@tangle-network/agent-eval 0.221.7 → 0.222.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +146 -0
- package/dist/adapters/http.d.ts +3 -5
- package/dist/adapters/http.d.ts.map +1 -1
- package/dist/adapters/http.js +1 -2
- package/dist/{adversarial-smnADNFS.d.ts → adversarial-DnQQFqTR.d.ts} +1 -3
- package/dist/adversarial-DnQQFqTR.d.ts.map +1 -0
- package/dist/{agent-profile-CNmwfaqq.d.ts → agent-profile-CoD1Dotf.d.ts} +4 -10
- package/dist/agent-profile-CoD1Dotf.d.ts.map +1 -0
- package/dist/agent-profile-Dngk3LZ0.js +153 -0
- package/dist/agent-profile-Dngk3LZ0.js.map +1 -0
- package/dist/{agent-profile-cell-BcpFbNd5.d.ts → agent-profile-cell-3kYxUsK2.d.ts} +3 -5
- package/dist/agent-profile-cell-3kYxUsK2.d.ts.map +1 -0
- package/dist/{agent-profile-cell-B1V1gUZs.js → agent-profile-cell-CT5JYueM.js} +5 -7
- package/dist/{agent-profile-cell-B1V1gUZs.js.map → agent-profile-cell-CT5JYueM.js.map} +1 -1
- package/dist/{agreement-irr-BhAqjluA.js → agreement-irr-BTjDN0i6.js} +4 -5
- package/dist/{agreement-irr-BhAqjluA.js.map → agreement-irr-BTjDN0i6.js.map} +1 -1
- package/dist/analyst/index.d.ts +95 -159
- package/dist/analyst/index.d.ts.map +1 -1
- package/dist/analyst/index.js +28 -39
- package/dist/analyst/index.js.map +1 -1
- package/dist/{analyze-runs-DfX6dHP0.d.ts → analyze-runs-C-pP1sK9.d.ts} +7 -13
- package/dist/analyze-runs-C-pP1sK9.d.ts.map +1 -0
- package/dist/{analyze-runs-CHmsykNJ.js → analyze-runs-CTheASYV.js} +23 -32
- package/dist/{analyze-runs-CHmsykNJ.js.map → analyze-runs-CTheASYV.js.map} +1 -1
- package/dist/{attribute-vocabulary-CcXqHODT.d.ts → attribute-vocabulary-5oo8ykIx.d.ts} +2 -4
- package/dist/attribute-vocabulary-5oo8ykIx.d.ts.map +1 -0
- package/dist/authenticity/index.d.ts +0 -2
- package/dist/authenticity/index.d.ts.map +1 -1
- package/dist/authenticity/index.js +1 -2
- package/dist/{benchmark-command-Cnjfo0Zh.js → benchmark-command-jYWR2gya.js} +61 -94
- package/dist/benchmark-command-jYWR2gya.js.map +1 -0
- package/dist/benchmarks/index.d.ts +5 -15
- package/dist/benchmarks/index.d.ts.map +1 -1
- package/dist/benchmarks/index.js +27 -38
- package/dist/benchmarks/index.js.map +1 -1
- package/dist/{bounded-process-Cj9YjXbE.js → bounded-process-BZrWXIKW.js} +2 -3
- package/dist/{bounded-process-Cj9YjXbE.js.map → bounded-process-BZrWXIKW.js.map} +1 -1
- package/dist/builder-eval/index.d.ts +5 -14
- package/dist/builder-eval/index.d.ts.map +1 -1
- package/dist/builder-eval/index.js +11 -17
- package/dist/builder-eval/index.js.map +1 -1
- package/dist/{calibration-CTwxpJ-t.js → calibration-6usfVYYo.js} +4 -5
- package/dist/{calibration-CTwxpJ-t.js.map → calibration-6usfVYYo.js.map} +1 -1
- package/dist/campaign/index.d.ts +9 -9
- package/dist/campaign/index.js +15 -14
- package/dist/{campaign-uddTlFg5.js → campaign-ddr1N3yD.js} +61 -58
- package/dist/campaign-ddr1N3yD.js.map +1 -0
- package/dist/{campaign-evidence-DunzcJGv.js → campaign-evidence-DN8rWuBj.js} +20 -322
- package/dist/campaign-evidence-DN8rWuBj.js.map +1 -0
- package/dist/{canonical-Dq2hOQ8L.js → canonical-CicJlINA.js} +3 -4
- package/dist/{canonical-Dq2hOQ8L.js.map → canonical-CicJlINA.js.map} +1 -1
- package/dist/{canonical-CFpojCN5.d.ts → canonical-gLeHiqzd.d.ts} +2 -4
- package/dist/canonical-gLeHiqzd.d.ts.map +1 -0
- package/dist/{cell-spend-DEWR3nj-.js → cell-spend-zxmzg4GH.js} +3 -4
- package/dist/{cell-spend-DEWR3nj-.js.map → cell-spend-zxmzg4GH.js.map} +1 -1
- package/dist/{chat-client-BNATB3rJ.js → chat-client-CO3uB33i.js} +3 -4
- package/dist/{chat-client-BNATB3rJ.js.map → chat-client-CO3uB33i.js.map} +1 -1
- package/dist/{chat-json-call-BzWYp18J.js → chat-json-call-8yUv4LFO.js} +4 -5
- package/dist/{chat-json-call-BzWYp18J.js.map → chat-json-call-8yUv4LFO.js.map} +1 -1
- package/dist/{claim-BPVuyHuy.js → claim-Dod_mjk3.js} +5 -6
- package/dist/{claim-BPVuyHuy.js.map → claim-Dod_mjk3.js.map} +1 -1
- package/dist/cli.js +179 -23
- package/dist/cli.js.map +1 -1
- package/dist/{client-hwo3cJjR.d.ts → client-BIFwUg7I.d.ts} +6 -12
- package/dist/client-BIFwUg7I.d.ts.map +1 -0
- package/dist/{cluster-bootstrap-DEwwhl0C.js → cluster-bootstrap-SbaeM44s.js} +4 -5
- package/dist/{cluster-bootstrap-DEwwhl0C.js.map → cluster-bootstrap-SbaeM44s.js.map} +1 -1
- package/dist/completion-verifier-7KEfroqx.d.ts +168 -0
- package/dist/completion-verifier-7KEfroqx.d.ts.map +1 -0
- package/dist/{completion-verifier-asaV1nSj.js → completion-verifier-rqbMh5-4.js} +9 -160
- package/dist/completion-verifier-rqbMh5-4.js.map +1 -0
- package/dist/contract/index.d.ts +13 -13
- package/dist/contract/index.js +41 -60
- package/dist/contract/index.js.map +1 -1
- package/dist/{cost-ledger-BGZB6yY2.js → cost-ledger-CS8R3IZj.js} +10 -17
- package/dist/{cost-ledger-BGZB6yY2.js.map → cost-ledger-CS8R3IZj.js.map} +1 -1
- package/dist/{cost-ledger-C1W3ICmP.d.ts → cost-ledger-Czro9wmo.d.ts} +1 -5
- package/dist/cost-ledger-Czro9wmo.d.ts.map +1 -0
- package/dist/{dataset-DQqhOCPt.d.ts → dataset-DThC1S6G.d.ts} +1 -3
- package/dist/dataset-DThC1S6G.d.ts.map +1 -0
- package/dist/{decision-point-0nJubuoM.js → decision-point-lhMgODOR.js} +14 -18
- package/dist/{decision-point-0nJubuoM.js.map → decision-point-lhMgODOR.js.map} +1 -1
- package/dist/{deep-freeze-CurcOQg3.js → deep-freeze-CK8-2W7p.js} +2 -3
- package/dist/{deep-freeze-CurcOQg3.js.map → deep-freeze-CK8-2W7p.js.map} +1 -1
- package/dist/{default-registry-CLG5YqGS.d.ts → default-registry-B1DstPYg.d.ts} +53 -61
- package/dist/default-registry-B1DstPYg.d.ts.map +1 -0
- package/dist/{default-registry-BBTDgkDM.js → default-registry-C1tPEjZ4.js} +29 -45
- package/dist/{default-registry-BBTDgkDM.js.map → default-registry-C1tPEjZ4.js.map} +1 -1
- package/dist/{define-agent-eval-CCbT7Kfr.js → define-agent-eval-BLoP4Lf9.js} +15 -18
- package/dist/{define-agent-eval-CCbT7Kfr.js.map → define-agent-eval-BLoP4Lf9.js.map} +1 -1
- package/dist/{define-agent-eval-JKt3gMQr.d.ts → define-agent-eval-CAffiFbN.d.ts} +19 -27
- package/dist/define-agent-eval-CAffiFbN.d.ts.map +1 -0
- package/dist/{descriptive-CVNy48rR.js → descriptive-cPwITVfy.js} +30 -8
- package/dist/descriptive-cPwITVfy.js.map +1 -0
- package/dist/diagnosis/index.d.ts +5 -15
- package/dist/diagnosis/index.d.ts.map +1 -1
- package/dist/diagnosis/index.js +13 -19
- package/dist/diagnosis/index.js.map +1 -1
- package/dist/{dspy-rlm-engine-DN1p_5jx.js → dspy-rlm-engine-jDiWkul6.js} +7 -10
- package/dist/{dspy-rlm-engine-DN1p_5jx.js.map → dspy-rlm-engine-jDiWkul6.js.map} +1 -1
- package/dist/{emitter-DbMXUIVP.js → emitter-DhQ65mza.js} +3 -4
- package/dist/{emitter-DbMXUIVP.js.map → emitter-DhQ65mza.js.map} +1 -1
- package/dist/{emitter-DABYYTcH.d.ts → emitter-mAXC5PSN.d.ts} +3 -5
- package/dist/emitter-mAXC5PSN.d.ts.map +1 -0
- package/dist/{engine-dtRVeuQs.d.ts → engine-CaHaCc3W.d.ts} +10 -18
- package/dist/engine-CaHaCc3W.d.ts.map +1 -0
- package/dist/{equal-terms-DbPGcVtI.js → equal-terms-K11VHmR3.js} +3 -4
- package/dist/{equal-terms-DbPGcVtI.js.map → equal-terms-K11VHmR3.js.map} +1 -1
- package/dist/{errors-BcOs5MKW.js → errors-DeEEV9Vx.js} +2 -3
- package/dist/{errors-BcOs5MKW.js.map → errors-DeEEV9Vx.js.map} +1 -1
- package/dist/{errors-DEE6u6ot.d.ts → errors-E9I2hldP.d.ts} +1 -3
- package/dist/errors-E9I2hldP.d.ts.map +1 -0
- package/dist/{eval-campaign-DoX5vz3E.js → eval-campaign-Cgixrvf7.js} +8 -9
- package/dist/{eval-campaign-DoX5vz3E.js.map → eval-campaign-Cgixrvf7.js.map} +1 -1
- package/dist/{evaluation-wdGyZ7jV.js → evaluation-B8H3GYej.js} +4 -6
- package/dist/{evaluation-wdGyZ7jV.js.map → evaluation-B8H3GYej.js.map} +1 -1
- package/dist/{evaluation-DblGquCB.d.ts → evaluation-BPqUvZ66.d.ts} +4 -6
- package/dist/evaluation-BPqUvZ66.d.ts.map +1 -0
- package/dist/evaluation.d.ts +1 -1
- package/dist/evaluation.js +1 -2
- package/dist/{exact-types-Do63KoKn.d.ts → exact-types-D98BKpxo.d.ts} +8 -10
- package/dist/exact-types-D98BKpxo.d.ts.map +1 -0
- package/dist/{exec-CpbtOXVt.js → exec-C0_AkLUv.js} +3 -5
- package/dist/{exec-CpbtOXVt.js.map → exec-C0_AkLUv.js.map} +1 -1
- package/dist/{execution-tracks-CpgFPpS5.js → execution-tracks-DaDXYwAq.js} +2 -3
- package/dist/{execution-tracks-CpgFPpS5.js.map → execution-tracks-DaDXYwAq.js.map} +1 -1
- package/dist/experiment/index.d.ts +9 -37
- package/dist/experiment/index.d.ts.map +1 -1
- package/dist/experiment/index.js +29 -40
- package/dist/experiment/index.js.map +1 -1
- package/dist/{exporters-Dk3oEbK_.d.ts → exporters-B5Ze_RKh.d.ts} +4 -4
- package/dist/exporters-B5Ze_RKh.d.ts.map +1 -0
- package/dist/{exporters-DWvM16LE.js → exporters-ByLcILYU.js} +23 -5
- package/dist/exporters-ByLcILYU.js.map +1 -0
- package/dist/{external-optimizer-process-DEzJ2CGX.js → external-optimizer-process-BRVmaDFe.js} +5 -6
- package/dist/{external-optimizer-process-DEzJ2CGX.js.map → external-optimizer-process-BRVmaDFe.js.map} +1 -1
- package/dist/{external-optimizer-subprocess-B_ukPtRD.js → external-optimizer-subprocess-U616hV_l.js} +24 -37
- package/dist/{external-optimizer-subprocess-B_ukPtRD.js.map → external-optimizer-subprocess-U616hV_l.js.map} +1 -1
- package/dist/{extract-usage-BZLU01kR.js → extract-usage-ISFqhfDz.js} +2 -3
- package/dist/{extract-usage-BZLU01kR.js.map → extract-usage-ISFqhfDz.js.map} +1 -1
- package/dist/{failure-cluster-DN0t63Z0.d.ts → failure-cluster-DOoa7CBg.d.ts} +3 -7
- package/dist/failure-cluster-DOoa7CBg.d.ts.map +1 -0
- package/dist/{failure-taxonomy-Bx5oWqaW.js → failure-taxonomy-9eQfXKoW.js} +2 -3
- package/dist/{failure-taxonomy-Bx5oWqaW.js.map → failure-taxonomy-9eQfXKoW.js.map} +1 -1
- package/dist/{feedback-trajectory-CYVJXolG.d.ts → feedback-trajectory-SBGncrCP.d.ts} +7 -13
- package/dist/feedback-trajectory-SBGncrCP.d.ts.map +1 -0
- package/dist/{finding-composition-8n0fFgVI.js → finding-composition-DlvE1Qul.js} +4 -6
- package/dist/{finding-composition-8n0fFgVI.js.map → finding-composition-DlvE1Qul.js.map} +1 -1
- package/dist/{findings-store-Bg4sdS-e.js → findings-store-B75Dspaq.js} +11 -14
- package/dist/{findings-store-Bg4sdS-e.js.map → findings-store-B75Dspaq.js.map} +1 -1
- package/dist/fuzz.d.ts +3 -17
- package/dist/fuzz.d.ts.map +1 -1
- package/dist/fuzz.js +30 -29
- package/dist/fuzz.js.map +1 -1
- package/dist/{hf-dataset-CKxIJQba.js → hf-dataset-DPVQXW5n.js} +362 -21
- package/dist/hf-dataset-DPVQXW5n.js.map +1 -0
- package/dist/hosted/index.d.ts +2 -6
- package/dist/hosted/index.d.ts.map +1 -1
- package/dist/hosted/index.js +1 -2
