openpond-sdk 0.12.2 → 0.12.4
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/dist/actions-local.js +2 -1
- package/dist/actions-local.js.map +1 -7
- package/dist/actions.js +2 -1
- package/dist/actions.js.map +1 -7
- package/dist/advanced-refiner-evaluations.js +13 -12
- package/dist/advanced-refiner-evaluations.js.map +1 -7
- package/dist/candidate-evaluations.js +9 -8
- package/dist/candidate-evaluations.js.map +1 -7
- package/dist/{chunk-MLFKQ7E7.js → chunk-24N6A6SF.js} +4 -4
- package/dist/chunk-24N6A6SF.js.map +1 -0
- package/dist/{chunk-YCQVVFE5.js → chunk-3B5KSH3V.js} +2 -2
- package/dist/chunk-3B5KSH3V.js.map +1 -0
- package/dist/{chunk-2DMYAJLS.js → chunk-3DWPLYFJ.js} +2 -2
- package/dist/chunk-3DWPLYFJ.js.map +1 -0
- package/dist/{chunk-B7VYE7YQ.js → chunk-53LLNDXF.js} +3 -3
- package/dist/chunk-53LLNDXF.js.map +1 -0
- package/dist/{chunk-VSRTJDVB.js → chunk-B5XTGEBE.js} +25 -13
- package/dist/chunk-B5XTGEBE.js.map +1 -0
- package/dist/{chunk-HKWX624G.js → chunk-BVMR5OFW.js} +5 -5
- package/dist/chunk-BVMR5OFW.js.map +1 -0
- package/dist/{chunk-3KXDH6OB.js → chunk-CL3IOOVL.js} +1 -1
- package/dist/chunk-CL3IOOVL.js.map +1 -0
- package/dist/chunk-CMWZZZGW.js +255 -0
- package/dist/chunk-CMWZZZGW.js.map +1 -0
- package/dist/{chunk-VOL3WES2.js → chunk-DF2QXVZ3.js} +14 -169
- package/dist/chunk-DF2QXVZ3.js.map +1 -0
- package/dist/{chunk-AG5DDL7T.js → chunk-DH6O4JFB.js} +2 -2
- package/dist/chunk-DH6O4JFB.js.map +1 -0
- package/dist/{chunk-7GYXC2AI.js → chunk-EFUT3XKW.js} +4 -4
- package/dist/chunk-EFUT3XKW.js.map +1 -0
- package/dist/{chunk-DBRRQASP.js → chunk-ETX25FTB.js} +3 -3
- package/dist/chunk-ETX25FTB.js.map +1 -0
- package/dist/{chunk-MEKWESBF.js → chunk-FBZTDDRC.js} +2 -2
- package/dist/chunk-FBZTDDRC.js.map +1 -0
- package/dist/{chunk-7AZDNZP6.js → chunk-GAIDBPFY.js} +353 -90
- package/dist/chunk-GAIDBPFY.js.map +1 -0
- package/dist/{chunk-YVUIHH6Z.js → chunk-GPQLBVQC.js} +1 -1
- package/dist/chunk-GPQLBVQC.js.map +1 -0
- package/dist/{chunk-KAE7AMYE.js → chunk-H3367ISL.js} +2 -2
- package/dist/chunk-H3367ISL.js.map +1 -0
- package/dist/{chunk-COPVBJPL.js → chunk-H3JZ3P5T.js} +2 -1
- package/dist/chunk-H3JZ3P5T.js.map +1 -0
- package/dist/{chunk-3LP7DNLQ.js → chunk-H4NTJFEE.js} +2 -2
- package/dist/chunk-H4NTJFEE.js.map +1 -0
- package/dist/{chunk-TCHAWJMB.js → chunk-HGXVOFCL.js} +2 -2
- package/dist/chunk-HGXVOFCL.js.map +1 -0
- package/dist/{chunk-CEZEVILU.js → chunk-HJ2AIAKX.js} +1 -1
- package/dist/chunk-HJ2AIAKX.js.map +1 -0
- package/dist/{chunk-ENLQCKY6.js → chunk-I54WQBXL.js} +8 -8
- package/dist/chunk-I54WQBXL.js.map +1 -0
- package/dist/{chunk-AV5JV26B.js → chunk-K5LWSWV7.js} +4 -4
- package/dist/chunk-K5LWSWV7.js.map +1 -0
- package/dist/{chunk-IWT6JUZD.js → chunk-L6EJX2VQ.js} +2 -2
- package/dist/chunk-L6EJX2VQ.js.map +1 -0
- package/dist/{chunk-WVISXXIL.js → chunk-LHX7ENKZ.js} +3 -3
- package/dist/chunk-LHX7ENKZ.js.map +1 -0
- package/dist/{chunk-WNH4DOEC.js → chunk-LR5CTOC5.js} +30 -299
- package/dist/chunk-LR5CTOC5.js.map +1 -0
- package/dist/chunk-MO4HAJZG.js +8709 -0
- package/dist/chunk-MO4HAJZG.js.map +1 -0
- package/dist/{chunk-T4APHDES.js → chunk-O6FKHAWN.js} +1 -1
- package/dist/chunk-O6FKHAWN.js.map +1 -0
- package/dist/{chunk-O5ARJVT5.js → chunk-OR3TMIJN.js} +5 -5
- package/dist/chunk-OR3TMIJN.js.map +1 -0
- package/dist/{chunk-P27DQ65Z.js → chunk-OVGDDKO2.js} +1 -1
- package/dist/chunk-OVGDDKO2.js.map +1 -0
- package/dist/{chunk-O4GE6VTR.js → chunk-PWW5QKZS.js} +2 -2
- package/dist/chunk-PWW5QKZS.js.map +1 -0
- package/dist/{chunk-DUKW3R7O.js → chunk-Q4TKI2MV.js} +2 -2
- package/dist/chunk-Q4TKI2MV.js.map +1 -0
- package/dist/{chunk-L5PQPAU6.js → chunk-QRS7Y3EM.js} +7 -3
- package/dist/chunk-QRS7Y3EM.js.map +1 -0
- package/dist/{chunk-YVFERCUC.js → chunk-RRKRA237.js} +83 -5
- package/dist/chunk-RRKRA237.js.map +1 -0
- package/dist/{chunk-2U54EOXJ.js → chunk-TK3DXH7A.js} +1 -1
- package/dist/chunk-TK3DXH7A.js.map +1 -0
- package/dist/{chunk-CK3ONRRC.js → chunk-U7XWIOSB.js} +3 -3
- package/dist/chunk-U7XWIOSB.js.map +1 -0
- package/dist/{chunk-DG5CRENU.js → chunk-VHGC44HA.js} +5 -5
- package/dist/chunk-VHGC44HA.js.map +1 -0
- package/dist/{chunk-7WUJBFCH.js → chunk-VN5V2E2S.js} +24 -5
- package/dist/chunk-VN5V2E2S.js.map +1 -0
- package/dist/{chunk-YSKUXQ2N.js → chunk-WMIKJGXW.js} +3 -3
- package/dist/chunk-WMIKJGXW.js.map +1 -0
- package/dist/{chunk-ROBAX5TN.js → chunk-WQFU635Q.js} +8 -6
- package/dist/chunk-WQFU635Q.js.map +1 -0
- package/dist/{chunk-D4WTE6ZA.js → chunk-XON75UFN.js} +3 -3
- package/dist/chunk-XON75UFN.js.map +1 -0
- package/dist/chunk-YLJ4XMA6.js +35 -0
- package/dist/chunk-YLJ4XMA6.js.map +1 -0
- package/dist/{chunk-K3JCP4FO.js → chunk-Z252BNAZ.js} +1 -1
- package/dist/chunk-Z252BNAZ.js.map +1 -0
- package/dist/{chunk-OEPYLZXK.js → chunk-ZHQK3MOM.js} +4 -4
- package/dist/chunk-ZHQK3MOM.js.map +1 -0
- package/dist/{chunk-6Q4I33CH.js → chunk-ZIFZFP5I.js} +5 -5
- package/dist/chunk-ZIFZFP5I.js.map +1 -0
- package/dist/connected-evidence.js +22 -21
- package/dist/connected-evidence.js.map +1 -7
- package/dist/dataset-commerce.js +2 -0
- package/dist/dataset-commerce.js.map +1 -7
- package/dist/dataset-marketplace.js +2 -0
- package/dist/dataset-marketplace.js.map +1 -7
- package/dist/dataset-workspaces.js +25 -23
- package/dist/dataset-workspaces.js.map +1 -7
- package/dist/experiment-evaluation-schedules.js +28 -27
- package/dist/experiment-evaluation-schedules.js.map +1 -7
- package/dist/experiment-improvements.js +12 -11
- package/dist/experiment-improvements.js.map +1 -7
- package/dist/experiments.js +31 -26
- package/dist/experiments.js.map +1 -7
- package/dist/human-review.js +3 -2
- package/dist/human-review.js.map +1 -7
- package/dist/index.js +98 -50
- package/dist/index.js.map +1 -7
- package/dist/learning.js +4 -3
- package/dist/learning.js.map +1 -7
- package/dist/model-batch-review.js +2 -1
- package/dist/model-batch-review.js.map +1 -7
- package/dist/model-projects.js +3 -2
- package/dist/model-projects.js.map +1 -7
- package/dist/model-starter-attempts.js +5 -4
- package/dist/model-starter-attempts.js.map +1 -7
- package/dist/model-starter-catalog.js +15 -14
- package/dist/model-starter-catalog.js.map +1 -7
- package/dist/model-starters.js +15 -14
- package/dist/model-starters.js.map +1 -7
- package/dist/model-taskset-authoring.js +2 -1
- package/dist/model-taskset-authoring.js.map +1 -7
- package/dist/model-taskset-runs.js +6 -5
- package/dist/model-taskset-runs.js.map +1 -7
- package/dist/post-training.js +9 -8
- package/dist/post-training.js.map +1 -7
- package/dist/profile-actions.js +3 -2
- package/dist/profile-actions.js.map +1 -7
- package/dist/project-actions.js +3 -2
- package/dist/project-actions.js.map +1 -7
- package/dist/refiner.js +3 -2
- package/dist/refiner.js.map +1 -7
- package/dist/taskset-catalog.js +2 -1
- package/dist/taskset-catalog.js.map +1 -7
- package/dist/taskset-drafts.js +27 -26
- package/dist/taskset-drafts.js.map +1 -7
- package/dist/taskset-packages.js +29 -26
- package/dist/taskset-packages.js.map +1 -7
- package/dist/training-bundle.js +195 -34
- package/dist/training-bundle.js.map +1 -7
- package/dist/training-projects.js +2 -1
- package/dist/training-projects.js.map +1 -7
- package/dist/training.js +17 -10
- package/dist/training.js.map +1 -7
- package/dist/types/packages/cloud/src/api/core.d.ts.map +1 -1
- package/dist/types/packages/cloud/src/sandbox/client-handles.d.ts.map +1 -1
- package/dist/types/packages/cloud/src/sandbox/client.d.ts.map +1 -1
- package/dist/types/packages/cloud/src/sandbox/sandbox-instance-client.d.ts.map +1 -1
- package/dist/types/packages/cloud/src/sandbox/types/runtime-profiles.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/advanced-refiner-evaluations.d.ts +487 -487
- package/dist/types/packages/sdk/src/advanced-refiner-evaluations.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/candidate-evaluation-client.d.ts +381 -348
- package/dist/types/packages/sdk/src/candidate-evaluation-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/candidate-evaluations.d.ts +86 -54
- package/dist/types/packages/sdk/src/candidate-evaluations.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/connected-collection-status.d.ts +5 -5
- package/dist/types/packages/sdk/src/connected-dataset-source.d.ts +10 -10
- package/dist/types/packages/sdk/src/connected-evidence-client.d.ts +464 -460
- package/dist/types/packages/sdk/src/connected-evidence-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/connected-evidence-contracts.d.ts +23 -23
- package/dist/types/packages/sdk/src/connected-recorded-execution.d.ts +77 -74
- package/dist/types/packages/sdk/src/connected-recorded-execution.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/connected-sync.d.ts +30 -30
- package/dist/types/packages/sdk/src/connected-sync.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-commerce-contracts.d.ts +14 -14
- package/dist/types/packages/sdk/src/dataset-commerce-files.d.ts +5 -5
- package/dist/types/packages/sdk/src/dataset-commerce-files.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-commerce-responses.d.ts +82 -82
- package/dist/types/packages/sdk/src/dataset-commerce.d.ts +58 -58
- package/dist/types/packages/sdk/src/dataset-commerce.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-marketplace.d.ts +4 -4
- package/dist/types/packages/sdk/src/dataset-marketplace.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-population.d.ts +10 -10
- package/dist/types/packages/sdk/src/dataset-population.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-preparation-authoring-contracts.d.ts +172 -172
- package/dist/types/packages/sdk/src/dataset-preparation-client.d.ts +47 -47
- package/dist/types/packages/sdk/src/dataset-preparation-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-preparation-contracts.d.ts +90 -90
- package/dist/types/packages/sdk/src/dataset-preparation-qualification-contracts.d.ts +7 -7
- package/dist/types/packages/sdk/src/dataset-preparation-research-contracts.d.ts +26 -26
- package/dist/types/packages/sdk/src/dataset-preparation-starting-state.d.ts +8 -8
- package/dist/types/packages/sdk/src/dataset-preparation-workspace.d.ts +111 -110
- package/dist/types/packages/sdk/src/dataset-preparation-workspace.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts +896 -889
- package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-workspace-contracts.d.ts +225 -223
- package/dist/types/packages/sdk/src/dataset-workspace-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-workspace-operations.d.ts +226 -224
- package/dist/types/packages/sdk/src/dataset-workspace-operations.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/evaluation-coordination-http.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/evaluation-project-context.d.ts +1 -1
- package/dist/types/packages/sdk/src/experiment-client.d.ts +2090 -1969
- package/dist/types/packages/sdk/src/experiment-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-contracts.d.ts +177 -113
- package/dist/types/packages/sdk/src/experiment-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-evaluation-schedules.d.ts +597 -505
- package/dist/types/packages/sdk/src/experiment-evaluation-schedules.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-execution-selection.d.ts +5 -5
- package/dist/types/packages/sdk/src/experiment-execution-selection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-field-mappings.d.ts +8 -8
- package/dist/types/packages/sdk/src/experiment-grading-contracts.d.ts +16 -16
