextToTraceAnalysisStore, pairArms, pairRunRecords, pairedBinaryScale, pairedBootstrap, pairedDeltaTieFraction, pairedEvalueSequence, pairedMde, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, pairedSignTest, pairedTTest, paretoChart, paretoFrontier, parseRunRecordSafe, partialCredit, passAtK, pearsonR, index_d_exports as profile, quotaExhaustedUntil, ranks, redTeamDataset, redTeamReport, redact, redactForShare, redactText, registerCalibrationSet, renderPreferenceMemoryMarkdown, requiredPairedSampleSize, requiredSampleSize, roundTripRunRecord, runAgentControlLoop, runCampaign, runCanaries, runCostFloorUsd, runCounterfactual, runEquivalenceCheck, runEvalCampaign, runSemanticConceptJudge, runTaskScore, runsForScenario, scoreRedTeamOutput, scoreTraceInsightReadiness, searchCellRunId, selfImprove, servedModelAcceptable, shareAllowed, snapshotFromJudgeAgreement, spearmanR, stripFencedJson, subjectiveEval, summarizeBackendIntegrity, summarizeNumberSeries, summarizePreferenceMemory, summaryTable, toAgentProfileJson, tokenizeDomainWords, toolSpansToTraceAnalysisStore, toolWasteView, transientDispatchFailure, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyAttestation, verifyCompletion, weightedComposite, weightedMean, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit };
|
|
3699
3699
|
//# sourceMappingURL=index.d.ts.map
|
package/dist/index.js
CHANGED
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
import { t as __exportAll } from "./rolldown-runtime-8H4AJuhK.js";
|
|
2
2
|
import { i as JudgeError, n as CaptureIntegrityError, o as NotFoundError, r as ConfigError, s as ValidationError, t as AgentEvalError } from "./errors-Dngq5h35.js";
|
|
3
3
|
import { a as hashCanonical, r as canonicalString } from "./canonical-BsZVgBhW.js";
|
|
4
|
-
import { a as CODING_HARNESSES, c as agentProfileId, d as harnessAxisOf, i as verifyCompletion, l as agentProfileModelId, n as completionVerdict, o as HARNESS_NATIVE_MODEL, r as createLlmCorrectnessChecker, s as agentProfileHash, t as extractProducedState, u as expandProfileAxes } from "./produced-state-
|
|
4
|
+
import { a as CODING_HARNESSES, c as agentProfileId, d as harnessAxisOf, i as verifyCompletion, l as agentProfileModelId, n as completionVerdict, o as HARNESS_NATIVE_MODEL, r as createLlmCorrectnessChecker, s as agentProfileHash, t as extractProducedState, u as expandProfileAxes } from "./produced-state-DLToUXrD.js";
|
|
5
5
|
import { c as groupRunsByAgentProfileCell, d as validateAgentProfileCell, f as verifyAgentProfileCell, h as manifestContentDigest, i as agentProfileCellKey, m as hashJson, r as agentProfileCellHashMaterial, s as buildAgentProfileCell, t as AGENT_PROFILE_KINDS, u as toAgentProfileJson } from "./agent-profile-cell-DcFW4vOc.js";
|
|
6
6
|
import { t as mulberry32 } from "./random-Dn5fPWkt.js";
|
|
7
7
|
import { a as spearmanR, c as weightedMean, i as ranks, n as partialCredit, o as summarizeNumberSeries, r as pearsonR, s as weightedComposite, t as confidenceInterval } from "./descriptive-D9LUFk1-.js";
|
|
@@ -23,14 +23,14 @@ import { r as observedSplitScore, t as isRealnessGated } from "./reward-nw2xZGZG
|
|
|
23
23
|
