openpond-sdk 0.9.7 → 0.9.8

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (58) hide show
  1. package/dist/{chunk-SR55WVPK.js → chunk-4HZORGK3.js} +2 -2
  2. package/dist/{chunk-XR6CLSCS.js → chunk-BGXWHDYM.js} +3 -2
  3. package/dist/chunk-BGXWHDYM.js.map +7 -0
  4. package/dist/{chunk-CNGJF6AM.js → chunk-RXKIOOFZ.js} +3 -3
  5. package/dist/{chunk-62LWLACP.js → chunk-YYJ4T4U4.js} +2 -2
  6. package/dist/dataset-marketplace.js +11 -0
  7. package/dist/dataset-marketplace.js.map +2 -2
  8. package/dist/dataset-workspaces.js +69 -20
  9. package/dist/dataset-workspaces.js.map +4 -4
  10. package/dist/experiments.js +60 -5
  11. package/dist/experiments.js.map +4 -4
  12. package/dist/index.js +4 -4
  13. package/dist/model-starter-attempts.js +1 -1
  14. package/dist/model-taskset-runs.js +2 -2
  15. package/dist/types/packages/sdk/src/dataset-marketplace.d.ts +9 -0
  16. package/dist/types/packages/sdk/src/dataset-marketplace.d.ts.map +1 -1
  17. package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts +1097 -30
  18. package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts.map +1 -1
  19. package/dist/types/packages/sdk/src/dataset-workspace-contracts.d.ts +12 -12
  20. package/dist/types/packages/sdk/src/dataset-workspace-operations.d.ts +1617 -0
  21. package/dist/types/packages/sdk/src/dataset-workspace-operations.d.ts.map +1 -0
  22. package/dist/types/packages/sdk/src/dataset-workspaces.d.ts +1 -0
  23. package/dist/types/packages/sdk/src/dataset-workspaces.d.ts.map +1 -1
  24. package/dist/types/packages/sdk/src/experiment-history.d.ts +266 -0
  25. package/dist/types/packages/sdk/src/experiment-history.d.ts.map +1 -0
  26. package/dist/types/packages/sdk/src/experiments.d.ts +1 -0
  27. package/dist/types/packages/sdk/src/experiments.d.ts.map +1 -1
  28. package/dist/types/packages/sdk/src/learning-batch-assets.d.ts +2 -2
  29. package/dist/types/packages/sdk/src/learning-client.d.ts +1 -1
  30. package/dist/types/packages/sdk/src/model-batch-review-preparation.d.ts +3 -3
  31. package/dist/types/packages/sdk/src/model-batch-review.d.ts +7 -7
  32. package/dist/types/packages/sdk/src/model-learning-policy.d.ts +2 -2
  33. package/dist/types/packages/sdk/src/model-starter-attempts.d.ts +2 -1
  34. package/dist/types/packages/sdk/src/model-starter-attempts.d.ts.map +1 -1
  35. package/dist/types/packages/sdk/src/model-starter-catalog.d.ts +1 -1
  36. package/dist/types/packages/sdk/src/model-starters.d.ts +1 -1
  37. package/dist/types/packages/sdk/src/task-inventory.d.ts +1 -1
  38. package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts +6 -6
  39. package/dist/types/packages/sdk/src/taskset-authored-files.d.ts +6 -6
  40. package/dist/types/packages/sdk/src/taskset-authored-portable-release.d.ts +2 -2
  41. package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts +12 -12
  42. package/dist/types/packages/sdk/src/taskset-draft-client.d.ts +36 -36
  43. package/dist/types/packages/sdk/src/taskset-draft-core.d.ts +6 -6
  44. package/dist/types/packages/sdk/src/taskset-draft-dataset-artifacts.d.ts +11 -11
  45. package/dist/types/packages/sdk/src/taskset-draft-document.d.ts +8 -8
  46. package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts +10 -10
  47. package/dist/types/packages/sdk/src/taskset-draft-source.d.ts +12 -12
  48. package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts +12 -12
  49. package/dist/types/packages/sdk/src/taskset-package-learning.d.ts +4 -4
  50. package/dist/types/packages/sdk/src/taskset-package-projection.d.ts +11 -11
  51. package/dist/types/packages/sdk/src/training-learning-batch.d.ts +14 -14
  52. package/dist/types/packages/sdk/src/training-project-client.d.ts +4 -4
  53. package/dist/types/packages/sdk/src/training-project-contracts.d.ts +5 -5
  54. package/package.json +1 -1
