openpond-sdk 0.12.3 → 0.12.5
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-77N3QEIL.js → chunk-25L5IXHA.js} +5 -5
- package/dist/chunk-25L5IXHA.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-ZJ3H3VS7.js → chunk-6GQ3GIAP.js} +7 -5
- package/dist/chunk-6GQ3GIAP.js.map +1 -0
- package/dist/{chunk-36WJTS5V.js → chunk-6KKM2XRR.js} +75 -15
- package/dist/chunk-6KKM2XRR.js.map +1 -0
- package/dist/{chunk-BSK3VLM3.js → chunk-6MFFKAU6.js} +3 -3
- package/dist/chunk-6MFFKAU6.js.map +1 -0
- package/dist/{chunk-UPS52OPZ.js → chunk-6Y67YHVH.js} +8 -8
- package/dist/chunk-6Y67YHVH.js.map +1 -0
- package/dist/{chunk-G6Y6CQYV.js → chunk-AUNNXJQE.js} +4 -4
- package/dist/chunk-AUNNXJQE.js.map +1 -0
- package/dist/{chunk-XYQD4LDI.js → chunk-BGNPKYBJ.js} +3 -3
- package/dist/chunk-BGNPKYBJ.js.map +1 -0
- package/dist/{chunk-K5WRVSOL.js → chunk-BJ5MKLWH.js} +8 -8
- package/dist/chunk-BJ5MKLWH.js.map +1 -0
- package/dist/{chunk-VY5AFE7A.js → chunk-BYNVSWIQ.js} +4 -4
- package/dist/chunk-BYNVSWIQ.js.map +1 -0
- package/dist/{chunk-TCHAWJMB.js → chunk-CAMLCKSF.js} +3 -3
- package/dist/chunk-CAMLCKSF.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-CTVKI3BV.js → chunk-DBKCSQVH.js} +3 -3
- package/dist/chunk-DBKCSQVH.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-EJZ7HOIS.js +8709 -0
- package/dist/chunk-EJZ7HOIS.js.map +1 -0
- package/dist/{chunk-QEIT5VXJ.js → chunk-F4GITADW.js} +4 -4
- package/dist/chunk-F4GITADW.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-FY7NHCEH.js → chunk-FLNWE5ZA.js} +5 -5
- package/dist/chunk-FLNWE5ZA.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-YUKV7RCD.js → chunk-GWGDWZVN.js} +5 -5
- package/dist/chunk-GWGDWZVN.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-CEZEVILU.js → chunk-HJ2AIAKX.js} +1 -1
- package/dist/chunk-HJ2AIAKX.js.map +1 -0
- package/dist/{chunk-BFR3VC25.js → chunk-HOOHJCE7.js} +18 -18
- package/dist/chunk-HOOHJCE7.js.map +1 -0
- package/dist/{chunk-TSMGMA6R.js → chunk-KKL6OICH.js} +3 -3
- package/dist/chunk-KKL6OICH.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-QTJQIHDZ.js → chunk-MJELLVQP.js} +4 -3
- package/dist/chunk-MJELLVQP.js.map +1 -0
- package/dist/{chunk-CK3ONRRC.js → chunk-O54KGALH.js} +3 -3
- package/dist/chunk-O54KGALH.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-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-HKWX624G.js → chunk-QFBSNSUL.js} +5 -5
- package/dist/chunk-QFBSNSUL.js.map +1 -0
- package/dist/{chunk-MLPPSCDQ.js → chunk-RE426Q7Y.js} +4 -4
- package/dist/chunk-RE426Q7Y.js.map +1 -0
- package/dist/{chunk-G53MA5XX.js → chunk-T3MFWGMX.js} +3 -3
- package/dist/chunk-T3MFWGMX.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-6Q4I33CH.js → chunk-VZPNRPXL.js} +5 -5
- package/dist/chunk-VZPNRPXL.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-TMUAKJWX.js → chunk-XACJ63U4.js} +14 -5
- package/dist/chunk-XACJ63U4.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-CWXM57QR.js → chunk-ZFYI4P6C.js} +5 -5
- package/dist/chunk-ZFYI4P6C.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 +22 -21
- 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 +26 -25
- 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 +154 -62
- 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 +19 -18
- package/dist/taskset-drafts.js.map +1 -7
- package/dist/taskset-packages.js +25 -22
- package/dist/taskset-packages.js.map +1 -7
- package/dist/training-bundle.js +44 -19
- 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 +11 -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/hosted-chat.d.ts +28 -0
- package/dist/types/packages/cloud/src/hosted-chat.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 +375 -351
- package/dist/types/packages/sdk/src/candidate-evaluation-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/candidate-evaluations.d.ts +80 -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 +2102 -2014
- package/dist/types/packages/sdk/src/experiment-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-contracts.d.ts +165 -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 +587 -516
- 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 +38 -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 +303 -243
- 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 +33 -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 +69 -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-bound-taskset-authoring.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 +342 -322
- 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 +179 -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 +865 -833
- package/dist/types/packages/sdk/src/model-taskset-runs.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/post-training.d.ts +444 -375
- 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 +745 -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 +4600 -4576
- 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 +14 -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 +2 -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 +75 -71
- package/dist/types/packages/sdk/src/training-qualification-preparation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training.d.ts +1442 -1416
- 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-36WJTS5V.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-77N3QEIL.js.map +0 -7
- package/dist/chunk-AG5DDL7T.js.map +0 -7
- package/dist/chunk-BFR3VC25.js.map +0 -7
- package/dist/chunk-BSK3VLM3.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-CTVKI3BV.js.map +0 -7
- package/dist/chunk-CWXM57QR.js.map +0 -7
- package/dist/chunk-DUKW3R7O.js.map +0 -7
- package/dist/chunk-FHCSDXIA.js +0 -1417
- package/dist/chunk-FHCSDXIA.js.map +0 -7
- package/dist/chunk-FY7NHCEH.js.map +0 -7
- package/dist/chunk-G53MA5XX.js.map +0 -7
- package/dist/chunk-G6Y6CQYV.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-K5WRVSOL.js.map +0 -7
- package/dist/chunk-KAE7AMYE.js.map +0 -7
- package/dist/chunk-MEKWESBF.js.map +0 -7
- package/dist/chunk-MLPPSCDQ.js.map +0 -7
- package/dist/chunk-O4GE6VTR.js.map +0 -7
- package/dist/chunk-P27DQ65Z.js.map +0 -7
- package/dist/chunk-QEIT5VXJ.js.map +0 -7
- package/dist/chunk-QTJQIHDZ.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-TMUAKJWX.js.map +0 -7
- package/dist/chunk-TSMGMA6R.js.map +0 -7
- package/dist/chunk-UPS52OPZ.js.map +0 -7
- package/dist/chunk-VY5AFE7A.js.map +0 -7
- package/dist/chunk-XYQD4LDI.js.map +0 -7
- package/dist/chunk-YCQVVFE5.js.map +0 -7
- package/dist/chunk-YUKV7RCD.js.map +0 -7
- package/dist/chunk-YVUIHH6Z.js.map +0 -7
- package/dist/chunk-ZJ3H3VS7.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>>>;
|
|
@@ -96,6 +96,19 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
96
96
|
maxOutputTokens: z.ZodOptional<z.ZodNumber>;
|
|
97
97
|
temperature: z.ZodOptional<z.ZodNumber>;
|
|
98
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.$strict>>;
|
|
99
112
|
candidate: z.ZodOptional<z.ZodObject<{
|
|
100
113
|
artifactId: z.ZodString;
|
|
101
114
|
contentHash: z.ZodString;
|
|
@@ -117,10 +130,10 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
117
130
|
reasoningEffort: z.ZodOptional<z.ZodEnum<{
|
|
118
131
|
high: "high";
|
|
119
132
|
low: "low";
|
|
133
|
+
max: "max";
|
|
120
134
|
medium: "medium";
|
|
121
|
-
xhigh: "xhigh";
|
|
122
135
|
off: "off";
|
|
123
|
-
|
|
136
|
+
xhigh: "xhigh";
|
|
124
137
|
}>>;
|
|
125
138
|
profileRepositoryId: z.ZodString;
|
|
126
139
|
source: z.ZodObject<{
|
|
@@ -198,10 +211,10 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
198
211
|
}, z.core.$strict>;
|
|
199
212
|
}, z.core.$strict>>;
|
|
200
213
|
split: z.ZodEnum<{
|
|
214
|
+
frozen_eval: "frozen_eval";
|
|
215
|
+
test: "test";
|
|
201
216
|
train: "train";
|
|
202
217
|
validation: "validation";
|
|
203
|
-
test: "test";
|
|
204
|
-
frozen_eval: "frozen_eval";
|
|
205
218
|
}>;
|
|
206
219
|
population: z.ZodArray<z.ZodObject<{
|
|
207
220
|
taskId: z.ZodString;
|
|