- package/dist/types/packages/sdk/src/experiment-improvement-client.d.ts +2524 -2524
- package/dist/types/packages/sdk/src/experiment-improvement-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-improvement-contracts.d.ts +1557 -1557
- package/dist/types/packages/sdk/src/experiment-improvement-hosted-adoption.d.ts +150 -150
- package/dist/types/packages/sdk/src/experiment-inspection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-run-contracts.d.ts +41 -25
- package/dist/types/packages/sdk/src/experiment-run-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-run-details.d.ts +315 -240
- package/dist/types/packages/sdk/src/experiment-run-details.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-scoring-contracts.d.ts +21 -21
- package/dist/types/packages/sdk/src/grader-inspection.d.ts +104 -104
- package/dist/types/packages/sdk/src/grader-inspection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/harness-experiment-policy.d.ts +50 -20
- package/dist/types/packages/sdk/src/harness-experiment-policy.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/human-review-client.d.ts +288 -288
- package/dist/types/packages/sdk/src/human-review-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/index.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/learning-batch-assets.d.ts +15 -13
- package/dist/types/packages/sdk/src/learning-batch-assets.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/learning-client.d.ts +5 -5
- package/dist/types/packages/sdk/src/learning-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/learning-credentials.d.ts +4 -4
- package/dist/types/packages/sdk/src/learning-credentials.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/managed-training-preparation.d.ts +71 -69
- package/dist/types/packages/sdk/src/managed-training-preparation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-batch-review-contracts.d.ts +65 -61
- package/dist/types/packages/sdk/src/model-batch-review-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-batch-review-preparation.d.ts +74 -72
- package/dist/types/packages/sdk/src/model-batch-review-preparation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-batch-review.d.ts +96 -92
- package/dist/types/packages/sdk/src/model-batch-review.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-learning-intent.d.ts +1 -1
- package/dist/types/packages/sdk/src/model-learning-policy.d.ts +247 -247
- package/dist/types/packages/sdk/src/model-learning-policy.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-projects.d.ts +324 -324
- package/dist/types/packages/sdk/src/model-projects.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-attempts.d.ts +187 -187
- package/dist/types/packages/sdk/src/model-starter-attempts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-catalog.d.ts +181 -179
- package/dist/types/packages/sdk/src/model-starter-catalog.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-execution.d.ts +40 -40
- package/dist/types/packages/sdk/src/model-starters.d.ts +291 -287
- package/dist/types/packages/sdk/src/model-starters.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-authoring-contracts.d.ts +3 -3
- package/dist/types/packages/sdk/src/model-taskset-derivation.d.ts +114 -112
- package/dist/types/packages/sdk/src/model-taskset-derivation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-resources.d.ts +18 -18
- package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts +191 -127
- package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-runs.d.ts +868 -824
- package/dist/types/packages/sdk/src/model-taskset-runs.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/post-training.d.ts +448 -372
- package/dist/types/packages/sdk/src/post-training.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/profile-actions.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/profile-evaluation-discovery.d.ts +150 -150
- package/dist/types/packages/sdk/src/profile-evaluation-discovery.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/profile-external-dataset-package.d.ts +254 -251
- package/dist/types/packages/sdk/src/profile-external-dataset-package.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/project-actions.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/protocol.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/run-diagnostics.d.ts +3 -3
- package/dist/types/packages/sdk/src/task-design-contracts.d.ts +50 -50
- package/dist/types/packages/sdk/src/task-inventory.d.ts +4 -4
- package/dist/types/packages/sdk/src/task-inventory.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts +190 -189
- package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-authored-files.d.ts +149 -148
- package/dist/types/packages/sdk/src/taskset-authored-files.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-authored-portable-release.d.ts +15 -6
- package/dist/types/packages/sdk/src/taskset-authored-portable-release.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-calibration-fixtures.d.ts +64 -63
- package/dist/types/packages/sdk/src/taskset-calibration-fixtures.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-catalog.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts +245 -243
- package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-authoring.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-client.d.ts +744 -738
- package/dist/types/packages/sdk/src/taskset-draft-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-core.d.ts +133 -131
- package/dist/types/packages/sdk/src/taskset-draft-core.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-dataset-artifacts.d.ts +80 -80
- package/dist/types/packages/sdk/src/taskset-draft-dataset-sources.d.ts +34 -34
- package/dist/types/packages/sdk/src/taskset-draft-document.d.ts +119 -117
- package/dist/types/packages/sdk/src/taskset-draft-document.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts +500 -492
- package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-source.d.ts +236 -234
- package/dist/types/packages/sdk/src/taskset-draft-source.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts +228 -224
- package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-intake.d.ts +1 -1
- package/dist/types/packages/sdk/src/taskset-package-client.d.ts +4589 -4577
- package/dist/types/packages/sdk/src/taskset-package-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-contracts.d.ts +2068 -2054
- package/dist/types/packages/sdk/src/taskset-package-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-files.d.ts +1 -1
- package/dist/types/packages/sdk/src/taskset-package-grader-projection.d.ts +21 -21
- package/dist/types/packages/sdk/src/taskset-package-learning.d.ts +88 -86
- package/dist/types/packages/sdk/src/taskset-package-learning.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-projection.d.ts +387 -383
- package/dist/types/packages/sdk/src/taskset-package-projection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-acceptance.d.ts +26 -26
- package/dist/types/packages/sdk/src/training-acceptance.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-bundle-contracts.d.ts +18 -10
- package/dist/types/packages/sdk/src/training-bundle-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-bundle.d.ts +6 -0
- package/dist/types/packages/sdk/src/training-bundle.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-candidate-decisions.d.ts +2 -2
- package/dist/types/packages/sdk/src/training-candidate-decisions.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-evaluation-results.d.ts +3 -3
- package/dist/types/packages/sdk/src/training-evaluation-source.d.ts +1 -1
- package/dist/types/packages/sdk/src/training-evaluation-source.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-grading-plan.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-harness-sources.d.ts +14 -14
- package/dist/types/packages/sdk/src/training-harness-sources.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-learning-batch.d.ts +344 -340
- package/dist/types/packages/sdk/src/training-learning-batch.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-preparations.d.ts +700 -700
- package/dist/types/packages/sdk/src/training-preparations.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-project-client.d.ts +272 -272
- package/dist/types/packages/sdk/src/training-project-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-project-contracts.d.ts +161 -161
- package/dist/types/packages/sdk/src/training-qualification-preparation.d.ts +368 -0
- package/dist/types/packages/sdk/src/training-qualification-preparation.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/training.d.ts +1574 -1387
- package/dist/types/packages/sdk/src/training.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/work.d.ts +4 -0
- package/dist/types/packages/sdk/src/work.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/workflows.d.ts.map +1 -1
- package/dist/workflows.js +3 -2
- package/dist/workflows.js.map +1 -7
- package/package.json +3 -3
- package/dist/chunk-2DMYAJLS.js.map +0 -7
- package/dist/chunk-2U54EOXJ.js.map +0 -7
- package/dist/chunk-3KXDH6OB.js.map +0 -7
- package/dist/chunk-3LP7DNLQ.js.map +0 -7
- package/dist/chunk-6Q4I33CH.js.map +0 -7
- package/dist/chunk-7AZDNZP6.js.map +0 -7
- package/dist/chunk-7GYXC2AI.js.map +0 -7
- package/dist/chunk-7WUJBFCH.js.map +0 -7
- package/dist/chunk-AG5DDL7T.js.map +0 -7
- package/dist/chunk-AV5JV26B.js.map +0 -7
- package/dist/chunk-B7VYE7YQ.js.map +0 -7
- package/dist/chunk-CEZEVILU.js.map +0 -7
- package/dist/chunk-CK3ONRRC.js.map +0 -7
- package/dist/chunk-COPVBJPL.js.map +0 -7
- package/dist/chunk-D4WTE6ZA.js.map +0 -7
- package/dist/chunk-DBRRQASP.js.map +0 -7
- package/dist/chunk-DG5CRENU.js.map +0 -7
- package/dist/chunk-DUKW3R7O.js.map +0 -7
- package/dist/chunk-ENLQCKY6.js.map +0 -7
- package/dist/chunk-FHCSDXIA.js +0 -1417
- package/dist/chunk-FHCSDXIA.js.map +0 -7
- package/dist/chunk-HKWX624G.js.map +0 -7
- package/dist/chunk-IWT6JUZD.js.map +0 -7
- package/dist/chunk-K3JCP4FO.js.map +0 -7
- package/dist/chunk-KAE7AMYE.js.map +0 -7
- package/dist/chunk-L5PQPAU6.js.map +0 -7
- package/dist/chunk-MEKWESBF.js.map +0 -7
- package/dist/chunk-MLFKQ7E7.js.map +0 -7
- package/dist/chunk-O4GE6VTR.js.map +0 -7
- package/dist/chunk-O5ARJVT5.js.map +0 -7
- package/dist/chunk-OEPYLZXK.js.map +0 -7
- package/dist/chunk-P27DQ65Z.js.map +0 -7
- package/dist/chunk-PGBFI5HC.js +0 -98
- package/dist/chunk-PGBFI5HC.js.map +0 -7
- package/dist/chunk-ROBAX5TN.js.map +0 -7
- package/dist/chunk-T4APHDES.js.map +0 -7
- package/dist/chunk-TCHAWJMB.js.map +0 -7
- package/dist/chunk-VOL3WES2.js.map +0 -7
- package/dist/chunk-VSRTJDVB.js.map +0 -7
- package/dist/chunk-WNH4DOEC.js.map +0 -7
- package/dist/chunk-WVISXXIL.js.map +0 -7
- package/dist/chunk-YCQVVFE5.js.map +0 -7
- package/dist/chunk-YSKUXQ2N.js.map +0 -7
- package/dist/chunk-YVFERCUC.js.map +0 -7
- package/dist/chunk-YVUIHH6Z.js.map +0 -7
|
@@ -79,9 +79,9 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
79
79
|
}, z.core.$strict>>;
|
|
80
80
|
messages: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
81
81
|
role: z.ZodEnum<{
|
|
82
|
+
assistant: "assistant";
|
|
82
83
|
system: "system";
|
|
83
84
|
user: "user";
|
|
84
|
-
assistant: "assistant";
|
|
85
85
|
}>;
|
|
86
86
|
content: z.ZodString;
|
|
87
87
|
}, z.core.$strict>>>;
|
|
@@ -93,6 +93,22 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
93
93
|
}, z.core.$strict>, z.ZodObject<{
|
|
94
94
|
kind: z.ZodLiteral<"hosted_harness">;
|
|
95
95
|