- package/dist/{index-CalAsXMR.d.ts → index-B1Bm9oJA.d.ts} +3 -15
- package/dist/index-B1Bm9oJA.d.ts.map +1 -0
- package/dist/{index-CdEVr8-O.d.ts → index-B7v2QHlD.d.ts} +24 -299
- package/dist/index-B7v2QHlD.d.ts.map +1 -0
- package/dist/{index-CTIRz-AB.d.ts → index-CUjeTDPv.d.ts} +14 -36
- package/dist/index-CUjeTDPv.d.ts.map +1 -0
- package/dist/{index-CfJULcW5.d.ts → index-DAk2quYu.d.ts} +18 -38
- package/dist/index-DAk2quYu.d.ts.map +1 -0
- package/dist/{index-D7fmC-1K.d.ts → index-DUn0GNiZ.d.ts} +183 -36
- package/dist/index-DUn0GNiZ.d.ts.map +1 -0
- package/dist/{index-CHxwfW5g.d.ts → index-DwUGehqX.d.ts} +3 -9
- package/dist/index-DwUGehqX.d.ts.map +1 -0
- package/dist/{index-jnuFLxOs.d.ts → index-aDPVYBIM.d.ts} +11 -39
- package/dist/index-aDPVYBIM.d.ts.map +1 -0
- package/dist/index.d.ts +55 -188
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +122 -195
- package/dist/index.js.map +1 -1
- package/dist/{integrity-D-s2iHiT.js → integrity-B1V6klyO.js} +12 -363
- package/dist/integrity-B1V6klyO.js.map +1 -0
- package/dist/{integrity-D9lBnv_K.d.ts → integrity-BAdVzT8u.d.ts} +4 -6
- package/dist/integrity-BAdVzT8u.d.ts.map +1 -0
- package/dist/{integrity-W4CADcFc.js → integrity-DOHt1mwU.js} +3 -4
- package/dist/{integrity-W4CADcFc.js.map → integrity-DOHt1mwU.js.map} +1 -1
- package/dist/{internal-BTK91Bqt.js → internal-v6XfOeJW.js} +6 -8
- package/dist/{internal-BTK91Bqt.js.map → internal-v6XfOeJW.js.map} +1 -1
- package/dist/{interval-Dq9gt7jx.js → interval-Bd2f4CRT.js} +6 -8
- package/dist/{interval-Dq9gt7jx.js.map → interval-Bd2f4CRT.js.map} +1 -1
- package/dist/{interval-ohuOJurX.d.ts → interval-CWyxqDpE.d.ts} +1 -5
- package/dist/interval-CWyxqDpE.d.ts.map +1 -0
- package/dist/{journal-C7om8Gum.js → journal-C1uLJi0w.js} +115 -121
- package/dist/{journal-C7om8Gum.js.map → journal-C1uLJi0w.js.map} +1 -1
- package/dist/{json-BhdsJaOQ.js → json-DtvZNAU9.js} +4 -6
- package/dist/{json-BhdsJaOQ.js.map → json-DtvZNAU9.js.map} +1 -1
- package/dist/json-recovery-DK4MDm5r.js +296 -0
- package/dist/json-recovery-DK4MDm5r.js.map +1 -0
- package/dist/{judge-calibration-BkXMIWA4.js → judge-calibration-DGqlfWW8.js} +5 -6
- package/dist/{judge-calibration-BkXMIWA4.js.map → judge-calibration-DGqlfWW8.js.map} +1 -1
- package/dist/{judge-calibration-DFtEMlde.d.ts → judge-calibration-DnrbQaXI.d.ts} +1 -3
- package/dist/judge-calibration-DnrbQaXI.d.ts.map +1 -0
- package/dist/{judge-gate-BVvmP00P.js → judge-gate-C2BCOt-M.js} +8 -12
- package/dist/{judge-gate-BVvmP00P.js.map → judge-gate-C2BCOt-M.js.map} +1 -1
- package/dist/judges-BLxg5p8E.js +20 -0
- package/dist/judges-BLxg5p8E.js.map +1 -0
- package/dist/{kind-factory-Dl0PmNQL.js → kind-factory-VKysp1ds.js} +23 -39
- package/dist/{kind-factory-Dl0PmNQL.js.map → kind-factory-VKysp1ds.js.map} +1 -1
- package/dist/ledger-core/index.d.ts +2 -2
- package/dist/ledger-core/index.js +3 -3
- package/dist/ledger-core/lock-holder.d.ts +1 -1
- package/dist/ledger-core/lock-holder.js +1 -2
- package/dist/{ledger-core-CAxUdDk0.js → ledger-core-CfT-FzPf.js} +5 -6
- package/dist/{ledger-core-CAxUdDk0.js.map → ledger-core-CfT-FzPf.js.map} +1 -1
- package/dist/{llm-client-b1ysZVIU.js → llm-client-kJ9EOqjt.js} +22 -30
- package/dist/{llm-client-b1ysZVIU.js.map → llm-client-kJ9EOqjt.js.map} +1 -1
- package/dist/{llm-judge-Cybq29bL.js → llm-judge-cgbhTKQp.js} +108 -164
- package/dist/llm-judge-cgbhTKQp.js.map +1 -0
- package/dist/{lock-holder-DaJbfpci.d.ts → lock-holder-DoxlZFkQ.d.ts} +1 -3
- package/dist/lock-holder-DoxlZFkQ.d.ts.map +1 -0
- package/dist/matrix/index.d.ts +3 -3
- package/dist/matrix/index.js +2 -2
- package/dist/{matrix-DtkZLD_3.d.ts → matrix-CBUzIZs4.d.ts} +4 -12
- package/dist/matrix-CBUzIZs4.d.ts.map +1 -0
- package/dist/{matrix-B8l-Nsxd.js → matrix-CgpX8wdm.js} +8 -13
- package/dist/matrix-CgpX8wdm.js.map +1 -0
- package/dist/{mean-BbA_JCzW.js → mean-CdDYG-us.js} +6 -9
- package/dist/{mean-BbA_JCzW.js.map → mean-CdDYG-us.js.map} +1 -1
- package/dist/meta-eval/index.d.ts +3 -3
- package/dist/meta-eval/index.js +17 -21
- package/dist/meta-eval/index.js.map +1 -1
- package/dist/{metrics-BKaPBtIG.js → metrics-DNCAUtuB.js} +2 -3
- package/dist/{metrics-BKaPBtIG.js.map → metrics-DNCAUtuB.js.map} +1 -1
- package/dist/{mint-CbdwjTPW.js → mint-D9ZysYtY.js} +7 -8
- package/dist/{mint-CbdwjTPW.js.map → mint-D9ZysYtY.js.map} +1 -1
- package/dist/{multi-layer-verifier-DEgMUk6v.d.ts → multi-layer-verifier-lM4aVN78.d.ts} +2 -4
- package/dist/multi-layer-verifier-lM4aVN78.d.ts.map +1 -0
- package/dist/{multiplicity-Sg0U7h8j.js → multiplicity-DAd4xaDC.js} +3 -4
- package/dist/{multiplicity-Sg0U7h8j.js.map → multiplicity-DAd4xaDC.js.map} +1 -1
- package/dist/{multiplicity-DIWHvysC.d.ts → multiplicity-D_SEsc5j.d.ts} +1 -3
- package/dist/multiplicity-D_SEsc5j.d.ts.map +1 -0
- package/dist/multishot/golden/index.d.ts +1 -17
- package/dist/multishot/golden/index.d.ts.map +1 -1
- package/dist/multishot/golden/index.js +10 -17
- package/dist/multishot/golden/index.js.map +1 -1
- package/dist/multishot/index.d.ts +2 -10
- package/dist/multishot/index.d.ts.map +1 -1
- package/dist/multishot/index.js +11 -18
- package/dist/multishot/index.js.map +1 -1
- package/dist/{nonnegative-number-DvdMxg_f.js → nonnegative-number-lEYbOsuh.js} +2 -3
- package/dist/{nonnegative-number-DvdMxg_f.js.map → nonnegative-number-lEYbOsuh.js.map} +1 -1
- package/dist/openapi.json +1 -1
- package/dist/{otel-spans-CdN909k1.js → otel-spans-BsGzTWIK.js} +5 -6
- package/dist/{otel-spans-CdN909k1.js.map → otel-spans-BsGzTWIK.js.map} +1 -1
- package/dist/{otlp-attributes-D9ptWaIJ.js → otlp-attributes-Boeb7gtH.js} +2 -3
- package/dist/{otlp-attributes-D9ptWaIJ.js.map → otlp-attributes-Boeb7gtH.js.map} +1 -1
- package/dist/{outcome-observations-FLmRpgYa.js → outcome-observations-D4snOtUb.js} +4 -5
- package/dist/{outcome-observations-FLmRpgYa.js.map → outcome-observations-D4snOtUb.js.map} +1 -1
- package/dist/{outcome-store-DUcMj-3D.js → outcome-store-TarrcfHG.js} +2 -3
- package/dist/{outcome-store-DUcMj-3D.js.map → outcome-store-TarrcfHG.js.map} +1 -1
- package/dist/{package-version-D7lQHt_-.js → package-version-CJl_pY9W.js} +2 -3
- package/dist/{package-version-D7lQHt_-.js.map → package-version-CJl_pY9W.js.map} +1 -1
- package/dist/{paired-arms-B2ttLc0_.js → paired-arms-D-54hO-k.js} +10 -14
- package/dist/{paired-arms-B2ttLc0_.js.map → paired-arms-D-54hO-k.js.map} +1 -1
- package/dist/{paired-binary-B_nejA44.js → paired-binary-BR0nwHOA.js} +3 -4
- package/dist/{paired-binary-B_nejA44.js.map → paired-binary-BR0nwHOA.js.map} +1 -1
- package/dist/{paired-promotion-decision-CQMJ5xGN.d.ts → paired-promotion-decision-BDnqepZ1.d.ts} +1 -5
- package/dist/paired-promotion-decision-BDnqepZ1.d.ts.map +1 -0
- package/dist/{paired-promotion-decision-D0SHb-ZL.js → paired-promotion-decision-FGvHA4p_.js} +5 -7
- package/dist/{paired-promotion-decision-D0SHb-ZL.js.map → paired-promotion-decision-FGvHA4p_.js.map} +1 -1
- package/dist/{pareto-BiBnQXAE.js → pareto-B9udAt2i.js} +28 -50
- package/dist/pareto-B9udAt2i.js.map +1 -0
- package/dist/pipelines/index.d.ts +5 -11
- package/dist/pipelines/index.d.ts.map +1 -1
- package/dist/pipelines/index.js +6 -8
- package/dist/pipelines/index.js.map +1 -1
- package/dist/{power-and-mde-CQX-0weG.js → power-and-mde-ClI6RWw4.js} +4 -5
- package/dist/{power-and-mde-CQX-0weG.js.map → power-and-mde-ClI6RWw4.js.map} +1 -1
- package/dist/{predicate-FdKmElA_.js → predicate-BYHA4MBn.js} +2 -3
- package/dist/{predicate-FdKmElA_.js.map → predicate-BYHA4MBn.js.map} +1 -1
- package/dist/{prime-bridge-transport-6feEglLf.d.ts → prime-bridge-transport-Cfqir_Im.d.ts} +1 -3
- package/dist/prime-bridge-transport-Cfqir_Im.d.ts.map +1 -0
- package/dist/{prime-bridge-transport-Z3cpKGoV.js → prime-bridge-transport-Csd5SQg7.js} +2 -3
- package/dist/{prime-bridge-transport-Z3cpKGoV.js.map → prime-bridge-transport-Csd5SQg7.js.map} +1 -1
- package/dist/{prime-protocol-BGxjfhM4.d.ts → prime-protocol-COFkxas6.d.ts} +4 -8
- package/dist/prime-protocol-COFkxas6.d.ts.map +1 -0
- package/dist/{prime-protocol-BFgGrEMP.js → prime-protocol-CrhpXRz9.js} +4 -6
- package/dist/{prime-protocol-BFgGrEMP.js.map → prime-protocol-CrhpXRz9.js.map} +1 -1
- package/dist/profile-cell.d.ts +1 -1
- package/dist/profile-cell.js +1 -2
- package/dist/{promotion-gate-CkVZHagC.d.ts → promotion-gate-CixshGI8.d.ts} +5 -11
- package/dist/promotion-gate-CixshGI8.d.ts.map +1 -0
- package/dist/{promotion-policy-6A8osGr_.js → promotion-policy-CBYeD4zi.js} +5 -6
- package/dist/{promotion-policy-6A8osGr_.js.map → promotion-policy-CBYeD4zi.js.map} +1 -1
- package/dist/{proposal-findings-DqLUWsAG.js → proposal-findings-WFQrxnfe.js} +3 -4
- package/dist/{proposal-findings-DqLUWsAG.js.map → proposal-findings-WFQrxnfe.js.map} +1 -1
- package/dist/{quantile--0yb07pD.js → quantile-CncXP-Ss.js} +2 -3
- package/dist/{quantile--0yb07pD.js.map → quantile-CncXP-Ss.js.map} +1 -1
- package/dist/{query-B_RZ36an.js → query-CGXzO9d4.js} +5 -6
- package/dist/{query-B_RZ36an.js.map → query-CGXzO9d4.js.map} +1 -1
- package/dist/{query-Cmv_U8vf.d.ts → query-UD0B08Sb.d.ts} +4 -6
- package/dist/query-UD0B08Sb.d.ts.map +1 -0
- package/dist/{random-DGEyr3zG.d.ts → random-BN150Q_X.d.ts} +1 -5
- package/dist/random-BN150Q_X.d.ts.map +1 -0
- package/dist/{random-DDDcC7YU.js → random-DrzcgC7F.js} +3 -4
- package/dist/{random-DDDcC7YU.js.map → random-DrzcgC7F.js.map} +1 -1
- package/dist/{raw-provider-sink-75w4FWBj.js → raw-provider-sink-DDLkP_eF.js} +3 -4
- package/dist/{raw-provider-sink-75w4FWBj.js.map → raw-provider-sink-DDLkP_eF.js.map} +1 -1
- package/dist/{record-id-DUgsK5qp.js → record-id-CWZXFh58.js} +2 -3
- package/dist/{record-id-DUgsK5qp.js.map → record-id-CWZXFh58.js.map} +1 -1
- package/dist/{redact-DXdlFb_U.d.ts → redact-C_wzC4c_.d.ts} +1 -3
- package/dist/redact-C_wzC4c_.d.ts.map +1 -0
- package/dist/{redact-mtkTiJVk.js → redact-D6APJS1d.js} +3 -5
- package/dist/{redact-mtkTiJVk.js.map → redact-D6APJS1d.js.map} +1 -1
- package/dist/{registry-BcdJmyoU.d.ts → registry-dhdDoAV8.d.ts} +5 -11
- package/dist/registry-dhdDoAV8.d.ts.map +1 -0
- package/dist/{release-confidence-CYIFH-fp.js → release-confidence-CTUJz5mM.js} +10 -16
- package/dist/release-confidence-CTUJz5mM.js.map +1 -0
- package/dist/{render-BhJqXoBH.js → render-pfAZ7IR4.js} +266 -33
- package/dist/render-pfAZ7IR4.js.map +1 -0
- package/dist/{reporting-html-Kvlc8oMt.js → reporting-html-DMXIjATy.js} +2 -3