@@ -218,8 +231,8 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
218
231
|
configurationHash: z.ZodString;
|
|
219
232
|
output: z.ZodEnum<{
|
|
220
233
|
boolean: "boolean";
|
|
221
|
-
score: "score";
|
|
222
234
|
category: "category";
|
|
235
|
+
score: "score";
|
|
223
236
|
}>;
|
|
224
237
|
categories: z.ZodArray<z.ZodString>;
|
|
225
238
|
}, z.core.$strict>>;
|
|
@@ -228,17 +241,17 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
228
241
|
graderId: z.ZodString;
|
|
229
242
|
fields: z.ZodArray<z.ZodObject<{
|
|
230
243
|
destination: z.ZodEnum<{
|
|
231
|
-
output: "output";
|
|
232
|
-
input: "input";
|
|
233
|
-
expectedOutput: "expectedOutput";
|
|
234
244
|
evaluatorContext: "evaluatorContext";
|
|
245
|
+
expectedOutput: "expectedOutput";
|
|
246
|
+
input: "input";
|
|
247
|
+
output: "output";
|
|
235
248
|
}>;
|
|
236
249
|
path: z.ZodArray<z.ZodString>;
|
|
237
250
|
source: z.ZodEnum<{
|
|
238
|
-
output: "output";
|
|
239
|
-
input: "input";
|
|
240
|
-
expectedOutput: "expectedOutput";
|
|
241
251
|
evaluatorContext: "evaluatorContext";
|
|
252
|
+
expectedOutput: "expectedOutput";
|
|
253
|
+
input: "input";
|
|
254
|
+
output: "output";
|
|
242
255
|
}>;
|
|
243
256
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
244
257
|
}, z.core.$strict>>;
|
|
@@ -316,10 +329,10 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
316
329
|
}, z.core.$strict>;
|
|
317
330
|
}, z.core.$strict>>;
|
|
318
331
|
split: z.ZodEnum<{
|
|
332
|
+
frozen_eval: "frozen_eval";
|
|
333
|
+
test: "test";
|
|
319
334
|
train: "train";
|
|
320
335
|
validation: "validation";
|
|
321
|
-
test: "test";
|
|
322
|
-
frozen_eval: "frozen_eval";
|
|
323
336
|
}>;
|
|
324
337
|
population: z.ZodArray<z.ZodObject<{
|
|
325
338
|
taskId: z.ZodString;
|
|
@@ -336,8 +349,8 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
336
349
|
configurationHash: z.ZodString;
|
|
337
350
|
output: z.ZodEnum<{
|
|
338
351
|
boolean: "boolean";
|
|
339
|
-
score: "score";
|
|
340
352
|
category: "category";
|
|
353
|
+
score: "score";
|
|
341
354
|
}>;
|
|
342
355
|
categories: z.ZodArray<z.ZodString>;
|
|
343
356
|
}, z.core.$strict>>;
|
|
@@ -346,17 +359,17 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
346
359
|
graderId: z.ZodString;
|
|
347
360
|
fields: z.ZodArray<z.ZodObject<{
|
|
348
361
|
destination: z.ZodEnum<{
|
|
349
|
-
output: "output";
|
|
350
|
-
input: "input";
|
|
351
|
-
expectedOutput: "expectedOutput";
|
|
352
362
|
evaluatorContext: "evaluatorContext";
|
|
363
|
+
expectedOutput: "expectedOutput";
|
|
364
|
+
input: "input";
|
|
365
|
+
output: "output";
|
|
353
366
|
}>;
|
|
354
367
|
path: z.ZodArray<z.ZodString>;
|
|
355
368
|
source: z.ZodEnum<{
|
|
356
|
-
output: "output";
|
|
357
|
-
input: "input";
|
|
358
|
-
expectedOutput: "expectedOutput";
|
|
359
369
|
evaluatorContext: "evaluatorContext";
|
|
370
|
+
expectedOutput: "expectedOutput";
|
|
371
|
+
input: "input";
|
|
372
|
+
output: "output";
|
|
360
373
|
}>;
|
|
361
374
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
362
375
|
}, z.core.$strict>>;
|
|
@@ -384,17 +397,17 @@ export declare const ExperimentEvaluationScheduleInputSchema: z.ZodObject<{
|
|
|
384
397
|
contentHash: z.ZodString;
|
|
385
398
|
mappings: z.ZodArray<z.ZodObject<{
|
|
386
399
|
destination: z.ZodEnum<{
|
|
400
|
+
evaluatorContext: "evaluatorContext";
|
|
401
|
+
expectedOutput: "expectedOutput";
|
|
387
402
|
input: "input";
|
|
388
403
|
output: "output";
|
|
389
|
-
expectedOutput: "expectedOutput";
|
|
390
|
-
evaluatorContext: "evaluatorContext";
|
|
391
404
|
}>;
|
|
392
405
|
path: z.ZodArray<z.ZodString>;
|
|
393
406
|
source: z.ZodEnum<{
|
|
407
|
+
evaluatorContext: "evaluatorContext";
|
|
408
|
+
expectedOutput: "expectedOutput";
|
|
394
409
|
input: "input";
|
|
395
410
|
output: "output";
|
|
396
|
-
expectedOutput: "expectedOutput";
|
|
397
|
-
evaluatorContext: "evaluatorContext";
|
|
398
411
|
}>;
|
|
399
412
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
400
413
|
}, z.core.$strict>>;
|
|
@@ -423,12 +436,12 @@ export declare const ExperimentEvaluationFireSchema: z.ZodObject<{
|
|
|
423
436
|
operationId: z.ZodString;
|
|
424
437
|
state: z.ZodEnum<{
|
|
425
438
|
cancelled: "cancelled";
|
|
426
|
-
|
|
427
|
-
running: "running";
|
|
439
|
+
cleaning: "cleaning";
|
|
428
440
|
completed: "completed";
|
|
429
441
|
dispatching: "dispatching";
|
|
430
|
-
|
|
442
|
+
failed: "failed";
|
|
431
443
|
reserved: "reserved";
|
|
444
|
+
running: "running";
|
|
432
445
|
uncertain: "uncertain";
|
|
433
446
|
}>;
|
|
434
447
|
executionId: z.ZodNullable<z.ZodString>;
|
|
@@ -518,9 +531,9 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
518
531
|
}, z.core.$strict>>;
|
|
519
532
|
messages: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
520
533
|
role: z.ZodEnum<{
|
|
534
|
+
assistant: "assistant";
|
|
521
535
|
system: "system";
|
|
522
536
|
user: "user";
|
|
523
|
-
assistant: "assistant";
|
|
524
537
|
}>;
|
|
525
538
|
content: z.ZodString;
|
|
526
539
|
}, z.core.$strict>>>;
|
|
@@ -535,6 +548,19 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
535
548
|
maxOutputTokens: z.ZodOptional<z.ZodNumber>;
|
|
536
549
|
temperature: z.ZodOptional<z.ZodNumber>;
|
|
537
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.$strict>>;
|
|
538
564
|
candidate: z.ZodOptional<z.ZodObject<{
|
|
539
565
|
artifactId: z.ZodString;
|
|
540
566
|
contentHash: z.ZodString;
|
|
@@ -556,10 +582,10 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
556
582
|
reasoningEffort: z.ZodOptional<z.ZodEnum<{
|
|
557
583
|
high: "high";
|
|
558
584
|
low: "low";
|
|
585
|
+
max: "max";
|
|
559
586
|
medium: "medium";
|
|
560
|
-
xhigh: "xhigh";
|
|
561
587
|
off: "off";
|
|
562
|
-
|
|
588
|
+
xhigh: "xhigh";
|
|
563
589
|
}>>;
|
|
564
590
|
profileRepositoryId: z.ZodString;
|
|
565
591
|
source: z.ZodObject<{
|
|
@@ -637,10 +663,10 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
637
663
|
}, z.core.$strict>;
|
|
638
664
|
}, z.core.$strict>>;
|
|
639
665
|
split: z.ZodEnum<{
|
|
666
|
+
frozen_eval: "frozen_eval";
|
|
667
|
+
test: "test";
|
|
640
668
|
train: "train";
|
|
641
669
|
validation: "validation";
|
|
642
|
-
test: "test";
|
|
643
|
-
frozen_eval: "frozen_eval";
|
|
644
670
|
}>;
|
|
645
671
|
population: z.ZodArray<z.ZodObject<{
|
|
646
672
|
taskId: z.ZodString;
|
|
@@ -657,8 +683,8 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
657
683
|
configurationHash: z.ZodString;
|
|
658
684
|
output: z.ZodEnum<{
|
|
659
685
|
boolean: "boolean";
|
|
660
|
-
score: "score";
|
|
661
686
|
category: "category";
|
|
687
|
+
score: "score";
|
|
662
688
|
}>;
|
|
663
689
|
categories: z.ZodArray<z.ZodString>;
|
|
664
690
|
}, z.core.$strict>>;
|
|
@@ -667,17 +693,17 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
667
693
|
graderId: z.ZodString;
|
|
668
694
|
fields: z.ZodArray<z.ZodObject<{
|
|
669
695
|
destination: z.ZodEnum<{
|
|
670
|
-
output: "output";
|
|
671
|
-
input: "input";
|
|
672
|
-
expectedOutput: "expectedOutput";
|
|
673
696
|
evaluatorContext: "evaluatorContext";
|
|
697
|
+
expectedOutput: "expectedOutput";
|
|
698
|
+
input: "input";
|
|
699
|
+
output: "output";
|
|
674
700
|
}>;
|
|
675
701
|
path: z.ZodArray<z.ZodString>;
|
|
676
702
|
source: z.ZodEnum<{
|
|
677
|
-
output: "output";
|
|
678
|
-
input: "input";
|
|
679
|
-
expectedOutput: "expectedOutput";
|
|
680
703
|
evaluatorContext: "evaluatorContext";
|
|
704
|
+
expectedOutput: "expectedOutput";
|
|
705
|
+
input: "input";
|
|
706
|
+
output: "output";
|
|
681
707
|
}>;
|
|
682
708
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
683
709
|
}, z.core.$strict>>;
|
|
@@ -755,10 +781,10 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
755
781
|
}, z.core.$strict>;
|
|
756
782
|
}, z.core.$strict>>;
|
|
757
783
|
split: z.ZodEnum<{
|
|
784
|
+
frozen_eval: "frozen_eval";
|
|
785
|
+
test: "test";
|
|
758
786
|
train: "train";
|
|
759
787
|