modelId: z.ZodString;
|
|
96
|
+
maxOutputTokens: z.ZodOptional<z.ZodNumber>;
|
|
97
|
+
temperature: z.ZodOptional<z.ZodNumber>;
|
|
98
|
+
topP: z.ZodOptional<z.ZodNumber>;
|
|
99
|
+
localRuntime: z.ZodOptional<z.ZodObject<{
|
|
100
|
+
providerId: z.ZodLiteral<"codex">;
|
|
101
|
+
configurationHash: z.ZodString;
|
|
102
|
+
accountHash: z.ZodString;
|
|
103
|
+
reasoningEffort: z.ZodEnum<{
|
|
104
|
+
high: "high";
|
|
105
|
+
low: "low";
|
|
106
|
+
medium: "medium";
|
|
107
|
+
xhigh: "xhigh";
|
|
108
|
+
}>;
|
|
109
|
+
maximumRequestCostUsd: z.ZodNumber;
|
|
110
|
+
requestTimeoutMs: z.ZodNumber;
|
|
111
|
+
}, z.core.$strict>>;
|
|
96
112
|
candidate: z.ZodOptional<z.ZodObject<{
|
|
97
113
|
artifactId: z.ZodString;
|
|
98
114
|
contentHash: z.ZodString;
|
|
@@ -114,10 +130,10 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
114
130
|
reasoningEffort: z.ZodOptional<z.ZodEnum<{
|
|
115
131
|
high: "high";
|
|
116
132
|
low: "low";
|
|
133
|
+
max: "max";
|
|
117
134
|
medium: "medium";
|
|
118
|
-
xhigh: "xhigh";
|
|
119
135
|
off: "off";
|
|
120
|
-
|
|
136
|
+
xhigh: "xhigh";
|
|
121
137
|
}>>;
|
|
122
138
|
profileRepositoryId: z.ZodString;
|
|
123
139
|
source: z.ZodObject<{
|
|
@@ -195,10 +211,10 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
195
211
|
}, z.core.$strict>;
|
|
196
212
|
}, z.core.$strict>>;
|
|
197
213
|
split: z.ZodEnum<{
|
|
214
|
+
frozen_eval: "frozen_eval";
|
|
215
|
+
test: "test";
|
|
198
216
|
train: "train";
|
|
199
217
|
validation: "validation";
|
|
200
|
-
test: "test";
|
|
201
|
-
frozen_eval: "frozen_eval";
|
|
202
218
|
}>;
|
|
203
219
|
population: z.ZodArray<z.ZodObject<{
|
|
204
220
|
taskId: z.ZodString;
|
|
@@ -215,8 +231,8 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
215
231
|
configurationHash: z.ZodString;
|
|
216
232
|
output: z.ZodEnum<{
|
|
217
233
|
boolean: "boolean";
|
|
218
|
-
score: "score";
|
|
219
234
|
category: "category";
|
|
235
|
+
score: "score";
|
|
220
236
|
}>;
|
|
221
237
|
categories: z.ZodArray<z.ZodString>;
|
|
222
238
|
}, z.core.$strict>>;
|
|
@@ -225,17 +241,17 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
225
241
|
graderId: z.ZodString;
|
|
226
242
|
fields: z.ZodArray<z.ZodObject<{
|
|
227
243
|
destination: z.ZodEnum<{
|
|
228
|
-
output: "output";
|
|
229
|
-
input: "input";
|
|
230
|
-
expectedOutput: "expectedOutput";
|
|
231
244
|
evaluatorContext: "evaluatorContext";
|
|
245
|
+
expectedOutput: "expectedOutput";
|
|
246
|
+
input: "input";
|
|
247
|
+
output: "output";
|
|
232
248
|
}>;
|
|
233
249
|
path: z.ZodArray<z.ZodString>;
|
|
234
250
|
source: z.ZodEnum<{
|
|
235
|
-
output: "output";
|
|
236
|
-
input: "input";
|
|
237
|
-
expectedOutput: "expectedOutput";
|
|
238
251
|
evaluatorContext: "evaluatorContext";
|
|
252
|
+
expectedOutput: "expectedOutput";
|
|
253
|
+
input: "input";
|
|
254
|
+
output: "output";
|
|
239
255
|
}>;
|
|
240
256
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
241
257
|
}, z.core.$strict>>;
|
|
@@ -313,10 +329,10 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
313
329
|
}, z.core.$strict>;
|
|
314
330
|
}, z.core.$strict>>;
|
|
315
331
|
split: z.ZodEnum<{
|
|
332
|
+
frozen_eval: "frozen_eval";
|
|
333
|
+
test: "test";
|
|
316
334
|
train: "train";
|
|
317
335
|
validation: "validation";
|
|
318
|
-
test: "test";
|
|
319
|
-
frozen_eval: "frozen_eval";
|
|
320
336
|
}>;
|
|
321
337
|
population: z.ZodArray<z.ZodObject<{
|
|
322
338
|
taskId: z.ZodString;
|
|
@@ -333,8 +349,8 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
333
349
|
configurationHash: z.ZodString;
|
|
334
350
|
output: z.ZodEnum<{
|
|
335
351
|
boolean: "boolean";
|
|
336
|
-
score: "score";
|
|
337
352
|
category: "category";
|
|
353
|
+
score: "score";
|
|
338
354
|
}>;
|
|
339
355
|
categories: z.ZodArray<z.ZodString>;
|
|
340
356
|
}, z.core.$strict>>;
|
|
@@ -343,17 +359,17 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
343
359
|
graderId: z.ZodString;
|
|
344
360
|
fields: z.ZodArray<z.ZodObject<{
|
|
345
361
|
destination: z.ZodEnum<{
|
|
346
|
-
output: "output";
|
|
347
|
-
input: "input";
|
|
348
|
-
expectedOutput: "expectedOutput";
|
|
349
362
|
evaluatorContext: "evaluatorContext";
|
|
363
|
+
expectedOutput: "expectedOutput";
|
|
364
|
+
input: "input";
|
|
365
|
+
output: "output";
|
|
350
366
|
}>;
|
|
351
367
|
path: z.ZodArray<z.ZodString>;
|
|
352
368
|
source: z.ZodEnum<{
|
|
353
|
-
output: "output";
|
|
354
|
-
input: "input";
|
|
355
|
-
expectedOutput: "expectedOutput";
|
|
356
369
|
evaluatorContext: "evaluatorContext";
|
|
370
|
+
expectedOutput: "expectedOutput";
|
|
371
|
+
input: "input";
|
|
372
|
+
output: "output";
|
|
357
373
|
}>;
|
|
358
374
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
359
375
|
}, z.core.$strict>>;
|
|
@@ -381,17 +397,17 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
381
397
|
contentHash: z.ZodString;
|
|
382
398
|
mappings: z.ZodArray<z.ZodObject<{
|
|
383
399
|
destination: z.ZodEnum<{
|
|
400
|
+
evaluatorContext: "evaluatorContext";
|
|
401
|
+
expectedOutput: "expectedOutput";
|
|
384
402
|
input: "input";
|
|
385
403
|
output: "output";
|
|
386
|
-
expectedOutput: "expectedOutput";
|
|
387
|
-
evaluatorContext: "evaluatorContext";
|
|
388
404
|
}>;
|
|
389
405
|
path: z.ZodArray<z.ZodString>;
|
|
390
406
|
source: z.ZodEnum<{
|
|
407
|
+
evaluatorContext: "evaluatorContext";
|
|
408
|
+
expectedOutput: "expectedOutput";
|
|
391
409
|
input: "input";
|
|
392
410
|
output: "output";
|
|
393
|
-
expectedOutput: "expectedOutput";
|
|
394
|
-
evaluatorContext: "evaluatorContext";
|
|
395
411
|
}>;
|
|
396
412
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
397
413
|
}, z.core.$strict>>;
|
|
@@ -420,12 +436,12 @@ export declare const ExperimentEvaluationFireSchema: z.ZodObject<{
|
|
|
420
436
|
operationId: z.ZodString;
|
|
421
437
|
state: z.ZodEnum<{
|
|
422
438
|
cancelled: "cancelled";
|
|
423
|
-
|
|
424
|
-
running: "running";
|
|
439
|
+
cleaning: "cleaning";
|
|
425
440
|
completed: "completed";
|
|
426
441
|
dispatching: "dispatching";
|
|
427
|
-
|
|
442
|
+
failed: "failed";
|
|
428
443
|
reserved: "reserved";
|
|
444
|
+
running: "running";
|
|
429
445
|
uncertain: "uncertain";
|
|
430
446
|
}>;
|
|
431
447
|
executionId: z.ZodNullable<z.ZodString>;
|
|
@@ -515,9 +531,9 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
515
531
|
}, z.core.$strict>>;
|
|
516
532
|
messages: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
517
533
|
role: z.ZodEnum<{
|
|
534
|
+
assistant: "assistant";
|
|
518
535
|
system: "system";
|
|
519
536
|
user: "user";
|
|
520
|
-
assistant: "assistant";
|
|
521
537
|
}>;
|
|
522
538
|
content: z.ZodString;
|
|
523
539
|
}, z.core.$strict>>>;
|
|
@@ -529,6 +545,22 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
529
545
|
}, z.core.$strict>, z.ZodObject<{
|
|
530
546
|
kind: z.ZodLiteral<"hosted_harness">;
|
|
531
547
|
modelId: z.ZodString;
|
|
548
|
+
maxOutputTokens: z.ZodOptional<z.ZodNumber>;
|
|
549
|
+
temperature: z.ZodOptional<z.ZodNumber>;
|
|
550
|
+
topP: z.ZodOptional<z.ZodNumber>;
|
|
551
|
+
localRuntime: z.ZodOptional<z.ZodObject<{
|
|
552
|
+
providerId: z.ZodLiteral<"codex">;
|
|
553
|
+
configurationHash: z.ZodString;
|
|
554
|
+
accountHash: z.ZodString;
|
|
555
|
+
reasoningEffort: z.ZodEnum<{
|
|
556
|
+
high: "high";
|
|
557
|
+
low: "low";
|
|
558
|
+
medium: "medium";
|
|
559
|
+
xhigh: "xhigh";
|
|
560
|
+
}>;
|
|
561
|
+
maximumRequestCostUsd: z.ZodNumber;
|
|
562
|
+
requestTimeoutMs: z.ZodNumber;
|
|
563
|
+
}, z.core.$strict>>;
|
|
532
564
|
candidate: z.ZodOptional<z.ZodObject<{
|
|
533
565
|
artifactId: z.ZodString;
|
|
534
566
|
contentHash: z.ZodString;
|
|
@@ -550,10 +582,10 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
550
582
|
reasoningEffort: z.ZodOptional<z.ZodEnum<{
|
|
551
583
|
high: "high";
|
|
552
584
|
low: "low";
|
|
585
|
+
max: "max";
|
|
553
586
|
medium: "medium";
|
|
554
|
-
xhigh: "xhigh";
|
|
555
587
|
off: "off";
|
|
556
|
-
|
|
588
|
+
xhigh: "xhigh";
|
|
557
589
|
}>>;
|
|
558
590
|
profileRepositoryId: z.ZodString;
|
|
559
591
|
source: z.ZodObject<{
|
|
@@ -631,10 +663,10 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
631
663
|
}, z.core.$strict>;
|
|
632
664
|
}, z.core.$strict>>;
|
|
633
665
|
split: z.ZodEnum<{
|
|
666
|
+
frozen_eval: "frozen_eval";
|
|
667
|
+
test: "test";
|
|
634
668
|
train: "train";
|
|
635
669
|
validation: "validation";
|
|
636
|
-
test: "test";
|
|
637
|
-
frozen_eval: "frozen_eval";
|
|
638
670
|
}>;
|
|
639
671
|
population: z.ZodArray<z.ZodObject<{
|
|
640
672
|
taskId: z.ZodString;
|
|
@@ -651,8 +683,8 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
651
683
|
configurationHash: z.ZodString;
|
|
652
684
|
output: z.ZodEnum<{
|
|
653
685
|
boolean: "boolean";
|
|
654
|
-
score: "score";
|
|
655
686
|
category: "category";
|
|
687
|
+
score: "score";
|
|
656
688
|
}>;
|
|
657
689
|
categories: z.ZodArray<z.ZodString>;
|
|
658
690
|
}, z.core.$strict>>;
|
|
@@ -661,17 +693,17 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
661
693
|
graderId: z.ZodString;
|
|
662
694
|
fields: z.ZodArray<z.ZodObject<{
|
|
663
695
|
destination: z.ZodEnum<{
|
|
664
|
-
output: "output";
|
|
665
|
-
input: "input";
|
|
666
|
-
expectedOutput: "expectedOutput";
|
|
667
696
|
evaluatorContext: "evaluatorContext";
|
|
697
|
+
expectedOutput: "expectedOutput";
|
|
698
|
+
input: "input";
|
|
699
|
+
output: "output";
|
|
668
700
|
}>;
|
|
669
701
|
path: z.ZodArray<z.ZodString>;
|
|
670
702
|
source: z.ZodEnum<{
|
|
671
|
-
output: "output";
|
|
672
|
-
input: "input";
|
|
673
|
-
expectedOutput: "expectedOutput";
|
|
674
703
|
evaluatorContext: "evaluatorContext";
|
|
704
|
+
expectedOutput: "expectedOutput";
|
|
705
|
+
input: "input";
|
|
706
|
+
output: "output";
|
|
675
707
|
}>;
|
|
676
708
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
677
709
|
}, z.core.$strict>>;
|
|
@@ -749,10 +781,10 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
749
781
|
}, z.core.$strict>;
|
|
750
782
|
}, z.core.$strict>>;
|
|
751
783
|
split: z.ZodEnum<{
|
|
784
|
+
frozen_eval: "frozen_eval";
|
|
785
|
+
test: "test";
|
|
752
786
|
train: "train";
|
|
753
787
|
validation: "validation";
|
|
754
|
-
test: "test";
|
|
755
|
-
frozen_eval: "frozen_eval";
|
|
756
788
|
}>;
|
|
757
789
|
population: z.ZodArray<z.ZodObject<{
|
|
758
790
|
taskId: z.ZodString;
|
|
@@ -769,8 +801,8 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
769
801
|
configurationHash: z.ZodString;
|
|
770
802
|
output: z.ZodEnum<{
|
|
771
803
|
boolean: "boolean";
|
|
772
|
-
score: "score";
|
|
773
804
|
category: "category";
|
|
805
|
+
score: "score";
|
|
774
806
|
}>;
|
|
775
807
|
categories: z.ZodArray<z.ZodString>;
|
|
776
808
|
}, z.core.$strict>>;
|
|
@@ -779,17 +811,17 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
779
811
|
graderId: z.ZodString;
|
|
780
812
|
fields: z.ZodArray<z.ZodObject<{
|
|
781
813
|
destination: z.ZodEnum<{
|
|
782
|
-
output: "output";
|
|
783
|
-
input: "input";
|
|
784
|
-
expectedOutput: "expectedOutput";
|
|
785
814
|
evaluatorContext: "evaluatorContext";
|
|
815
|
+
expectedOutput: "expectedOutput";
|
|
816
|
+
input: "input";
|
|
817
|
+
output: "output";
|
|
786
818
|
}>;
|
|
787
819
|
path: z.ZodArray<z.ZodString>;
|
|
788
820
|
source: z.ZodEnum<{
|
|
789
|
-
output: "output";
|
|
790
|
-
input: "input";
|
|
791
|
-
expectedOutput: "expectedOutput";
|
|
792
821
|
evaluatorContext: "evaluatorContext";
|
|
822
|
+
expectedOutput: "expectedOutput";