  55. package/dist/chunk-XR6CLSCS.js.map +0 -7
  56. /package/dist/{chunk-SR55WVPK.js.map → chunk-4HZORGK3.js.map} +0 -0
  57. /package/dist/{chunk-CNGJF6AM.js.map → chunk-RXKIOOFZ.js.map} +0 -0
  58. /package/dist/{chunk-62LWLACP.js.map → chunk-YYJ4T4U4.js.map} +0 -0
package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts CHANGED
@@ -1,5 +1,6 @@
1
1
  import { DatasetWorkspaceBeginVersionSchema, DatasetWorkspacePublishSchema, DatasetWorkspaceWriteSchema } from "./dataset-workspace-contracts.js";
2
2
  import type { z } from "zod";
3
+ import { DatasetWorkspaceOperationKindSchema } from "./dataset-workspace-operations.js";
3
4
  /** Hosted independent authoring. Writes use explicit revision/operation identity;
4
5
  * callers retain the operation ID when retrying an uncertain response. */
5
6
  export declare class OpenPondDatasetWorkspaceClient {
@@ -93,8 +94,8 @@ export declare class OpenPondDatasetWorkspaceClient {
93
94
  releaseId: string;
94
95
  releaseHash: string;
95
96
  managedReleasePath: string;
96
- adaptationSplit: "train" | "validation" | "frozen_eval" | "test";
97
- evaluationSplit: "train" | "validation" | "frozen_eval" | "test";
97
+ adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
98
+ evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
98
99
  primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
99
100
  qualityGate: "none" | "non_regression" | "all_pass";
100
101
  source: "builtin" | "imported";
@@ -312,7 +313,7 @@ export declare class OpenPondDatasetWorkspaceClient {
312
313
  schemaVersion: "openpond.taskData.v1";
313
314
  id: string;
314
315
  clusterKey: string;
315
- split: "train" | "validation" | "frozen_eval" | "test";
316
+ split: "train" | "validation" | "test" | "frozen_eval";
316
317
  input: Record<string, unknown>;
317
318
  expectedOutput: Record<string, unknown> | null;
318
319
  policyVisibleContext: Record<string, unknown>;
@@ -328,7 +329,7 @@ export declare class OpenPondDatasetWorkspaceClient {
328
329
  mediaType: string;
329
330
  sha256: string;
330
331
  sizeBytes: number;
331
- split: "train" | "validation" | "frozen_eval" | "test";
332
+ split: "train" | "validation" | "test" | "frozen_eval";
332
333
  metadata: Record<string, unknown>;
333
334
  }[] | undefined;
334
335
  resourceRefs?: string[] | undefined;
@@ -528,7 +529,7 @@ export declare class OpenPondDatasetWorkspaceClient {
528
529
  };
529
530
  shards: {
530
531
  id: string;
531
- split: "train" | "validation" | "frozen_eval" | "test";
532
+ split: "train" | "validation" | "test" | "frozen_eval";
532
533
  path: string;
533
534
  contentHash: string;
534
535
  schemaHash: string;
@@ -537,7 +538,7 @@ export declare class OpenPondDatasetWorkspaceClient {
537
538
  rowGroupCount: number;
538
539
  }[];
539
540
  rowCount: number;
540
- splitCounts: Record<"train" | "validation" | "frozen_eval" | "test", number>;
541
+ splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
541
542
  sourceReceiptRefs: string[];
542
543
  mappingHash: string;
543
544
  qualityReportHash: string;
@@ -564,6 +565,1072 @@ export declare class OpenPondDatasetWorkspaceClient {
564
565
  originProjectId?: string | undefined;
565
566
  ownerScope?: "workspace" | "personal" | undefined;
566
567
  }>;
568
+ /** Recover only this actor's retained operation; base bytes allow adapters to
569
+ * reconstruct and verify the identical original CAS request before replay. */
570
+ operationResult(operationId: string, options: {
571
+ datasetId?: string;
572
+ kind: z.infer<typeof DatasetWorkspaceOperationKindSchema>;
573
+ requestHash?: string;
574
+ signal?: AbortSignal;
575
+ }): Promise<{
576
+ operationId: string;
577
+ kind: "create" | "save" | "publish" | "begin_version";
578
+ requestHash: string;
579
+ receipt: {
580
+ schemaVersion: "openpond.datasetWorkspaceReceipt.v1";