validation: "validation";
|
|
760
|
-
test: "test";
|
|
761
|
-
frozen_eval: "frozen_eval";
|
|
762
788
|
}>;
|
|
763
789
|
population: z.ZodArray<z.ZodObject<{
|
|
764
790
|
taskId: z.ZodString;
|
|
@@ -775,8 +801,8 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
775
801
|
configurationHash: z.ZodString;
|
|
776
802
|
output: z.ZodEnum<{
|
|
777
803
|
boolean: "boolean";
|
|
778
|
-
score: "score";
|
|
779
804
|
category: "category";
|
|
805
|
+
score: "score";
|
|
780
806
|
}>;
|
|
781
807
|
categories: z.ZodArray<z.ZodString>;
|
|
782
808
|
}, z.core.$strict>>;
|
|
@@ -785,17 +811,17 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
785
811
|
graderId: z.ZodString;
|
|
786
812
|
fields: z.ZodArray<z.ZodObject<{
|
|
787
813
|
destination: z.ZodEnum<{
|
|
788
|
-
output: "output";
|
|
789
|
-
input: "input";
|
|
790
|
-
expectedOutput: "expectedOutput";
|
|
791
814
|
evaluatorContext: "evaluatorContext";
|
|
815
|
+
expectedOutput: "expectedOutput";
|
|
816
|
+
input: "input";
|
|
817
|
+
output: "output";
|
|
792
818
|
}>;
|
|
793
819
|
path: z.ZodArray<z.ZodString>;
|
|
794
820
|
source: z.ZodEnum<{
|
|
795
|
-
output: "output";
|
|
796
|
-
input: "input";
|
|
797
|
-
expectedOutput: "expectedOutput";
|
|
798
821
|
evaluatorContext: "evaluatorContext";
|
|
822
|
+
expectedOutput: "expectedOutput";
|
|
823
|
+
input: "input";
|
|
824
|
+
output: "output";
|
|
799
825
|
}>;
|
|
800
826
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
801
827
|
}, z.core.$strict>>;
|
|
@@ -823,17 +849,17 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
823
849
|
contentHash: z.ZodString;
|
|
824
850
|
mappings: z.ZodArray<z.ZodObject<{
|
|
825
851
|
destination: z.ZodEnum<{
|
|
852
|
+
evaluatorContext: "evaluatorContext";
|
|
853
|
+
expectedOutput: "expectedOutput";
|
|
826
854
|
input: "input";
|
|
827
855
|
output: "output";
|
|
828
|
-
expectedOutput: "expectedOutput";
|
|
829
|
-
evaluatorContext: "evaluatorContext";
|
|
830
856
|
}>;
|
|
831
857
|
path: z.ZodArray<z.ZodString>;
|
|
832
858
|
source: z.ZodEnum<{
|
|
859
|
+
evaluatorContext: "evaluatorContext";
|
|
860
|
+
expectedOutput: "expectedOutput";
|
|
833
861
|
input: "input";
|
|
834
862
|
output: "output";
|
|
835
|
-
expectedOutput: "expectedOutput";
|
|
836
|
-
evaluatorContext: "evaluatorContext";
|
|
837
863
|
}>;
|
|
838
864
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
839
865
|
}, z.core.$strict>>;
|
|
@@ -853,8 +879,8 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
853
879
|
}, z.core.$strict>], "kind">;
|
|
854
880
|
state: z.ZodEnum<{
|
|
855
881
|
blocked: "blocked";
|
|
856
|
-
scheduled: "scheduled";
|
|
857
882
|
disabled: "disabled";
|
|
883
|
+
scheduled: "scheduled";
|
|
858
884
|
}>;
|
|
859
885
|
nextRunAt: z.ZodNullable<z.ZodString>;
|
|
860
886
|
maximumTotalSpendUsd: z.ZodNumber;
|
|
@@ -873,12 +899,12 @@ export declare const ExperimentEvaluationScheduleSchema: z.ZodObject<{
|
|
|
873
899
|
operationId: z.ZodString;
|
|
874
900
|
state: z.ZodEnum<{
|
|
875
901
|
cancelled: "cancelled";
|
|
876
|
-
|
|
877
|
-
running: "running";
|
|
902
|
+
cleaning: "cleaning";
|
|
878
903
|
completed: "completed";
|
|
879
904
|
dispatching: "dispatching";
|
|
880
|
-
|
|
905
|
+
failed: "failed";
|
|
881
906
|
reserved: "reserved";
|
|
907
|
+
running: "running";
|
|
882
908
|
uncertain: "uncertain";
|
|
883
909
|
}>;
|
|
884
910
|
executionId: z.ZodNullable<z.ZodString>;
|
|
@@ -913,6 +939,13 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
913
939
|
operationId: string;
|
|
914
940
|
teamId: string;
|
|
915
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;
|
|
916
949
|
taskset: {
|
|
917
950
|
id: string;
|
|
918
951
|
contentHash: string;
|
|
@@ -921,10 +954,12 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
921
954
|
policy: {
|
|
922
955
|
kind: "hosted_chat";
|
|
923
956
|
modelId: string;
|
|
924
|
-
maxOutputTokens: number;
|
|
925
|
-
temperature: number;
|
|
926
|
-
topP: number;
|
|
927
957
|
localRuntime?: {
|
|
958
|
+
providerId: "custom-openai-compatible";
|
|
959
|
+
configurationHash: string;
|
|
960
|
+
maximumRequestCostUsd: number;
|
|
961
|
+
requestTimeoutMs: number;
|
|
962
|
+
} | {
|
|
928
963
|
providerId: "claude-code";
|
|
929
964
|
configurationHash: string;
|
|
930
965
|
executableHash: string;
|
|
@@ -941,11 +976,6 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
941
976
|
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
942
977
|
maximumRequestCostUsd: number;
|
|
943
978
|
requestTimeoutMs: number;
|
|
944
|
-
} | {
|
|
945
|
-
providerId: "custom-openai-compatible";
|
|
946
|
-
configurationHash: string;
|
|
947
|
-
maximumRequestCostUsd: number;
|
|
948
|
-
requestTimeoutMs: number;
|
|
949
979
|
} | undefined;
|
|
950
980
|
candidate?: {
|
|
951
981
|
artifactId: string;
|
|
@@ -967,12 +997,47 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
967
997
|
sourcePackageHash: string;
|
|
968
998
|
} | undefined;
|
|
969
999
|
messages?: {
|
|
970
|
-
role: "
|
|
1000
|
+
role: "assistant" | "system" | "user";
|
|
971
1001
|
content: string;
|
|
972
1002
|
}[] | undefined;
|
|
1003
|
+
maxOutputTokens: number;
|
|
1004
|
+
temperature: number;
|
|
1005
|
+
topP: number;
|
|
1006
|
+
} | {
|
|
1007
|
+
kind: "fixture";
|
|
973
1008
|
} | {
|
|
974
1009
|
kind: "hosted_harness";
|
|
975
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;
|
|
976
1041
|
profileRepositoryId: string;
|
|
977
1042
|
source: {
|
|
978
1043
|
profileId: string;
|
|
@@ -1020,7 +1085,35 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1020
1085
|
};
|
|
1021
1086
|
declaredProfileCatalogHash: string | null;
|
|
1022
1087
|
protectedProfileClosureHash: string;
|
|
1023
|
-
|
|
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";
|
|
1024
1117
|
population: {
|
|
1025
1118
|
taskId: string;
|
|
1026
1119
|
seed: string;
|
|
@@ -1034,16 +1127,16 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1034
1127
|
};
|
|
1035
1128
|
feedbackKey: string;
|
|
1036
1129
|
configurationHash: string;
|
|
1037
|
-
output: "boolean" | "
|
|
1130
|
+
output: "boolean" | "category" | "score";
|
|
1038
1131
|
categories: string[];
|
|
1039
1132
|
}[];
|
|
1040
1133
|
fieldMappingsHash: string;
|
|
1041
1134
|
fieldMappings: {
|
|
1042
1135
|
graderId: string;
|
|
1043
1136
|
fields: {
|
|
1044
|
-
destination: "
|
|
1137
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1045
1138
|
path: string[];
|
|
1046
|
-
source: "
|
|
1139
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1047
1140
|
sourcePath: string[];
|
|
1048
1141
|
}[];
|
|
1049
1142
|
}[];
|
|
@@ -1054,34 +1147,6 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1054
1147
|
requireComplete: boolean;
|
|
1055
1148
|
};
|
|
1056
1149
|
contentHash: string;
|
|
1057
|
-
recordedOrigin?: {
|
|
1058
|
-
execution: {
|
|
1059
|
-
id: string;
|
|
1060
|
-
contentHash: string;
|
|
1061
|
-
};
|
|
1062
|
-
originalEvidence: {
|
|
1063
|
-
id: string;
|
|
1064
|
-
contentHash: string;
|
|
1065
|
-
};
|
|
1066
|
-
members: {
|
|
1067
|
-
caseId: string;
|
|
1068
|
-
seed: string;
|
|
1069
|
-
fixtureId: string;
|
|
1070
|
-
sourceId: string;
|
|
1071
|
-
snapshotHash: string;
|
|
1072
|
-
boundaryRevisionHash: string;
|
|
1073
|
-
}[];
|
|
1074
|
-
} | undefined;
|
|
1075
|
-
gradingSource?: {
|
|
1076
|
-
pass: {
|
|
1077
|
-
id: string;
|
|
1078
|
-
contentHash: string;
|
|
1079
|
-
};
|
|
1080
|
-
evidence: {
|
|
1081
|
-
id: string;
|
|
1082
|
-
contentHash: string;
|
|
1083
|
-
};
|
|
1084
|
-
} | undefined;
|
|
1085
1150
|
} | undefined;
|
|
1086
1151
|
sourceCandidate?: {
|
|
1087
1152
|
candidateId: string;
|
|
@@ -1096,28 +1161,6 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1096
1161
|
};
|
|
1097
1162
|
modelConfigurationHash: string;
|
|
1098
1163
|
packageHash: string;
|
|
1099
|
-
maxOutputTokens?: number | undefined;
|
|
1100
|
-
temperature?: number | undefined;
|
|
1101
|
-
topP?: number | undefined;
|
|
1102