import { a as parseRunRecordSafe, c as runTaskScore, i as modelHasSnapshot, l as searchCellRunId, n as UNKNOWN_MODEL, o as roundTripRunRecord, r as isRunRecord, s as runCostFloorUsd, t as RunRecordValidationError, u as validateRunRecord } from "./run-record-BRlM99rX.js";
|
|
24
24
|
import { a as summaryTable, n as gainHistogram, r as paretoChart } from "./summary-report-DLzptLSA.js";
|
|
25
25
|
import { n as contentHash, r as fileVerdictCache, t as canonicalJson } from "./verdict-cache-BQEsVy9n.js";
|
|
26
|
-
import { J as redTeamDataset, X as scoreRedTeamOutput, Y as redTeamReport, Z as runCanaries, a as transientDispatchFailure, i as quotaExhaustedUntil, o as aggregateRunScore, q as DEFAULT_RED_TEAM_CORPUS, rt as runCampaign, s as clamp01, t as llmJudge } from "./llm-judge-
|
|
26
|
+
import { J as redTeamDataset, X as scoreRedTeamOutput, Y as redTeamReport, Z as runCanaries, a as transientDispatchFailure, i as quotaExhaustedUntil, o as aggregateRunScore, q as DEFAULT_RED_TEAM_CORPUS, rt as runCampaign, s as clamp01, t as llmJudge } from "./llm-judge-DyZdrp6I.js";
|
|
27
27
|
import { a as resolveModelPricing, i as isModelPriced, n as estimateCost, r as estimateTokens, t as MODEL_PRICING } from "./metrics-DOMpkoeb.js";
|
|
28
28
|
import { a as CostLedgerPersistenceError, c as costForTokenPricing, i as CostLedger, l as costForUsage, n as CostCallConflictError, o as CostReceiptCaptureError, r as CostCeilingReachedError, s as CostReservationExceededError, t as CostAccountingIncompleteError, u as modelPriceKey } from "./cost-ledger-_IbkejCF.js";
|
|
29
29
|
import { B as assertRealBackend, F as surfaceContentHash, T as parseReflectionResponse, V as summarizeBackendIntegrity, c as ATTESTATION_ALGORITHM, l as attest, u as verifyAttestation, w as buildReflectionPrompt, z as BackendIntegrityError } from "./campaign-evidence-Dt_rWPk2.js";
|
|
30
30
|
import { d as redactForShare, f as redactText, n as REDACTION_VERSION, p as shareAllowed, r as assessShareSafety, u as redact } from "./redact-Dk1ZdJEm.js";
|
|
31
31
|
import { D as paretoFrontier, E as dominates, d as pairedPromotionPower, f as requiredPairsForPairedPromotion } from "./search-kernel-Dpniloks.js";
|
|
32
32
|
import { S as judgeFamily, _ as assertServedModels, b as CrossFamilyError, d as stripFencedJson, f as ModelSubstitutionError, h as assertServedModel, i as backoffMs, l as isTransientLlmError, m as assertCrossFamilyServed, o as costReceiptFromLlm, p as ServedCrossFamilyError, r as LlmResponseError, s as costReceiptFromLlmError, t as LlmCallError, u as maximumChargeForLlmRequest, v as checkServedModel, x as assertCrossFamily, y as servedModelAcceptable } from "./llm-client-DmCNzrc9.js";
|
|
33
|
-
import { r as selfImprove, t as defineAgentEval } from "./define-agent-eval-
|
|
33
|
+
import { r as selfImprove, t as defineAgentEval } from "./define-agent-eval-B25Cc3ki.js";
|
|
34
34
|
import { n as equivalenceVerdict, t as certificationEvidenceDigest } from "./verdict-DhPoXApk.js";
|
|
35
35
|
import { t as TraceEmitter } from "./emitter-DbMXUIVP.js";
|
|
36
36
|
import { t as buildTrajectory } from "./trajectory-D_7rLrvE.js";
|
|
@@ -54,7 +54,7 @@ import { n as paidJsonChat } from "./chat-json-call-Digk-pfF.js";
|
|
|
54
54
|