581
+ teamId: string;
582
+ datasetId: string;
583
+ revision: number;
584
+ workspace: {
585
+ schemaVersion: "openpond.tasksetDraftWorkspace.v1";
586
+ draft: {
587
+ schemaVersion: "openpond.tasksetDraft.v1";
588
+ id: string;
589
+ revision: number;
590
+ profileId: string;
591
+ modelScope: {
592
+ modelId: string;
593
+ expectedModelRevision: number;
594
+ source?: {
595
+ schemaVersion: "openpond.modelTasksetDraftPreparation.v1";
596
+ draftId: string;
597
+ requestHash: string;
598
+ sourcePackageHash: string;
599
+ sourceTasksetRef: {
600
+ id: string;
601
+ contentHash: string;
602
+ revision: number;
603
+ };
604
+ tasksetId: string;
605
+ tasksetRevision: number;
606
+ lineage: {
607
+ schemaVersion: "openpond.modelTasksetAuthoring.v1";
608
+ owner: {
609
+ scopeId: string;
610
+ modelId: string;
611
+ };
612
+ root: {
613
+ id: string;
614
+ contentHash: string;
615
+ revision: number;
616
+ };
617
+ parent: {
618
+ id: string;
619
+ contentHash: string;
620
+ revision: number;
621
+ };
622
+ };
623
+ authoringGraph?: "bound" | undefined;
624
+ } | undefined;
625
+ } | null;
626
+ name: string;
627
+ objective: string;
628
+ purpose: "general" | "benchmark";
629
+ benchmark: {
630
+ schemaVersion: "openpond.tasksetBenchmark.v1";
631
+ definitionId: string;
632
+ releaseId: string;
633
+ releaseHash: string;
634
+ managedReleasePath: string;
635
+ adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
636
+ evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
637
+ primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
638
+ qualityGate: "none" | "non_regression" | "all_pass";
639
+ source: "builtin" | "imported";
640
+ metadata: Record<string, unknown>;
641
+ } | null;
642
+ preferenceComparison: {
643
+ schemaVersion: "openpond.tasksetPreferenceComparisonBinding.v1";
644
+ releaseId: string;
645
+ releaseHash: string;
646
+ publishedAt: string;
647
+ metadata: Record<string, unknown>;
648
+ } | null;
649
+ status: "draft" | "validating" | "needs_review" | "published";
650
+ sourceRefs: ({
651
+ id: string;
652
+ profileId: string;
653
+ title: string;
654
+ sourceHash: string;
655
+ occurredAt: string;
656
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
657
+ secretScanStatus: "passed" | "pending" | "blocked";
658
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
659
+ metadata: Record<string, unknown>;
660
+ schemaVersion: "openpond.huggingFaceDatasetSource.v1";
661
+ kind: "huggingface";
662
+ repositoryId: string;
663
+ repositoryUrl: string;
664
+ revision: string;
665
+ configuration: string;
666
+ upstreamSplits: string[];
667
+ gated: boolean;
668
+ private: boolean;
669
+ declaredLicense: string | null;
670
+ sourceFileHashes: string[];
671
+ } | {
672
+ id: string;
673
+ profileId: string;
674
+ title: string;
675
+ sourceHash: string;
676
+ occurredAt: string;
677
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
678
+ secretScanStatus: "passed" | "pending" | "blocked";
679
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
680
+ metadata: Record<string, unknown>;
681
+ schemaVersion: "openpond.uploadedFileDatasetSource.v1";
682
+ kind: "uploaded_file";
683
+ originalFileNames: string[];
684
+ mediaTypes: string[];
685
+ sourceFileHashes: string[];
686
+ totalBytes: number;
687
+ parserVersion: string;
688
+ } | {
689
+ id: string;
690
+ profileId: string;
691
+ title: string;
692
+ sourceHash: string;
693
+ occurredAt: string;
694
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
695
+ secretScanStatus: "passed" | "pending" | "blocked";
696
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
697
+ metadata: Record<string, unknown>;
698
+ schemaVersion: "openpond.generatedDatasetSource.v1";
699
+ kind: "generated";
700
+ generatorId: string;
701
+ generatorVersion: string;
702