|
-
candidate?: {
|
|
1103
|
-
artifactId: string;
|
|
1104
|
-
contentHash: string;
|
|
1105
|
-
preparationId: string;
|
|
1106
|
-
dispatchId: string;
|
|
1107
|
-
baseProfileId: string;
|
|
1108
|
-
workerImageDigest: string;
|
|
1109
|
-
} | undefined;
|
|
1110
|
-
sourceCandidate?: {
|
|
1111
|
-
candidateId: string;
|
|
1112
|
-
frozenRevision: number;
|
|
1113
|
-
frozenHash: string;
|
|
1114
|
-
harnessRelease: {
|
|
1115
|
-
id: string;
|
|
1116
|
-
contentHash: string;
|
|
1117
|
-
};
|
|
1118
|
-
protectedClosureHash: string;
|
|
1119
|
-
} | undefined;
|
|
1120
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
1121
1164
|
externalDatasetBinding?: {
|
|
1122
1165
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
1123
1166
|
dataset: {
|
|
@@ -1141,7 +1184,35 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1141
1184
|
};
|
|
1142
1185
|
declaredProfileCatalogHash: string | null;
|
|
1143
1186
|
protectedProfileClosureHash: string;
|
|
1144
|
-
|
|
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";
|
|
1145
1216
|
population: {
|
|
1146
1217
|
taskId: string;
|
|
1147
1218
|
seed: string;
|
|
@@ -1155,16 +1226,16 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1155
1226
|
};
|
|
1156
1227
|
feedbackKey: string;
|
|
1157
1228
|
configurationHash: string;
|
|
1158
|
-
output: "boolean" | "
|
|
1229
|
+
output: "boolean" | "category" | "score";
|
|
1159
1230
|
categories: string[];
|
|
1160
1231
|
}[];
|
|
1161
1232
|
fieldMappingsHash: string;
|
|
1162
1233
|
fieldMappings: {
|
|
1163
1234
|
graderId: string;
|
|
1164
1235
|
fields: {
|
|
1165
|
-
destination: "
|
|
1236
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1166
1237
|
path: string[];
|
|
1167
|
-
source: "
|
|
1238
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1168
1239
|
sourcePath: string[];
|
|
1169
1240
|
}[];
|
|
1170
1241
|
}[];
|
|
@@ -1175,37 +1246,7 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1175
1246
|
requireComplete: boolean;
|
|
1176
1247
|
};
|
|
1177
1248
|
contentHash: string;
|
|
1178
|
-
recordedOrigin?: {
|
|
1179
|
-
execution: {
|
|
1180
|
-
id: string;
|
|
1181
|
-
contentHash: string;
|
|
1182
|
-
};
|
|
1183
|
-
originalEvidence: {
|
|
1184
|
-
id: string;
|
|
1185
|
-
contentHash: string;
|
|
1186
|
-
};
|
|
1187
|
-
members: {
|
|
1188
|
-
caseId: string;
|
|
1189
|
-
seed: string;
|
|
1190
|
-
fixtureId: string;
|
|
1191
|
-
sourceId: string;
|
|
1192
|
-
snapshotHash: string;
|
|
1193
|
-
boundaryRevisionHash: string;
|
|
1194
|
-
}[];
|
|
1195
|
-
} | undefined;
|
|
1196
|
-
gradingSource?: {
|
|
1197
|
-
pass: {
|
|
1198
|
-
id: string;
|
|
1199
|
-
contentHash: string;
|
|
1200
|
-
};
|
|
1201
|
-
evidence: {
|
|
1202
|
-
id: string;
|
|
1203
|
-
contentHash: string;
|
|
1204
|
-
};
|
|
1205
|
-
} | undefined;
|
|
1206
1249
|
} | undefined;
|
|
1207
|
-
} | {
|
|
1208
|
-
kind: "fixture";
|
|
1209
1250
|
};
|
|
1210
1251
|
population: {
|
|
1211
1252
|
receiptId: string;
|
|
@@ -1213,13 +1254,6 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1213
1254
|
seed: string;
|
|
1214
1255
|
fixtureId: string | null;
|
|
1215
1256
|
}[];
|
|
1216
|
-
name?: string | undefined;
|
|
1217
|
-
project?: {
|
|
1218
|
-
id: string;
|
|
1219
|
-
revision: number;
|
|
1220
|
-
contentHash: string;
|
|
1221
|
-
targetId: string | null;
|
|
1222
|
-
} | undefined;
|
|
1223
1257
|
};
|
|
1224
1258
|
maximumCostUsd: number;
|
|
1225
1259
|
graders?: {
|
|
@@ -1227,9 +1261,9 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1227
1261
|
version: string;
|
|
1228
1262
|
contentHash: string;
|
|
1229
1263
|
mappings: {
|
|
1230
|
-
destination: "
|
|
1264
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1231
1265
|
path: string[];
|
|
1232
|
-
source: "
|
|
1266
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1233
1267
|
sourcePath: string[];
|
|
1234
1268
|
}[];
|
|
1235
1269
|
}[] | undefined;
|
|
@@ -1246,7 +1280,7 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1246
1280
|
timeZone: string;
|
|
1247
1281
|
};
|
|
1248
1282
|
};
|
|
1249
|
-
state: "blocked" | "
|
|
1283
|
+
state: "blocked" | "disabled" | "scheduled";
|
|
1250
1284
|
nextRunAt: string | null;
|
|
1251
1285
|
maximumTotalSpendUsd: number;
|
|
1252
1286
|
maximumAttempts: number;
|
|
@@ -1262,7 +1296,7 @@ export declare function verifyExperimentEvaluationSchedule(raw: unknown): {
|
|
|
1262
1296
|
coalescedCount: number;
|
|
1263
1297
|
configurationHash: string;
|
|
1264
1298
|
operationId: string;
|
|
1265
|
-
state: "cancelled" | "
|
|
1299
|
+
state: "cancelled" | "cleaning" | "completed" | "dispatching" | "failed" | "reserved" | "running" | "uncertain";
|
|
1266
1300
|
executionId: string | null;
|
|
1267
1301
|
attempt: number;
|
|
1268
1302
|
maximumCostUsd: number;
|
|
@@ -1291,6 +1325,13 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1291
1325
|
operationId: string;
|
|
1292
1326
|
teamId: string;
|
|
1293
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;
|
|
1294
1335
|
taskset: {
|
|
1295
1336
|
id: string;
|
|
1296
1337
|
contentHash: string;
|
|
@@ -1299,10 +1340,12 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1299
1340
|
policy: {
|
|
1300
1341
|
kind: "hosted_chat";
|
|
1301
1342
|
modelId: string;
|
|
1302
|
-
maxOutputTokens: number;
|
|
1303
|
-
temperature: number;
|
|
1304
|
-
topP: number;
|
|
1305
1343
|
localRuntime?: {
|
|
1344
|
+
providerId: "custom-openai-compatible";
|
|
1345
|
+
configurationHash: string;
|
|
1346
|
+
maximumRequestCostUsd: number;
|
|
1347
|
+
requestTimeoutMs: number;
|
|
1348
|
+
} | {
|
|
1306
1349
|
providerId: "claude-code";
|
|
1307
1350
|
configurationHash: string;
|
|
1308
1351
|
executableHash: string;
|
|
@@ -1319,11 +1362,6 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1319
1362
|
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
1320
1363
|
maximumRequestCostUsd: number;
|
|
1321
1364
|
requestTimeoutMs: number;
|
|
1322
|
-
} | {
|
|
1323
|
-
providerId: "custom-openai-compatible";
|
|
1324
|
-
configurationHash: string;
|
|
1325
|
-
maximumRequestCostUsd: number;
|
|
1326
|
-
requestTimeoutMs: number;
|
|
1327
1365
|
} | undefined;
|
|
1328
1366
|
candidate?: {
|
|
1329
1367
|
artifactId: string;
|
|
@@ -1345,12 +1383,47 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1345
1383
|
sourcePackageHash: string;
|
|
1346
1384
|
} | undefined;
|
|
1347
1385
|
messages?: {
|
|
1348
|
-
role: "
|
|
1386
|
+
role: "assistant" | "system" | "user";
|
|
1349
1387
|
content: string;
|
|
1350
1388
|
}[] | undefined;
|
|
1389
|
+
maxOutputTokens: number;
|
|
1390
|
+
temperature: number;
|
|
1391
|
+
topP: number;
|
|
1392
|
+
} | {
|
|
1393
|
+
kind: "fixture";
|
|
1351
1394
|
} | {
|
|
1352
1395
|
kind: "hosted_harness";
|
|
1353
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;
|
|
1354
1427
|
profileRepositoryId: string;
|
|
1355
1428
|
source: {
|
|
1356
1429
|
profileId: string;
|
|
@@ -1398,7 +1471,35 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1398
1471
|
};
|
|
1399
1472
|
declaredProfileCatalogHash: string | null;
|
|
1400
1473
|
protectedProfileClosureHash: string;
|
|
1401
|
-
|
|
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";
|
|
1402
1503
|
population: {
|
|
1403
1504
|
taskId: string;
|
|
1404
1505
|
seed: string;
|
|
@@ -1412,16 +1513,16 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1412
1513
|
};
|
|
1413
1514
|
feedbackKey: string;
|
|
1414
1515
|
configurationHash: string;
|
|
1415
|
-
output: "boolean" | "
|
|
1516
|
+
output: "boolean" | "category" | "score";
|
|
1416
1517
|
categories: string[];
|
|
1417
1518
|
}[];
|
|
1418
1519
|
fieldMappingsHash: string;
|
|
1419
1520
|
fieldMappings: {
|
|
1420
1521
|
graderId: string;
|
|
1421
1522
|
fields: {
|
|
1422
|
-
destination: "
|
|
1523
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1423
1524
|
path: string[];
|
|
1424
|
-
source: "
|
|
1525
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1425
1526
|