|
|
823
|
+
input: "input";
|
|
824
|
+
output: "output";
|
|
793
825
|
}>;
|
|
794
826
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
795
827
|
}, z.core.$strict>>;
|
|
@@ -817,17 +849,17 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
817
849
|
contentHash: z.ZodString;
|
|
818
850
|
mappings: z.ZodArray<z.ZodObject<{
|
|
819
851
|
destination: z.ZodEnum<{
|
|
852
|
+
evaluatorContext: "evaluatorContext";
|
|
853
|
+
expectedOutput: "expectedOutput";
|
|
820
854
|
input: "input";
|
|
821
855
|
output: "output";
|
|
822
|
-
expectedOutput: "expectedOutput";
|
|
823
|
-
evaluatorContext: "evaluatorContext";
|
|
824
856
|
}>;
|
|
825
857
|
path: z.ZodArray<z.ZodString>;
|
|
826
858
|
source: z.ZodEnum<{
|
|
859
|
+
evaluatorContext: "evaluatorContext";
|
|
860
|
+
expectedOutput: "expectedOutput";
|
|
827
861
|
input: "input";
|
|
828
862
|
output: "output";
|
|
829
|
-
expectedOutput: "expectedOutput";
|
|
830
|
-
evaluatorContext: "evaluatorContext";
|
|
831
863
|
}>;
|
|
832
864
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
833
865
|
}, z.core.$strict>>;
|
|
@@ -847,8 +879,8 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
847
879
|
}, z.core.$strict>], "kind">;
|
|
848
880
|
state: z.ZodEnum<{
|
|
849
881
|
blocked: "blocked";
|
|
850
|
-
scheduled: "scheduled";
|
|
851
882
|
disabled: "disabled";
|
|
883
|
+
scheduled: "scheduled";
|
|
852
884
|
}>;
|
|
853
885
|
nextRunAt: z.ZodNullable<z.ZodString>;
|
|
854
886
|
maximumTotalSpendUsd: z.ZodNumber;
|
|
@@ -867,12 +899,12 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
867
899
|
operationId: z.ZodString;
|
|
868
900
|
state: z.ZodEnum<{
|
|
869
901
|
cancelled: "cancelled";
|
|
870
|
-
|
|
871
|
-
running: "running";
|
|
902
|
+
cleaning: "cleaning";
|
|
872
903
|
completed: "completed";
|
|
873
904
|
dispatching: "dispatching";
|
|
874
|
-
|
|
905
|
+
failed: "failed";
|
|
875
906
|
reserved: "reserved";
|
|
907
|
+
running: "running";
|
|
876
908
|
uncertain: "uncertain";
|
|
877
909
|
}>;
|
|
878
910
|
executionId: z.ZodNullable<z.ZodString>;
|
|
@@ -907,6 +939,13 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
907
939
|
operationId: string;
|
|
908
940
|
teamId: string;
|
|
909
941
|
modelProjectId: string | null;
|
|
942
|
+
name?: string | undefined;
|
|
943
|
+
project?: {
|
|
944
|
+
id: string;
|
|
945
|
+
revision: number;
|
|
946
|
+
contentHash: string;
|
|
947
|
+
targetId: string | null;
|
|
948
|
+
} | undefined;
|
|
910
949
|
taskset: {
|
|
911
950
|
id: string;
|
|
912
951
|
contentHash: string;
|
|
@@ -915,10 +954,12 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
915
954
|
policy: {
|
|
916
955
|
kind: "hosted_chat";
|
|
917
956
|
modelId: string;
|
|
918
|
-
maxOutputTokens: number;
|
|
919
|
-
temperature: number;
|
|
920
|
-
topP: number;
|
|
921
957
|
localRuntime?: {
|
|
958
|
+
providerId: "custom-openai-compatible";
|
|
959
|
+
configurationHash: string;
|
|
960
|
+
maximumRequestCostUsd: number;
|
|
961
|
+
requestTimeoutMs: number;
|
|
962
|
+
} | {
|
|
922
963
|
providerId: "claude-code";
|
|
923
964
|
configurationHash: string;
|
|
924
965
|
executableHash: string;
|
|
@@ -935,11 +976,6 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
935
976
|
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
936
977
|
maximumRequestCostUsd: number;
|
|
937
978
|
requestTimeoutMs: number;
|
|
938
|
-
} | {
|
|
939
|
-
providerId: "custom-openai-compatible";
|
|
940
|
-
configurationHash: string;
|
|
941
|
-
maximumRequestCostUsd: number;
|
|
942
|
-
requestTimeoutMs: number;
|
|
943
979
|
} | undefined;
|
|
944
980
|
candidate?: {
|
|
945
981
|
artifactId: string;
|
|
@@ -961,12 +997,47 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
961
997
|
sourcePackageHash: string;
|
|
962
998
|
} | undefined;
|
|
963
999
|
messages?: {
|
|
964
|
-
role: "
|
|
1000
|
+
role: "assistant" | "system" | "user";
|
|
965
1001
|
content: string;
|
|
966
1002
|
}[] | undefined;
|
|
1003
|
+
maxOutputTokens: number;
|
|
1004
|
+
temperature: number;
|
|
1005
|
+
topP: number;
|
|
1006
|
+
} | {
|
|
1007
|
+
kind: "fixture";
|
|
967
1008
|
} | {
|
|
968
1009
|
kind: "hosted_harness";
|
|
969
1010
|
modelId: string;
|
|
1011
|
+
maxOutputTokens?: number | undefined;
|
|
1012
|
+
temperature?: number | undefined;
|
|
1013
|
+
topP?: number | undefined;
|
|
1014
|
+
localRuntime?: {
|
|
1015
|
+
providerId: "codex";
|
|
1016
|
+
configurationHash: string;
|
|
1017
|
+
accountHash: string;
|
|
1018
|
+
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
1019
|
+
maximumRequestCostUsd: number;
|
|
1020
|
+
requestTimeoutMs: number;
|
|
1021
|
+
} | undefined;
|
|
1022
|
+
candidate?: {
|
|
1023
|
+
artifactId: string;
|
|
1024
|
+
contentHash: string;
|
|
1025
|
+
preparationId: string;
|
|
1026
|
+
dispatchId: string;
|
|
1027
|
+
baseProfileId: string;
|
|
1028
|
+
workerImageDigest: string;
|
|
1029
|
+
} | undefined;
|
|
1030
|
+
sourceCandidate?: {
|
|
1031
|
+
candidateId: string;
|
|
1032
|
+
frozenRevision: number;
|
|
1033
|
+
frozenHash: string;
|
|
1034
|
+
harnessRelease: {
|
|
1035
|
+
id: string;
|
|
1036
|
+
contentHash: string;
|
|
1037
|
+
};
|
|
1038
|
+
protectedClosureHash: string;
|
|
1039
|
+
} | undefined;
|
|
1040
|
+
reasoningEffort?: "high" | "low" | "max" | "medium" | "off" | "xhigh" | undefined;
|
|
970
1041
|
profileRepositoryId: string;
|
|
971
1042
|
source: {
|
|
972
1043
|
profileId: string;
|
|
@@ -1014,7 +1085,35 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1014
1085
|
};
|
|
1015
1086
|
declaredProfileCatalogHash: string | null;
|
|
1016
1087
|
protectedProfileClosureHash: string;
|
|
1017
|
-
|
|
1088
|
+
recordedOrigin?: {
|
|
1089
|
+
execution: {
|
|
1090
|
+
id: string;
|
|
1091
|
+
contentHash: string;
|
|
1092
|
+
};
|
|
1093
|
+
originalEvidence: {
|
|
1094
|
+
id: string;
|
|
1095
|
+
contentHash: string;
|
|
1096
|
+
};
|
|
1097
|
+
members: {
|
|
1098
|
+
caseId: string;
|
|
1099
|
+
seed: string;
|
|
1100
|
+
fixtureId: string;
|
|
1101
|
+
sourceId: string;
|
|
1102
|
+
snapshotHash: string;
|
|
1103
|
+
boundaryRevisionHash: string;
|
|
1104
|
+
}[];
|
|
1105
|
+
} | undefined;
|
|
1106
|
+
gradingSource?: {
|
|
1107
|
+
pass: {
|
|
1108
|
+
id: string;
|
|
1109
|
+
contentHash: string;
|
|
1110
|
+
};
|
|
1111
|
+
evidence: {
|
|
1112
|
+
id: string;
|
|
1113
|
+
contentHash: string;
|
|
1114
|
+
};
|
|
1115
|
+
} | undefined;
|
|
1116
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1018
1117
|
population: {
|
|
1019
1118
|
taskId: string;
|
|
1020
1119
|
seed: string;
|
|
@@ -1028,16 +1127,16 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1028
1127
|
};
|
|
1029
1128
|
feedbackKey: string;
|
|
1030
1129
|
configurationHash: string;
|
|
1031
|
-
output: "boolean" | "
|
|
1130
|
+
output: "boolean" | "category" | "score";
|
|
1032
1131
|
categories: string[];
|
|
1033
1132
|
}[];
|
|
1034
1133
|
fieldMappingsHash: string;
|
|
1035
1134
|
fieldMappings: {
|
|
1036
1135
|
graderId: string;
|
|
1037
1136
|
fields: {
|
|
1038
|
-
destination: "
|
|
1137
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1039
1138
|
path: string[];
|
|
1040
|
-
source: "
|
|
1139
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1041
1140
|
sourcePath: string[];
|
|
1042
1141
|
}[];
|
|
1043
1142
|
}[];
|
|
@@ -1048,34 +1147,6 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1048
1147
|
requireComplete: boolean;
|
|
1049
1148
|
};
|
|
1050
1149
|
contentHash: string;
|
|
1051
|
-
recordedOrigin?: {
|
|
1052
|
-
execution: {
|
|
1053
|
-
id: string;
|
|
1054
|
-
contentHash: string;
|
|
1055
|
-
};
|
|
1056
|
-
originalEvidence: {
|
|
1057
|
-
id: string;
|
|
1058
|
-
contentHash: string;
|
|
1059
|
-
};
|
|
1060
|
-
members: {
|
|
1061
|
-
caseId: string;
|
|
1062
|
-
seed: string;
|
|
1063
|
-
fixtureId: string;
|
|
1064
|
-
sourceId: string;
|
|
1065
|
-
snapshotHash: string;
|
|
1066
|
-
boundaryRevisionHash: string;
|
|
1067
|
-
}[];
|
|
1068
|
-
} | undefined;
|
|
1069
|
-
gradingSource?: {
|
|
1070
|
-
pass: {
|
|
1071
|
-
id: string;
|
|
1072
|
-
contentHash: string;
|
|
1073
|
-
};
|
|
1074
|
-
evidence: {
|
|
1075
|
-
id: string;
|
|
1076
|
-
contentHash: string;
|
|
1077
|
-
};
|
|
1078
|
-
} | undefined;
|
|
1079
1150
|
} | undefined;
|
|
1080
1151
|
sourceCandidate?: {
|
|
1081
1152
|
candidateId: string;
|
|
@@ -1090,25 +1161,6 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1090
1161
|
};
|
|
1091
1162
|
modelConfigurationHash: string;
|
|
1092
1163
|
packageHash: string;
|
|
1093
|
-
candidate?: {
|
|
1094
|
-
artifactId: string;
|
|
1095
|
-
contentHash: string;
|
|
1096
|
-
preparationId: string;
|
|
1097
|
-
dispatchId: string;
|
|
1098
|
-
baseProfileId: string;
|
|
1099
|
-
workerImageDigest: string;
|
|
1100
|
-
} | undefined;
|
|
1101
|
-
sourceCandidate?: {
|
|
1102
|
-
candidateId: string;
|
|
1103
|
-
frozenRevision: number;
|
|
1104
|
-
frozenHash: string;
|
|
1105
|
-
harnessRelease: {
|
|
1106
|
-
id: string;
|
|
1107
|
-
contentHash: string;
|
|
1108
|
-
};
|
|
1109
|
-
protectedClosureHash: string;
|
|
1110
|
-
} | undefined;
|
|
1111
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
1112
1164
|
externalDatasetBinding?: {
|
|
1113
1165
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
1114
1166
|
dataset: {
|
|
@@ -1132,7 +1184,35 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1132
1184
|
};
|
|
1133
1185
|
declaredProfileCatalogHash: string | null;
|
|
1134
1186
|
protectedProfileClosureHash: string;
|
|
1135
|
-
|
|
1187
|
+
recordedOrigin?: {
|
|
1188
|
+
execution: {
|
|
1189
|
+
id: string;
|
|
1190
|
+
contentHash: string;
|
|
1191
|
+
};
|
|
1192
|
+
originalEvidence: {
|
|
1193
|
+
id: string;
|
|
1194
|
+
contentHash: string;
|
|
1195
|
+
};
|
|
1196
|
+
members: {
|
|
1197
|
+
caseId: string;
|
|
1198
|
+
seed: string;
|
|
1199
|
+
fixtureId: string;
|
|
1200
|
+
sourceId: string;
|
|
1201
|
+
snapshotHash: string;
|
|
1202
|
+
boundaryRevisionHash: string;
|
|
1203
|
+
}[];
|
|
1204
|
+
} | undefined;
|
|
1205
|
+
gradingSource?: {
|
|
1206
|
+
pass: {
|
|
1207
|
+
id: string;
|
|
1208
|
+
contentHash: string;
|
|
1209
|
+
};
|
|
1210
|
+
evidence: {
|
|
1211
|
+
id: string;
|
|
1212
|
+
contentHash: string;
|
|
1213
|
+
};
|
|
1214
|
+
} | undefined;
|
|
1215
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1136
1216
|
population: {
|
|
1137
1217
|
taskId: string;
|
|
1138
1218
|
seed: string;
|
|
@@ -1146,16 +1226,16 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1146
1226
|
};
|
|
1147
1227
|
feedbackKey: string;
|
|
1148
1228
|
configurationHash: string;
|
|
1149
|
-
output: "boolean" | "
|
|
1229
|
+
output: "boolean" | "category" | "score";
|
|
1150
1230
|
categories: string[];
|
|
1151
1231
|
}[];
|
|
1152
1232
|
fieldMappingsHash: string;
|
|
1153
1233
|
fieldMappings: {
|
|
1154
1234
|
graderId: string;
|
|
1155
1235
|
fields: {
|
|
1156
|
-
destination: "
|
|
1236
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1157
1237
|
path: string[];
|
|
1158
|
-
source: "
|
|
1238
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1159
1239
|
sourcePath: string[];
|
|
1160
1240
|
}[];
|
|
1161
1241
|
}[];
|
|
@@ -1166,37 +1246,7 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1166
1246
|
requireComplete: boolean;
|
|
1167
1247
|
};
|
|
1168
1248
|
contentHash: string;
|
|
1169
|
-
recordedOrigin?: {
|
|
1170
|
-
execution: {
|
|
1171
|
-
id: string;
|
|
1172
|
-
contentHash: string;
|
|
1173
|
-
};
|
|
1174
|
-
originalEvidence: {
|
|
1175
|
-
id: string;