+ seed: number;
703
+ generatorHash: string;
704
+ } | {
705
+ id: string;
706
+ profileId: string;
707
+ title: string;
708
+ sourceHash: string;
709
+ occurredAt: string;
710
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
711
+ secretScanStatus: "passed" | "pending" | "blocked";
712
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
713
+ metadata: Record<string, unknown>;
714
+ schemaVersion: "openpond.learningBatchDatasetSource.v1";
715
+ kind: "learning_batch";
716
+ batch: {
717
+ id: string;
718
+ revision: number;
719
+ contentHash: string;
720
+ };
721
+ taskDefinition: {
722
+ id: string;
723
+ revision: number;
724
+ contentHash: string;
725
+ };
726
+ admittedBy: string;
727
+ } | {
728
+ schemaVersion: "openpond.trainingSource.v1";
729
+ id: string;
730
+ profileId: string;
731
+ sessionId: string;
732
+ turnIds: string[];
733
+ workspaceId: string | null;
734
+ sourceHash: string;
735
+ clusterKey: string;
736
+ title: string;
737
+ occurredAt: string;
738
+ consent: {
739
+ status: "pending" | "granted" | "denied" | "revoked";
740
+ scope: "metadata_only" | "selected_turns" | "full_session";
741
+ grantedBy: string | null;
742
+ grantedAt: string | null;
743
+ purpose: "task_authoring_and_evaluation";
744
+ };
745
+ connectedAppIds: string[];
746
+ secretScanStatus: "passed" | "pending" | "blocked";
747
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
748
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
749
+ metadata: Record<string, unknown>;
750
+ })[];
751
+ policy: {
752
+ policyVisibleFields: string[];
753
+ privilegedFields: string[];
754
+ hiddenGraderRefs: string[];
755
+ connectedAppScopes: string[];
756
+ };
757
+ environment: {
758
+ protocolVersion: "openpond.taskEnvironment.v1";
759
+ kind: "chat" | "agent" | "program" | "stateful_harness" | "work";
760
+ entrypoint: string;
761
+ stateful: boolean;
762
+ deterministicSeeds: boolean;
763
+ toolNames: string[];
764
+ lifecycle: ("grade" | "create" | "reset" | "step" | "cleanup")[];
765
+ defaultTimeoutMs: number;
766
+ networkPolicy: "none" | "declared_read_only" | "declared_scoped";
767
+ metadata: Record<string, unknown>;
768
+ actionBindings?: {
769
+ schemaVersion: "openpond.harnessActionBinding.v1";
770
+ actionId: string;
771
+ modelToolName: string;
772
+ description: string;
773
+ inputSchema: Record<string, unknown>;
774
+ actionSchemaHash: string;
775
+ agentRelease: {
776
+ id: string;
777
+ contentHash: string;
778
+ };
779
+ implementationHash: string;
780
+ runtimeBindingId: string;
781
+ capabilityReceiptHash: string;
782
+ sideEffect: "read" | "write";
783
+ studentVisible: boolean;
784
+ timeoutMs: number;
785
+ episodeArgumentBindings: {
786
+ argument: string;
787
+ source: "case_id";
788
+ }[];
789
+ }[] | undefined;
790
+ resources?: {
791
+ id: string;
792
+ kind: "file" | "configuration" | "catalog" | "code_module";
793
+ path: string;
794
+ visibility: "privileged" | "policy_visible" | "policy_hidden";
795
+ required: boolean;
796
+ metadata: Record<string, unknown>;
797
+ mediaType?: string | null | undefined;
798
+ }[] | undefined;
799
+ };
800
+ output: {
801
+ mode: "artifacts" | "text" | "structured_json";
802
+ jsonSchema: Record<string, unknown> | null;
803
+ renderer: {
804
+ module: string;
805
+ exportName: string;
806
+ configRef: string | null;
807
+ } | null;
808
+ };
809
+ capabilities: {
810
+ schemaVersion: "openpond.tasksetCapabilities.v1";
811
+ taskKind: "chat" | "single_agent" | "multi_agent" | "custom_program";
812
+ supportedSignals: ("label" | "reward" | "preference" | "feedback" | "demonstration" | "correction")[];
813
+ compatibleMethods: ("none" | "sft" | "dpo" | "grpo" | "ppo" | "sdft" | "opd" | "opsd" | "sdpo" | "retrieval")[];
814
+ rewardKinds: ("none" | "exact" | "deterministic" | "model_judge" | "human")[];