- package/dist/{reporting-html-Kvlc8oMt.js.map → reporting-html-DMXIjATy.js.map} +1 -1
- package/dist/reporting.d.ts +6 -6
- package/dist/reporting.js +7 -7
- package/dist/{researcher-DKWtoMWk.d.ts → researcher-CLnNJ7Or.d.ts} +8 -12
- package/dist/researcher-CLnNJ7Or.d.ts.map +1 -0
- package/dist/{reward-nw2xZGZG.js → reward-P0VD_Aft.js} +2 -3
- package/dist/{reward-nw2xZGZG.js.map → reward-P0VD_Aft.js.map} +1 -1
- package/dist/{reward-BeUgVVC_.d.ts → reward-ST4PXNXF.d.ts} +2 -4
- package/dist/reward-ST4PXNXF.d.ts.map +1 -0
- package/dist/{reward-hacking-WTxUbtbb.d.ts → reward-hacking-C9hPVT0Q.d.ts} +4 -8
- package/dist/reward-hacking-C9hPVT0Q.d.ts.map +1 -0
- package/dist/{reward-hacking-ByuETaIb.js → reward-hacking-DZzHZlS1.js} +5 -7
- package/dist/{reward-hacking-ByuETaIb.js.map → reward-hacking-DZzHZlS1.js.map} +1 -1
- package/dist/rl.d.ts +26 -63
- package/dist/rl.d.ts.map +1 -1
- package/dist/rl.js +35 -52
- package/dist/rl.js.map +1 -1
- package/dist/{rolldown-runtime-8H4AJuhK.js → rolldown-runtime-BJohB4LM.js} +1 -2
- package/dist/rollout/index.d.ts +6 -6
- package/dist/rollout/index.js +7 -7
- package/dist/{rollout-Dd1ucT_M.js → rollout-BhkMpAJT.js} +279 -16
- package/dist/rollout-BhkMpAJT.js.map +1 -0
- package/dist/{terminal-record-CB9BFtjS.js → rollout-nodes-BHR9Q5jP.js} +351 -11
- package/dist/rollout-nodes-BHR9Q5jP.js.map +1 -0
- package/dist/{round-decimal-COrKglMh.js → round-decimal-ZmjAQPo1.js} +2 -3
- package/dist/{round-decimal-COrKglMh.js.map → round-decimal-ZmjAQPo1.js.map} +1 -1
- package/dist/{rubric-predictive-validity-Cd5SCtAf.d.ts → rubric-predictive-validity-B3RzbwOx.d.ts} +2 -8
- package/dist/rubric-predictive-validity-B3RzbwOx.d.ts.map +1 -0
- package/dist/{rubric-predictive-validity-CeKW2GQN.js → rubric-predictive-validity-CQp_M0vy.js} +3 -4
- package/dist/{rubric-predictive-validity-CeKW2GQN.js.map → rubric-predictive-validity-CQp_M0vy.js.map} +1 -1
- package/dist/{run-record-CysKLB93.js → run-record-CZy4-TnZ.js} +4 -5
- package/dist/{run-record-CysKLB93.js.map → run-record-CZy4-TnZ.js.map} +1 -1
- package/dist/{run-record-C3awiXPf.d.ts → run-record-D2BFMPGl.d.ts} +5 -7
- package/dist/run-record-D2BFMPGl.d.ts.map +1 -0
- package/dist/{run-record-DhEJgoPz.js → run-record-DEjx43cI.js} +6 -7
- package/dist/{run-record-DhEJgoPz.js.map → run-record-DEjx43cI.js.map} +1 -1
- package/dist/{sandbox-harness-BlSOu4LX.d.ts → sandbox-harness-B3Kx1fOA.d.ts} +1 -3
- package/dist/sandbox-harness-B3Kx1fOA.d.ts.map +1 -0
- package/dist/{schema-D6EVKzTB.d.ts → schema-CkIvnmLU.d.ts} +2 -43
- package/dist/schema-CkIvnmLU.d.ts.map +1 -0
- package/dist/{schema-Bt2vLFm3.js → schema-D-xu3MU_.js} +2 -3
- package/dist/{schema-Bt2vLFm3.js.map → schema-D-xu3MU_.js.map} +1 -1
- package/dist/{schema-C-ziyJD7.d.ts → schema-DkFJr3Y8.d.ts} +111 -48
- package/dist/schema-DkFJr3Y8.d.ts.map +1 -0
- package/dist/{schema-CvhfLjtZ.js → schema-YUIXT8iW.js} +37 -7
- package/dist/schema-YUIXT8iW.js.map +1 -0
- package/dist/search/index.d.ts +3 -35
- package/dist/search/index.d.ts.map +1 -1
- package/dist/search/index.js +4 -5
- package/dist/{search-kernel-CctpiDd2.d.ts → search-kernel-CBdwFZ05.d.ts} +10 -36
- package/dist/search-kernel-CBdwFZ05.d.ts.map +1 -0
- package/dist/{search-ledger-CZRqtFbW.js → search-ledger-BsVi-Ngv.js} +17 -26
- package/dist/{search-ledger-CZRqtFbW.js.map → search-ledger-BsVi-Ngv.js.map} +1 -1
- package/dist/{search-ledger-file-Bd5Tioq6.js → search-ledger-file-DTBD6X-y.js} +6 -8
- package/dist/{search-ledger-file-Bd5Tioq6.js.map → search-ledger-file-DTBD6X-y.js.map} +1 -1
- package/dist/{search-ledger-types-p2xKfiDJ.d.ts → search-ledger-types-qiYRUo_N.d.ts} +5 -9
- package/dist/search-ledger-types-qiYRUo_N.d.ts.map +1 -0
- package/dist/{search-shipper-Dn3NLsuH.js → search-shipper-pdg4viK7.js} +8 -13
- package/dist/{search-shipper-Dn3NLsuH.js.map → search-shipper-pdg4viK7.js.map} +1 -1
- package/dist/{search-summary-D4bwYnAE.js → search-summary-Ct7SrmVc.js} +6 -7
- package/dist/{search-summary-D4bwYnAE.js.map → search-summary-Ct7SrmVc.js.map} +1 -1
- package/dist/{sequential-CimQrbjv.js → sequential-B1I0tRfL.js} +6 -7
- package/dist/{sequential-CimQrbjv.js.map → sequential-B1I0tRfL.js.map} +1 -1
- package/dist/{sequential-KwGspfeR.d.ts → sequential-Bb_pdTqG.d.ts} +1 -3
- package/dist/sequential-Bb_pdTqG.d.ts.map +1 -0
- package/dist/{sequential-BVv1sjsw.js → sequential-Bjs5roiK.js} +2 -3
- package/dist/{sequential-BVv1sjsw.js.map → sequential-Bjs5roiK.js.map} +1 -1
- package/dist/{sequential-eprocess-DodI6RoF.js → sequential-eprocess-C1YLwFaz.js} +3 -4
- package/dist/{sequential-eprocess-DodI6RoF.js.map → sequential-eprocess-C1YLwFaz.js.map} +1 -1
- package/dist/{server-ChhXPCgX.js → server-DA6h3co5.js} +11 -17
- package/dist/{server-ChhXPCgX.js.map → server-DA6h3co5.js.map} +1 -1
- package/dist/{single-run-lock-Dekd8_Wv.js → single-run-lock-BGlW3OSk.js} +3 -4
- package/dist/{single-run-lock-Dekd8_Wv.js.map → single-run-lock-BGlW3OSk.js.map} +1 -1
- package/dist/{skillopt-optimization-method-sB_-yi1G.js → skillopt-optimization-method-qB1M554K.js} +38 -57
- package/dist/{skillopt-optimization-method-sB_-yi1G.js.map → skillopt-optimization-method-qB1M554K.js.map} +1 -1
- package/dist/{spans-CVngGij3.d.ts → spans-vZoW__3q.d.ts} +2 -4
- package/dist/spans-vZoW__3q.d.ts.map +1 -0
- package/dist/{statistical-heldout-BjzSOmLH.d.ts → statistical-heldout-Cei036nE.d.ts} +9 -37
- package/dist/statistical-heldout-Cei036nE.d.ts.map +1 -0
- package/dist/{statistical-heldout-Bksv92Th.js → statistical-heldout-TP6ndL0B.js} +4 -5
- package/dist/{statistical-heldout-Bksv92Th.js.map → statistical-heldout-TP6ndL0B.js.map} +1 -1
- package/dist/{steps-DlepJPwD.d.ts → steps-B8fYY_Vo.d.ts} +1 -3
- package/dist/steps-B8fYY_Vo.d.ts.map +1 -0
- package/dist/{store-B3nWRy3B.d.ts → store-CMs5wGXP.d.ts} +2 -4
- package/dist/store-CMs5wGXP.d.ts.map +1 -0
- package/dist/{store-BhNuLdJw.js → store-ZykIyQmy.js} +2 -3
- package/dist/{store-BhNuLdJw.js.map → store-ZykIyQmy.js.map} +1 -1
- package/dist/{store-otlp-CFf27E8_.js → store-otlp-DmLm65XW.js} +8 -13
- package/dist/{store-otlp-CFf27E8_.js.map → store-otlp-DmLm65XW.js.map} +1 -1
- package/dist/{store-tool-spans-DxsFwtdr.d.ts → store-tool-spans-B3i3fl0A.d.ts} +11 -29
- package/dist/store-tool-spans-B3i3fl0A.d.ts.map +1 -0
- package/dist/{store-tool-spans-BK_8FjGG.js → store-tool-spans-C2g8UYm-.js} +17 -41
- package/dist/{store-tool-spans-BK_8FjGG.js.map → store-tool-spans-C2g8UYm-.js.map} +1 -1
- package/dist/storyboard/index.d.ts +1 -5
- package/dist/storyboard/index.d.ts.map +1 -1
- package/dist/storyboard/index.js +3 -5
- package/dist/storyboard/index.js.map +1 -1
- package/dist/{summary-report-C8V49Jt3.d.ts → summary-report-0La5vejY.d.ts} +5 -13
- package/dist/summary-report-0La5vejY.d.ts.map +1 -0
- package/dist/{summary-report-DHDWxe6p.js → summary-report-CPfsrLac.js} +21 -30
- package/dist/summary-report-CPfsrLac.js.map +1 -0
- package/dist/supervisor-run/index.d.ts +30 -29
- package/dist/supervisor-run/index.d.ts.map +1 -1
- package/dist/supervisor-run/index.js +6 -7
- package/dist/{systemone-CHnqwmYg.d.ts → systemone-Cw2txK5m.d.ts} +12 -24
- package/dist/systemone-Cw2txK5m.d.ts.map +1 -0
- package/dist/{systemone-protocol-Bl2EI2KQ.d.ts → systemone-protocol-C4t7JH4Y.d.ts} +1 -3
- package/dist/systemone-protocol-C4t7JH4Y.d.ts.map +1 -0
- package/dist/systemone-protocol.d.ts +1 -1
- package/dist/systemone-protocol.js +2 -3
- package/dist/{systemone-review-cQ9acAsb.d.ts → systemone-review-BLgxt-4Q.d.ts} +8 -10
- package/dist/systemone-review-BLgxt-4Q.d.ts.map +1 -0
- package/dist/{systemone-review-DFGtIUcf.js → systemone-review-S49gsy_y.js} +5 -6
- package/dist/{systemone-review-DFGtIUcf.js.map → systemone-review-S49gsy_y.js.map} +1 -1
- package/dist/systemone.d.ts +4 -4
- package/dist/systemone.js +3 -3
- package/dist/{task-failure-attributes-BMsblaDh.js → task-failure-attributes-CC6VuIpk.js} +12 -11
- package/dist/task-failure-attributes-CC6VuIpk.js.map +1 -0
- package/dist/{tool-groups-b3oU_D88.d.ts → tool-groups-C3K-d3gd.d.ts} +3 -5
- package/dist/tool-groups-C3K-d3gd.d.ts.map +1 -0
- package/dist/{tool-waste-CkjzWmeY.d.ts → tool-waste-BFG1vwTV.d.ts} +4 -16
- package/dist/tool-waste-BFG1vwTV.d.ts.map +1 -0
- package/dist/{tool-waste-DLsjS_p-.js → tool-waste-CLJ46AbT.js} +10 -26
- package/dist/{tool-waste-DLsjS_p-.js.map → tool-waste-CLJ46AbT.js.map} +1 -1
- package/dist/trace-attributes.d.ts +1 -1
- package/dist/trace-attributes.js +1 -2
- package/dist/{trace-questions-BmT8M6UQ.d.ts → trace-questions-ej5AEJ5X.d.ts} +4 -6
- package/dist/trace-questions-ej5AEJ5X.d.ts.map +1 -0
- package/dist/trace-repair/index.d.ts +46 -92
- package/dist/trace-repair/index.d.ts.map +1 -1
- package/dist/trace-repair/index.js +32 -54
- package/dist/trace-repair/index.js.map +1 -1
- package/dist/traces.d.ts +13 -29
- package/dist/traces.d.ts.map +1 -1
- package/dist/traces.js +18 -22
- package/dist/traces.js.map +1 -1
- package/dist/{trajectory-D_7rLrvE.js → trajectory-O3GDHmev.js} +2 -3
- package/dist/{trajectory-D_7rLrvE.js.map → trajectory-O3GDHmev.js.map} +1 -1
- package/dist/trajectory-replay/index.d.ts +2 -22
- package/dist/trajectory-replay/index.d.ts.map +1 -1
- package/dist/trajectory-replay/index.js +19 -30
- package/dist/trajectory-replay/index.js.map +1 -1
- package/dist/{tree-DSqwJpL9.js → tree-DrIChF4a.js} +39 -68
- package/dist/{tree-DSqwJpL9.js.map → tree-DrIChF4a.js.map} +1 -1
- package/dist/{types-B_IFcKp5.d.ts → types-0KglnQ9o.d.ts} +3 -15
- package/dist/types-0KglnQ9o.d.ts.map +1 -0
- package/dist/{types-DZ0mYVw9.d.ts → types-0oMg7zee.d.ts} +3 -5
- package/dist/types-0oMg7zee.d.ts.map +1 -0
- package/dist/{types-CnWmj8NW.js → types-BK0blWhe.js} +3 -4
- package/dist/{types-CnWmj8NW.js.map → types-BK0blWhe.js.map} +1 -1
- package/dist/{types-CmXY0PQf.d.ts → types-C-aOqgt2.d.ts} +2 -5
- package/dist/types-C-aOqgt2.d.ts.map +1 -0
- package/dist/{types-M6eF9m-r.d.ts → types-C54Uk_a1.d.ts} +84 -24
- package/dist/types-C54Uk_a1.d.ts.map +1 -0
- package/dist/{types-BXgASBCL.js → types-DK4vNv5q.js} +2 -3
- package/dist/{types-BXgASBCL.js.map → types-DK4vNv5q.js.map} +1 -1
- package/dist/{types-6Lfdlshp.d.ts → types-DzCO-S1m.d.ts} +50 -62
- package/dist/types-DzCO-S1m.d.ts.map +1 -0
- package/dist/{verdict-BGQmxQ8A.js → verdict-C8TcdwfD.js} +3 -4
- package/dist/{verdict-BGQmxQ8A.js.map → verdict-C8TcdwfD.js.map} +1 -1