import { i as makeProposalFinding, n as computeFindingId, r as makeFinding } from "./types-CW5b90JH.js";
|
|
55
55
|
import { t as createDspyRlmTraceEngine } from "./dspy-rlm-engine-D0h-P2Mz.js";
|
|
56
56
|
import { a as diffFindings, n as runSemanticConceptJudge, o as createChatTraceEngine, r as FindingsStore, t as SEMANTIC_CONCEPT_JUDGE_VERSION } from "./semantic-concept-judge-BgWxa4L_.js";
|
|
57
|
-
import {
|
|
57
|
+
import { _ as analyzeSeries, a as judgeGateDecision, c as snapshotFromJudgeAgreement, i as assertJudgeMayGate, n as JudgeGateRefusedError, o as measureJudgeAgreement, r as addCalibrationVerdicts, s as registerCalibrationSet, t as DEFAULT_JUDGE_GATE_POLICY } from "./judge-gate-B1qYnkpi.js";
|
|
58
58
|
import { i as otlpTextToTraceAnalysisStore, n as OtlpFileTraceStore } from "./store-otlp-yd9lDziN.js";
|
|
59
59
|
import { n as evaluateReleaseConfidence, r as bootstrapCi } from "./release-confidence-CiFg1EPq.js";
|
|
60
60
|
import { t as createChatClient } from "./chat-client-5LLl4Ctl.js";
|
|
@@ -9694,6 +9694,6 @@ function rankRows(rows, weights) {
|
|
|
9694
9694
|
})).sort((a, b) => b.mean - a.mean);
|
|
9695
9695
|
}
|
|
9696
9696
|
//#endregion
|
|
9697
|
-
export { AGENT_PROFILE_KINDS, ATTESTATION_ALGORITHM, AgentEvalError, AnalystRegistry, BOOTSTRAP_GATE_MIN_N, BackendIntegrityError, BudgetBreachError, BudgetGuard, CODING_HARNESSES, ConfigError, CostAccountingIncompleteError, CostCallConflictError, CostCeilingReachedError, CostLedger, CostLedgerPersistenceError, CostReceiptCaptureError, CostReservationExceededError, CostTracker, CrossFamilyError, DECISION_PAIRED_DELTA_STATISTIC, DEFAULT_BOUNDED_PROCESS_TIMEOUT_MS, DEFAULT_FAILURE_REASON_RULES, DEFAULT_PERMUTATIONS, DEFAULT_RED_TEAM_CORPUS, DEFAULT_TRACE_ANALYST_BUDGETS, DEFAULT_TRACE_ANALYST_KINDS, DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS, ERROR_COUNT_PATTERNS, EquivalenceProtocolError, FAILURE_BLAME, FAILURE_CLASSES, FAILURE_MODE_KIND_SPEC, FileSystemFeedbackTrajectoryStore, FileSystemRawProviderSink, FileSystemTraceStore, FindingsStore, HARNESS_NATIVE_MODEL, HeldOutGate, INFRA_FAILURE_BLAMES, InMemoryFeedbackTrajectoryStore, InMemoryRawProviderSink, InMemoryTraceStore, JudgeError, LOCAL_COMMAND_RUNNER_DEFAULT_CAP_MS, LlmCallError, LlmResponseError, MANN_WHITNEY_EXACT_MAX_STATES, MANN_WHITNEY_EXACT_MAX_WORK, MODEL_PRICING, ModelSubstitutionError, MultiLayerVerifier, NoopRawProviderSink, NotFoundError, OUTPUT_VALUE, OtlpFileTraceStore, POLICY_EDIT_AXES, POLICY_EDIT_CANDIDATE_RECORD_SCHEMA, POLICY_EDIT_TARGET_SURFACES, PairwiseSteeringOptimizer, PolicyEditValidationError, ProductClient, PromptRegistry, REDACTION_VERSION, RunIntegrityError, RunRecordValidationError, SEMANTIC_CONCEPT_JUDGE_VERSION, ServedCrossFamilyError, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, TraceEmitter, UNKNOWN_MODEL, VERIFICATION_STRATEGIES, VERIFICATION_STRATEGY_SOURCES, ValidationError, WILCOXON_EXACT_MAX_N, acquisitionPlansForKnowledgeGaps, admitPolicyEdit, agentProfileCellHashMaterial, agentProfileCellKey, agentProfileHash, agentProfileId, aggregateJudgeVerdicts, aggregateRunScore, analystFindingDigest, analystRunDigest, analystRunToFeedbackTrajectory, analystRunToReviewRequests, analyzeAntiSlop, analyzeRuns, analyzeSeries, analyzeTraces, applyPolicyEditToSurface, argHash, assertCapabilityHeadroom, assertCrossFamily, assertCrossFamilyServed, assertNoHiddenLeak, assertNoJudgeVerdict, assertProductBenchmarkRun, assertRealBackend, assertRunCaptured, assertServedModel, assertServedModels, assertSingleBackend, assessShareSafety, assignFeedbackSplit, attest, benjaminiHochberg, blendHeldout, blockingKnowledgeEval, bonferroni, bootstrapCi, budgetBreachView, buildAgentProfileCell, buildDefaultAnalystRegistry, buildEquivalenceRecord, buildReflectionPrompt, buildTraceInsightContext, buildTraceInsightPrompt, buildTrajectory, calibrateJudge, calibrateJudgeContinuous, canonicalJson, capabilityHeadroom, captureFetchToRawSink, certificationEvidenceDigest, checkCanaries, checkServedModel, checkTraceContracts, clamp01, classifyFailure, classifyFailureReason, cliffsDelta, cohensD, comparePairedArms, compileTraceContractSpec, completionVerdict, computeExperimentStats, computeFindingId, computePolicyEditId, computeToolUseMetrics, confidenceInterval, contentHash, continuousAgreement, controlRunToFeedbackTrajectory, corpusInterRaterAgreement, corpusInterRaterAgreementFromJudgeScores, costForTokenPricing, costForUsage, costReceiptFromLlm, costReceiptFromLlmError, createAntiSlopJudge, createBoundedTraceAnalysisStore, createChatClient, createChatTraceEngine, createDspyRlmTraceEngine, createFeedbackTrajectory, createLlmCorrectnessChecker, createLlmReviewer, createTraceAnalyst, decidePairedPromotion, defaultBlendWeights, defineAgentEval, defineEquivalenceCheck, deployGateLayer, describeTraceInsightScope, diffFindings, diffScorecard, discoverPersonas, domainEvidencePattern, dominates, eProcess, ensembleJudge, equivalenceVerdict, errorStreakDetector, estimateCost, estimateTokens, evaluateActionPolicy, evaluateInterimReleaseConfidence, evaluateOracles, evaluateReleaseConfidence, evaluateTraceContract, expandProfileAxes, explainTraceContract, exportProductBenchmark, exportProductBenchmarkRuns, exportRunAsOtlp, extractErrorCount, extractProducedState, extractUsage, extractUsageFromSse, failureBlame, failureClusterView, feedbackTrajectoriesToDatasetScenarios, feedbackTrajectoriesToOptimizerRows, feedbackTrajectoryToOptimizerRow, fileVerdictCache, formatScorecardDiff, gainHistogram, gateTreatmentApplied, gradeOnHidden, gradeSemanticStatus, groupRunsByAgentProfileCell, harnessAxisOf, hashContent, hashJson, hiddenGrade, holm, improvementVerdict, inMemoryReviewStore, inferDomainKeywords, interRaterReliability, interpretCliffs, iqr, isBinaryOutcomeVector, isJudgeSpan, isLlmSpan, isModelPriced, isPolicyEdit, isRunRecord, isToolSpan, isTransientLlmError, jsonShape, jsonlReviewStore, jsonlRunRecordBackend, judgeAgreementView, judgeFamily, judgeSpans, knowledgeReadinessTracePayload, leaderboard, lintTraceContractSpec, llmJudge, loadScorecard, localCommandRunner, makeFinding, makePolicyEdit, makePolicyEditCandidateRecord, makeProposalFinding, manifestContentDigest, mannWhitneyU, maximumChargeForLlmRequest, mcnemar, mcnemarPower, mcnemarRequiredN, minimumPairsForPairedDeltaTest, mintRolloutRows, modelHasSnapshot, modelPriceKey, mulberry32, notBlocked, objectiveEval, observeAll, otlpTextToTraceAnalysisStore, pairArms, pairRunRecords, pairedBinaryScale, pairedBootstrap, pairedCohensDz, pairedDeltaTest, pairedDeltaTieFraction, pairedEvalueSequence, pairedMde, pairedPromotionPower, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, pairedSignTest, pairedTTest, paretoChart, paretoFrontier, parseReflectionResponse, parseRunRecordSafe, partialCredit, partitionHeldOut, partitionTrainSelectionTest, passAtK, pearsonR, policyEditFromFinding, policyEditsFromFindings, preflightModels, productBenchmarkRepoIdentity, profile_exports