815
+ requiresTools: boolean;
816
+ requiresState: boolean;
817
+ requiresPrivilegedGrading: boolean;
818
+ environmentPlacements: ("local" | "remote" | "colocated" | "provider_native")[];
819
+ exportable: boolean;
820
+ portabilityBlockers: string[];
821
+ };
822
+ metrics: {
823
+ schemaVersion: "openpond.tasksetMetricPolicy.v1";
824
+ primaryMetric: string;
825
+ aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
826
+ missingReward: "zero" | "exclude";
827
+ customAggregator: {
828
+ module: string;
829
+ exportName: string;
830
+ contentHash: string;
831
+ timeoutMs: number;
832
+ networkPolicy: "none";
833
+ } | null;
834
+ taskWeights?: Record<string, number> | undefined;
835
+ };
836
+ review: {
837
+ enabled: boolean;
838
+ candidateCount: number;
839
+ minimumSamples: number;
840
+ allowTies: boolean;
841
+ allowRejectAll: boolean;
842
+ rubric: string;
843
+ criteria: {
844
+ id: string;
845
+ label: string;
846
+ description: string;
847
+ weight: number;
848
+ }[];
849
+ };
850
+ tasks: {
851
+ schemaVersion: "openpond.taskData.v1";
852
+ id: string;
853
+ clusterKey: string;
854
+ split: "train" | "validation" | "test" | "frozen_eval";
855
+ input: Record<string, unknown>;
856
+ expectedOutput: Record<string, unknown> | null;
857
+ policyVisibleContext: Record<string, unknown>;
858
+ privilegedContextRef: string | null;
859
+ tags: string[];
860
+ metadata: Record<string, unknown>;
861
+ sourceRefs: string[];
862
+ assets?: {
863
+ id: string;
864
+ sourceRefId: string;
865
+ artifactRef: string;
866
+ fileName: string;
867
+ mediaType: string;
868
+ sha256: string;
869
+ sizeBytes: number;
870
+ split: "train" | "validation" | "test" | "frozen_eval";
871
+ metadata: Record<string, unknown>;
872
+ }[] | undefined;
873
+ resourceRefs?: string[] | undefined;
874
+ requiredOutputs?: {
875
+ path: string;
876
+ mediaType: string;
877
+ metadata: Record<string, unknown>;
878
+ schemaRef?: string | null | undefined;
879
+ maxBytes?: number | undefined;
880
+ }[] | undefined;
881
+ }[];
882
+ graders: ({
883
+ id: string;
884
+ version: string;
885
+ label: string;
886
+ weight: number;
887
+ hardGate: boolean;
888
+ rewardEligible: boolean;
889
+ privileged: boolean;
890
+ metadata: Record<string, unknown>;
891
+ kind: "file" | "content" | "state" | "test" | "schema" | "diff" | "runtime_event";
892
+ config: Record<string, unknown>;
893
+ } | {
894
+ id: string;
895
+ version: string;
896
+ label: string;
897
+ weight: number;
898
+ hardGate: boolean;
899
+ rewardEligible: boolean;
900
+ privileged: boolean;
901
+ metadata: Record<string, unknown>;
902
+ kind: "model_judge";
903
+ rubric: string;
904
+ judge: {
905
+ providerId: "openpond" | "codex" | "anthropic" | "openai" | "xai" | "google" | "openrouter" | "deepseek" | "zai" | "moonshot" | "together" | "groq" | "custom-openai-compatible";
906
+ modelId: string;
907
+ };
908
+ calibrationFixtureRefs: string[];
909
+ calibrationStatus: "passed" | "failed" | "pending";
910
+ temperature: number;
911
+ } | {
912
+ id: string;
913
+ version: string;
914
+ label: string;
915
+ weight: number;
916
+ hardGate: boolean;
917
+ rewardEligible: boolean;
918
+ privileged: boolean;
919
+ metadata: Record<string, unknown>;
920
+ kind: "human";
921
+ rubric: string;
922
+ reviewerRole: string;
923
+ } | {
924
+ id: string;
925
+ version: string;
926
+ label: string;
927
+ weight: number;
928
+ hardGate: boolean;
929
+ rewardEligible: boolean;
930
+ privileged: boolean;
931
+ metadata: Record<string, unknown>;
932
+ kind: "custom_verifier";
933
+ module: string;
934
+ exportName: string;
935
+ timeoutMs: number;
936
+ networkPolicy: "none";
937
+ runtime?: "isolated_javascript" | "sandbox_process" | undefined;
938
+ })[];
939
+ graderFixtures: {
940
+ id: string;
941
+ taskId: string;
942