- package/dist/{verdict-C24NELtN.d.ts → verdict-D_qiBQ6_.d.ts} +1 -5
- package/dist/verdict-D_qiBQ6_.d.ts.map +1 -0
- package/dist/{verdict-cache-CL_7tePD.js → verdict-cache-BVH-fl18.js} +3 -4
- package/dist/{verdict-cache-CL_7tePD.js.map → verdict-cache-BVH-fl18.js.map} +1 -1
- package/dist/wire/index.d.ts +43 -51
- package/dist/wire/index.d.ts.map +1 -1
- package/dist/wire/index.js +1 -2
- package/docs/concepts.md +1 -0
- package/docs/eval-surface-map.md +15 -0
- package/docs/rollout.md +132 -6
- package/package.json +18 -18
- package/dist/adversarial-smnADNFS.d.ts.map +0 -1
- package/dist/agent-profile-CNmwfaqq.d.ts.map +0 -1
- package/dist/agent-profile-cell-BcpFbNd5.d.ts.map +0 -1
- package/dist/analyze-runs-DfX6dHP0.d.ts.map +0 -1
- package/dist/attribute-vocabulary-CcXqHODT.d.ts.map +0 -1
- package/dist/benchmark-command-Cnjfo0Zh.js.map +0 -1
- package/dist/campaign-evidence-DunzcJGv.js.map +0 -1
- package/dist/campaign-uddTlFg5.js.map +0 -1
- package/dist/canonical-CFpojCN5.d.ts.map +0 -1
- package/dist/client-hwo3cJjR.d.ts.map +0 -1
- package/dist/completion-verifier-asaV1nSj.js.map +0 -1
- package/dist/cost-ledger-C1W3ICmP.d.ts.map +0 -1
- package/dist/dataset-DQqhOCPt.d.ts.map +0 -1
- package/dist/default-registry-CLG5YqGS.d.ts.map +0 -1
- package/dist/define-agent-eval-JKt3gMQr.d.ts.map +0 -1
- package/dist/descriptive-B2iPaT9J.d.ts +0 -89
- package/dist/descriptive-B2iPaT9J.d.ts.map +0 -1
- package/dist/descriptive-CVNy48rR.js.map +0 -1
- package/dist/emitter-DABYYTcH.d.ts.map +0 -1
- package/dist/engine-dtRVeuQs.d.ts.map +0 -1
- package/dist/errors-DEE6u6ot.d.ts.map +0 -1
- package/dist/evaluation-DblGquCB.d.ts.map +0 -1
- package/dist/exact-types-Do63KoKn.d.ts.map +0 -1
- package/dist/exporters-DWvM16LE.js.map +0 -1
- package/dist/exporters-Dk3oEbK_.d.ts.map +0 -1
- package/dist/failure-cluster-DN0t63Z0.d.ts.map +0 -1
- package/dist/feedback-trajectory-CYVJXolG.d.ts.map +0 -1
- package/dist/hf-dataset-CKxIJQba.js.map +0 -1
- package/dist/index-CHxwfW5g.d.ts.map +0 -1
- package/dist/index-CTIRz-AB.d.ts.map +0 -1
- package/dist/index-CalAsXMR.d.ts.map +0 -1
- package/dist/index-CdEVr8-O.d.ts.map +0 -1
- package/dist/index-CfJULcW5.d.ts.map +0 -1
- package/dist/index-D7fmC-1K.d.ts.map +0 -1
- package/dist/index-jnuFLxOs.d.ts.map +0 -1
- package/dist/integrity-D-s2iHiT.js.map +0 -1
- package/dist/integrity-D9lBnv_K.d.ts.map +0 -1
- package/dist/interval-ohuOJurX.d.ts.map +0 -1
- package/dist/judge-calibration-DFtEMlde.d.ts.map +0 -1
- package/dist/llm-judge-Cybq29bL.js.map +0 -1
- package/dist/lock-holder-DaJbfpci.d.ts.map +0 -1
- package/dist/matrix-B8l-Nsxd.js.map +0 -1
- package/dist/matrix-DtkZLD_3.d.ts.map +0 -1
- package/dist/mint-CkzpVPdz.d.ts +0 -81
- package/dist/mint-CkzpVPdz.d.ts.map +0 -1
- package/dist/multi-layer-verifier-DEgMUk6v.d.ts.map +0 -1
- package/dist/multiplicity-DIWHvysC.d.ts.map +0 -1
- package/dist/paired-promotion-decision-CQMJ5xGN.d.ts.map +0 -1
- package/dist/pareto-BiBnQXAE.js.map +0 -1
- package/dist/prime-bridge-transport-6feEglLf.d.ts.map +0 -1
- package/dist/prime-protocol-BGxjfhM4.d.ts.map +0 -1
- package/dist/promotion-gate-CkVZHagC.d.ts.map +0 -1
- package/dist/query-Cmv_U8vf.d.ts.map +0 -1
- package/dist/random-DGEyr3zG.d.ts.map +0 -1
- package/dist/redact-DXdlFb_U.d.ts.map +0 -1
- package/dist/registry-BcdJmyoU.d.ts.map +0 -1
- package/dist/release-confidence-CYIFH-fp.js.map +0 -1
- package/dist/render-BhJqXoBH.js.map +0 -1
- package/dist/researcher-DKWtoMWk.d.ts.map +0 -1
- package/dist/reward-BeUgVVC_.d.ts.map +0 -1
- package/dist/reward-hacking-WTxUbtbb.d.ts.map +0 -1
- package/dist/rollout-Dd1ucT_M.js.map +0 -1
- package/dist/rubric-predictive-validity-Cd5SCtAf.d.ts.map +0 -1
- package/dist/run-record-C3awiXPf.d.ts.map +0 -1
- package/dist/sandbox-harness-BlSOu4LX.d.ts.map +0 -1
- package/dist/schema-C-ziyJD7.d.ts.map +0 -1
- package/dist/schema-CvhfLjtZ.js.map +0 -1
- package/dist/schema-D6EVKzTB.d.ts.map +0 -1
- package/dist/search-kernel-CctpiDd2.d.ts.map +0 -1
- package/dist/search-ledger-types-p2xKfiDJ.d.ts.map +0 -1
- package/dist/sequential-KwGspfeR.d.ts.map +0 -1
- package/dist/spans-CVngGij3.d.ts.map +0 -1
- package/dist/statistical-heldout-BjzSOmLH.d.ts.map +0 -1
- package/dist/steps-DlepJPwD.d.ts.map +0 -1
- package/dist/store-B3nWRy3B.d.ts.map +0 -1
- package/dist/store-tool-spans-DxsFwtdr.d.ts.map +0 -1
- package/dist/summary-report-C8V49Jt3.d.ts.map +0 -1
- package/dist/summary-report-DHDWxe6p.js.map +0 -1
- package/dist/systemone-CHnqwmYg.d.ts.map +0 -1
- package/dist/systemone-protocol-Bl2EI2KQ.d.ts.map +0 -1
- package/dist/systemone-review-cQ9acAsb.d.ts.map +0 -1
- package/dist/task-failure-attributes-BMsblaDh.js.map +0 -1
- package/dist/terminal-record-CB9BFtjS.js.map +0 -1
- package/dist/tool-groups-b3oU_D88.d.ts.map +0 -1
- package/dist/tool-waste-CkjzWmeY.d.ts.map +0 -1
- package/dist/trace-questions-BmT8M6UQ.d.ts.map +0 -1
- package/dist/types-6Lfdlshp.d.ts.map +0 -1
- package/dist/types-B_IFcKp5.d.ts.map +0 -1
- package/dist/types-CmXY0PQf.d.ts.map +0 -1
- package/dist/types-DZ0mYVw9.d.ts.map +0 -1
- package/dist/types-M6eF9m-r.d.ts.map +0 -1
- package/dist/verdict-C24NELtN.d.ts.map +0 -1
|
@@ -1,11 +1,10 @@
|
|
|
1
|
-
import { m as CustomTokenPricing, o as CostLedgerHandle, u as CostProvenance } from "../cost-ledger-
|
|
2
|
-
import { r as CaptureIntegrityError } from "../errors-
|
|
3
|
-
import { d as RunTokenUsage } from "../run-record-
|
|
4
|
-
import { t as DefaultVerdict } from "../verdict-
|
|
5
|
-
import { i as TraceAnalystLimits, t as TraceAnalysisEngine } from "../engine-
|
|
6
|
-
import { t as PrimeBridgeTransport } from "../prime-bridge-transport-
|
|
7
|
-
import { i as RecordedTrajectoryStep } from "../steps-
|
|
8
|
-
//#region src/trace-repair/mini-swe-scaffold.d.ts
|
|
1
|
+
import { m as CustomTokenPricing, o as CostLedgerHandle, u as CostProvenance } from "../cost-ledger-Czro9wmo.js";
|
|
2
|
+
import { r as CaptureIntegrityError } from "../errors-E9I2hldP.js";
|
|
3
|
+
import { d as RunTokenUsage } from "../run-record-D2BFMPGl.js";
|
|
4
|
+
import { t as DefaultVerdict } from "../verdict-D_qiBQ6_.js";
|
|
5
|
+
import { i as TraceAnalystLimits, t as TraceAnalysisEngine } from "../engine-CaHaCc3W.js";
|
|
6
|
+
import { t as PrimeBridgeTransport } from "../prime-bridge-transport-Cfqir_Im.js";
|
|
7
|
+
import { i as RecordedTrajectoryStep } from "../steps-B8fYY_Vo.js";
|
|
9
8
|
type ParsedAction = {
|
|
10
9
|
kind: 'action';
|
|
11
10
|
command: string;
|
|
@@ -25,8 +24,6 @@ interface CommandOutput {
|
|
|
25
24
|
/** Set when the environment itself failed rather than the command. */
|
|
26
25
|
exceptionInfo?: string;
|
|
27
26
|
}
|
|
28
|
-
//#endregion
|
|
29
|
-
//#region src/trace-repair/action-budget.d.ts
|
|
30
27
|
/**
|
|
31
28
|
* The action budget an intervention must fit inside.
|
|
32
29
|
*
|
|
@@ -80,8 +77,6 @@ type BudgetCheck = {
|
|
|
80
77
|
readonly detail: string;
|
|
81
78
|
readonly measurement: BudgetMeasurement;
|
|
82
79
|
};
|
|
83
|
-
//#endregion
|
|
84
|
-
//#region src/trace-repair/continuation-records.d.ts
|
|
85
80
|
/** Which arm produced a rollout. The policy is identical across all three. */
|
|
86
81
|
type ContinuationArm = 'intervention' | 'no-fix-control' | 'no-op-control';
|
|
87
82
|
/** Chat message in the scaffold's own vocabulary. */
|
|
@@ -199,8 +194,6 @@ export declare function rolloutRecordedSteps(rollout: ContinuationRollout): Reco
|
|
|
199
194
|
* seconds. Use it to assert determinism, never to assert equal latency.
|
|
200
195
|
*/
|
|
201
196
|
export declare function rolloutDigest(rollout: ContinuationRollout): string;
|
|
202
|
-
//#endregion
|
|
203
|
-
//#region src/trace-repair/control-policy.d.ts
|
|
204
197
|
/** The declared control cannot produce the outcome the criteria screen for. */
|
|
205
198
|
export declare class UncalibratedControlError extends CaptureIntegrityError {}
|
|
206
199
|
/**
|
|
@@ -261,8 +254,6 @@ export declare function defineControlPolicy(input: ControlPolicyInput): ControlP
|
|
|
261
254
|
* behind a label that says nothing was screened.
|
|
262
255
|
*/
|
|
263
256
|
export declare function assertControlCalibrated(policy: ControlCapability, screening: ControlScreening): void;
|
|
264
|
-
//#endregion
|
|
265
|
-
//#region src/trace-repair/admission-records.d.ts
|
|
266
257
|
/**
|
|
267
258
|
* One corpus row, described by what the recording holds and nothing else.
|
|
268
259
|
*
|
|
@@ -418,8 +409,6 @@ interface DenominatorChainArtifact {
|
|
|
418
409
|
export declare function buildDenominatorChain(verdicts: readonly AdmissionRowVerdict[], admitStrata: readonly AdmissionStratum[]): DenominatorChainArtifact;
|
|
419
410
|
/** A chain that does not add up is a broken denominator, so this throws. */
|
|
420
411
|
export declare function assertChainReconciles(artifact: DenominatorChainArtifact): void;
|
|
421
|
-
//#endregion
|
|
422
|
-
//#region src/trace-repair/continuation-policy.d.ts
|
|
423
412
|
interface PinnedContinuationPolicy {
|
|
424
413
|
/** Stable name for the frozen configuration, recorded on every rollout. */
|
|
425
414
|
readonly id: string;
|
|
@@ -556,8 +545,6 @@ export declare function totalCost(steps: readonly ContinuationStepRecord[]): Cos
|
|
|
556
545
|
* measured difference includes a policy change, not only the intervention.
|
|
557
546
|
*/
|
|
558
547
|
export declare function assertArmSymmetry(rollouts: readonly ContinuationRollout[]): void;
|
|
559
|
-
//#endregion
|
|
560
|
-
//#region src/trace-repair/oracle-determinism.d.ts
|
|
561
548
|
/** The two states certification produces: the published image, and that image
|
|
562
549
|
* after the task's own reference solution ran. */
|
|
563
550
|
type OracleStateLabel = 'unsolved' | 'solved';
|
|
@@ -672,8 +659,6 @@ interface TaskOracleRegistryDocument {
|
|
|
672
659
|
export declare function parseTaskOracleRegistry(document: unknown): TaskOracleRegistry;
|
|
673
660
|
/** Throws unless the task graded the same bytes the same way every replicate. */
|
|
674
661
|
export declare function assertDeterministicOracle(verdict: OracleDeterminismVerdict): void;
|
|
675
|
-
//#endregion
|
|
676
|
-
//#region src/trace-repair/admission.d.ts
|
|
677
662
|
/**
|
|
678
663
|
* Result of an external call. Callers read `succeeded` before `value`: a
|
|
679
664
|
* boundary failure is an exclusion with its own reason, never a false verdict.