|
|
1176
|
-
contentHash: string;
|
|
1177
|
-
};
|
|
1178
|
-
members: {
|
|
1179
|
-
caseId: string;
|
|
1180
|
-
seed: string;
|
|
1181
|
-
fixtureId: string;
|
|
1182
|
-
sourceId: string;
|
|
1183
|
-
snapshotHash: string;
|
|
1184
|
-
boundaryRevisionHash: string;
|
|
1185
|
-
}[];
|
|
1186
|
-
} | undefined;
|
|
1187
|
-
gradingSource?: {
|
|
1188
|
-
pass: {
|
|
1189
|
-
id: string;
|
|
1190
|
-
contentHash: string;
|
|
1191
|
-
};
|
|
1192
|
-
evidence: {
|
|
1193
|
-
id: string;
|
|
1194
|
-
contentHash: string;
|
|
1195
|
-
};
|
|
1196
|
-
} | undefined;
|
|
1197
1249
|
} | undefined;
|
|
1198
|
-
} | {
|
|
1199
|
-
kind: "fixture";
|
|
1200
1250
|
};
|
|
1201
1251
|
population: {
|
|
1202
1252
|
receiptId: string;
|
|
@@ -1204,13 +1254,6 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1204
1254
|
seed: string;
|
|
1205
1255
|
fixtureId: string | null;
|
|
1206
1256
|
}[];
|
|
1207
|
-
name?: string | undefined;
|
|
1208
|
-
project?: {
|
|
1209
|
-
id: string;
|
|
1210
|
-
revision: number;
|
|
1211
|
-
contentHash: string;
|
|
1212
|
-
targetId: string | null;
|
|
1213
|
-
} | undefined;
|
|
1214
1257
|
};
|
|
1215
1258
|
maximumCostUsd: number;
|
|
1216
1259
|
graders?: {
|
|
@@ -1218,9 +1261,9 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1218
1261
|
version: string;
|
|
1219
1262
|
contentHash: string;
|
|
1220
1263
|
mappings: {
|
|
1221
|
-
destination: "
|
|
1264
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1222
1265
|
path: string[];
|
|
1223
|
-
source: "
|
|
1266
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1224
1267
|
sourcePath: string[];
|
|
1225
1268
|
}[];
|
|
1226
1269
|
}[] | undefined;
|
|
@@ -1237,7 +1280,7 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1237
1280
|
timeZone: string;
|
|
1238
1281
|
};
|
|
1239
1282
|
};
|
|
1240
|
-
state: "blocked" | "
|
|
1283
|
+
state: "blocked" | "disabled" | "scheduled";
|
|
1241
1284
|
nextRunAt: string | null;
|
|
1242
1285
|
maximumTotalSpendUsd: number;
|
|
1243
1286
|
maximumAttempts: number;
|
|
@@ -1253,7 +1296,7 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1253
1296
|
coalescedCount: number;
|
|
1254
1297
|
configurationHash: string;
|
|
1255
1298
|
operationId: string;
|
|
1256
|
-
state: "cancelled" | "
|
|
1299
|
+
state: "cancelled" | "cleaning" | "completed" | "dispatching" | "failed" | "reserved" | "running" | "uncertain";
|
|
1257
1300
|
executionId: string | null;
|
|
1258
1301
|
attempt: number;
|
|
1259
1302
|
maximumCostUsd: number;
|
|
@@ -1282,6 +1325,13 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1282
1325
|
operationId: string;
|
|
1283
1326
|
teamId: string;
|
|
1284
1327
|
modelProjectId: string | null;
|
|
1328
|
+
name?: string | undefined;
|
|
1329
|
+
project?: {
|
|
1330
|
+
id: string;
|
|
1331
|
+
revision: number;
|
|
1332
|
+
contentHash: string;
|
|
1333
|
+
targetId: string | null;
|
|
1334
|
+
} | undefined;
|
|
1285
1335
|
taskset: {
|
|
1286
1336
|
id: string;
|
|
1287
1337
|
contentHash: string;
|
|
@@ -1290,10 +1340,12 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1290
1340
|
policy: {
|
|
1291
1341
|
kind: "hosted_chat";
|
|
1292
1342
|
modelId: string;
|
|
1293
|
-
maxOutputTokens: number;
|
|
1294
|
-
temperature: number;
|
|
1295
|
-
topP: number;
|
|
1296
1343
|
localRuntime?: {
|
|
1344
|
+
providerId: "custom-openai-compatible";
|
|
1345
|
+
configurationHash: string;
|
|
1346
|
+
maximumRequestCostUsd: number;
|
|
1347
|
+
requestTimeoutMs: number;
|
|
1348
|
+
} | {
|
|
1297
1349
|
providerId: "claude-code";
|
|
1298
1350
|
configurationHash: string;
|
|
1299
1351
|
executableHash: string;
|
|
@@ -1310,11 +1362,6 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1310
1362
|
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
1311
1363
|
maximumRequestCostUsd: number;
|
|
1312
1364
|
requestTimeoutMs: number;
|
|
1313
|
-
} | {
|
|
1314
|
-
providerId: "custom-openai-compatible";
|
|
1315
|
-
configurationHash: string;
|
|
1316
|
-
maximumRequestCostUsd: number;
|
|
1317
|
-
requestTimeoutMs: number;
|
|
1318
1365
|
} | undefined;
|
|
1319
1366
|
candidate?: {
|
|
1320
1367
|
artifactId: string;
|
|
@@ -1336,12 +1383,47 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1336
1383
|
sourcePackageHash: string;
|
|
1337
1384
|
} | undefined;
|
|
1338
1385
|
messages?: {
|
|
1339
|
-
role: "
|
|
1386
|
+
role: "assistant" | "system" | "user";
|
|
1340
1387
|
content: string;
|
|
1341
1388
|
}[] | undefined;
|
|
1389
|
+
maxOutputTokens: number;
|
|
1390
|
+
temperature: number;
|
|
1391
|
+
topP: number;
|
|
1392
|
+
} | {
|
|
1393
|
+
kind: "fixture";
|
|
1342
1394
|
} | {
|
|
1343
1395
|
kind: "hosted_harness";
|
|
1344
1396
|
modelId: string;
|
|
1397
|
+
maxOutputTokens?: number | undefined;
|
|
1398
|
+
temperature?: number | undefined;
|
|
1399
|
+
topP?: number | undefined;
|
|
1400
|
+
localRuntime?: {
|
|
1401
|
+
providerId: "codex";
|
|
1402
|
+
configurationHash: string;
|
|
1403
|
+
accountHash: string;
|
|
1404
|
+
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
1405
|
+
maximumRequestCostUsd: number;
|
|
1406
|
+
requestTimeoutMs: number;
|
|
1407
|
+
} | undefined;
|
|
1408
|
+
candidate?: {
|
|
1409
|
+
artifactId: string;
|
|
1410
|
+
contentHash: string;
|
|
1411
|
+
preparationId: string;
|
|
1412
|
+
dispatchId: string;
|
|
1413
|
+
baseProfileId: string;
|
|
1414
|
+
workerImageDigest: string;
|
|
1415
|
+
} | undefined;
|
|
1416
|
+
sourceCandidate?: {
|
|
1417
|
+
candidateId: string;
|
|
1418
|
+
frozenRevision: number;
|
|
1419
|
+
frozenHash: string;
|
|
1420
|
+
harnessRelease: {
|
|
1421
|
+
id: string;
|
|
1422
|
+
contentHash: string;
|
|
1423
|
+
};
|
|
1424
|
+
protectedClosureHash: string;
|
|
1425
|
+
} | undefined;
|
|
1426
|
+
reasoningEffort?: "high" | "low" | "max" | "medium" | "off" | "xhigh" | undefined;
|
|
1345
1427
|
profileRepositoryId: string;
|
|
1346
1428
|
source: {
|
|
1347
1429
|
profileId: string;
|
|
@@ -1389,7 +1471,35 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1389
1471
|
};
|
|
1390
1472
|
declaredProfileCatalogHash: string | null;
|
|
1391
1473
|
protectedProfileClosureHash: string;
|
|
1392
|
-
|
|
1474
|
+
recordedOrigin?: {
|
|
1475
|
+
execution: {
|
|
1476
|
+
id: string;
|
|
1477
|
+
contentHash: string;
|
|
1478
|
+
};
|
|
1479
|
+
originalEvidence: {
|
|
1480
|
+
id: string;
|
|
1481
|
+
contentHash: string;
|
|
1482
|
+
};
|
|
1483
|
+
members: {
|
|
1484
|
+
caseId: string;
|
|
1485
|
+
seed: string;
|
|
1486
|
+
fixtureId: string;
|
|
1487
|
+
sourceId: string;
|
|
1488
|
+
snapshotHash: string;
|
|
1489
|
+
boundaryRevisionHash: string;
|
|
1490
|
+
}[];
|
|
1491
|
+
} | undefined;
|
|
1492
|
+
gradingSource?: {
|
|
1493
|
+
pass: {
|
|
1494
|
+
id: string;
|
|
1495
|
+
contentHash: string;
|
|
1496
|
+
};
|
|
1497
|
+
evidence: {
|
|
1498
|
+
id: string;
|
|
1499
|
+
contentHash: string;
|
|
1500
|
+
};
|
|
1501
|
+
} | undefined;
|
|
1502
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1393
1503
|
population: {
|
|
1394
1504
|
taskId: string;
|
|
1395
1505
|
seed: string;
|
|
@@ -1403,16 +1513,16 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1403
1513
|
};
|
|
1404
1514
|
feedbackKey: string;
|
|
1405
1515
|
configurationHash: string;
|
|
1406
|
-
output: "boolean" | "
|
|
1516
|
+
output: "boolean" | "category" | "score";
|
|
1407
1517
|
categories: string[];
|
|
1408
1518
|
}[];
|
|
1409
1519
|
fieldMappingsHash: string;
|
|
1410
1520
|
fieldMappings: {
|
|
1411
1521
|
graderId: string;
|
|
1412
1522
|
fields: {
|
|
1413
|
-
destination: "
|
|
1523
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1414
1524
|
path: string[];
|
|
1415
|
-
source: "
|
|
1525
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1416
1526
|
sourcePath: string[];
|
|
1417
1527
|
}[];
|
|
1418
1528
|
}[];
|
|
@@ -1423,34 +1533,6 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1423
1533
|
requireComplete: boolean;
|
|
1424
1534
|
};
|
|
1425
1535
|
contentHash: string;
|
|
1426
|
-
recordedOrigin?: {
|
|
1427
|
-
execution: {
|
|
1428
|
-
id: string;
|
|
1429
|
-
contentHash: string;
|
|
1430
|
-
};
|
|
1431
|
-
originalEvidence: {
|
|
1432
|
-
id: string;
|
|
1433
|
-
contentHash: string;
|
|
1434
|
-
};
|
|
1435
|
-
members: {
|
|
1436
|
-
caseId: string;
|
|
1437
|
-
seed: string;
|
|
1438
|
-
fixtureId: string;
|
|
1439
|
-
sourceId: string;
|
|
1440
|
-
snapshotHash: string;
|
|
1441
|
-
boundaryRevisionHash: string;
|
|
1442
|
-
}[];
|
|
1443
|
-
} | undefined;
|
|
1444
|
-
gradingSource?: {
|
|
1445
|
-
pass: {
|
|
1446
|
-
id: string;
|
|
1447
|
-
contentHash: string;
|
|
1448
|
-
};
|
|
1449
|
-
evidence: {
|
|
1450
|
-
id: string;
|
|
1451
|
-
contentHash: string;
|
|
1452
|
-
};
|
|
1453
|
-
} | undefined;
|
|
1454
1536
|
} | undefined;
|
|
1455
1537
|
sourceCandidate?: {
|
|
1456
1538
|
candidateId: string;
|
|
@@ -1465,25 +1547,6 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1465
1547
|
};
|
|
1466
1548
|
modelConfigurationHash: string;
|
|
1467
1549
|
packageHash: string;
|
|
1468
|
-
candidate?: {
|
|
1469
|
-
artifactId: string;
|
|
1470
|
-
contentHash: string;
|
|
1471
|
-
preparationId: string;
|
|
1472
|
-
dispatchId: string;
|
|
1473
|
-
baseProfileId: string;
|
|
1474
|
-
workerImageDigest: string;
|
|
1475
|
-
} | undefined;
|
|
1476
|
-
sourceCandidate?: {
|
|
1477
|
-
candidateId: string;
|
|
1478
|
-
frozenRevision: number;
|
|
1479
|
-
frozenHash: string;
|
|
1480
|
-
harnessRelease: {
|
|
1481
|
-
id: string;
|
|
1482
|
-
contentHash: string;
|
|
1483
|
-
};
|
|
1484
|
-
protectedClosureHash: string;
|
|
1485
|
-
} | undefined;
|
|
1486
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
1487
1550
|
externalDatasetBinding?: {
|
|
1488
1551
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
1489
1552
|
dataset: {
|
|
@@ -1507,7 +1570,35 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1507
1570
|
};
|
|
1508
1571
|
declaredProfileCatalogHash: string | null;
|
|
1509
1572
|
protectedProfileClosureHash: string;
|
|
1510
|
-
|
|
1573
|
+
recordedOrigin?: {
|
|
1574
|
+
execution: {
|
|
1575
|
+
id: string;
|
|
1576
|
+
contentHash: string;
|
|
1577
|
+
};
|
|
1578
|
+
originalEvidence: {
|
|
1579
|
+
id: string;
|
|
1580
|
+
contentHash: string;
|
|
1581
|
+
};
|
|
1582
|
+
members: {
|
|
1583
|
+
caseId: string;
|
|
1584
|
+
seed: string;
|
|
1585
|
+
fixtureId: string;
|
|
1586
|
+
sourceId: string;
|
|
1587
|
+
snapshotHash: string;
|
|
1588
|
+
boundaryRevisionHash: string;
|
|
1589
|
+
}[];
|
|
1590
|
+
} | undefined;
|
|
1591
|
+
gradingSource?: {
|
|
1592
|
+
pass: {
|
|
1593
|
+
id: string;
|
|
1594
|
+
contentHash: string;
|
|
1595
|
+
};
|
|
1596
|
+
evidence: {
|
|
1597
|
+
id: string;
|
|
1598
|
+
contentHash: string;
|
|
1599
|
+
};
|
|
1600
|
+
} | undefined;
|
|
1601
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
1511
1602
|
population: {
|
|
1512
1603
|
taskId: string;
|
|
1513
1604
|
seed: string;
|
|
@@ -1521,16 +1612,16 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1521
1612
|
};
|
|
1522
1613
|
feedbackKey: string;
|
|
1523
1614
|
configurationHash: string;
|
|
1524
|
-
output: "boolean" | "
|
|
1615
|
+
output: "boolean" | "category" | "score";
|
|
1525
1616
|
categories: string[];
|
|
1526
1617
|
}[];
|
|
1527
1618
|
fieldMappingsHash: string;
|
|
1528
1619
|
fieldMappings: {
|
|
1529
1620
|
graderId: string;
|
|
1530
1621
|
fields: {
|
|
1531
|
-
destination: "
|
|
1622
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1532
1623
|
path: string[];
|
|
1533
|
-