sourcePath: string[];
|
|
1426
1527
|
}[];
|
|
1427
1528
|
}[];
|
|
@@ -1432,34 +1533,6 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1432
1533
|
requireComplete: boolean;
|
|
1433
1534
|
};
|
|
1434
1535
|
contentHash: string;
|
|
1435
|
-
recordedOrigin?: {
|
|
1436
|
-
execution: {
|
|
1437
|
-
id: string;
|
|
1438
|
-
contentHash: string;
|
|
1439
|
-
};
|
|
1440
|
-
originalEvidence: {
|
|
1441
|
-
id: string;
|
|
1442
|
-
contentHash: string;
|
|
1443
|
-
};
|
|
1444
|
-
members: {
|
|
1445
|
-
caseId: string;
|
|
1446
|
-
seed: string;
|
|
1447
|
-
fixtureId: string;
|
|
1448
|
-
sourceId: string;
|
|
1449
|
-
snapshotHash: string;
|
|
1450
|
-
boundaryRevisionHash: string;
|
|
1451
|
-
}[];
|
|
1452
|
-
} | undefined;
|
|
1453
|
-
gradingSource?: {
|
|
1454
|
-
pass: {
|
|
1455
|
-
id: string;
|
|
1456
|
-
contentHash: string;
|
|
1457
|
-
};
|
|
1458
|
-
evidence: {
|
|
1459
|
-
id: string;
|
|
1460
|
-
contentHash: string;
|
|
1461
|
-
};
|
|
1462
|
-
} | undefined;
|
|
1463
1536
|
} | undefined;
|
|
1464
1537
|
sourceCandidate?: {
|
|
1465
1538
|
candidateId: string;
|
|
@@ -1474,28 +1547,6 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1474
1547
|
};
|
|
1475
1548
|
modelConfigurationHash: string;
|
|
1476
1549
|
packageHash: string;
|
|
1477
|
-
maxOutputTokens?: number | undefined;
|
|
1478
|
-
temperature?: number | undefined;
|
|
1479
|
-
topP?: number | undefined;
|
|
1480
|
-
candidate?: {
|
|
1481
|
-
artifactId: string;
|
|
1482
|
-
contentHash: string;
|
|
1483
|
-
preparationId: string;
|
|
1484
|
-
dispatchId: string;
|
|
1485
|
-
baseProfileId: string;
|
|
1486
|
-
workerImageDigest: string;
|
|
1487
|
-
} | undefined;
|
|
1488
|
-
sourceCandidate?: {
|
|
1489
|
-
candidateId: string;
|
|
1490
|
-
frozenRevision: number;
|
|
1491
|
-
frozenHash: string;
|
|
1492
|
-
harnessRelease: {
|
|
1493
|
-
id: string;
|
|
1494
|
-
contentHash: string;
|
|
1495
|
-
};
|
|
1496
|
-
protectedClosureHash: string;
|
|
1497
|
-
} | undefined;
|
|
1498
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
1499
1550
|
externalDatasetBinding?: {
|
|
1500
1551
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
1501
1552
|
dataset: {
|
|
@@ -1519,7 +1570,35 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1519
1570
|
};
|
|
1520
1571
|
declaredProfileCatalogHash: string | null;
|
|
1521
1572
|
protectedProfileClosureHash: string;
|
|
1522
|
-
|
|
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";
|
|
1523
1602
|
population: {
|
|
1524
1603
|
taskId: string;
|
|
1525
1604
|
seed: string;
|
|
@@ -1533,16 +1612,16 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1533
1612
|
};
|
|
1534
1613
|
feedbackKey: string;
|
|
1535
1614
|
configurationHash: string;
|
|
1536
|
-
output: "boolean" | "
|
|
1615
|
+
output: "boolean" | "category" | "score";
|
|
1537
1616
|
categories: string[];
|
|
1538
1617
|
}[];
|
|
1539
1618
|
fieldMappingsHash: string;
|
|
1540
1619
|
fieldMappings: {
|
|
1541
1620
|
graderId: string;
|
|
1542
1621
|
fields: {
|
|
1543
|
-
destination: "
|
|
1622
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1544
1623
|
path: string[];
|
|
1545
|
-
source: "
|
|
1624
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1546
1625
|
sourcePath: string[];
|
|
1547
1626
|
}[];
|
|
1548
1627
|
}[];
|
|
@@ -1553,37 +1632,7 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1553
1632
|
requireComplete: boolean;
|
|
1554
1633
|
};
|
|
1555
1634
|
contentHash: string;
|
|
1556
|
-
recordedOrigin?: {
|
|
1557
|
-
execution: {
|
|
1558
|
-
id: string;
|
|
1559
|
-
contentHash: string;
|
|
1560
|
-
};
|
|
1561
|
-
originalEvidence: {
|
|
1562
|
-
id: string;
|
|
1563
|
-
contentHash: string;
|
|
1564
|
-
};
|
|
1565
|
-
members: {
|
|
1566
|
-
caseId: string;
|
|
1567
|
-
seed: string;
|
|
1568
|
-
fixtureId: string;
|
|
1569
|
-
sourceId: string;
|
|
1570
|
-
snapshotHash: string;
|
|
1571
|
-
boundaryRevisionHash: string;
|
|
1572
|
-
}[];
|
|
1573
|
-
} | undefined;
|
|
1574
|
-
gradingSource?: {
|
|
1575
|
-
pass: {
|
|
1576
|
-
id: string;
|
|
1577
|
-
contentHash: string;
|
|
1578
|
-
};
|
|
1579
|
-
evidence: {
|
|
1580
|
-
id: string;
|
|
1581
|
-
contentHash: string;
|
|
1582
|
-
};
|
|
1583
|
-
} | undefined;
|
|
1584
1635
|
} | undefined;
|
|
1585
|
-
} | {
|
|
1586
|
-
kind: "fixture";
|
|
1587
1636
|
};
|
|
1588
1637
|
population: {
|
|
1589
1638
|
receiptId: string;
|
|
@@ -1591,13 +1640,6 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1591
1640
|
seed: string;
|
|
1592
1641
|
fixtureId: string | null;
|
|
1593
1642
|
}[];
|
|
1594
|
-
name?: string | undefined;
|
|
1595
|
-
project?: {
|
|
1596
|
-
id: string;
|
|
1597
|
-
revision: number;
|
|
1598
|
-
contentHash: string;
|
|
1599
|
-
targetId: string | null;
|
|
1600
|
-
} | undefined;
|
|
1601
1643
|
};
|
|
1602
1644
|
maximumCostUsd: number;
|
|
1603
1645
|
graders?: {
|
|
@@ -1605,9 +1647,9 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1605
1647
|
version: string;
|
|
1606
1648
|
contentHash: string;
|
|
1607
1649
|
mappings: {
|
|
1608
|
-
destination: "
|
|
1650
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1609
1651
|
path: string[];
|
|
1610
|
-
source: "
|
|
1652
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
1611
1653
|
sourcePath: string[];
|
|
1612
1654
|
}[];
|
|
1613
1655
|
}[] | undefined;
|
|
@@ -1624,7 +1666,7 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1624
1666
|
timeZone: string;
|
|
1625
1667
|
};
|
|
1626
1668
|
};
|
|
1627
|
-
state: "blocked" | "
|
|
1669
|
+
state: "blocked" | "disabled" | "scheduled";
|
|
1628
1670
|
nextRunAt: string | null;
|
|
1629
1671
|
maximumTotalSpendUsd: number;
|
|
1630
1672
|
maximumAttempts: number;
|
|
@@ -1640,7 +1682,7 @@ export declare function sealExperimentEvaluationSchedule(value: Omit<ExperimentE
|
|
|
1640
1682
|
coalescedCount: number;
|
|
1641
1683
|
configurationHash: string;
|
|
1642
1684
|
operationId: string;
|
|
1643
|
-
state: "cancelled" | "
|
|
1685
|
+
state: "cancelled" | "cleaning" | "completed" | "dispatching" | "failed" | "reserved" | "running" | "uncertain";
|
|
1644
1686
|
executionId: string | null;
|
|
1645
1687
|
attempt: number;
|
|
1646
1688
|
maximumCostUsd: number;
|
|
@@ -1659,8 +1701,8 @@ export declare const ExperimentEvaluationScheduleControlSchema: z.ZodObject<{
|
|
|
1659
1701
|
operationId: z.ZodString;
|
|
1660
1702
|
expectedRevision: z.ZodNumber;
|
|
1661
1703
|
action: z.ZodEnum<{
|
|
1662
|
-
pause: "pause";
|
|
1663
1704
|
cancel_active: "cancel_active";
|
|
1705
|
+
pause: "pause";
|
|
1664
1706
|
retry_active: "retry_active";
|
|
1665
1707
|
}>;
|
|
1666
1708
|
}, z.core.$strict>;
|
|
@@ -1744,9 +1786,9 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1744
1786
|
}, z.core.$strict>>;
|
|
1745
1787
|
messages: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
1746
1788
|
role: z.ZodEnum<{
|
|
1789
|
+
assistant: "assistant";
|
|
1747
1790
|
system: "system";
|
|
1748
1791
|
user: "user";
|
|
1749
|
-
assistant: "assistant";
|
|
1750
1792
|
}>;
|
|
1751
1793
|
content: z.ZodString;
|
|
1752
1794
|
}, z.core.$strict>>>;
|
|
@@ -1761,6 +1803,19 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1761
1803
|
maxOutputTokens: z.ZodOptional<z.ZodNumber>;
|
|
1762
1804
|
temperature: z.ZodOptional<z.ZodNumber>;
|
|
1763
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.$strict>>;
|
|
1764
1819
|
candidate: z.ZodOptional<z.ZodObject<{
|
|
1765
1820
|
artifactId: z.ZodString;
|
|
1766
1821
|
contentHash: z.ZodString;
|
|
@@ -1782,10 +1837,10 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1782
1837
|
reasoningEffort: z.ZodOptional<z.ZodEnum<{
|
|
1783
1838
|
high: "high";
|
|
1784
1839
|
low: "low";
|
|
1840
|
+
max: "max";
|
|
1785
1841
|
medium: "medium";
|
|
1786
|
-
xhigh: "xhigh";
|
|
1787
1842
|
off: "off";
|
|
1788
|
-
|
|
1843
|
+
xhigh: "xhigh";
|
|
1789
1844
|
}>>;
|
|
1790
1845
|
profileRepositoryId: z.ZodString;
|
|
1791
1846
|
source: z.ZodObject<{
|
|