|
|
@@ -840,8 +825,6 @@ interface DenominatorIntactInput {
|
|
|
840
825
|
* with no outcome is an error, not a smaller `n`.
|
|
841
826
|
*/
|
|
842
827
|
export declare function assertDenominatorIntact(input: DenominatorIntactInput): void;
|
|
843
|
-
//#endregion
|
|
844
|
-
//#region src/trace-repair/admission-contract.d.ts
|
|
845
828
|
/** Pre-registered admission thresholds. */
|
|
846
829
|
interface AdmissionCriteria {
|
|
847
830
|
/** Share of replayed prefix steps allowed to diverge from their recorded
|
|
@@ -964,8 +947,6 @@ type AdmissionDecision = {
|
|
|
964
947
|
* that reads as if a check had been applied.
|
|
965
948
|
*/
|
|
966
949
|
export declare function admitRow(evidence: AdmissionEvidence, criteria?: AdmissionCriteria): AdmissionDecision;
|
|
967
|
-
//#endregion
|
|
968
|
-
//#region src/trace-repair/admission-report.d.ts
|
|
969
950
|
interface AdmissionArtifact {
|
|
970
951
|
version: 1;
|
|
971
952
|
kind: 'tb-repair-admission';
|
|
@@ -992,8 +973,6 @@ interface RenderAdmissionOptions {
|
|
|
992
973
|
}
|
|
993
974
|
/** Markdown for the campaign report. Every number here is also in the artifact. */
|
|
994
975
|
export declare function renderAdmissionReport(artifact: AdmissionArtifact, options?: RenderAdmissionOptions): string;
|
|
995
|
-
//#endregion
|
|
996
|
-
//#region src/trace-repair/analyst-response.d.ts
|
|
997
976
|
interface RepairIntervention {
|
|
998
977
|
/** What the action is: a shell command, or an edit that authors file content. */
|
|
999
978
|
readonly kind: ActionPayloadKind;
|
|
@@ -1032,8 +1011,6 @@ type ParseAnalystResponseOutcome = {
|
|
|
1032
1011
|
* missing `k` is a parse failure, never step 1.
|
|
1033
1012
|
*/
|
|
1034
1013
|
export declare function parseAnalystResponse(reply: string): ParseAnalystResponseOutcome;
|
|
1035
|
-
//#endregion
|
|
1036
|
-
//#region src/trace-repair/blinding.d.ts
|
|
1037
1014
|
interface BlindedStep {
|
|
1038
1015
|
readonly step_id: number;
|
|
1039
1016
|
readonly action: string;
|
|
@@ -1056,8 +1033,6 @@ interface BlindTrajectoryOptions {
|
|
|
1056
1033
|
readonly throughStep?: number;
|
|
1057
1034
|
}
|
|
1058
1035
|
export declare function blindTrajectory(row: AdmittedRow, options?: BlindTrajectoryOptions): BlindedTrajectoryPrefix;
|
|
1059
|
-
//#endregion
|
|
1060
|
-
//#region src/trace-repair/analyst-arm.d.ts
|
|
1061
1036
|
/**
|
|
1062
1037
|
* Something an arm can do that another arm cannot.
|
|
1063
1038
|
*
|
|
@@ -1270,8 +1245,6 @@ interface RepairArmAsymmetryReport {
|
|
|
1270
1245
|
export declare function repairArmAsymmetries(arms: readonly RepairAnalystArm[], options?: {
|
|
1271
1246
|
readonly budget?: InterventionBudget;
|
|
1272
1247
|
}): RepairArmAsymmetryReport;
|
|
1273
|
-
//#endregion
|
|
1274
|
-
//#region src/trace-repair/arm-completion.d.ts
|
|
1275
1248
|
interface CompletionRepairArmOptions {
|
|
1276
1249
|
readonly id: string;
|
|
1277
1250
|
/** What actually answers the POST, in one plain sentence. */
|
|
@@ -1291,8 +1264,6 @@ interface CompletionRepairArmOptions {
|
|
|
1291
1264
|
readonly certification: RepairArmCertification;
|
|
1292
1265
|
}
|
|
1293
1266
|
export declare function createCompletionRepairArm(options: CompletionRepairArmOptions): RepairAnalystArm;
|
|
1294
|
-
//#endregion
|
|
1295
|
-
//#region src/trace-repair/arm-dspy.d.ts
|
|
1296
1267
|
/** Signature the bridge reports for a repair analysis. */
|
|
1297
1268
|
export declare const DSPY_REPAIR_SIGNATURE = "tb-repair-typed-v1";
|
|
1298
1269
|
interface DspyRepairArmOptions {
|
|
@@ -1315,8 +1286,6 @@ interface ReadRepairPayloadOptions {
|
|
|
1315
1286
|
* grader would refuse as out of range. */
|
|
1316
1287
|
readonly validSteps: readonly number[];
|
|
1317
1288
|
}
|
|
1318
|
-
//#endregion
|
|
1319
|
-
//#region src/trace-repair/degenerate-strategies.d.ts
|
|
1320
1289
|
/**
|
|
1321
1290
|
* The strategies that would score without repairing anything, named here so a
|
|
1322
1291
|
* reviewer can check each one against the mechanism that defeats it and a
|
|
@@ -1343,57 +1312,55 @@ interface DegenerateStrategy {
|
|
|
1343
1312
|
readonly enforcedIn: string;
|
|
1344
1313
|
}
|
|
1345
1314
|
export declare const DEGENERATE_STRATEGIES: readonly [{
|
|
1346
|
-
readonly id:
|
|
1347
|
-
readonly strategy:
|
|
1348
|
-
readonly defeat:
|
|
1349
|
-
readonly defeatKind:
|
|
1350
|
-
readonly enforcedIn:
|
|
1315
|
+
readonly id: "point-at-any-nonzero-exit-step";
|
|
1316
|
+
readonly strategy: "Name the first step with a nonzero returncode, claim it failed, and let the reproduction gate confirm it.";
|
|
1317
|
+
readonly defeat: "Reproduction is a gate, not a tier that pays. The credit vector has no term for it, so a reproduced step with no working intervention scores exactly what an unreproduced one scores: nothing.";
|
|
1318
|
+
readonly defeatKind: "gate";
|
|
1319
|
+
readonly enforcedIn: "funnel.ts (repairCredit has no reproduction term)";
|
|
1351
1320
|
}, {
|
|
1352
|
-
readonly id:
|
|
1353
|
-
readonly strategy:
|
|
1354
|
-
readonly defeat:
|
|
1355
|
-
readonly defeatKind:
|
|
1356
|
-
readonly enforcedIn:
|
|
1321
|
+
readonly id: "propose-the-recorded-command-again";
|
|
1322
|
+
readonly strategy: "Return the action the agent already ran at step k, so the arm reproduces the recorded state and looks like a faithful replay.";
|
|
1323
|
+
readonly defeat: "The intervention is compared against the recorded action at k after whitespace normalisation and rejected before a container opens.";
|
|
1324
|
+
readonly defeatKind: "gate";
|
|
1325
|
+
readonly enforcedIn: "grade.ts (recorded-action-reproposed rejection)";
|
|
1357
1326
|
}, {
|
|
1358
|
-
readonly id:
|
|
1359
|
-
readonly strategy:
|
|
1360
|
-
readonly defeat:
|
|
1361
|
-
readonly defeatKind:
|
|
1362
|
-
readonly enforcedIn:
|
|
1327
|
+
readonly id: "propose-a-no-op";
|
|
1328
|
+
readonly strategy: "Return an action that changes nothing, so the arm inherits whatever the trajectory would have done anyway.";
|
|
1329
|
+
readonly defeat: "Literal no-ops are rejected at the budget check. A semantic no-op cannot be detected syntactically, so it runs and measures at the no-op control floor: its paired delta is zero.";
|
|
1330
|
+
readonly defeatKind: "measurement";
|
|
1331
|
+
readonly enforcedIn: "action-budget.ts (NO_OP_ACTIONS) and delta-repair.ts (paired delta)";
|
|
1363
1332
|
}, {
|
|
1364
|
-
readonly id:
|
|
1365
|
-
readonly strategy:
|
|
1366
|
-
readonly defeat:
|
|
1367
|
-
readonly defeatKind:
|
|
1368
|
-
readonly enforcedIn:
|
|
1333
|
+
readonly id: "submit-instead-of-repair";
|
|
1334
|
+
readonly strategy: "Return the submit sentinel so the run ends immediately and the arm terminates cleanly.";
|
|
1335
|
+
readonly defeat: "An action carrying the submit sentinel is rejected at the budget check. Ending the run is what the recorded agent already did and the tests already failed on it.";
|
|
1336
|
+
readonly defeatKind: "gate";
|
|
1337
|
+
readonly enforcedIn: "action-budget.ts (SUBMIT_SENTINEL rejection)";
|
|
1369
1338
|
}, {
|
|
1370
|
-
readonly id:
|
|
1339
|
+
readonly id: "touch-the-test-suite";
|
|
1371
1340
|
readonly strategy: "Write a passing suite, or a reward file, at the path the grader will read, so the oracle grades the trajectory's own artifact.";
|
|
1372
|
-
readonly defeat:
|
|
1373
|
-
readonly defeatKind:
|
|
1374
|
-
readonly enforcedIn:
|
|
1341
|
+
readonly defeat: "The oracle purges the suite root and uploads the held-out suite from outside the session at grade time, then verifies the bytes it reads back. A planted suite is overwritten; a session that refuses the overwrite raises a tamper error instead of returning a pass.";
|
|
1342
|
+
readonly defeatKind: "gate";
|
|
1343
|
+
readonly enforcedIn: "test-oracle.ts (purge, upload, read-back digest)";
|
|
1375
1344
|
}, {
|
|
1376
|
-
readonly id:
|
|
1377
|
-
readonly strategy:
|
|
1378
|
-
readonly defeat:
|
|
1379
|
-
readonly defeatKind:
|
|
1380
|
-
readonly enforcedIn:
|
|
1345
|
+
readonly id: "buy-a-bigger-action";
|
|
1346
|
+
readonly strategy: "Return a multi-command script or a whole-file rewrite that does far more than one scaffold turn could.";
|
|
1347
|
+
readonly defeat: "The budget counts top-level statements, heredocs and bytes. More than one action, more than one authored file, or more than 4 KB is rejected before a container opens.";
|
|
1348
|
+
readonly defeatKind: "gate";
|
|
1349
|
+
readonly enforcedIn: "action-budget.ts (checkInterventionBudget)";
|
|
1381
1350
|
}, {
|
|
1382
|
-
readonly id:
|
|
1383
|
-
readonly strategy:
|
|
1384
|
-
readonly defeat:
|
|
1385
|
-
readonly defeatKind:
|
|
1386
|
-
readonly enforcedIn:
|
|
1351
|
+
readonly id: "decline-every-hard-row";
|
|
1352
|
+
readonly strategy: "Answer no-decisive-failure on everything except the rows that are obviously repairable, so the reported rate is computed on an easy subset.";
|
|
1353
|
+
readonly defeat: "A declined row keeps its cell in the funnel and stays in the denominator with a paired delta of zero, because its intervention arm is definitionally its control arm. Declining cannot raise the headline; it can only dilute it.";
|
|
1354
|
+
readonly defeatKind: "measurement";
|
|
1355
|
+
readonly enforcedIn: "grade.ts (declined outcome) and delta-repair.ts (full admitted denominator)";
|
|
1387
1356
|
}, {
|
|
1388
|
-
readonly id:
|
|
1389
|
-
readonly strategy:
|
|
1390
|
-
readonly defeat:
|
|
1391
|
-
readonly defeatKind:
|
|
1392
|
-
readonly enforcedIn:
|
|
1357
|
+
readonly id: "repair-somewhere-other-than-k";
|
|
1358
|
+
readonly strategy: "Name a plausible-looking k, then submit an intervention that fixes the task from any state, so the answer scores without localising anything.";
|
|
1359
|
+
readonly defeat: "The intervention is executed at the k the analyst named, on the state produced by replaying steps 1..k-1. There is no separate localisation credit to win and no label the grader reads, so a wrong k is only penalised through the repair failing to work there.";
|
|
1360
|
+
readonly defeatKind: "measurement";
|
|
1361
|
+
readonly enforcedIn: "grade.ts (the intervention is applied at the named k only)";
|
|
1393
1362
|
}];
|
|
1394
1363
|
type DegenerateStrategyId = (typeof DEGENERATE_STRATEGIES)[number]['id'];
|
|
1395
|
-
//#endregion
|
|
1396
|
-
//#region src/trace-repair/ports.d.ts
|
|
1397
1364
|
/**
|
|
1398
1365
|
* Which arm a session serves. The arm names the container state under test,
|
|
1399
1366
|
* never a different policy: `Delta-repair` only measures the intervention when
|
|
@@ -1491,8 +1458,6 @@ interface RepairContinuationOutcome {
|
|
|
1491
1458
|
}
|
|
1492
1459
|
/** Runs the pinned continuation policy forward from a prepared session. */
|
|
1493
1460
|
type RepairContinuationRunner = (request: RepairContinuationRequest) => Promise<RepairContinuationOutcome>;
|
|
1494
|
-
//#endregion
|
|
1495
|
-
//#region src/trace-repair/funnel.d.ts
|
|
1496
1461
|
/** Why an answer never reached the reproduction gate. */
|
|
1497
1462
|
type RepairRejection = {
|
|
1498
1463
|
readonly source: 'parse';
|
|
@@ -1638,8 +1603,6 @@ interface RepairFunnelCounts {
|
|
|
1638
1603
|
/** Rows where every repair rollout passed. */
|
|
1639
1604
|
readonly t4RepairFlipAll: number;
|
|
1640
1605
|
}
|
|
1641
|
-
//#endregion
|
|
1642
|
-
//#region src/trace-repair/grade.d.ts
|
|
1643
1606
|
interface GradeRepairOptions {
|
|
1644
1607
|
readonly row: AdmittedRow;
|
|
1645
1608
|
readonly response: AnalystResponse;
|
|
@@ -1689,8 +1652,6 @@ interface RepairRowResult extends DefaultVerdict {
|
|
|
1689
1652
|
readonly wallMs: number;
|
|
1690
1653
|
}
|
|
1691
1654
|
export declare function gradeRepairRow(options: GradeRepairOptions): Promise<RepairRowResult>;
|
|
1692
|
-
//#endregion
|
|
1693
|
-
//#region src/trace-repair/delta-repair.d.ts
|
|
1694
1655
|
interface DeltaRepairOptions {
|
|
1695
1656
|
/** Deterministic resampling seed. Derived from the deltas when omitted. */
|
|
1696
1657
|
readonly seed?: number;
|
|
@@ -1735,8 +1696,6 @@ export declare function deltaRepair(rowResults: readonly RepairRowResult[], opti
|
|
|
1735
1696
|
/** Markdown report: provenance, the funnel, every per-row column, the
|
|
1736
1697
|
* distribution, and the threats. */
|
|
1737
1698
|
export declare function renderDeltaRepairReport(report: DeltaRepairReport): string;
|
|
1738
|
-
//#endregion
|
|
1739
|
-
//#region src/trace-repair/docker-environment.d.ts
|
|
1740
1699
|
interface ProcessRequest {
|
|
1741
1700
|
argv: string[];
|
|
1742
1701
|
/** Kills the process group when exceeded. */
|
|
@@ -1773,8 +1732,6 @@ interface DockerContinuationEnvironmentOptions {
|
|
|
1773
1732
|
removeOnDispose?: boolean;
|
|
1774
1733
|
}
|
|
1775
1734
|
export declare function createDockerContinuationEnvironment(options: DockerContinuationEnvironmentOptions): ContinuationEnvironment;
|
|
1776
|
-
//#endregion
|
|
1777
|
-
//#region src/trace-repair/repair-prompt.d.ts
|
|
1778
1735
|
/**
|
|
1779
1736
|
* Digest of the question and task policy every arm shares.