source: "
|
|
1624
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1534
1625
|
sourcePath: string[];
|
|
1535
1626
|
}[];
|
|
1536
1627
|
}[];
|
|
@@ -1541,37 +1632,7 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1541
1632
|
requireComplete: boolean;
|
|
1542
1633
|
};
|
|
1543
1634
|
contentHash: string;
|
|
1544
|
-
recordedOrigin?: {
|
|
1545
|
-
execution: {
|
|
1546
|
-
id: string;
|
|
1547
|
-
contentHash: string;
|
|
1548
|
-
};
|
|
1549
|
-
originalEvidence: {
|
|
1550
|
-
id: string;
|
|
1551
|
-
contentHash: string;
|
|
1552
|
-
};
|
|
1553
|
-
members: {
|
|
1554
|
-
caseId: string;
|
|
1555
|
-
seed: string;
|
|
1556
|
-
fixtureId: string;
|
|
1557
|
-
sourceId: string;
|
|
1558
|
-
snapshotHash: string;
|
|
1559
|
-
boundaryRevisionHash: string;
|
|
1560
|
-
}[];
|
|
1561
|
-
} | undefined;
|
|
1562
|
-
gradingSource?: {
|
|
1563
|
-
pass: {
|
|
1564
|
-
id: string;
|
|
1565
|
-
contentHash: string;
|
|
1566
|
-
};
|
|
1567
|
-
evidence: {
|
|
1568
|
-
id: string;
|
|
1569
|
-
contentHash: string;
|
|
1570
|
-
};
|
|
1571
|
-
} | undefined;
|
|
1572
1635
|
} | undefined;
|
|
1573
|
-
} | {
|
|
1574
|
-
kind: "fixture";
|
|
1575
1636
|
};
|
|
1576
1637
|
population: {
|
|
1577
1638
|
receiptId: string;
|
|
@@ -1579,13 +1640,6 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1579
1640
|
seed: string;
|
|
1580
1641
|
fixtureId: string | null;
|
|
1581
1642
|
}[];
|
|
1582
|
-
name?: string | undefined;
|
|
1583
|
-
project?: {
|
|
1584
|
-
id: string;
|
|
1585
|
-
revision: number;
|
|
1586
|
-
contentHash: string;
|
|
1587
|
-
targetId: string | null;
|
|
1588
|
-
} | undefined;
|
|
1589
1643
|
};
|
|
1590
1644
|
maximumCostUsd: number;
|
|
1591
1645
|
graders?: {
|
|
@@ -1593,9 +1647,9 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1593
1647
|
version: string;
|
|
1594
1648
|
contentHash: string;
|
|
1595
1649
|
mappings: {
|
|
1596
|
-
destination: "
|
|
1650
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1597
1651
|
path: string[];
|
|
1598
|
-
source: "
|
|
1652
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1599
1653
|
sourcePath: string[];
|
|
1600
1654
|
}[];
|
|
1601
1655
|
}[] | undefined;
|
|
@@ -1612,7 +1666,7 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1612
1666
|
timeZone: string;
|
|
1613
1667
|
};
|
|
1614
1668
|
};
|
|
1615
|
-
state: "blocked" | "
|
|
1669
|
+
state: "blocked" | "disabled" | "scheduled";
|
|
1616
1670
|
nextRunAt: string | null;
|
|
1617
1671
|
maximumTotalSpendUsd: number;
|
|
1618
1672
|
maximumAttempts: number;
|
|
@@ -1628,7 +1682,7 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1628
1682
|
coalescedCount: number;
|
|
1629
1683
|
configurationHash: string;
|
|
1630
1684
|
operationId: string;
|
|
1631
|
-
state: "cancelled" | "
|
|
1685
|
+
state: "cancelled" | "cleaning" | "completed" | "dispatching" | "failed" | "reserved" | "running" | "uncertain";
|
|
1632
1686
|
executionId: string | null;
|
|
1633
1687
|
attempt: number;
|
|
1634
1688
|
maximumCostUsd: number;
|
|
@@ -1647,8 +1701,8 @@ export declare const ExperimentEvaluationScheduleControlSchema: z.ZodObject<{
|
|
|
1647
1701
|
operationId: z.ZodString;
|
|
1648
1702
|
expectedRevision: z.ZodNumber;
|
|
1649
1703
|
action: z.ZodEnum<{
|
|
1650
|
-
pause: "pause";
|
|
1651
1704
|
cancel_active: "cancel_active";
|
|
1705
|
+
pause: "pause";
|
|
1652
1706
|
retry_active: "retry_active";
|
|
1653
1707
|
}>;
|
|
1654
1708
|
}, z.core.$strict>;
|
|
@@ -1732,9 +1786,9 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1732
1786
|
}, z.core.$strict>>;
|
|
1733
1787
|
messages: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
1734
1788
|
role: z.ZodEnum<{
|
|
1789
|
+
assistant: "assistant";
|
|
1735
1790
|
system: "system";
|
|
1736
1791
|
user: "user";
|
|
1737
|
-
assistant: "assistant";
|
|
1738
1792
|
}>;
|
|
1739
1793
|
content: z.ZodString;
|
|
1740
1794
|
}, z.core.$strict>>>;
|
|
@@ -1746,6 +1800,22 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1746
1800
|
}, z.core.$strict>, z.ZodObject<{
|
|
1747
1801
|
kind: z.ZodLiteral<"hosted_harness">;
|
|
1748
1802
|
modelId: z.ZodString;
|
|
1803
|
+
maxOutputTokens: z.ZodOptional<z.ZodNumber>;
|
|
1804
|
+
temperature: z.ZodOptional<z.ZodNumber>;
|
|
1805
|
+
topP: z.ZodOptional<z.ZodNumber>;
|
|
1806
|
+
localRuntime: z.ZodOptional<z.ZodObject<{
|
|
1807
|
+
providerId: z.ZodLiteral<"codex">;
|
|
1808
|
+
configurationHash: z.ZodString;
|
|
1809
|
+
accountHash: z.ZodString;
|
|
1810
|
+
reasoningEffort: z.ZodEnum<{
|
|
1811
|
+
high: "high";
|
|
1812
|
+
low: "low";
|
|
1813
|
+
medium: "medium";
|
|
1814
|
+
xhigh: "xhigh";
|
|
1815
|
+
}>;
|
|
1816
|
+
maximumRequestCostUsd: z.ZodNumber;
|
|
1817
|
+
requestTimeoutMs: z.ZodNumber;
|
|
1818
|
+
}, z.core.$strict>>;
|
|
1749
1819
|
candidate: z.ZodOptional<z.ZodObject<{
|
|
1750
1820
|
artifactId: z.ZodString;
|
|
1751
1821
|
contentHash: z.ZodString;
|
|
@@ -1767,10 +1837,10 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1767
1837
|
reasoningEffort: z.ZodOptional<z.ZodEnum<{
|
|
1768
1838
|
high: "high";
|
|
1769
1839
|
low: "low";
|
|
1840
|
+
max: "max";
|
|
1770
1841
|
medium: "medium";
|
|
1771
|
-
xhigh: "xhigh";
|
|
1772
1842
|
off: "off";
|
|
1773
|
-
|
|
1843
|
+
xhigh: "xhigh";
|
|
1774
1844
|
}>>;
|
|
1775
1845
|
profileRepositoryId: z.ZodString;
|
|
1776
1846
|
source: z.ZodObject<{
|
|
@@ -1848,10 +1918,10 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1848
1918
|
}, z.core.$strict>;
|
|
1849
1919
|
}, z.core.$strict>>;
|
|
1850
1920
|
split: z.ZodEnum<{
|
|
1921
|
+
frozen_eval: "frozen_eval";
|
|
1922
|
+
test: "test";
|
|
1851
1923
|
train: "train";
|
|
1852
1924
|
validation: "validation";
|
|
1853
|
-
test: "test";
|
|
1854
|
-
frozen_eval: "frozen_eval";
|
|
1855
1925
|
}>;
|
|
1856
1926
|
population: z.ZodArray<z.ZodObject<{
|
|
1857
1927
|
taskId: z.ZodString;
|
|
@@ -1868,8 +1938,8 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1868
1938
|
configurationHash: z.ZodString;
|
|
1869
1939
|
output: z.ZodEnum<{
|
|
1870
1940
|
boolean: "boolean";
|
|
1871
|
-
score: "score";
|
|
1872
1941
|
category: "category";
|
|
1942
|
+
score: "score";
|
|
1873
1943
|
}>;
|
|
1874
1944
|
categories: z.ZodArray<z.ZodString>;
|
|
1875
1945
|
}, z.core.$strict>>;
|
|
@@ -1878,17 +1948,17 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1878
1948
|
graderId: z.ZodString;
|
|
1879
1949
|
fields: z.ZodArray<z.ZodObject<{
|
|
1880
1950
|
destination: z.ZodEnum<{
|
|
1881
|
-
output: "output";
|
|
1882
|
-
input: "input";
|
|
1883
|
-
expectedOutput: "expectedOutput";
|
|
1884
1951
|
evaluatorContext: "evaluatorContext";
|
|
1952
|
+
expectedOutput: "expectedOutput";
|
|
1953
|
+
input: "input";
|
|
1954
|
+
output: "output";
|
|
1885
1955
|
}>;
|
|
1886
1956
|
path: z.ZodArray<z.ZodString>;
|
|
1887
1957
|
source: z.ZodEnum<{
|
|
1888
|
-
output: "output";
|
|
1889
|
-
input: "input";
|
|
1890
|
-
expectedOutput: "expectedOutput";
|
|
1891
1958
|
evaluatorContext: "evaluatorContext";
|
|
1959
|
+
expectedOutput: "expectedOutput";
|
|
1960
|
+
input: "input";
|
|
1961
|
+
output: "output";
|
|
1892
1962
|
}>;
|
|
1893
1963
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
1894
1964
|
}, z.core.$strict>>;
|
|
@@ -1966,10 +2036,10 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1966
2036
|
}, z.core.$strict>;
|
|
1967
2037
|
}, z.core.$strict>>;
|
|
1968
2038
|
split: z.ZodEnum<{
|
|
2039
|
+
frozen_eval: "frozen_eval";
|
|
2040
|
+
test: "test";
|
|
1969
2041
|
train: "train";
|
|
1970
2042
|
validation: "validation";
|
|
1971
|
-
test: "test";
|
|
1972
|
-
frozen_eval: "frozen_eval";
|
|
1973
2043
|
}>;
|
|
1974
2044
|
population: z.ZodArray<z.ZodObject<{
|
|
1975
2045
|
taskId: z.ZodString;
|
|
@@ -1986,8 +2056,8 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1986
2056
|
configurationHash: z.ZodString;
|
|
1987
2057
|
output: z.ZodEnum<{
|
|
1988
2058
|
boolean: "boolean";
|
|
1989
|
-
score: "score";
|
|
1990
2059
|
category: "category";
|
|
2060
|
+
score: "score";
|
|
1991
2061
|
}>;
|
|
1992
2062
|
categories: z.ZodArray<z.ZodString>;
|
|
1993
2063
|
}, z.core.$strict>>;
|
|
@@ -1996,17 +2066,17 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1996
2066
|
graderId: z.ZodString;
|
|
1997
2067
|
fields: z.ZodArray<z.ZodObject<{
|
|
1998
2068
|
destination: z.ZodEnum<{
|
|
1999
|
-
output: "output";
|
|
2000
|
-
input: "input";
|
|
2001
|
-
expectedOutput: "expectedOutput";
|
|
2002
2069
|
evaluatorContext: "evaluatorContext";
|
|
2070
|
+
expectedOutput: "expectedOutput";
|
|
2071
|
+
input: "input";
|
|
2072
|
+
output: "output";
|
|
2003
2073
|
}>;
|
|
2004
2074
|
path: z.ZodArray<z.ZodString>;
|
|
2005
2075
|
source: z.ZodEnum<{
|
|
2006
|
-
output: "output";
|
|
2007
|
-
input: "input";
|
|
2008
|
-
expectedOutput: "expectedOutput";
|
|
2009
2076
|
evaluatorContext: "evaluatorContext";
|
|
2077
|
+
expectedOutput: "expectedOutput";
|
|
2078
|
+
input: "input";
|
|
2079
|
+
output: "output";
|
|
2010
2080
|
}>;
|
|
2011
2081
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
2012
2082
|
}, z.core.$strict>>;
|
|
@@ -2034,17 +2104,17 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
2034
2104
|
contentHash: z.ZodString;
|
|
2035
2105
|
mappings: z.ZodArray<z.ZodObject<{
|
|
2036
2106
|
destination: z.ZodEnum<{
|
|
2107
|
+
evaluatorContext: "evaluatorContext";
|
|
2108
|
+
expectedOutput: "expectedOutput";
|
|
2037
2109
|
input: "input";
|
|
2038
2110
|
output: "output";
|
|
2039
|
-
expectedOutput: "expectedOutput";
|
|
2040
|
-
evaluatorContext: "evaluatorContext";
|
|
2041
2111
|
}>;
|
|
2042
2112
|
path: z.ZodArray<z.ZodString>;
|
|
2043
2113
|
source: z.ZodEnum<{
|
|
2114
|
+
evaluatorContext: "evaluatorContext";
|
|
2115
|
+
expectedOutput: "expectedOutput";
|
|
2044
2116
|
input: "input";
|
|
2045
2117
|
output: "output";
|
|
2046
|
-
expectedOutput: "expectedOutput";
|
|
2047
|
-
evaluatorContext: "evaluatorContext";
|
|
2048
2118
|
}>;
|
|
2049
2119
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
2050
2120
|
}, z.core.$strict>>;
|
|
@@ -2072,8 +2142,8 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
2072
2142
|
operationId: z.ZodString;
|
|
2073
2143
|
expectedRevision: z.ZodNumber;
|
|
2074
2144
|
action: z.ZodEnum<{
|
|
2075
|
-
pause: "pause";
|
|
2076
2145
|
cancel_active: "cancel_active";
|
|
2146
|
+
pause: "pause";
|
|
2077
2147
|
retry_active: "retry_active";
|
|
2078
2148
|
}>;
|
|
2079
2149
|
}, z.core.$strict>;
|
|
@@ -2104,6 +2174,13 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2104
2174
|
operationId: string;
|
|
2105
2175
|
teamId: string;
|
|
2106
2176
|
modelProjectId: string | null;
|
|
2177
|
+
name?: string | undefined;
|
|
2178
|
+
project?: {
|
|
2179
|
+
id: string;
|
|
2180
|
+
revision: number;
|
|
2181
|
+
contentHash: string;
|
|
2182
|
+
targetId: string | null;
|
|
2183
|
+
} | undefined;
|
|
2107
2184
|
taskset: {
|
|
2108
2185
|
id: string;
|
|
2109
2186
|
contentHash: string;
|
|
@@ -2112,10 +2189,12 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2112
2189
|
policy: {
|
|
2113
2190
|
kind: "hosted_chat";
|
|
2114
2191
|
modelId: string;
|
|
2115
|
-
maxOutputTokens: number;
|
|
2116
|
-
temperature: number;
|
|
2117
|
-
topP: number;
|
|
2118
2192
|
localRuntime?: {
|
|
2193
|
+
providerId: "custom-openai-compatible";
|
|
2194
|
+
configurationHash: string;
|
|
2195
|
+
maximumRequestCostUsd: number;
|
|
2196
|
+
requestTimeoutMs: number;
|
|
2197
|
+
} | {
|
|
2119
2198
|
providerId: "claude-code";
|
|
2120
2199
|
configurationHash: string;