@@ -1863,10 +1918,10 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1863
1918
|
}, z.core.$strict>;
|
|
1864
1919
|
}, z.core.$strict>>;
|
|
1865
1920
|
split: z.ZodEnum<{
|
|
1921
|
+
frozen_eval: "frozen_eval";
|
|
1922
|
+
test: "test";
|
|
1866
1923
|
train: "train";
|
|
1867
1924
|
validation: "validation";
|
|
1868
|
-
test: "test";
|
|
1869
|
-
frozen_eval: "frozen_eval";
|
|
1870
1925
|
}>;
|
|
1871
1926
|
population: z.ZodArray<z.ZodObject<{
|
|
1872
1927
|
taskId: z.ZodString;
|
|
@@ -1883,8 +1938,8 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1883
1938
|
configurationHash: z.ZodString;
|
|
1884
1939
|
output: z.ZodEnum<{
|
|
1885
1940
|
boolean: "boolean";
|
|
1886
|
-
score: "score";
|
|
1887
1941
|
category: "category";
|
|
1942
|
+
score: "score";
|
|
1888
1943
|
}>;
|
|
1889
1944
|
categories: z.ZodArray<z.ZodString>;
|
|
1890
1945
|
}, z.core.$strict>>;
|
|
@@ -1893,17 +1948,17 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1893
1948
|
graderId: z.ZodString;
|
|
1894
1949
|
fields: z.ZodArray<z.ZodObject<{
|
|
1895
1950
|
destination: z.ZodEnum<{
|
|
1896
|
-
output: "output";
|
|
1897
|
-
input: "input";
|
|
1898
|
-
expectedOutput: "expectedOutput";
|
|
1899
1951
|
evaluatorContext: "evaluatorContext";
|
|
1952
|
+
expectedOutput: "expectedOutput";
|
|
1953
|
+
input: "input";
|
|
1954
|
+
output: "output";
|
|
1900
1955
|
}>;
|
|
1901
1956
|
path: z.ZodArray<z.ZodString>;
|
|
1902
1957
|
source: z.ZodEnum<{
|
|
1903
|
-
output: "output";
|
|
1904
|
-
input: "input";
|
|
1905
|
-
expectedOutput: "expectedOutput";
|
|
1906
1958
|
evaluatorContext: "evaluatorContext";
|
|
1959
|
+
expectedOutput: "expectedOutput";
|
|
1960
|
+
input: "input";
|
|
1961
|
+
output: "output";
|
|
1907
1962
|
}>;
|
|
1908
1963
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
1909
1964
|
}, z.core.$strict>>;
|
|
@@ -1981,10 +2036,10 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
1981
2036
|
}, z.core.$strict>;
|
|
1982
2037
|
}, z.core.$strict>>;
|
|
1983
2038
|
split: z.ZodEnum<{
|
|
2039
|
+
frozen_eval: "frozen_eval";
|
|
2040
|
+
test: "test";
|
|
1984
2041
|
train: "train";
|
|
1985
2042
|
validation: "validation";
|
|
1986
|
-
test: "test";
|
|
1987
|
-
frozen_eval: "frozen_eval";
|
|
1988
2043
|
}>;
|
|
1989
2044
|
population: z.ZodArray<z.ZodObject<{
|
|
1990
2045
|
taskId: z.ZodString;
|
|
@@ -2001,8 +2056,8 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
2001
2056
|
configurationHash: z.ZodString;
|
|
2002
2057
|
output: z.ZodEnum<{
|
|
2003
2058
|
boolean: "boolean";
|
|
2004
|
-
score: "score";
|
|
2005
2059
|
category: "category";
|
|
2060
|
+
score: "score";
|
|
2006
2061
|
}>;
|
|
2007
2062
|
categories: z.ZodArray<z.ZodString>;
|
|
2008
2063
|
}, z.core.$strict>>;
|
|
@@ -2011,17 +2066,17 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
2011
2066
|
graderId: z.ZodString;
|
|
2012
2067
|
fields: z.ZodArray<z.ZodObject<{
|
|
2013
2068
|
destination: z.ZodEnum<{
|
|
2014
|
-
output: "output";
|
|
2015
|
-
input: "input";
|
|
2016
|
-
expectedOutput: "expectedOutput";
|
|
2017
2069
|
evaluatorContext: "evaluatorContext";
|
|
2070
|
+
expectedOutput: "expectedOutput";
|
|
2071
|
+
input: "input";
|
|
2072
|
+
output: "output";
|
|
2018
2073
|
}>;
|
|
2019
2074
|
path: z.ZodArray<z.ZodString>;
|
|
2020
2075
|
source: z.ZodEnum<{
|
|
2021
|
-
output: "output";
|
|
2022
|
-
input: "input";
|
|
2023
|
-
expectedOutput: "expectedOutput";
|
|
2024
2076
|
evaluatorContext: "evaluatorContext";
|
|
2077
|
+
expectedOutput: "expectedOutput";
|
|
2078
|
+
input: "input";
|
|
2079
|
+
output: "output";
|
|
2025
2080
|
}>;
|
|
2026
2081
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
2027
2082
|
}, z.core.$strict>>;
|
|
@@ -2049,17 +2104,17 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
2049
2104
|
contentHash: z.ZodString;
|
|
2050
2105
|
mappings: z.ZodArray<z.ZodObject<{
|
|
2051
2106
|
destination: z.ZodEnum<{
|
|
2107
|
+
evaluatorContext: "evaluatorContext";
|
|
2108
|
+
expectedOutput: "expectedOutput";
|
|
2052
2109
|
input: "input";
|
|
2053
2110
|
output: "output";
|
|
2054
|
-
expectedOutput: "expectedOutput";
|
|
2055
|
-
evaluatorContext: "evaluatorContext";
|
|
2056
2111
|
}>;
|
|
2057
2112
|
path: z.ZodArray<z.ZodString>;
|
|
2058
2113
|
source: z.ZodEnum<{
|
|
2114
|
+
evaluatorContext: "evaluatorContext";
|
|
2115
|
+
expectedOutput: "expectedOutput";
|
|
2059
2116
|
input: "input";
|
|
2060
2117
|
output: "output";
|
|
2061
|
-
expectedOutput: "expectedOutput";
|
|
2062
|
-
evaluatorContext: "evaluatorContext";
|
|
2063
2118
|
}>;
|
|
2064
2119
|
sourcePath: z.ZodArray<z.ZodString>;
|
|
2065
2120
|
}, z.core.$strict>>;
|
|
@@ -2087,8 +2142,8 @@ export declare const ExperimentEvaluationScheduleCommandSchema: z.ZodDiscriminat
|
|
|
2087
2142
|
operationId: z.ZodString;
|
|
2088
2143
|
expectedRevision: z.ZodNumber;
|
|
2089
2144
|
action: z.ZodEnum<{
|
|
2090
|
-
pause: "pause";
|
|
2091
2145
|
cancel_active: "cancel_active";
|
|
2146
|
+
pause: "pause";
|
|
2092
2147
|
retry_active: "retry_active";
|
|
2093
2148
|
}>;
|
|
2094
2149
|
}, z.core.$strict>;
|
|
@@ -2119,6 +2174,13 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2119
2174
|
operationId: string;
|
|
2120
2175
|
teamId: string;
|
|
2121
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;
|
|
2122
2184
|
taskset: {
|
|
2123
2185
|
id: string;
|
|
2124
2186
|
contentHash: string;
|
|
@@ -2127,10 +2189,12 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2127
2189
|
policy: {
|
|
2128
2190
|
kind: "hosted_chat";
|
|
2129
2191
|
modelId: string;
|
|
2130
|
-
maxOutputTokens: number;
|
|
2131
|
-
temperature: number;
|
|
2132
|
-
topP: number;
|
|
2133
2192
|
localRuntime?: {
|
|
2193
|
+
providerId: "custom-openai-compatible";
|
|
2194
|
+
configurationHash: string;
|
|
2195
|
+
maximumRequestCostUsd: number;
|
|
2196
|
+
requestTimeoutMs: number;
|
|
2197
|
+
} | {
|
|
2134
2198
|
providerId: "claude-code";
|
|
2135
2199
|
configurationHash: string;
|
|
2136
2200
|
executableHash: string;
|
|
@@ -2147,11 +2211,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2147
2211
|
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
2148
2212
|
maximumRequestCostUsd: number;
|
|
2149
2213
|
requestTimeoutMs: number;
|
|
2150
|
-
} | {
|
|
2151
|
-
providerId: "custom-openai-compatible";
|
|
2152
|
-
configurationHash: string;
|
|
2153
|
-
maximumRequestCostUsd: number;
|
|
2154
|
-
requestTimeoutMs: number;
|
|
2155
2214
|
} | undefined;
|
|
2156
2215
|
candidate?: {
|
|
2157
2216
|
artifactId: string;
|
|
@@ -2173,12 +2232,47 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2173
2232
|
sourcePackageHash: string;
|
|
2174
2233
|
} | undefined;
|
|
2175
2234
|
messages?: {
|
|
2176
|
-
role: "
|
|
2235
|
+
role: "assistant" | "system" | "user";
|
|
2177
2236
|
content: string;
|
|
2178
2237
|
}[] | undefined;
|
|
2238
|
+
maxOutputTokens: number;
|
|
2239
|
+
temperature: number;
|
|
2240
|
+
topP: number;
|
|
2241
|
+
} | {
|
|
2242
|
+
kind: "fixture";
|
|
2179
2243
|
} | {
|
|
2180
2244
|
kind: "hosted_harness";
|
|
2181
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;
|
|
2182
2276
|
profileRepositoryId: string;
|
|
2183
2277
|
source: {
|
|
2184
2278
|
profileId: string;
|
|
@@ -2226,7 +2320,35 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2226
2320
|
};
|
|
2227
2321
|
declaredProfileCatalogHash: string | null;
|
|
2228
2322
|
protectedProfileClosureHash: string;
|
|
2229
|
-
|
|
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";
|
|
2230
2352
|
population: {
|
|
2231
2353
|
taskId: string;
|
|
2232
2354
|
seed: string;
|
|
@@ -2240,16 +2362,16 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2240
2362
|
};
|
|
2241
2363
|
feedbackKey: string;
|
|
2242
2364
|
configurationHash: string;
|
|
2243
|
-
output: "boolean" | "
|
|
2365
|
+
output: "boolean" | "category" | "score";
|
|
2244
2366
|
categories: string[];
|
|
2245
2367
|
}[];
|
|
2246
2368
|