as profile, projectRuntimeTrajectoryEvidence, proposeSynthesisTargets, quotaExhaustedUntil, ranks, readProductBenchmarkManifest, recordRuns, recordRunsToScorecard, redTeamDataset, redTeamReport, redact, redactForShare, redactText, regexMatches, renderPreferenceMemoryMarkdown, repeatedActionDetector, requiredPairedSampleSize, requiredPairsForPairedPromotion, requiredSampleSize, resolveModelPricing, resolveSeat, roundTripRunRecord, routeFields, runAgentControlLoop, runBoundedProcess, runCampaign, runCanaries, runCostFloorUsd, runCounterfactual, runEquivalenceCheck, runEvalCampaign, runIntentMatchJudge, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runProposeReview, runProposeReviewAsControlLoop, runSemanticConceptJudge, runTaskScore, runsForScenario, scoreKnowledgeReadiness, scorePolicyEditReadiness, scoreRedTeamOutput, scoreTraceInsightReadiness, searchCellRunId, seatPresets, selfImprove, servedModelAcceptable, shareAllowed, spearmanR, stripFencedJson, subjectiveEval, summarizeBackendIntegrity, summarizeNumberSeries, summarizePreferenceMemory, summaryTable, textInSnapshot, toAgentProfileJson, tokenizeDomainWords, toolSpansToTraceAnalysisStore, toolWasteView, traceContract, transientDispatchFailure, urlContains, userQuestionsForKnowledgeGaps, validatePolicyEdit, validatePolicyEditCandidateRecord, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyAttestation, verifyCompletion, viteDeployRunner, weightedComposite, weightedMean, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit, withHeldoutBlend, withJudgeRetry, wranglerDeployRunner };
|
|
9697
|
+
export { AGENT_PROFILE_KINDS, ATTESTATION_ALGORITHM, AgentEvalError, AnalystRegistry, BOOTSTRAP_GATE_MIN_N, BackendIntegrityError, BudgetBreachError, BudgetGuard, CODING_HARNESSES, ConfigError, CostAccountingIncompleteError, CostCallConflictError, CostCeilingReachedError, CostLedger, CostLedgerPersistenceError, CostReceiptCaptureError, CostReservationExceededError, CostTracker, CrossFamilyError, DECISION_PAIRED_DELTA_STATISTIC, DEFAULT_BOUNDED_PROCESS_TIMEOUT_MS, DEFAULT_FAILURE_REASON_RULES, DEFAULT_JUDGE_GATE_POLICY, DEFAULT_PERMUTATIONS, DEFAULT_RED_TEAM_CORPUS, DEFAULT_TRACE_ANALYST_BUDGETS, DEFAULT_TRACE_ANALYST_KINDS, DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS, ERROR_COUNT_PATTERNS, EquivalenceProtocolError, FAILURE_BLAME, FAILURE_CLASSES, FAILURE_MODE_KIND_SPEC, FileSystemFeedbackTrajectoryStore, FileSystemRawProviderSink, FileSystemTraceStore, FindingsStore, HARNESS_NATIVE_MODEL, HeldOutGate, INFRA_FAILURE_BLAMES, InMemoryFeedbackTrajectoryStore, InMemoryRawProviderSink, InMemoryTraceStore, JudgeError, JudgeGateRefusedError, LOCAL_COMMAND_RUNNER_DEFAULT_CAP_MS, LlmCallError, LlmResponseError, MANN_WHITNEY_EXACT_MAX_STATES