|
|
2121
2200
|
executableHash: string;
|
|
@@ -2132,11 +2211,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2132
2211
|
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
2133
2212
|
maximumRequestCostUsd: number;
|
|
2134
2213
|
requestTimeoutMs: number;
|
|
2135
|
-
} | {
|
|
2136
|
-
providerId: "custom-openai-compatible";
|
|
2137
|
-
configurationHash: string;
|
|
2138
|
-
maximumRequestCostUsd: number;
|
|
2139
|
-
requestTimeoutMs: number;
|
|
2140
2214
|
} | undefined;
|
|
2141
2215
|
candidate?: {
|
|
2142
2216
|
artifactId: string;
|
|
@@ -2158,12 +2232,47 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2158
2232
|
sourcePackageHash: string;
|
|
2159
2233
|
} | undefined;
|
|
2160
2234
|
messages?: {
|
|
2161
|
-
role: "
|
|
2235
|
+
role: "assistant" | "system" | "user";
|
|
2162
2236
|
content: string;
|
|
2163
2237
|
}[] | undefined;
|
|
2238
|
+
maxOutputTokens: number;
|
|
2239
|
+
temperature: number;
|
|
2240
|
+
topP: number;
|
|
2241
|
+
} | {
|
|
2242
|
+
kind: "fixture";
|
|
2164
2243
|
} | {
|
|
2165
2244
|
kind: "hosted_harness";
|
|
2166
2245
|
modelId: string;
|
|
2246
|
+
maxOutputTokens?: number | undefined;
|
|
2247
|
+
temperature?: number | undefined;
|
|
2248
|
+
topP?: number | undefined;
|
|
2249
|
+
localRuntime?: {
|
|
2250
|
+
providerId: "codex";
|
|
2251
|
+
configurationHash: string;
|
|
2252
|
+
accountHash: string;
|
|
2253
|
+
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
2254
|
+
maximumRequestCostUsd: number;
|
|
2255
|
+
requestTimeoutMs: number;
|
|
2256
|
+
} | undefined;
|
|
2257
|
+
candidate?: {
|
|
2258
|
+
artifactId: string;
|
|
2259
|
+
contentHash: string;
|
|
2260
|
+
preparationId: string;
|
|
2261
|
+
dispatchId: string;
|
|
2262
|
+
baseProfileId: string;
|
|
2263
|
+
workerImageDigest: string;
|
|
2264
|
+
} | undefined;
|
|
2265
|
+
sourceCandidate?: {
|
|
2266
|
+
candidateId: string;
|
|
2267
|
+
frozenRevision: number;
|
|
2268
|
+
frozenHash: string;
|
|
2269
|
+
harnessRelease: {
|
|
2270
|
+
id: string;
|
|
2271
|
+
contentHash: string;
|
|
2272
|
+
};
|
|
2273
|
+
protectedClosureHash: string;
|
|
2274
|
+
} | undefined;
|
|
2275
|
+
reasoningEffort?: "high" | "low" | "max" | "medium" | "off" | "xhigh" | undefined;
|
|
2167
2276
|
profileRepositoryId: string;
|
|
2168
2277
|
source: {
|
|
2169
2278
|
profileId: string;
|
|
@@ -2211,7 +2320,35 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2211
2320
|
};
|
|
2212
2321
|
declaredProfileCatalogHash: string | null;
|
|
2213
2322
|
protectedProfileClosureHash: string;
|
|
2214
|
-
|
|
2323
|
+
recordedOrigin?: {
|
|
2324
|
+
execution: {
|
|
2325
|
+
id: string;
|
|
2326
|
+
contentHash: string;
|
|
2327
|
+
};
|
|
2328
|
+
originalEvidence: {
|
|
2329
|
+
id: string;
|
|
2330
|
+
contentHash: string;
|
|
2331
|
+
};
|
|
2332
|
+
members: {
|
|
2333
|
+
caseId: string;
|
|
2334
|
+
seed: string;
|
|
2335
|
+
fixtureId: string;
|
|
2336
|
+
sourceId: string;
|
|
2337
|
+
snapshotHash: string;
|
|
2338
|
+
boundaryRevisionHash: string;
|
|
2339
|
+
}[];
|
|
2340
|
+
} | undefined;
|
|
2341
|
+
gradingSource?: {
|
|
2342
|
+
pass: {
|
|
2343
|
+
id: string;
|
|
2344
|
+
contentHash: string;
|
|
2345
|
+
};
|
|
2346
|
+
evidence: {
|
|
2347
|
+
id: string;
|
|
2348
|
+
contentHash: string;
|
|
2349
|
+
};
|
|
2350
|
+
} | undefined;
|
|
2351
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
2215
2352
|
population: {
|
|
2216
2353
|
taskId: string;
|
|
2217
2354
|
seed: string;
|
|
@@ -2225,16 +2362,16 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2225
2362
|
};
|
|
2226
2363
|
feedbackKey: string;
|
|
2227
2364
|
configurationHash: string;
|
|
2228
|
-
output: "boolean" | "
|
|
2365
|
+
output: "boolean" | "category" | "score";
|
|
2229
2366
|
categories: string[];
|
|
2230
2367
|
}[];
|
|
2231
2368
|
fieldMappingsHash: string;
|
|
2232
2369
|
fieldMappings: {
|
|
2233
2370
|
graderId: string;
|
|
2234
2371
|
fields: {
|
|
2235
|
-
destination: "
|
|
2372
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2236
2373
|
path: string[];
|
|
2237
|
-
source: "
|
|
2374
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2238
2375
|
sourcePath: string[];
|
|
2239
2376
|
}[];
|
|
2240
2377
|
}[];
|
|
@@ -2245,34 +2382,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2245
2382
|
requireComplete: boolean;
|
|
2246
2383
|
};
|
|
2247
2384
|
contentHash: string;
|
|
2248
|
-
recordedOrigin?: {
|
|
2249
|
-
execution: {
|
|
2250
|
-
id: string;
|
|
2251
|
-
contentHash: string;
|
|
2252
|
-
};
|
|
2253
|
-
originalEvidence: {
|
|
2254
|
-
id: string;
|
|
2255
|
-
contentHash: string;
|
|
2256
|
-
};
|
|
2257
|
-
members: {
|
|
2258
|
-
caseId: string;
|
|
2259
|
-
seed: string;
|
|
2260
|
-
fixtureId: string;
|
|
2261
|
-
sourceId: string;
|
|
2262
|
-
snapshotHash: string;
|
|
2263
|
-
boundaryRevisionHash: string;
|
|
2264
|
-
}[];
|
|
2265
|
-
} | undefined;
|
|
2266
|
-
gradingSource?: {
|
|
2267
|
-
pass: {
|
|
2268
|
-
id: string;
|
|
2269
|
-
contentHash: string;
|
|
2270
|
-
};
|
|
2271
|
-
evidence: {
|
|
2272
|
-
id: string;
|
|
2273
|
-
contentHash: string;
|
|
2274
|
-
};
|
|
2275
|
-
} | undefined;
|
|
2276
2385
|
} | undefined;
|
|
2277
2386
|
sourceCandidate?: {
|
|
2278
2387
|
candidateId: string;
|
|
@@ -2286,26 +2395,7 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2286
2395
|
} | undefined;
|
|
2287
2396
|
};
|
|
2288
2397
|
modelConfigurationHash: string;
|
|
2289
|
-
packageHash: string;
|
|
2290
|
-
candidate?: {
|
|
2291
|
-
artifactId: string;
|
|
2292
|
-
contentHash: string;
|
|
2293
|
-
preparationId: string;
|
|
2294
|
-
dispatchId: string;
|
|
2295
|
-
baseProfileId: string;
|
|
2296
|
-
workerImageDigest: string;
|
|
2297
|
-
} | undefined;
|
|
2298
|
-
sourceCandidate?: {
|
|
2299
|
-
candidateId: string;
|
|
2300
|
-
frozenRevision: number;
|
|
2301
|
-
frozenHash: string;
|
|
2302
|
-
harnessRelease: {
|
|
2303
|
-
id: string;
|
|
2304
|
-
contentHash: string;
|
|
2305
|
-
};
|
|
2306
|
-
protectedClosureHash: string;
|
|
2307
|
-
} | undefined;
|
|
2308
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
2398
|
+
packageHash: string;
|
|
2309
2399
|
externalDatasetBinding?: {
|
|
2310
2400
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
2311
2401
|
dataset: {
|
|
@@ -2329,7 +2419,35 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2329
2419
|
};
|
|
2330
2420
|
declaredProfileCatalogHash: string | null;
|
|
2331
2421
|
protectedProfileClosureHash: string;
|
|
2332
|
-
|
|
2422
|
+
recordedOrigin?: {
|
|
2423
|
+
execution: {
|
|
2424
|
+
id: string;
|
|
2425
|
+
contentHash: string;
|
|
2426
|
+
};
|
|
2427
|
+
originalEvidence: {
|
|
2428
|
+
id: string;
|
|
2429
|
+
contentHash: string;
|
|
2430
|
+
};
|
|
2431
|
+
members: {
|
|
2432
|
+
caseId: string;
|
|
2433
|
+
seed: string;
|
|
2434
|
+
fixtureId: string;
|
|
2435
|
+
sourceId: string;
|
|
2436
|
+
snapshotHash: string;
|
|
2437
|
+
boundaryRevisionHash: string;
|
|
2438
|
+
}[];
|
|
2439
|
+
} | undefined;
|
|
2440
|
+
gradingSource?: {
|
|
2441
|
+
pass: {
|
|
2442
|
+
id: string;
|
|
2443
|
+
contentHash: string;
|
|
2444
|
+
};
|
|
2445
|
+
evidence: {
|
|
2446
|
+
id: string;
|
|
2447
|
+
contentHash: string;
|
|
2448
|
+
};
|
|
2449
|
+
} | undefined;
|
|
2450
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
2333
2451
|
population: {
|
|
2334
2452
|
taskId: string;
|
|
2335
2453
|
seed: string;
|
|
@@ -2343,16 +2461,16 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2343
2461
|
};
|
|
2344
2462
|
feedbackKey: string;
|
|
2345
2463
|
configurationHash: string;
|
|
2346
|
-
output: "boolean" | "
|
|
2464
|
+
output: "boolean" | "category" | "score";
|
|
2347
2465
|
categories: string[];
|
|
2348
2466
|
}[];
|
|
2349
2467
|
fieldMappingsHash: string;
|
|
2350
2468
|
fieldMappings: {
|
|
2351
2469
|
graderId: string;
|
|
2352
2470
|
fields: {
|
|
2353
|
-
destination: "
|
|
2471
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2354
2472
|
path: string[];
|
|
2355
|
-
source: "
|
|
2473
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2356
2474
|
sourcePath: string[];
|
|
2357
2475
|
}[];
|
|
2358
2476
|
}[];
|
|
@@ -2363,37 +2481,7 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2363
2481
|
requireComplete: boolean;
|
|
2364
2482
|
};
|
|
2365
2483
|
contentHash: string;
|
|
2366
|
-
recordedOrigin?: {
|
|
2367
|
-
execution: {
|
|
2368
|
-
id: string;
|
|
2369
|
-
contentHash: string;
|
|
2370
|
-
};
|
|
2371
|
-
originalEvidence: {
|
|
2372
|
-
id: string;
|
|
2373
|
-
contentHash: string;
|
|
2374
|
-
};
|
|
2375
|
-
members: {
|
|
2376
|
-
caseId: string;
|
|
2377
|
-
seed: string;
|
|
2378
|
-
fixtureId: string;
|
|
2379
|
-
sourceId: string;
|
|
2380
|
-
snapshotHash: string;
|
|
2381
|
-
boundaryRevisionHash: string;
|
|
2382
|
-
}[];
|
|
2383
|
-
} | undefined;
|
|
2384
|
-
gradingSource?: {
|
|
2385
|
-
pass: {
|
|
2386
|
-
id: string;
|
|
2387
|
-
contentHash: string;
|
|
2388
|
-
};
|
|
2389
|
-
evidence: {
|
|
2390
|
-
id: string;
|
|
2391
|
-
contentHash: string;
|
|
2392
|
-
};
|
|
2393
|
-
} | undefined;
|
|
2394
2484
|
} | undefined;
|
|
2395
|
-
} | {
|
|
2396
|
-
kind: "fixture";
|
|
2397
2485
|
};
|
|
2398
2486
|
population: {
|
|
2399
2487
|
receiptId: string;
|
|
@@ -2401,13 +2489,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2401
2489
|
seed: string;
|
|
2402
2490
|
fixtureId: string | null;
|
|
2403
2491
|
}[];
|
|
2404
|
-
name?: string | undefined;
|
|
2405
|
-
project?: {
|
|
2406
|
-
id: string;
|
|
2407
|
-
revision: number;
|
|
2408
|
-
contentHash: string;
|
|
2409
|
-
targetId: string | null;
|
|
2410
|
-
} | undefined;
|
|
2411
2492
|
};
|
|
2412
2493
|
maximumCostUsd: number;
|
|
2413
2494
|
graders?: {
|
|
@@ -2415,9 +2496,9 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2415
2496
|
version: string;
|
|
2416
2497
|
contentHash: string;
|
|
2417
2498
|
mappings: {
|
|
2418
|
-
destination: "
|
|
2499
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2419
2500
|
path: string[];
|
|
2420
|
-
source: "
|
|
2501
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2421
2502
|
sourcePath: string[];
|
|
2422
2503
|
}[];
|
|
2423
2504
|
}[] | undefined;
|
|
@@ -2434,7 +2515,7 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2434
2515
|
timeZone: string;
|
|
2435
2516
|
};
|
|
2436
2517
|
};
|
|
2437
|
-
state: "blocked" | "
|
|
2518
|
+
state: "blocked" | "disabled" | "scheduled";
|
|
2438
2519
|
nextRunAt: string | null;
|
|
2439
2520
|
maximumTotalSpendUsd: number;
|
|
2440
2521
|
maximumAttempts: number;
|
|
@@ -2450,7 +2531,7 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2450
2531
|
coalescedCount: number;
|
|
2451
2532
|
configurationHash: string;
|
|
2452
2533
|
operationId: string;
|
|
2453
|
-
state: "cancelled" | "
|
|
2534
|
+
state: "cancelled" | "cleaning" | "completed" | "dispatching" | "failed" | "reserved" | "running" | "uncertain";
|
|
2454
2535
|
executionId: string | null;
|
|
2455
2536
|
attempt: number;
|
|
2456
2537
|
maximumCostUsd: number;
|
|
@@ -2479,6 +2560,13 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2479
2560
|
operationId: string;
|
|
2480
2561
|
teamId: string;
|
|
2481
2562
|
modelProjectId: string | null;
|
|
2563
|
+
name?: string | undefined;
|
|
2564
|
+
project?: {
|
|
2565
|
+
id: string;
|
|
2566
|
+
revision: number;
|
|
2567
|
+
contentHash: string;
|
|
2568
|
+
targetId: string | null;
|
|
2569
|
+
} | undefined;
|
|
2482
2570
|
taskset: {
|
|
2483
2571
|
id: string;
|
|
2484
2572
|
contentHash: string;
|
|