fieldMappingsHash: string;
|
|
2247
2369
|
fieldMappings: {
|
|
2248
2370
|
graderId: string;
|
|
2249
2371
|
fields: {
|
|
2250
|
-
destination: "
|
|
2372
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2251
2373
|
path: string[];
|
|
2252
|
-
source: "
|
|
2374
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2253
2375
|
sourcePath: string[];
|
|
2254
2376
|
}[];
|
|
2255
2377
|
}[];
|
|
@@ -2260,34 +2382,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2260
2382
|
requireComplete: boolean;
|
|
2261
2383
|
};
|
|
2262
2384
|
contentHash: string;
|
|
2263
|
-
recordedOrigin?: {
|
|
2264
|
-
execution: {
|
|
2265
|
-
id: string;
|
|
2266
|
-
contentHash: string;
|
|
2267
|
-
};
|
|
2268
|
-
originalEvidence: {
|
|
2269
|
-
id: string;
|
|
2270
|
-
contentHash: string;
|
|
2271
|
-
};
|
|
2272
|
-
members: {
|
|
2273
|
-
caseId: string;
|
|
2274
|
-
seed: string;
|
|
2275
|
-
fixtureId: string;
|
|
2276
|
-
sourceId: string;
|
|
2277
|
-
snapshotHash: string;
|
|
2278
|
-
boundaryRevisionHash: string;
|
|
2279
|
-
}[];
|
|
2280
|
-
} | undefined;
|
|
2281
|
-
gradingSource?: {
|
|
2282
|
-
pass: {
|
|
2283
|
-
id: string;
|
|
2284
|
-
contentHash: string;
|
|
2285
|
-
};
|
|
2286
|
-
evidence: {
|
|
2287
|
-
id: string;
|
|
2288
|
-
contentHash: string;
|
|
2289
|
-
};
|
|
2290
|
-
} | undefined;
|
|
2291
2385
|
} | undefined;
|
|
2292
2386
|
sourceCandidate?: {
|
|
2293
2387
|
candidateId: string;
|
|
@@ -2302,28 +2396,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2302
2396
|
};
|
|
2303
2397
|
modelConfigurationHash: string;
|
|
2304
2398
|
packageHash: string;
|
|
2305
|
-
maxOutputTokens?: number | undefined;
|
|
2306
|
-
temperature?: number | undefined;
|
|
2307
|
-
topP?: number | undefined;
|
|
2308
|
-
candidate?: {
|
|
2309
|
-
artifactId: string;
|
|
2310
|
-
contentHash: string;
|
|
2311
|
-
preparationId: string;
|
|
2312
|
-
dispatchId: string;
|
|
2313
|
-
baseProfileId: string;
|
|
2314
|
-
workerImageDigest: string;
|
|
2315
|
-
} | undefined;
|
|
2316
|
-
sourceCandidate?: {
|
|
2317
|
-
candidateId: string;
|
|
2318
|
-
frozenRevision: number;
|
|
2319
|
-
frozenHash: string;
|
|
2320
|
-
harnessRelease: {
|
|
2321
|
-
id: string;
|
|
2322
|
-
contentHash: string;
|
|
2323
|
-
};
|
|
2324
|
-
protectedClosureHash: string;
|
|
2325
|
-
} | undefined;
|
|
2326
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
2327
2399
|
externalDatasetBinding?: {
|
|
2328
2400
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
2329
2401
|
dataset: {
|
|
@@ -2347,7 +2419,35 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2347
2419
|
};
|
|
2348
2420
|
declaredProfileCatalogHash: string | null;
|
|
2349
2421
|
protectedProfileClosureHash: string;
|
|
2350
|
-
|
|
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";
|
|
2351
2451
|
population: {
|
|
2352
2452
|
taskId: string;
|
|
2353
2453
|
seed: string;
|
|
@@ -2361,16 +2461,16 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2361
2461
|
};
|
|
2362
2462
|
feedbackKey: string;
|
|
2363
2463
|
configurationHash: string;
|
|
2364
|
-
output: "boolean" | "
|
|
2464
|
+
output: "boolean" | "category" | "score";
|
|
2365
2465
|
categories: string[];
|
|
2366
2466
|
}[];
|
|
2367
2467
|
fieldMappingsHash: string;
|
|
2368
2468
|
fieldMappings: {
|
|
2369
2469
|
graderId: string;
|
|
2370
2470
|
fields: {
|
|
2371
|
-
destination: "
|
|
2471
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2372
2472
|
path: string[];
|
|
2373
|
-
source: "
|
|
2473
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2374
2474
|
sourcePath: string[];
|
|
2375
2475
|
}[];
|
|
2376
2476
|
}[];
|
|
@@ -2381,37 +2481,7 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2381
2481
|
requireComplete: boolean;
|
|
2382
2482
|
};
|
|
2383
2483
|
contentHash: string;
|
|
2384
|
-
recordedOrigin?: {
|
|
2385
|
-
execution: {
|
|
2386
|
-
id: string;
|
|
2387
|
-
contentHash: string;
|
|
2388
|
-
};
|
|
2389
|
-
originalEvidence: {
|
|
2390
|
-
id: string;
|
|
2391
|
-
contentHash: string;
|
|
2392
|
-
};
|
|
2393
|
-
members: {
|
|
2394
|
-
caseId: string;
|
|
2395
|
-
seed: string;
|
|
2396
|
-
fixtureId: string;
|
|
2397
|
-
sourceId: string;
|
|
2398
|
-
snapshotHash: string;
|
|
2399
|
-
boundaryRevisionHash: string;
|
|
2400
|
-
}[];
|
|
2401
|
-
} | undefined;
|
|
2402
|
-
gradingSource?: {
|
|
2403
|
-
pass: {
|
|
2404
|
-
id: string;
|
|
2405
|
-
contentHash: string;
|
|
2406
|
-
};
|
|
2407
|
-
evidence: {
|
|
2408
|
-
id: string;
|
|
2409
|
-
contentHash: string;
|
|
2410
|
-
};
|
|
2411
|
-
} | undefined;
|
|
2412
2484
|
} | undefined;
|
|
2413
|
-
} | {
|
|
2414
|
-
kind: "fixture";
|
|
2415
2485
|
};
|
|
2416
2486
|
population: {
|
|
2417
2487
|
receiptId: string;
|
|
@@ -2419,13 +2489,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2419
2489
|
seed: string;
|
|
2420
2490
|
fixtureId: string | null;
|
|
2421
2491
|
}[];
|
|
2422
|
-
name?: string | undefined;
|
|
2423
|
-
project?: {
|
|
2424
|
-
id: string;
|
|
2425
|
-
revision: number;
|
|
2426
|
-
contentHash: string;
|
|
2427
|
-
targetId: string | null;
|
|
2428
|
-
} | undefined;
|
|
2429
2492
|
};
|
|
2430
2493
|
maximumCostUsd: number;
|
|
2431
2494
|
graders?: {
|
|
@@ -2433,9 +2496,9 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2433
2496
|
version: string;
|
|
2434
2497
|
contentHash: string;
|
|
2435
2498
|
mappings: {
|
|
2436
|
-
destination: "
|
|
2499
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2437
2500
|
path: string[];
|
|
2438
|
-
source: "
|
|
2501
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2439
2502
|
sourcePath: string[];
|
|
2440
2503
|
}[];
|
|
2441
2504
|
}[] | undefined;
|
|
@@ -2452,7 +2515,7 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2452
2515
|
timeZone: string;
|
|
2453
2516
|
};
|
|
2454
2517
|
};
|
|
2455
|
-
state: "blocked" | "
|
|
2518
|
+
state: "blocked" | "disabled" | "scheduled";
|
|
2456
2519
|
nextRunAt: string | null;
|
|
2457
2520
|
maximumTotalSpendUsd: number;
|
|
2458
2521
|
maximumAttempts: number;
|
|
@@ -2468,7 +2531,7 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2468
2531
|
coalescedCount: number;
|
|
2469
2532
|
configurationHash: string;
|
|
2470
2533
|
operationId: string;
|
|
2471
|
-
state: "cancelled" | "
|
|
2534
|
+
state: "cancelled" | "cleaning" | "completed" | "dispatching" | "failed" | "reserved" | "running" | "uncertain";
|
|
2472
2535
|
executionId: string | null;
|
|
2473
2536
|
attempt: number;
|
|
2474
2537
|
maximumCostUsd: number;
|
|
@@ -2497,6 +2560,13 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2497
2560
|
operationId: string;
|
|
2498
2561
|
teamId: string;
|
|
2499
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;
|
|
2500
2570
|
taskset: {
|
|
2501
2571
|
id: string;
|
|
2502
2572
|
contentHash: string;
|
|
@@ -2505,10 +2575,12 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2505
2575
|
policy: {
|
|
2506
2576
|
kind: "hosted_chat";
|
|
2507
2577
|
modelId: string;
|
|
2508
|
-
maxOutputTokens: number;
|
|
2509
|
-
temperature: number;
|
|
2510
|
-
topP: number;
|
|
2511
2578
|
localRuntime?: {
|
|
2579
|
+
providerId: "custom-openai-compatible";
|
|
2580
|
+
configurationHash: string;
|
|
2581
|
+
maximumRequestCostUsd: number;
|
|
2582
|
+
requestTimeoutMs: number;
|
|
2583
|
+
} | {
|
|
2512
2584
|
providerId: "claude-code";
|
|
2513
2585
|
configurationHash: string;
|
|
2514
2586
|
executableHash: string;
|
|
@@ -2525,11 +2597,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2525
2597
|
reasoningEffort: "high" | "low" | "medium" | "xhigh";
|
|
2526
2598
|
maximumRequestCostUsd: number;
|
|
2527
2599
|
requestTimeoutMs: number;
|
|
2528
|
-
} | {
|
|
2529
|
-
providerId: "custom-openai-compatible";
|
|
2530
|
-
configurationHash: string;
|
|
2531
|
-
maximumRequestCostUsd: number;
|
|
2532
|
-
requestTimeoutMs: number;
|
|
2533
2600
|
} | undefined;
|
|
2534
2601
|
candidate?: {
|
|
2535
2602
|
artifactId: string;
|
|