, MANN_WHITNEY_EXACT_MAX_WORK, MODEL_PRICING, ModelSubstitutionError, MultiLayerVerifier, NoopRawProviderSink, NotFoundError, OUTPUT_VALUE, OtlpFileTraceStore, POLICY_EDIT_AXES, POLICY_EDIT_CANDIDATE_RECORD_SCHEMA, POLICY_EDIT_TARGET_SURFACES, PairwiseSteeringOptimizer, PolicyEditValidationError, ProductClient, PromptRegistry, REDACTION_VERSION, RunIntegrityError, RunRecordValidationError, SEMANTIC_CONCEPT_JUDGE_VERSION, ServedCrossFamilyError, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, TraceEmitter, UNKNOWN_MODEL, VERIFICATION_STRATEGIES, VERIFICATION_STRATEGY_SOURCES, ValidationError, WILCOXON_EXACT_MAX_N, acquisitionPlansForKnowledgeGaps, addCalibrationVerdicts, admitPolicyEdit, agentProfileCellHashMaterial, agentProfileCellKey, agentProfileHash, agentProfileId, aggregateJudgeVerdicts, aggregateRunScore, analystFindingDigest, analystRunDigest, analystRunToFeedbackTrajectory, analystRunToReviewRequests, analyzeAntiSlop, analyzeRuns, analyzeSeries, analyzeTraces, applyPolicyEditToSurface, argHash, assertCapabilityHeadroom, assertCrossFamily, assertCrossFamilyServed, assertJudgeMayGate, assertNoHiddenLeak, assertNoJudgeVerdict, assertProductBenchmarkRun, assertRealBackend, assertRunCaptured, assertServedModel, assertServedModels, assertSingleBackend, assessShareSafety, assignFeedbackSplit, attest, benjaminiHochberg, blendHeldout, blockingKnowledgeEval, bonferroni, bootstrapCi, budgetBreachView, buildAgentProfileCell, buildDefaultAnalystRegistry, buildEquivalenceRecord, buildReflectionPrompt, buildTraceInsightContext, buildTraceInsightPrompt, buildTrajectory, calibrateJudge, calibrateJudgeContinuous, canonicalJson, capabilityHeadroom, captureFetchToRawSink, certificationEvidenceDigest, checkCanaries, checkServedModel, checkTraceContracts, clamp01, classifyFailure, classifyFailureReason, cliffsDelta, cohensD, comparePairedArms, compileTraceContractSpec, completionVerdict, computeExperimentStats, computeFindingId, computePolicyEditId, computeToolUseMetrics, confidenceInterval, contentHash, continuousAgreement, controlRunToFeedbackTrajectory, corpusInterRaterAgreement, corpusInterRaterAgreementFromJudgeScores, costForTokenPricing, costForUsage, costReceiptFromLlm, costReceiptFromLlmError, createAntiSlopJudge, createBoundedTraceAnalysisStore, createChatClient, createChatTraceEngine, createDspyRlmTraceEngine, createFeedbackTrajectory, createLlmCorrectnessChecker, createLlmReviewer, createTraceAnalyst, decidePairedPromotion, defaultBlendWeights, defineAgentEval, defineEquivalenceCheck, deployGateLayer, describeTraceInsightScope, diffFindings, diffScorecard, discoverPersonas, domainEvidencePattern, dominates, eProcess, ensembleJudge, equivalenceVerdict, errorStreakDetector, estimateCost, estimateTokens, evaluateActionPolicy, evaluateInterimReleaseConfidence, evaluateOracles, evaluateReleaseConfidence, evaluateTraceContract, expandProfileAxes, explainTraceContract, exportProductBenchmark, exportProductBenchmarkRuns, exportRunAsOtlp, extractErrorCount, extractProducedState, extractUsage, extractUsageFromSse, failureBlame, failureClusterView, feedbackTrajectoriesToDatasetScenarios, feedbackTrajectoriesToOptimizerRows, feedbackTrajectoryToOptimizerRow, fileVerdictCache