|
|
1780
1737
|
*
|
|
@@ -1793,8 +1750,6 @@ export declare function repairQuestionSha256(budget?: InterventionBudget): strin
|
|
|
1793
1750
|
* that ask the identical composed question share one.
|
|
1794
1751
|
*/
|
|
1795
1752
|
export declare function repairArmPromptSha256(budget: InterventionBudget, contract: readonly string[]): string;
|
|
1796
|
-
//#endregion
|
|
1797
|
-
//#region src/trace-repair/test-oracle.d.ts
|
|
1798
1753
|
/** The suite the oracle read back is not the suite it uploaded. */
|
|
1799
1754
|
export declare class TestSuiteTamperedError extends CaptureIntegrityError {}
|
|
1800
1755
|
interface TestSuiteFile {
|
|
@@ -1816,6 +1771,5 @@ interface InjectedTestOracleOptions {
|
|
|
1816
1771
|
readonly commandTimeoutMs?: number;
|
|
1817
1772
|
}
|
|
1818
1773
|
export declare function injectedTestOracle(options: InjectedTestOracleOptions): TestOracle;
|
|
1819
|
-
//#endregion
|
|
1820
1774
|
export type { ActionPayloadKind, AdmissionArtifact, AdmissionCheckRecord, AdmissionConfig, AdmissionConfigInput, AdmissionControlArm, AdmissionControlObservation, AdmissionControlRequest, AdmissionControlRunner, AdmissionCriteria, AdmissionDecision, AdmissionDivergence, AdmissionEndStateOracle, AdmissionEvidence, AdmissionExclusionReason, AdmissionNoOpInjection, AdmissionOutcome, AdmissionPrefixReplay, AdmissionPrefixReplayer, AdmissionProvenance, AdmissionRejection, AdmissionReport, AdmissionRolloutSummary, AdmissionRow, AdmissionRowVerdict, AdmissionScreeningRecord, AdmissionStratum, AdmissionTestVerdict, AdmittedRow, AnalystResponse, AskRepairArmOptions, BlindTrajectoryOptions, BlindedStep, BlindedTrajectoryPrefix, BudgetCheck, BudgetMeasurement, BudgetViolation, CommandOutput, CompletionRepairArmOptions, ContinuationArm, ContinuationEnvironment, ContinuationEnvironmentDescription, ContinuationEnvironmentFactory, ContinuationEnvironmentRequest, ContinuationExecRecord, ContinuationExecResult, ContinuationExitStatus, ContinuationMessage, ContinuationModel, ContinuationModelCall, ContinuationModelRequest, ContinuationModelResponse, ContinuationRollout, ContinuationStepRecord, ContinuationUsageTotals, ControlArmEvidence, ControlCapability, ControlPolicy, ControlPolicyInput, ControlScreening, DefineContinuationPolicyInput, DegenerateDefeatKind, DegenerateStrategy, DegenerateStrategyId, DeltaRepairInterval, DeltaRepairOptions, DeltaRepairReport, DenominatorChain, DenominatorChainArtifact, DenominatorIntactInput, DenominatorStage, DockerContinuationEnvironmentOptions, DspyRepairArmOptions, GradeRepairOptions, InjectedAction, InjectedTestOracleOptions, InterventionBudget, InterventionExecution, OracleAssertionResult, OracleDeterminismEvidence, OracleDeterminismVerdict, OracleFlippedUnit, OracleLoad, OracleReplicate, OracleReplicateGroup, OracleStateLabel, OracleStateVerdict, ParseAnalystResponseOutcome, ParsedAction, PinnedContinuationPolicy, PrefixFidelityEvidence, PrefixReplayEvidence, ProcessRequest, ProcessResult, ProcessRunner, ReadRepairPayloadOptions, RecordedStep, RecordedToolCall, RenderAdmissionOptions, RepairAnalystArm, RepairArm, RepairArmAffordance, RepairArmAnswer, RepairArmAsymmetry, RepairArmAsymmetryReport, RepairArmBudgetRecord, RepairArmCertification, RepairArmDeclaration, RepairArmEvidence, RepairArmRejectedRow, RepairArmRepairTurn, RepairArmReply, RepairArmRequest, RepairArmUsage, RepairContinuationOutcome, RepairContinuationRequest, RepairContinuationRunner, RepairCredit, RepairDeclined, RepairExecResult, RepairFinding, RepairFunnelCounts, RepairGrade, RepairIntervention, RepairRejection, RepairRolloutEvidence, RepairRowResult, RepairSession, RepairSessionFactory, RepairSessionRequest, RepairThreat, RepairThreatId, ReproductionEvidence, ResponseParseFailure, RunAdmissionOptions, RunContinuationOptions, TaskOracleRegistry, TaskOracleRegistryDocument, TestOracle, TestOracleContext, TestOracleOutcome, TestRunEvidence, TestSuiteFile };
|
|
1821
1775
|
//# sourceMappingURL=index.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/trace-repair/mini-swe-scaffold.ts","../../src/trace-repair/action-budget.ts","../../src/trace-repair/continuation-records.ts","../../src/trace-repair/control-policy.ts","../../src/trace-repair/admission-records.ts","../../src/trace-repair/continuation-policy.ts","../../src/trace-repair/oracle-determinism.ts","../../src/trace-repair/admission.ts","../../src/trace-repair/admission-contract.ts","../../src/trace-repair/admission-report.ts","../../src/trace-repair/analyst-response.ts","../../src/trace-repair/blinding.ts","../../src/trace-repair/analyst-arm.ts","../../src/trace-repair/arm-completion.ts","../../src/trace-repair/arm-dspy.ts","../../src/trace-repair/degenerate-strategies.ts","../../src/trace-repair/ports.ts","../../src/trace-repair/funnel.ts","../../src/trace-repair/grade.ts","../../src/trace-repair/delta-repair.ts","../../src/trace-repair/docker-environment.ts","../../src/trace-repair/repair-prompt.ts","../../src/trace-repair/test-oracle.ts"],"mappings":"
|
|
1
|
+
{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/trace-repair/mini-swe-scaffold.ts","../../src/trace-repair/action-budget.ts","../../src/trace-repair/continuation-records.ts","../../src/trace-repair/control-policy.ts","../../src/trace-repair/admission-records.ts","../../src/trace-repair/continuation-policy.ts","../../src/trace-repair/oracle-determinism.ts","../../src/trace-repair/admission.ts","../../src/trace-repair/admission-contract.ts","../../src/trace-repair/admission-report.ts","../../src/trace-repair/analyst-response.ts","../../src/trace-repair/blinding.ts","../../src/trace-repair/analyst-arm.ts","../../src/trace-repair/arm-completion.ts","../../src/trace-repair/arm-dspy.ts","../../src/trace-repair/degenerate-strategies.ts","../../src/trace-repair/ports.ts","../../src/trace-repair/funnel.ts","../../src/trace-repair/grade.ts","../../src/trace-repair/delta-repair.ts","../../src/trace-repair/docker-environment.ts","../../src/trace-repair/repair-prompt.ts","../../src/trace-repair/test-oracle.ts"],"mappings":";;;;;;;KA8IY;EACN;EAAgB;;EAChB;EAAsB;;;;;;;wBASZ,YAAY,2BAA2B;UAMtC;EACf;EACA;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;KCxIU;UAEK;;WAEN;;WAEA;;WAEA;;;;KAoUC;UAQK;EACf;EACA;EACA;;EAEA,SAAS;;EAET,UAAU;;KAGA;WACG;WAA2B,aAAa;;WAExC;WACA,WAAW;WACX;WACA,aAAa;;;KC/WhB;;UAGK;EACf;EACA;;;KAIU;UAOK;EACf;EACA;;EAEA;EACA;EACA;;UAGe;;EAEf;EACA;EACA;;EAEA,OAAO;;EAEP;EACA;EACA;;UAGe;;EAEf;EACA;;EAEA;;;;;;EAMA;;EAEA,WAAW;EACX,OAAO;;EAEP;;UAGe;;;;;EAKf;;EAEA;;EAEA;EACA;EACA;EACA;EACA;EACA;;UAGe;;EAEf;EACA;;UAGe;EACf;EACA,KAAK;;EAEL;;EAEA;;EAEA;;EAEA;EACA;EACA;EACA,aAAa;EACb,OAAO;EACP,YAAY;;EAEZ;EACA,OAAO;EACP,gBAAgB;EAChB;EACA;EACA;;EAEA;;;UAIe;EACf;EACA;;;UAIe;EACf;EACA;EACA,OAAO;EACP;;;wBAMc,qBAAqB,SAAS,sBAAsB;;;;;;;;;wBAiBpD,cAAc,SAAS;;qBCpI1B,iCAAiC;;;;;;;;;;;KAYlC;UAIK;;WAEN;;WAEA;;WAEA;;WAEA;;WAEA;;;;;;;;;;UAWM;WACN;WACA;;WAEA;;UAGM,sBAAsB,oBAAoB;;;WAGhD;;;;;;WAMA;;wBAYK,oBAAoB,OAAO,qBAAqB;;;;;;;;;wBAiDhD,wBACd,QAAQ,mBACR,WAAW;;;;;;;;UChHI;;EAEf;;EAEA;;EAEA;;;;;EAKA;;;;;;EAMA;;;;;;;;;wBAkBc,yBAAyB,eAAe;;;;;;;;;;;;KAyB5C;;;;;;;;KAwBA;KAoEA;;UAGK;;EAEf;EACA;;KAGU;EACN;EAAkB,SAAS;;EAE3B;EACA;EACA;EACA;;EAEA;EAAsB;EAAiB;EAAkB;;EACzD;EAA0B;EAAiB;;EAE3C;EACA,KAAK;EACL;EACA;EACA,YAAY;;;UAID;EACf,KAAK;EACL;EACA;EACA;EACA,YAAY;EACZ;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA,SAAS;;;;;EAKT;EACA,kBAAkB;;;EAGlB;EACA;;EAEA,YAAY;;EAEZ;;EAEA,QAAQ;EACR,UAAU;;UAKK;EACf,QAAQ;EACR;EACA;EACA;;UAGe;;EAEf,eAAe;EACf;EACA,QAAQ;EACR;;UAGe;EACf;EACA,SAAS;EACT,WAAW;EACX,cAAc,OAAO;;EAErB;;EAEA,sBAAsB;;;;;;;;;wBAUR,sBACd,mBAAmB,uBACnB,sBAAsB,qBACrB;;wBAsDa,sBAAsB,UAAU;UCxS/B;;WAEN;;WAEA;;WAEA;;WAEA;WACA;WACA;;WAEA;;WAEA;;WAEA;WACA;;UAqBM,sCACP,QAAQ,KAAK;;EAErB;;EAEA;;wBAGc,+BACd,OAAO,gCACN;;;;;;wBA+Ba,yBAAyB,QAAQ;;;;;wBAejC,iBAAiB,oBAAoB,eAAe;UAYnD;EACf;EACA,UAAU;EACV;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA,OAAO;;EAEP;EACA;;KAGU,qBACV,SAAS,6BACN,QAAQ;UAEI;EACf;EACA;;EAEA;;EAEA;;UAGe;;EAEf;EACA,YAAY,QAAQ;EACpB,KAAK,iBAAiB;IAAW;MAA2B,QAAQ;EACpE,WAAW;;UAGI;EACf;EACA,KAAK;EACL;;UAGe;EACf;;;;;;EAMA,OAAO,SAAS,iCAAiC,QAAQ;;UAG1C;EACf,QAAQ;EACR,KAAK;;EAEL;;;;;;EAMA,iBAAiB;;EAEjB;;;;;;;EAOA;EACA,OAAO;EACP,cAAc;;EAEd;;;;;;;;;wBAUoB,gBACpB,SAAS,yBACR,QAAQ;;;;;;wBAmOK,WAAW,gBAAgB,2BAA2B;;;;;wBA8CtD,UAAU,gBAAgB,2BAA2B;;;;;;wBAerD,kBAAkB,mBAAmB;;;KC1ezC;;KAGA;UAKK;;WAEN;WACA;;UAGM;;WAEN;;;;;;WAMA;WACA;WACA;;;WAGA,qBAAqB;;UAGf;WACN,OAAO;WACP,MAAM;;;;;WAKN,qBAAqB;;UAGf;WACN;;;;WAIA;;;;;;;WAOA;WACA,iBAAiB;WACjB;;;UAIM;;WAEN;WACA;WACA;WACA;;UAGM;WACN,OAAO;WACP;;WAEA;WACA;;WAEA;;WAEA;WACA,kBAAkB;;;WAGlB;;WAEA;;WAEA;;UAGM;WACN;WACA;WACA;;WAEA;WACA;;WAEA;WACA,kBAAkB;WAClB;;WAEA;;;;;;;;;KA4IC,qBAAqB,oBAAoB;;;;;;;;;UAUpC;WACN;WACA,uBAAuB;;wBAGlB,wBAAwB,oBAAoB;;wBA8B5C,0BAA0B,SAAS;;;;;KC3PvC,iBAAiB;EACvB;EAAiB,OAAO;;EACxB;EAAkB;;UAEP;;EAEf;EACA;EACA;;UAGe;;EAEf;;EAEA,4BAA4B;;UAGb;;EAEf;;EAEA,OAAO,KAAK,eAAe,QAAQ,iBAAiB;;UAGrC;EACf;;EAEA;;UAGe;EACf;;EAEA,MAAM,KAAK,eAAe,QAAQ,iBAAiB;;UAGpC;EACf,KAAK;EACL,KAAK;;EAEL;;EAEA,WAAW;;UAGI;;EAEf,OAAO;;EAEP,SAAS;;UAGM;EACf;;;;;EAKA,IAAI,SAAS,0BAA0B,QAAQ,iBAAiB;;UAKjD;;EAEf;;EAEA;;;;;;EAMA,uBAAuB;;EAEvB;;;;;EAKA,mBAAmB;;EAEnB;;UAGe;WACN;WACA;WACA,sBAAsB;WACtB;WACA,kBAAkB;WAClB;;qBAGE,2BAA2B;UA2EvB;EACf;EACA;EACA;EACA;EACA;EACA;EACA;;;EAGA;EACA,kBAAkB;;;EAGlB,gBAAgB,SAAS;;UAGV;EACf,QAAQ;EACR,YAAY;;EAEZ,eAAe;;EAEf,QAAQ,SAAS,OAAO;EACxB,OAAO;;EAEP,aAAa;;EAEb;EACA;;UAGe;EACf,eAAe;;EAEf,QAAQ;EACR,UAAU;EACV,QAAQ;EACR,UAAU;;;;;;;EAOV,aAAa;EACb,SAAS;;EAET;;;;;;;;;wBAUoB,aAAa,SAAS,sBAAsB,QAAQ;UAuVzD;EACf,QAAQ;;EAER,iBAAiB;;EAEjB;;EAEA;;;;;;;;;;;wBAYc,wBAAwB,OAAO;;UCtmB9B;;;;WAIN;;WAEA;;WAEA,kBAAkB;;UASZ;;WAEN;;WAEA;;UAGM;WACN;;WAEA;;;WAGA;;UAGM;WACN;;;WAGA;;WAEA;;WAEA;;WAEA;WACA,gBAAgB;;;;;;WAMhB,mBAAmB;;WAEnB,eAAe;WACf,gBAAgB;;WAEhB;;WAEA;WACA,cAAc;WACd,aAAa;;;KAIZ;cAWE;;;;;;;;;;UAWG;YACL;WACD;WACA;WACA;WACA;WACA;WACA,gBAAgB;WAChB,UAAU;WACV;WACA;;;WAGA,eAAe;WACf,kBAAkB;;;WAGlB;;;WAGA;WACA;;;;UAKM;WACN,eAAe;WACf,kBAAkB;WAClB;WACA;WACA;WACA;;KAGC;WAEG;WACA,WAAW;WACX,KAAK;;WAGL;WACA;WACA,WAAW;WACX,WAAW;WACX;;;;;;;;;;;;;;wBAeC,SACd,UAAU,mBACV,WAAU,oBACT;UC7Kc;EACf;EACA;EACA;;EAEA;EACA,QAAQ;EACR,YAAY;EACZ,OAAO;;EAEP,UAAU,OAAO;EACjB;IAAe;IAAc;;;EAE7B,eAAe;;;wBAID,kBAAkB,QAAQ,kBAAkB;UAoB3C;;EAEf;;;wBAIc,sBACd,UAAU,mBACV,UAAS;UC9CM;;WAEN,MAAM;;WAEN;;UAGM;WACN;;;WAGA;;;WAGA;WACA,cAAc;;UAGR;WACN;;KAGC,kBAAkB,gBAAgB;;KAGlC;KAQA;WACG;WAA0B,OAAO;;WACjC;WAA2B,SAAS;WAA+B;;;;;;;;;wBASlE,qBAAqB,gBAAgB;UC5CpC;WACN;WACA;WACA;;UAGM;WACN;;WAEA;WACA,gBAAgB;;;WAGhB;;WAEA;;UAGM;;;WAGN;;wBAGK,gBACd,KAAK,aACL,UAAS,yBACR;;;;;;;;;KCGS;;;;;;;;;;;;;;;;;;KAmBA;WACG;WAAuB;;WAEvB;;WAEA;;WAEA;;WAEA;;UAGE;WACN;;;WAGA;WACA,eAAe;;WAEf,QAAQ;;WAER;WACA,sBAAsB;;;;;WAKtB;;;;;UAMM;WACN,QAAQ;WACR,SAAS;;;;UAKH;WACN;WACA;WACA;WACA;;UAGM;WACN;;WAEA;;;UAIM;WACN;WACA;;;;;;;;;;KAWC;WAEG;WACA;WACA;WACA,cAAc;WACd;WACA;WACA,uBAAuB;WACvB,QAAQ;WACR,OAAO;;WAGP;WACA;WACA;WACA,uBAAuB;WACvB,QAAQ;WACR,OAAO;;WAGP;WACA;WACA;WACA,uBAAuB;WACvB,QAAQ;WACR,OAAO;;UAGL;WACN,aAAa;EACtB,IAAI,SAAS,mBAAmB,QAAQ;;;;;;;;;;KAW9B;WACG;WAA2B,aAAa;;WAExC;WACA,WAAW;WACX;WACA,aAAa;;UAGX;WACN;WACA;;WAEA;WACA,OAAO;;WAEP,QAAQ;WACR;;UAGM;WACN,KAAK;WACL,KAAK;;WAEL;WACA,SAAS;;WAET;;;;;;;;wBASW,aAAa,SAAS,sBAAsB,QAAQ;;;;;;;;wBAuC1D,kBAAkB,QAAQ,kBAAkB;;UAc3C;WACN;;WAEA,2BAA2B;;WAE3B,6BAA6B;WAC7B,eAAe;;;;WAIf;;UAGM;WACN;;;WAGA,4BAA4B;WAC5B,sBAAsB;;;WAGtB;WACA,QAAQ;WACR;;;WAGA;;;;;;;;;;;;;;;wBAgBK,qBACd,eAAe,oBACf;WAAoB,SAAS;IAC5B;UCpQc;WACN;;WAEA;;WAEA;WACA;;WAEA,WAAW;;WAEX;;WAEA,SAAS;WACT,SAAS;;WAET,sBAAsB;WACtB,eAAe;;wBAGV,0BAA0B,SAAS,6BAA6B;;qBCZnE;UAmCI;;WAEN,QAAQ;WACR,QAAQ;WACR,YAAY;WACZ;WACA;WACA,WAAW;WACX,SAAS;WACT;WACA,OAAO,iBAAiB,SAAS;;wBAG5B,oBAAoB,SAAS,uBAAuB;UAuJnD;;;;;WAKN;;;;;;;;;;;;;;;;KCzPC;UAEK;WACN;;WAEA;;WAEA;WACA,YAAY;;WAEZ;;qBAGE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KA2ED,+BAA+B;;;;;;KCjF/B;UAQK;EACf;EACA;EACA;;EAEA;;UAGe;;WAEN;EACT,KAAK,iBAAiB,oBAAoB,QAAQ;EAClD,SAAS;;UAGM;EACf;;;EAGA;EACA,KAAK;;EAEL;;UAGe;;EAEf,KAAK,SAAS,uBAAuB,QAAQ;;UAG9B;EACf;EACA,KAAK;EACL;;UAGe;;EAEf;EACA;EACA;;EAEA;EACA;;;;;;;;;;UAWe;EACf,MAAM,SAAS,eAAe,SAAS,oBAAoB,QAAQ;;UAGpD;EACf;EACA,KAAK;EACL;;EAEA,SAAS;;EAET,gBAAgB;;;EAGhB;;;;;;;EAOA,UAAU;EACV;;UAGe;EACf;EACA;EACA;EACA;;UAGe;;EAEf;;;EAGA;;EAEA;;EAEA;EACA;;;KAIU,4BACV,SAAS,8BACN,QAAQ;;KCtGD;WACG;WAA0B,QAAQ;WAA+B;;WAEjE;WACA,QAAQ;WACR;WACA,aAAa;;WAGb;WACA;WACA;;UAGE;;WAEN;;WAEA;WACA;;;;;;;;;KAUC;;;WAIG;WACA;;WAGA;WACA;WACA;WACA;WACA;WACA;WACA,QAAQ;;UAGN;WACN;WACA;WACA;WACA;WACA;WACA,QAAQ;;UAGF;WACN;WACA;WACA;;WAEA;;;;;;;;;;KAWC;WAEG;WACA;WACA;WACA,cAAc;WACd,OAAO;;WAGP;WACA;WACA;WACA;;UAGE;;WAEN;;WAEA;;WAEA;;;WAGA;WACA,0BAA0B;;;;;;;;KASzB;WACG;WAA8B,WAAW;;WACzC;;WAEA;WACA;WACA,cAAc;;WAGd;WACA;WACA,cAAc;WACd,WAAW;;WAGX;WACA;WACA,cAAc;WACd,WAAW;WACX,WAAW;WACX,QAAQ;;;;;;;;;;UAWN;WACN;WACA;;WAEA;;UA0CM;WACN;WACA;WACA;WACA;WACA;WACA;WACA;;WAEA;;WAEA;;UChKM;WACN,KAAK;WACL,UAAU;WACV,UAAU;WACV,QAAQ;WACR,cAAc;;;WAGd;WACA,SAAS;;WAET;;;;;;;;WAQA;WACA,cAAc;;;;;;;;;;UAWR,wBAAwB;WAC9B;WACA,OAAO;WACP,QAAQ;;;WAGR;WACA;WACA;;;WAGA,kBAAkB;WAClB;WACA;;WAEA;WACA;;wBAMW,eAAe,SAAS,qBAAqB,QAAQ;UCtF1D;;WAEN;WACA;WACA;;UAGM;WACN;WACA;WACA;WACA;WACA;WACA;WACA;;;WAGA;;KAGC;UAUK;WACN,IAAI;WACJ;;WAEA;;UAGM;WACN;WACA,QAAQ;;WAER;WACA;WACA,aAAa;;;;WAIb,cAAc;WACd;WACA,qBAAqB;WACrB,kBAAkB;;wBAGb,YACd,qBAAqB,mBACrB,UAAS,qBACR;;;wBAoKa,wBAAwB,QAAQ;UCnO/B;EACf;;EAEA;;UAGe;;EAEf;EACA;;EAEA;;;KAIU,iBAAiB,SAAS,mBAAmB,QAAQ;;;;;;qBAOpD,mBAAmB;UAuDf;;EAEf;;EAEA;;EAEA,MAAM;;EAEN;;EAEA;EACA,YAAY;;EAEZ;;wBAiBc,oCACd,SAAS,uCACR;;;;;;;;;wBCjBa,qBACd,SAAQ;;;;;;;;;wBAsBM,sBACd,QAAQ,oBACR;;qBC5GW,+BAA+B;UAK3B;;WAEN;WACA;;WAEA;;UAGM;;WAEN,gBAAgB;;WAEhB;;;WAGA;WACA;WACA;;wBAqBK,mBAAmB,SAAS,4BAA4B"}
|
|
@@ -1,17 +1,17 @@
|
|
|
1
|
-
import { c as errorMessage, n as CaptureIntegrityError, s as ValidationError } from "../errors-
|
|
2
|
-
import { i as compareCodeUnits } from "../canonical-
|
|
3
|
-
import { i as pairedBootstrap, t as mean } from "../mean-
|
|
4
|
-
import { n as contentHash } from "../verdict-cache-
|
|
5
|
-
import { n as isRecord } from "../json-
|
|
6
|
-
import { t as certificationEvidenceDigest } from "../verdict-
|
|
7
|
-
import { t as packageVersion } from "../package-version-
|
|
8
|
-
import { t as assertEqualDeclarativeTerms } from "../equal-terms-
|
|
9
|
-
import { c as runPrimeExchange, n as buildPrimePrompt, t as analystUsageReceiptFromPrimeUsage } from "../prime-protocol-
|
|
10
|
-
import { i as TIMEOUT_OBSERVATION_MARKER, l as isRecordedTimeout, n as FORMAT_ERROR_OBSERVATION_PREFIX, p as parseRecordedReturncode, r as SUBMIT_ACTION_SIGNATURE, s as deriveFailureSignature, t as wrapActionForExec } from "../exec-
|
|
1
|
+
import { c as errorMessage, n as CaptureIntegrityError, s as ValidationError } from "../errors-DeEEV9Vx.js";
|
|
2
|
+
import { i as compareCodeUnits } from "../canonical-CicJlINA.js";
|
|
3
|
+
import { i as pairedBootstrap, t as mean } from "../mean-CdDYG-us.js";
|
|
4
|
+
import { n as contentHash } from "../verdict-cache-BVH-fl18.js";
|
|
5
|
+
import { n as isRecord } from "../json-DtvZNAU9.js";
|
|
6
|
+
import { t as certificationEvidenceDigest } from "../verdict-C8TcdwfD.js";
|
|
7
|
+
import { t as packageVersion } from "../package-version-CJl_pY9W.js";
|
|
8
|
+
import { t as assertEqualDeclarativeTerms } from "../equal-terms-K11VHmR3.js";
|
|
9
|
+
import { c as runPrimeExchange, n as buildPrimePrompt, t as analystUsageReceiptFromPrimeUsage } from "../prime-protocol-CrhpXRz9.js";
|
|
10
|
+
import { i as TIMEOUT_OBSERVATION_MARKER, l as isRecordedTimeout, n as FORMAT_ERROR_OBSERVATION_PREFIX, p as parseRecordedReturncode, r as SUBMIT_ACTION_SIGNATURE, s as deriveFailureSignature, t as wrapActionForExec } from "../exec-C0_AkLUv.js";
|
|
11
11
|
import { createHash } from "node:crypto";
|
|
12
12
|
import { constants } from "node:os";
|
|
13
13
|
import { spawn } from "node:child_process";
|
|
14
|
-
|
|
14
|
+
// src/trace-repair/admission-records.ts
|
|
15
15
|
/**
|
|
16
16
|
* The vocabulary of the TB-Repair admission pre-pass: what a corpus row is,
|
|
17
17
|
* which failure population it belongs to, why it can leave the funnel, and the
|
|
@@ -156,8 +156,7 @@ function assertChainReconciles(artifact) {
|
|
|
156
156
|
const stratumInputs = artifact.byStratum.reduce((total, chain) => total + chain.input, 0);
|
|
157
157
|
if (stratumInputs + artifact.unstratified !== artifact.overall.input) throw new AdmissionDenominatorError(`stratum inputs ${stratumInputs} plus ${artifact.unstratified} unstratified do not cover ${artifact.overall.input} input rows`);
|
|
158
158
|
}
|
|
159
|
-
|
|
160
|
-
//#region src/trace-repair/mini-swe-scaffold.ts
|
|
159
|
+
// src/trace-repair/mini-swe-scaffold.ts
|
|
161
160
|
/**
|
|
162
161
|
* The mini-swe-agent scaffold as the Terminal-Bench-2 trajectory corpus
|
|
163
162
|
* recorded it: one bash block per turn, one observation per command, and a
|
|
@@ -342,8 +341,7 @@ function submissionOf(output) {
|
|
|
342
341
|
if (first === void 0 || first.trim() !== SUBMIT_SENTINEL) return null;
|
|
343
342
|
return lines.slice(1).join("");
|
|
344
343
|
}
|
|
345
|
-
|
|
346
|
-
//#region src/trace-repair/continuation-policy.ts
|
|
344
|
+
// src/trace-repair/continuation-policy.ts
|
|
347
345
|
/** A rollout ran outside the pinned policy, so its evidence cannot be used. */
|
|
348
346
|
var ContinuationPolicyViolationError = class extends CaptureIntegrityError {};
|
|
349
347
|
/** Two arms did not run the same policy, so their difference is not the intervention. */
|
|
@@ -707,8 +705,7 @@ function assertArmSymmetry(rollouts) {
|
|
|
707
705
|
if (seen.seed !== rollout.seed) throw new ContinuationSymmetryError(`paired rollouts ${key} drew different seeds: ${seen.arm}=${seen.seed}, ${rollout.arm}=${rollout.seed}`);
|
|
708
706
|
}
|
|
709
707
|
}
|
|
710
|
-
|
|
711
|
-
//#region src/trace-repair/control-policy.ts
|
|
708
|
+
// src/trace-repair/control-policy.ts
|
|
712
709
|
/**
|
|
713
710
|
* The control policy admission screens under, declared rather than assumed.