@@ -2487,10 +2575,12 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2487
2575
|
policy: {
|
|
2488
2576
|
kind: "hosted_chat";
|
|
2489
2577
|
modelId: string;
|
|
2490
|
-
maxOutputTokens: number;
|
|
2491
|
-
temperature: number;
|
|
2492
|
-
topP: number;
|
|
2493
2578
|
localRuntime?: {
|
|
2579
|
+
providerId: "custom-openai-compatible";
|
|
2580
|
+
configurationHash: string;
|
|
2581
|
+
maximumRequestCostUsd: number;
|
|
2582
|
+
requestTimeoutMs: number;
|
|
2583
|
+
} | {
|
|
2494
2584
|
providerId: "claude-code";
|
|
2495
2585
|
configurationHash: string;
|
|
2496
2586
|
executableHash: string;
|
|
@@ -2507,11 +2597,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2507
2597
|
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
2508
2598
|
maximumRequestCostUsd: number;
|
|
2509
2599
|
requestTimeoutMs: number;
|
|
2510
|
-
} | {
|
|
2511
|
-
providerId: "custom-openai-compatible";
|
|
2512
|
-
configurationHash: string;
|
|
2513
|
-
maximumRequestCostUsd: number;
|
|
2514
|
-
requestTimeoutMs: number;
|
|
2515
2600
|
} | undefined;
|
|
2516
2601
|
candidate?: {
|
|
2517
2602
|
artifactId: string;
|
|
@@ -2533,12 +2618,47 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2533
2618
|
sourcePackageHash: string;
|
|
2534
2619
|
} | undefined;
|
|
2535
2620
|
messages?: {
|
|
2536
|
-
role: "
|
|
2621
|
+
role: "assistant" | "system" | "user";
|
|
2537
2622
|
content: string;
|
|
2538
2623
|
}[] | undefined;
|
|
2624
|
+
maxOutputTokens: number;
|
|
2625
|
+
temperature: number;
|
|
2626
|
+
topP: number;
|
|
2627
|
+
} | {
|
|
2628
|
+
kind: "fixture";
|
|
2539
2629
|
} | {
|
|
2540
2630
|
kind: "hosted_harness";
|
|
2541
2631
|
modelId: string;
|
|
2632
|
+
maxOutputTokens?: number | undefined;
|
|
2633
|
+
temperature?: number | undefined;
|
|
2634
|
+
topP?: number | undefined;
|
|
2635
|
+
localRuntime?: {
|
|
2636
|
+
providerId: "codex";
|
|
2637
|
+
configurationHash: string;
|
|
2638
|
+
accountHash: string;
|
|
2639
|
+
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
2640
|
+
maximumRequestCostUsd: number;
|
|
2641
|
+
requestTimeoutMs: number;
|
|
2642
|
+
} | undefined;
|
|
2643
|
+
candidate?: {
|
|
2644
|
+
artifactId: string;
|
|
2645
|
+
contentHash: string;
|
|
2646
|
+
preparationId: string;
|
|
2647
|
+
dispatchId: string;
|
|
2648
|
+
baseProfileId: string;
|
|
2649
|
+
workerImageDigest: string;
|
|
2650
|
+
} | undefined;
|
|
2651
|
+
sourceCandidate?: {
|
|
2652
|
+
candidateId: string;
|
|
2653
|
+
frozenRevision: number;
|
|
2654
|
+
frozenHash: string;
|
|
2655
|
+
harnessRelease: {
|
|
2656
|
+
id: string;
|
|
2657
|
+
contentHash: string;
|
|
2658
|
+
};
|
|
2659
|
+
protectedClosureHash: string;
|
|
2660
|
+
} | undefined;
|
|
2661
|
+
reasoningEffort?: "high" | "low" | "max" | "medium" | "off" | "xhigh" | undefined;
|
|
2542
2662
|
profileRepositoryId: string;
|
|
2543
2663
|
source: {
|
|
2544
2664
|
profileId: string;
|
|
@@ -2586,7 +2706,35 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2586
2706
|
};
|
|
2587
2707
|
declaredProfileCatalogHash: string | null;
|
|
2588
2708
|
protectedProfileClosureHash: string;
|
|
2589
|
-
|
|
2709
|
+
recordedOrigin?: {
|
|
2710
|
+
execution: {
|
|
2711
|
+
id: string;
|
|
2712
|
+
contentHash: string;
|
|
2713
|
+
};
|
|
2714
|
+
originalEvidence: {
|
|
2715
|
+
id: string;
|
|
2716
|
+
contentHash: string;
|
|
2717
|
+
};
|
|
2718
|
+
members: {
|
|
2719
|
+
caseId: string;
|
|
2720
|
+
seed: string;
|
|
2721
|
+
fixtureId: string;
|
|
2722
|
+
sourceId: string;
|
|
2723
|
+
snapshotHash: string;
|
|
2724
|
+
boundaryRevisionHash: string;
|
|
2725
|
+
}[];
|
|
2726
|
+
} | undefined;
|
|
2727
|
+
gradingSource?: {
|
|
2728
|
+
pass: {
|
|
2729
|
+
id: string;
|
|
2730
|
+
contentHash: string;
|
|
2731
|
+
};
|
|
2732
|
+
evidence: {
|
|
2733
|
+
id: string;
|
|
2734
|
+
contentHash: string;
|
|
2735
|
+
};
|
|
2736
|
+
} | undefined;
|
|
2737
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
2590
2738
|
population: {
|
|
2591
2739
|
taskId: string;
|
|
2592
2740
|
seed: string;
|
|
@@ -2600,16 +2748,16 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2600
2748
|
};
|
|
2601
2749
|
feedbackKey: string;
|
|
2602
2750
|
configurationHash: string;
|
|
2603
|
-
output: "boolean" | "
|
|
2751
|
+
output: "boolean" | "category" | "score";
|
|
2604
2752
|
categories: string[];
|
|
2605
2753
|
}[];
|
|
2606
2754
|
fieldMappingsHash: string;
|
|
2607
2755
|
fieldMappings: {
|
|
2608
2756
|
graderId: string;
|
|
2609
2757
|
fields: {
|
|
2610
|
-
destination: "
|
|
2758
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2611
2759
|
path: string[];
|
|
2612
|
-
source: "
|
|
2760
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2613
2761
|
sourcePath: string[];
|
|
2614
2762
|
}[];
|
|
2615
2763
|
}[];
|
|
@@ -2620,34 +2768,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2620
2768
|
requireComplete: boolean;
|
|
2621
2769
|
};
|
|
2622
2770
|
contentHash: string;
|
|
2623
|
-
recordedOrigin?: {
|
|
2624
|
-
execution: {
|
|
2625
|
-
id: string;
|
|
2626
|
-
contentHash: string;
|
|
2627
|
-
};
|
|
2628
|
-
originalEvidence: {
|
|
2629
|
-
id: string;
|
|
2630
|
-
contentHash: string;
|
|
2631
|
-
};
|
|
2632
|
-
members: {
|
|
2633
|
-
caseId: string;
|
|
2634
|
-
seed: string;
|
|
2635
|
-
fixtureId: string;
|
|
2636
|
-
sourceId: string;
|
|
2637
|
-
snapshotHash: string;
|
|
2638
|
-
boundaryRevisionHash: string;
|
|
2639
|
-
}[];
|
|
2640
|
-
} | undefined;
|
|
2641
|
-
gradingSource?: {
|
|
2642
|
-
pass: {
|
|
2643
|
-
id: string;
|
|
2644
|
-
contentHash: string;
|
|
2645
|
-
};
|
|
2646
|
-
evidence: {
|
|
2647
|
-
id: string;
|
|
2648
|
-
contentHash: string;
|
|
2649
|
-
};
|
|
2650
|
-
} | undefined;
|
|
2651
2771
|
} | undefined;
|
|
2652
2772
|
sourceCandidate?: {
|
|
2653
2773
|
candidateId: string;
|
|
@@ -2662,25 +2782,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2662
2782
|
};
|
|
2663
2783
|
modelConfigurationHash: string;
|
|
2664
2784
|
packageHash: string;
|
|
2665
|
-
candidate?: {
|
|
2666
|
-
artifactId: string;
|
|
2667
|
-
contentHash: string;
|
|
2668
|
-
preparationId: string;
|
|
2669
|
-
dispatchId: string;
|
|
2670
|
-
baseProfileId: string;
|
|
2671
|
-
workerImageDigest: string;
|
|
2672
|
-
} | undefined;
|
|
2673
|
-
sourceCandidate?: {
|
|
2674
|
-
candidateId: string;
|
|
2675
|
-
frozenRevision: number;
|
|
2676
|
-
frozenHash: string;
|
|
2677
|
-
harnessRelease: {
|
|
2678
|
-
id: string;
|
|
2679
|
-
contentHash: string;
|
|
2680
|
-
};
|
|
2681
|
-
protectedClosureHash: string;
|
|
2682
|
-
} | undefined;
|
|
2683
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
2684
2785
|
externalDatasetBinding?: {
|
|
2685
2786
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
2686
2787
|
dataset: {
|
|
@@ -2704,7 +2805,35 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2704
2805
|
};
|
|
2705
2806
|
declaredProfileCatalogHash: string | null;
|
|
2706
2807
|
protectedProfileClosureHash: string;
|
|
2707
|
-
|
|
2808
|
+
recordedOrigin?: {
|
|
2809
|
+
execution: {
|
|
2810
|
+
id: string;
|
|
2811
|
+
contentHash: string;
|
|
2812
|
+
};
|
|
2813
|
+
originalEvidence: {
|
|
2814
|
+
id: string;
|
|
2815
|
+
contentHash: string;
|
|
2816
|
+
};
|
|
2817
|
+
members: {
|
|
2818
|
+
caseId: string;
|
|
2819
|
+
seed: string;
|
|
2820
|
+
fixtureId: string;
|
|
2821
|
+
sourceId: string;
|
|
2822
|
+
snapshotHash: string;
|
|
2823
|
+
boundaryRevisionHash: string;
|
|
2824
|
+
}[];
|
|
2825
|
+
} | undefined;
|
|
2826
|
+
gradingSource?: {
|
|
2827
|
+
pass: {
|
|
2828
|
+
id: string;
|
|
2829
|
+
contentHash: string;
|
|
2830
|
+
};
|
|
2831
|
+
evidence: {
|
|
2832
|
+
id: string;
|
|
2833
|
+
contentHash: string;
|
|
2834
|
+
};
|
|
2835
|
+
} | undefined;
|
|
2836
|
+
split: "frozen_eval" | "test" | "train" | "validation";
|
|
2708
2837
|
population: {
|
|
2709
2838
|
taskId: string;
|
|
2710
2839
|
seed: string;
|
|
@@ -2718,16 +2847,16 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2718
2847
|
};
|
|
2719
2848
|
feedbackKey: string;
|
|
2720
2849
|
configurationHash: string;
|
|
2721
|
-
output: "boolean" | "
|
|
2850
|
+
output: "boolean" | "category" | "score";
|
|
2722
2851
|
categories: string[];
|
|
2723
2852
|
}[];
|
|
2724
2853
|
fieldMappingsHash: string;
|
|
2725
2854
|
fieldMappings: {
|
|
2726
2855
|
graderId: string;
|
|
2727
2856
|
fields: {
|
|
2728
|
-
destination: "
|
|
2857
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2729
2858
|
path: string[];
|
|
2730
|
-
source: "
|
|
2859
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2731
2860
|
sourcePath: string[];
|
|
2732
2861
|
}[];
|
|
2733
2862
|
}[];
|
|
@@ -2738,37 +2867,7 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2738
2867
|
requireComplete: boolean;
|
|
2739
2868
|
};
|
|
2740
2869
|
contentHash: string;
|
|
2741
|
-
recordedOrigin?: {
|
|
2742
|
-
execution: {
|
|
2743
|
-
id: string;
|
|
2744
|
-
contentHash: string;
|
|
2745
|
-
};
|
|
2746
|
-
originalEvidence: {
|
|
2747
|
-
id: string;
|
|
2748
|
-
contentHash: string;
|
|
2749
|
-
};
|
|
2750
|
-
members: {
|
|
2751
|
-
caseId: string;
|
|
2752
|
-
seed: string;
|
|
2753
|
-
fixtureId: string;
|
|
2754
|
-
sourceId: string;
|
|
2755
|
-
snapshotHash: string;
|
|
2756
|
-
boundaryRevisionHash: string;
|
|
2757
|
-
}[];
|
|
2758
|
-
} | undefined;
|
|
2759
|
-
gradingSource?: {
|
|
2760
|
-
pass: {
|
|
2761
|
-
id: string;
|
|
2762
|
-
contentHash: string;
|
|
2763
|
-
};
|
|
2764
|
-
evidence: {
|
|
2765
|
-
id: string;
|
|
2766
|
-
contentHash: string;
|
|
2767
|
-
};
|
|
2768
|
-
} | undefined;
|
|
2769
2870
|
} | undefined;
|
|
2770
|
-
} | {
|
|
2771
|
-
kind: "fixture";
|
|
2772
2871
|
};
|
|
2773
2872
|
population: {
|
|
2774
2873
|
receiptId: string;
|
|
@@ -2776,13 +2875,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2776
2875
|
seed: string;
|
|
2777
2876
|
fixtureId: string | null;
|
|
2778
2877
|
}[];
|
|
2779
|
-
name?: string | undefined;
|
|
2780
|
-
project?: {
|
|
2781
|
-
id: string;
|
|
2782
|
-
revision: number;
|
|
2783
|
-
contentHash: string;
|
|
2784
|
-
targetId: string | null;
|
|
2785
|
-
} | undefined;
|
|
2786
2878
|
};
|
|
2787
2879
|
maximumCostUsd: number;
|
|
2788
2880
|
graders?: {
|
|
@@ -2790,9 +2882,9 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2790
2882
|
version: string;
|
|
2791
2883
|
contentHash: string;
|
|
2792
2884
|
mappings: {
|
|
2793
|
-
destination: "
|
|
2885
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2794
2886
|
path: string[];
|
|
2795
|
-
source: "
|
|
2887
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2796
2888
|
sourcePath: string[];
|
|
2797
2889
|
}[];
|
|
2798
2890
|
}[] | undefined;
|
|
@@ -2809,7 +2901,7 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2809
2901
|
timeZone: string;
|
|
2810
2902
|
};
|
|
2811
2903
|
};
|
|
2812
|
-
state: "blocked" | "
|
|
2904
|
+
state: "blocked" | "disabled" | "scheduled";
|
|
2813
2905
|
nextRunAt: string | null;
|
|
2814
2906
|
maximumTotalSpendUsd: number;
|
|
2815
2907
|
maximumAttempts: number;
|
|
@@ -2825,7 +2917,7 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2825
2917
|
coalescedCount: number;
|
|
2826
2918
|
configurationHash: string;
|
|
2827
2919
|
operationId: string;
|
|
2828
|
-
state: "cancelled" | "
|
|
2920
|
+
state: "cancelled" | "cleaning" | "completed" | "dispatching" | "failed" | "reserved" | "running" | "uncertain";
|
|
2829
2921
|
executionId: string | null;
|
|
2830
2922
|
attempt: number;
|
|
2831
2923
|
maximumCostUsd: number;
|