, formatScorecardDiff, gainHistogram, gateTreatmentApplied, gradeOnHidden, gradeSemanticStatus, groupRunsByAgentProfileCell, harnessAxisOf, hashContent, hashJson, hiddenGrade, holm, improvementVerdict, inMemoryReviewStore, inferDomainKeywords, interRaterReliability, interpretCliffs, iqr, isBinaryOutcomeVector, isJudgeSpan, isLlmSpan, isModelPriced, isPolicyEdit, isRunRecord, isToolSpan, isTransientLlmError, jsonShape, jsonlReviewStore, jsonlRunRecordBackend, judgeAgreementView, judgeFamily, judgeGateDecision, judgeSpans, knowledgeReadinessTracePayload, leaderboard, lintTraceContractSpec, llmJudge, loadScorecard, localCommandRunner, makeFinding, makePolicyEdit, makePolicyEditCandidateRecord, makeProposalFinding, manifestContentDigest, mannWhitneyU, maximumChargeForLlmRequest, mcnemar, mcnemarPower, mcnemarRequiredN, measureJudgeAgreement, minimumPairsForPairedDeltaTest, mintRolloutRows, modelHasSnapshot, modelPriceKey, mulberry32, notBlocked, objectiveEval, observeAll, otlpTextToTraceAnalysisStore, pairArms, pairRunRecords, pairedBinaryScale, pairedBootstrap, pairedCohensDz, pairedDeltaTest, pairedDeltaTieFraction, pairedEvalueSequence, pairedMde, pairedPromotionPower, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, pairedSignTest, pairedTTest, paretoChart, paretoFrontier, parseReflectionResponse, parseRunRecordSafe, partialCredit, partitionHeldOut, partitionTrainSelectionTest, passAtK, pearsonR, policyEditFromFinding, policyEditsFromFindings, preflightModels, productBenchmarkRepoIdentity, profile_exports as profile, projectRuntimeTrajectoryEvidence, proposeSynthesisTargets, quotaExhaustedUntil, ranks, readProductBenchmarkManifest, recordRuns, recordRunsToScorecard, redTeamDataset, redTeamReport, redact, redactForShare, redactText, regexMatches, registerCalibrationSet, renderPreferenceMemoryMarkdown, repeatedActionDetector, requiredPairedSampleSize, requiredPairsForPairedPromotion, requiredSampleSize, resolveModelPricing, resolveSeat, roundTripRunRecord, routeFields, runAgentControlLoop, runBoundedProcess, runCampaign, runCanaries, runCostFloorUsd, runCounterfactual, runEquivalenceCheck, runEvalCampaign, runIntentMatchJudge, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runProposeReview, runProposeReviewAsControlLoop, runSemanticConceptJudge, runTaskScore, runsForScenario, scoreKnowledgeReadiness, scorePolicyEditReadiness, scoreRedTeamOutput, scoreTraceInsightReadiness, searchCellRunId, seatPresets, selfImprove, servedModelAcceptable, shareAllowed, snapshotFromJudgeAgreement, spearmanR, stripFencedJson, subjectiveEval, summarizeBackendIntegrity, summarizeNumberSeries, summarizePreferenceMemory, summaryTable, textInSnapshot, toAgentProfileJson, tokenizeDomainWords, toolSpansToTraceAnalysisStore, toolWasteView, traceContract, transientDispatchFailure, urlContains, userQuestionsForKnowledgeGaps, validatePolicyEdit, validatePolicyEditCandidateRecord, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyAttestation, verifyCompletion, viteDeployRunner, weightedComposite, weightedMean, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit, withHeldoutBlend, withJudgeRetry, wranglerDeployRunner };
|
|
9698
9698
|
|
|
9699
9699
|
//# sourceMappingURL=index.js.map
|