@@ -2551,12 +2618,47 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2551
2618
|
sourcePackageHash: string;
|
|
2552
2619
|
} | undefined;
|
|
2553
2620
|
messages?: {
|
|
2554
|
-
role: "
|
|
2621
|
+
role: "assistant" | "system" | "user";
|
|
2555
2622
|
content: string;
|
|
2556
2623
|
}[] | undefined;
|
|
2624
|
+
maxOutputTokens: number;
|
|
2625
|
+
temperature: number;
|
|
2626
|
+
topP: number;
|
|
2627
|
+
} | {
|
|
2628
|
+
kind: "fixture";
|
|
2557
2629
|
} | {
|
|
2558
2630
|
kind: "hosted_harness";
|
|
2559
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;
|
|
2560
2662
|
profileRepositoryId: string;
|
|
2561
2663
|
source: {
|
|
2562
2664
|
profileId: string;
|
|
@@ -2604,7 +2706,35 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2604
2706
|
};
|
|
2605
2707
|
declaredProfileCatalogHash: string | null;
|
|
2606
2708
|
protectedProfileClosureHash: string;
|
|
2607
|
-
|
|
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";
|
|
2608
2738
|
population: {
|
|
2609
2739
|
taskId: string;
|
|
2610
2740
|
seed: string;
|
|
@@ -2618,16 +2748,16 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2618
2748
|
};
|
|
2619
2749
|
feedbackKey: string;
|
|
2620
2750
|
configurationHash: string;
|
|
2621
|
-
output: "boolean" | "
|
|
2751
|
+
output: "boolean" | "category" | "score";
|
|
2622
2752
|
categories: string[];
|
|
2623
2753
|
}[];
|
|
2624
2754
|
fieldMappingsHash: string;
|
|
2625
2755
|
fieldMappings: {
|
|
2626
2756
|
graderId: string;
|
|
2627
2757
|
fields: {
|
|
2628
|
-
destination: "
|
|
2758
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2629
2759
|
path: string[];
|
|
2630
|
-
source: "
|
|
2760
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2631
2761
|
sourcePath: string[];
|
|
2632
2762
|
}[];
|
|
2633
2763
|
}[];
|
|
@@ -2638,34 +2768,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2638
2768
|
requireComplete: boolean;
|
|
2639
2769
|
};
|
|
2640
2770
|
contentHash: string;
|
|
2641
|
-
recordedOrigin?: {
|
|
2642
|
-
execution: {
|
|
2643
|
-
id: string;
|
|
2644
|
-
contentHash: string;
|
|
2645
|
-
};
|
|
2646
|
-
originalEvidence: {
|
|
2647
|
-
id: string;
|
|
2648
|
-
contentHash: string;
|
|
2649
|
-
};
|
|
2650
|
-
members: {
|
|
2651
|
-
caseId: string;
|
|
2652
|
-
seed: string;
|
|
2653
|
-
fixtureId: string;
|
|
2654
|
-
sourceId: string;
|
|
2655
|
-
snapshotHash: string;
|
|
2656
|
-
boundaryRevisionHash: string;
|
|
2657
|
-
}[];
|
|
2658
|
-
} | undefined;
|
|
2659
|
-
gradingSource?: {
|
|
2660
|
-
pass: {
|
|
2661
|
-
id: string;
|
|
2662
|
-
contentHash: string;
|
|
2663
|
-
};
|
|
2664
|
-
evidence: {
|
|
2665
|
-
id: string;
|
|
2666
|
-
contentHash: string;
|
|
2667
|
-
};
|
|
2668
|
-
} | undefined;
|
|
2669
2771
|
} | undefined;
|
|
2670
2772
|
sourceCandidate?: {
|
|
2671
2773
|
candidateId: string;
|
|
@@ -2680,28 +2782,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2680
2782
|
};
|
|
2681
2783
|
modelConfigurationHash: string;
|
|
2682
2784
|
packageHash: string;
|
|
2683
|
-
maxOutputTokens?: number | undefined;
|
|
2684
|
-
temperature?: number | undefined;
|
|
2685
|
-
topP?: number | undefined;
|
|
2686
|
-
candidate?: {
|
|
2687
|
-
artifactId: string;
|
|
2688
|
-
contentHash: string;
|
|
2689
|
-
preparationId: string;
|
|
2690
|
-
dispatchId: string;
|
|
2691
|
-
baseProfileId: string;
|
|
2692
|
-
workerImageDigest: string;
|
|
2693
|
-
} | undefined;
|
|
2694
|
-
sourceCandidate?: {
|
|
2695
|
-
candidateId: string;
|
|
2696
|
-
frozenRevision: number;
|
|
2697
|
-
frozenHash: string;
|
|
2698
|
-
harnessRelease: {
|
|
2699
|
-
id: string;
|
|
2700
|
-
contentHash: string;
|
|
2701
|
-
};
|
|
2702
|
-
protectedClosureHash: string;
|
|
2703
|
-
} | undefined;
|
|
2704
|
-
reasoningEffort?: "high" | "low" | "medium" | "xhigh" | "off" | "max" | undefined;
|
|
2705
2785
|
externalDatasetBinding?: {
|
|
2706
2786
|
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
2707
2787
|
dataset: {
|
|
@@ -2725,7 +2805,35 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2725
2805
|
};
|
|
2726
2806
|
declaredProfileCatalogHash: string | null;
|
|
2727
2807
|
protectedProfileClosureHash: string;
|
|
2728
|
-
|
|
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";
|
|
2729
2837
|
population: {
|
|
2730
2838
|
taskId: string;
|
|
2731
2839
|
seed: string;
|
|
@@ -2739,16 +2847,16 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2739
2847
|
};
|
|
2740
2848
|
feedbackKey: string;
|
|
2741
2849
|
configurationHash: string;
|
|
2742
|
-
output: "boolean" | "
|
|
2850
|
+
output: "boolean" | "category" | "score";
|
|
2743
2851
|
categories: string[];
|
|
2744
2852
|
}[];
|
|
2745
2853
|
fieldMappingsHash: string;
|
|
2746
2854
|
fieldMappings: {
|
|
2747
2855
|
graderId: string;
|
|
2748
2856
|
fields: {
|
|
2749
|
-
destination: "
|
|
2857
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2750
2858
|
path: string[];
|
|
2751
|
-
source: "
|
|
2859
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2752
2860
|
sourcePath: string[];
|
|
2753
2861
|
}[];
|
|
2754
2862
|
}[];
|
|
@@ -2759,37 +2867,7 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2759
2867
|
requireComplete: boolean;
|
|
2760
2868
|
};
|
|
2761
2869
|
contentHash: string;
|
|
2762
|
-
recordedOrigin?: {
|
|
2763
|
-
execution: {
|
|
2764
|
-
id: string;
|
|
2765
|
-
contentHash: string;
|
|
2766
|
-
};
|
|
2767
|
-
originalEvidence: {
|
|
2768
|
-
id: string;
|
|
2769
|
-
contentHash: string;
|
|
2770
|
-
};
|
|
2771
|
-
members: {
|
|
2772
|
-
caseId: string;
|
|
2773
|
-
seed: string;
|
|
2774
|
-
fixtureId: string;
|
|
2775
|
-
sourceId: string;
|
|
2776
|
-
snapshotHash: string;
|
|
2777
|
-
boundaryRevisionHash: string;
|
|
2778
|
-
}[];
|
|
2779
|
-
} | undefined;
|
|
2780
|
-
gradingSource?: {
|
|
2781
|
-
pass: {
|
|
2782
|
-
id: string;
|
|
2783
|
-
contentHash: string;
|
|
2784
|
-
};
|
|
2785
|
-
evidence: {
|
|
2786
|
-
id: string;
|
|
2787
|
-
contentHash: string;
|
|
2788
|
-
};
|
|
2789
|
-
} | undefined;
|
|
2790
2870
|
} | undefined;
|
|
2791
|
-
} | {
|
|
2792
|
-
kind: "fixture";
|
|
2793
2871
|
};
|
|
2794
2872
|
population: {
|
|
2795
2873
|
receiptId: string;
|
|
@@ -2797,13 +2875,6 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2797
2875
|
seed: string;
|
|
2798
2876
|
fixtureId: string | null;
|
|
2799
2877
|
}[];
|
|
2800
|
-
name?: string | undefined;
|
|
2801
|
-
project?: {
|
|
2802
|
-
id: string;
|
|
2803
|
-
revision: number;
|
|
2804
|
-
contentHash: string;
|
|
2805
|
-
targetId: string | null;
|
|
2806
|
-
} | undefined;
|
|
2807
2878
|
};
|
|
2808
2879
|
maximumCostUsd: number;
|
|
2809
2880
|
graders?: {
|
|
@@ -2811,9 +2882,9 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2811
2882
|
version: string;
|
|
2812
2883
|
contentHash: string;
|
|
2813
2884
|
mappings: {
|
|
2814
|
-
destination: "
|
|
2885
|
+
destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2815
2886
|
path: string[];
|
|
2816
|
-
source: "
|
|
2887
|
+
source: "evaluatorContext" | "expectedOutput" | "input" | "output";
|
|
2817
2888
|
sourcePath: string[];
|
|
2818
2889
|
}[];
|
|
2819
2890
|
}[] | undefined;
|
|
@@ -2830,7 +2901,7 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2830
2901
|
timeZone: string;
|
|
2831
2902
|
};
|
|
2832
2903
|
};
|
|
2833
|
-
state: "blocked" | "
|
|
2904
|
+
state: "blocked" | "disabled" | "scheduled";
|
|
2834
2905
|
nextRunAt: string | null;
|
|
2835
2906
|
maximumTotalSpendUsd: number;
|
|
2836
2907
|
maximumAttempts: number;
|
|
@@ -2846,7 +2917,7 @@ export declare class OpenPondExperimentEvaluationScheduleClient {
|
|
|
2846
2917
|
coalescedCount: number;
|
|
2847
2918
|
configurationHash: string;
|
|
2848
2919
|
operationId: string;
|
|
2849
|
-
state: "cancelled" | "
|
|
2920
|
+
state: "cancelled" | "cleaning" | "completed" | "dispatching" | "failed" | "reserved" | "running" | "uncertain";
|
|
2850
2921
|
executionId: string | null;
|
|
2851
2922
|
attempt: number;
|
|
2852
2923
|
maximumCostUsd: number;
|