|
|
714
711
|
*
|
|
@@ -774,8 +771,7 @@ function assertControlCalibrated(policy, screening) {
|
|
|
774
771
|
if (screening === "enforced" && !canRescue) throw new UncalibratedControlError(`control policy ${policy.id} (digest ${policy.digest}) allows ${policy.stepBudget} model call(s) per rollout, so a control rollout executes no command and grades the same bytes the end-state check already graded as failing. Under it conditions 3 and 4 can only fire on a grader that disagrees with itself, so they screen nothing. Give the control a step budget of at least 1, or set controlScreening to 'declared-inert' and read a control pass as the oracle flip it is.`);
|
|
775
772
|
if (screening === "declared-inert" && canRescue) throw new UncalibratedControlError(`control policy ${policy.id} (digest ${policy.digest}) allows ${policy.stepBudget} model call(s) per rollout, so it can rescue a row, but the criteria declare it inert. Set controlScreening to 'enforced' so a control pass is recorded as a rescue.`);
|
|
776
773
|
}
|
|
777
|
-
|
|
778
|
-
//#region src/trace-repair/admission.ts
|
|
774
|
+
// src/trace-repair/admission.ts
|
|
779
775
|
const ADMISSION_CONFIG_DEFAULTS = Object.freeze({
|
|
780
776
|
maxPrefixDivergence: .1,
|
|
781
777
|
controlRollouts: 3,
|
|
@@ -1154,8 +1150,7 @@ function preview(rowIds) {
|
|
|
1154
1150
|
const shown = rowIds.slice(0, 5).join(", ");
|
|
1155
1151
|
return rowIds.length > 5 ? `${shown}, +${rowIds.length - 5} more` : shown;
|
|
1156
1152
|
}
|
|
1157
|
-
|
|
1158
|
-
//#region src/trace-repair/admission-contract.ts
|
|
1153
|
+
// src/trace-repair/admission-contract.ts
|
|
1159
1154
|
/**
|
|
1160
1155
|
* Admission: the executed checks a row must pass before an analyst is allowed
|
|
1161
1156
|
* to see it, and the brand that proves it did.
|
|
@@ -1274,8 +1269,7 @@ function assertCriteria(criteria) {
|
|
|
1274
1269
|
if (!(maxPrefixDivergenceRatio >= 0 && maxPrefixDivergenceRatio <= 1)) throw new ValidationError(`admission maxPrefixDivergenceRatio must be within [0,1], got ${maxPrefixDivergenceRatio}`);
|
|
1275
1270
|
if (!Number.isInteger(controlRollouts) || controlRollouts <= 0) throw new ValidationError(`admission controlRollouts must be a positive integer, got ${controlRollouts}`);
|
|
1276
1271
|
}
|
|
1277
|
-
|
|
1278
|
-
//#region src/trace-repair/admission-report.ts
|
|
1272
|
+
// src/trace-repair/admission-report.ts
|
|
1279
1273
|
/** Plain JSON. `JSON.stringify` of this object is the machine-readable artifact. */
|
|
1280
1274
|
function admissionArtifact(report) {
|
|
1281
1275
|
assertChainReconciles(report.chain);
|
|
@@ -1403,8 +1397,7 @@ function rate(part, total) {
|
|
|
1403
1397
|
function formatShare(share) {
|
|
1404
1398
|
return `${(share * 100).toFixed(2)}%`;
|
|
1405
1399
|
}
|
|
1406
|
-
|
|
1407
|
-
//#region src/trace-repair/action-budget.ts
|
|
1400
|
+
// src/trace-repair/action-budget.ts
|
|
1408
1401
|
/**
|
|
1409
1402
|
* The action budget an intervention must fit inside.
|
|
1410
1403
|
*
|
|
@@ -1765,8 +1758,7 @@ function assertBudget(budget) {
|
|
|
1765
1758
|
function normalizeActionForComparison(action) {
|
|
1766
1759
|
return action.split("\n").map((line) => line.trimEnd()).filter((line, index, lines) => line.length > 0 || index > 0 && index < lines.length - 1).join("\n").trim();
|
|
1767
1760
|
}
|
|
1768
|
-
|
|
1769
|
-
//#region src/trace-repair/blinding.ts
|
|
1761
|
+
// src/trace-repair/blinding.ts
|
|
1770
1762
|
/**
|
|
1771
1763
|
* What the analyst is shown.
|
|
1772
1764
|
*
|
|
@@ -1812,8 +1804,7 @@ Object.freeze([
|
|
|
1812
1804
|
"image",
|
|
1813
1805
|
"cwd"
|
|
1814
1806
|
]);
|
|
1815
|
-
|
|
1816
|
-
//#region src/trace-repair/repair-prompt.ts
|
|
1807
|
+
// src/trace-repair/repair-prompt.ts
|
|
1817
1808
|
/**
|
|
1818
1809
|
* The one question every repair arm answers.
|
|
1819
1810
|
*
|
|
@@ -1932,8 +1923,7 @@ function repairArmPromptSha256(budget, contract) {
|
|
|
1932
1923
|
contract
|
|
1933
1924
|
})).digest("hex");
|
|
1934
1925
|
}
|
|
1935
|
-
|
|
1936
|
-
//#region src/trace-repair/analyst-arm.ts
|
|
1926
|
+
// src/trace-repair/analyst-arm.ts
|
|
1937
1927
|
/**
|
|
1938
1928
|
* What an analyst arm is, and what every arm owes the comparison.
|
|
1939
1929
|
*
|
|
@@ -2074,8 +2064,7 @@ function assertReplyWithinDeclaration(declaration, reply) {
|
|
|
2074
2064
|
if (reply.status === "failed" && !reply.failure.trim()) throw new ValidationError(`arm '${declaration.id}' failed without stating why`);
|
|
2075
2065
|
if (reply.status === "finding" && !reply.intervention.action) throw new ValidationError(`arm '${declaration.id}' returned a finding with an empty intervention`);
|
|
2076
2066
|
}
|
|
2077
|
-
|
|
2078
|
-
//#region src/trace-repair/analyst-response.ts
|
|
2067
|
+
// src/trace-repair/analyst-response.ts
|
|
2079
2068
|
/** The literal an analyst returns when no single step carries the failure. */
|
|
2080
2069
|
const NO_DECISIVE_FAILURE = "no-decisive-failure";
|
|
2081
2070
|
/**
|
|
@@ -2221,8 +2210,7 @@ function topLevelObjectSlices(text) {
|
|
|
2221
2210
|
}
|
|
2222
2211
|
return slices;
|
|
2223
2212
|
}
|
|
2224
|
-
|
|
2225
|
-
//#region src/trace-repair/arm-completion.ts
|
|
2213
|
+
// src/trace-repair/arm-completion.ts
|
|
2226
2214
|
function createCompletionRepairArm(options) {
|
|
2227
2215
|
if (!Number.isSafeInteger(options.timeoutMs) || options.timeoutMs <= 0) throw new ValidationError(`repair arm '${options.id}' timeoutMs must be a positive safe integer, got ${options.timeoutMs}`);
|
|
2228
2216
|
const budget = options.budget ?? SCAFFOLD_INTERVENTION_BUDGET;
|
|
@@ -2361,8 +2349,7 @@ function describe$1(value) {
|
|
|
2361
2349
|
if (value === void 0) return "undefined";
|
|
2362
2350
|
return typeof value === "string" ? JSON.stringify(value) : String(value);
|
|
2363
2351
|
}
|
|
2364
|
-
|
|
2365
|
-
//#region src/trace-repair/arm-dspy.ts
|
|
2352
|
+
// src/trace-repair/arm-dspy.ts
|
|
2366
2353
|
/**
|
|
2367
2354
|
* Token the bridge reads to select the typed repair signature.
|
|
2368
2355
|
*
|
|
@@ -2669,8 +2656,7 @@ function describe(value) {
|
|
|
2669
2656
|
function assertDspyRepairEngine(engine) {
|
|
2670
2657
|
if (engine.id !== "dspy-rlm") throw new ValidationError(`the DSPy repair arm needs the dspy-rlm engine, got '${engine.id}'; another engine does not run the typed repair signature and would answer a different contract`);
|
|
2671
2658
|
}
|
|
2672
|
-
|
|
2673
|
-
//#region src/trace-repair/continuation-records.ts
|
|
2659
|
+
// src/trace-repair/continuation-records.ts
|
|
2674
2660
|
/** Corpus steps for the continuation alone, excluding the prefix it inherited. */
|
|
2675
2661
|
function rolloutRecordedSteps(rollout) {
|
|
2676
2662
|
return rollout.steps.map((step) => ({
|
|
@@ -2717,8 +2703,7 @@ function rolloutDigest(rollout) {
|
|
|
2717
2703
|
}))
|
|
2718
2704
|
});
|
|
2719
2705
|
}
|
|
2720
|
-
|
|
2721
|
-
//#region src/trace-repair/degenerate-strategies.ts
|
|
2706
|
+
// src/trace-repair/degenerate-strategies.ts
|
|
2722
2707
|
const DEGENERATE_STRATEGIES = [
|
|
2723
2708
|
{
|
|
2724
2709
|
id: "point-at-any-nonzero-exit-step",
|
|
@@ -2777,8 +2762,7 @@ const DEGENERATE_STRATEGIES = [
|
|
|
2777
2762
|
enforcedIn: "grade.ts (the intervention is applied at the named k only)"
|
|
2778
2763
|
}
|
|
2779
2764
|
];
|
|
2780
|
-
|
|
2781
|
-
//#region src/trace-repair/funnel.ts
|
|
2765
|
+
// src/trace-repair/funnel.ts
|
|
2782
2766
|
const ZERO_CREDIT = Object.freeze({
|
|
2783
2767
|
executes: 0,
|
|
2784
2768
|
localFlip: 0,
|
|
@@ -2833,8 +2817,7 @@ function countFunnel(grades) {
|
|
|
2833
2817
|
t4RepairFlipAll: t4All
|
|
2834
2818
|
};
|
|
2835
2819
|
}
|
|
2836
|
-
|
|
2837
|
-
//#region src/trace-repair/delta-repair.ts
|
|
2820
|
+
// src/trace-repair/delta-repair.ts
|
|
2838
2821
|
/**
|
|
2839
2822
|
* The headline metric.
|
|
2840
2823
|
*
|
|
@@ -3039,8 +3022,7 @@ function signed(value) {
|
|
|
3039
3022
|
const points = value * 100;
|
|
3040
3023
|
return `${points >= 0 ? "+" : ""}${points.toFixed(1)} pp`;
|
|
3041
3024
|
}
|
|
3042
|
-
|
|
3043
|
-
//#region src/trace-repair/docker-environment.ts
|
|
3025
|
+
// src/trace-repair/docker-environment.ts
|
|
3044
3026
|
/**
|
|
3045
3027
|
* Docker-backed continuation environment.
|
|
3046
3028
|
*
|
|
@@ -3187,8 +3169,7 @@ function createDockerContinuationEnvironment(options) {
|
|
|
3187
3169
|
}
|
|
3188
3170
|
};
|
|
3189
3171
|
}
|
|
3190
|
-
|
|
3191
|
-
//#region src/trace-repair/grade.ts
|
|
3172
|
+
// src/trace-repair/grade.ts
|
|
3192
3173
|
/**
|
|
3193
3174
|
* Grade one analyst answer about one admitted row.
|
|
3194
3175
|
*
|
|
@@ -3515,8 +3496,7 @@ function excerpt(text) {
|
|
|
3515
3496
|
const half = Math.floor(OUTPUT_EXCERPT_CHARS / 2);
|
|
3516
3497
|
return `${text.slice(0, half)}\n… [${text.length - OUTPUT_EXCERPT_CHARS} chars elided] …\n${text.slice(-2e3)}`;
|
|
3517
3498
|
}
|
|
3518
|
-
|
|
3519
|
-
//#region src/trace-repair/oracle-determinism.ts
|
|
3499
|
+
// src/trace-repair/oracle-determinism.ts
|
|
3520
3500
|
/**
|
|
3521
3501
|
* Whether a task's own grader is a function of the container state.
|
|
3522
3502
|
*
|
|
@@ -3658,8 +3638,7 @@ function assertDeterministicOracle(verdict) {
|
|
|
3658
3638
|
if (verdict.stable) return;
|
|
3659
3639
|
throw new NondeterministicOracleError(`task ${verdict.taskName} graded byte-identical state inconsistently: flip rate ${(verdict.flipRate * 100).toFixed(1)} % over ${verdict.replicates} replicates (${verdict.detail}). Its verdict is not a function of the state, so it cannot carry ground truth for an intervention study.`);
|
|
3660
3640
|
}
|
|
3661
|
-
|
|
3662
|
-
//#region src/trace-repair/test-oracle.ts
|
|
3641
|
+
// src/trace-repair/test-oracle.ts
|
|
3663
3642
|
/**
|
|
3664
3643
|
* The held-out suite, injected from outside the box at grade time.
|
|
3665
3644
|
*
|
|
@@ -3763,7 +3742,6 @@ function parentDirectory(path) {
|
|
|
3763
3742
|
function shellQuote(value) {
|
|
3764
3743
|
return `'${value.replaceAll("'", `'\\''`)}'`;
|
|
3765
3744
|
}
|
|
3766
|
-
//#endregion
|
|
3767
3745
|
export { ADMISSION_CONFIG_DEFAULTS, DEGENERATE_STRATEGIES, DSPY_REPAIR_SIGNATURE, TestSuiteTamperedError, UncalibratedControlError, admissionArtifact, admitRow, askRepairArm, assertAnalystIndependent, assertArmSymmetry, assertChainReconciles, assertControlCalibrated, assertDenominatorIntact, assertDeterministicOracle, blindTrajectory, buildDenominatorChain, continuationPolicyDigest, continuationSeed, createCompletionRepairArm, createDockerContinuationEnvironment, createDspyRepairArm, defineControlPolicy, definePinnedContinuationPolicy, deltaRepair, gradeRepairRow, injectedTestOracle, nodeProcessRunner, parseAction, parseAnalystResponse, parseTaskOracleRegistry, renderAdmissionReport, renderDeltaRepairReport, repairArmAsymmetries, repairArmPromptSha256, repairArmResponse, repairQuestionSha256, rolloutDigest, rolloutRecordedSteps, runAdmission, runContinuation, totalCost, totalUsage };
|
|
3768
3746
|
|
|
3769
3747
|
//# sourceMappingURL=index.js.map
|