+ label: "infrastructure_failure" | "positive" | "negative" | "boundary" | "adversarial" | "prompt_injection";
943
+ output: Record<string, unknown>;
944
+ infrastructureError: string | null;
945
+ expectedPassed: boolean;
946
+ expectedRewardEligible: boolean;
947
+ metadata: Record<string, unknown>;
948
+ }[];
949
+ learningSignals: {
950
+ demonstrations: {
951
+ id: string;
952
+ taskId: string | null;
953
+ sourceRefs: string[];
954
+ artifactRef: string;
955
+ approved: boolean;
956
+ confidence: number;
957
+ metadata: Record<string, unknown>;
958
+ kind: "demonstration";
959
+ prompt: string | null;
960
+ response: string | null;
961
+ }[];
962
+ preferences: {
963
+ id: string;
964
+ taskId: string | null;
965
+ sourceRefs: string[];
966
+ artifactRef: string;
967
+ approved: boolean;
968
+ confidence: number;
969
+ metadata: Record<string, unknown>;
970
+ kind: "preference";
971
+ prompt: string;
972
+ chosen: string;
973
+ rejected: string;
974
+ rationale: string | null;
975
+ }[];
976
+ corrections: {
977
+ id: string;
978
+ taskId: string | null;
979
+ sourceRefs: string[];
980
+ artifactRef: string;
981
+ approved: boolean;
982
+ confidence: number;
983
+ metadata: Record<string, unknown>;
984
+ kind: "correction";
985
+ original: string;
986
+ corrected: string;
987
+ rationale: string | null;
988
+ }[];
989
+ feedback: {
990
+ id: string;
991
+ taskId: string | null;
992
+ sourceRefs: string[];
993
+ artifactRef: string;
994
+ approved: boolean;
995
+ confidence: number;
996
+ metadata: Record<string, unknown>;
997
+ kind: "feedback";
998
+ feedback: string;
999
+ polarity: "neutral" | "positive" | "negative" | "mixed";
1000
+ }[];
1001
+ rewards: {
1002
+ id: string;
1003
+ taskId: string | null;
1004
+ sourceRefs: string[];
1005
+ artifactRef: string;
1006
+ approved: boolean;
1007
+ confidence: number;
1008
+ metadata: Record<string, unknown>;
1009
+ kind: "reward";
1010
+ task: string;
1011
+ rules: {
1012
+ id: string;
1013
+ points: number;
1014
+ condition: string;
1015
+ }[];
1016
+ otherwisePoints: number;
1017
+ executable: boolean;
1018
+ }[];
1019
+ labels: {
1020
+ id: string;
1021
+ taskId: string | null;
1022
+ sourceRefs: string[];
1023
+ artifactRef: string;
1024
+ approved: boolean;
1025
+ confidence: number;
1026
+ metadata: Record<string, unknown>;
1027
+ kind: "label";
1028
+ labelKind: "rubric";
1029
+ task: string;
1030
+ criteria: {
1031
+ id: string;
1032
+ label: string;
1033
+ description: string;
1034
+ }[];
1035
+ calibrationExamples: {
1036
+ positive: string;
1037
+ negative: string;
1038
+ boundary: string;
1039
+ };
1040
+ }[];
1041
+ };
1042
+ publishedTasksetRef: {
1043
+ id: string;
1044
+ revision: number;
1045
+ contentHash: string;
1046
+ } | null;
1047
+ createdAt: string;
1048
+ updatedAt: string;
1049
+ metadata: Record<string, unknown>;
1050
+ datasetArtifact?: {
1051
+ schemaVersion: "openpond.datasetArtifact.v1";
1052
+ id: string;
1053
+ tasksetId: string;
1054
+ tasksetRevision: number;
1055
+ contentHash: string;
1056
+ format: "parquet";
1057
+ schema: {
1058
+ schemaVersion: "openpond.datasetSemanticSchema.v1";
1059
+ fields: {
1060
+ name: string;
1061
+ semanticRole: "metadata" | "reward" | "feedback" | "messages" | "row_id" | "cluster_id" | "prompt" | "demonstration" | "chosen" | "rejected" | "expected_output" | "privileged_context" | "tag";
1062
+ logicalType: "string" | "boolean" | "messages" | "integer" | "float" | "json";
1063
+ nullable: boolean;
1064
+ policy: "metadata" | "visible" | "privileged";
1065
+ }[];
1066
+ schemaHash: string;
1067
+ };
1068
+ shards: {
1069
+ id: string;
1070
+ split: "train" | "validation" | "test" | "frozen_eval";
1071
+ path: string;
1072
+ contentHash: string;
1073
+ schemaHash: string;
1074
+ sizeBytes: number;
1075
+ rowCount: number;
1076
+ rowGroupCount: number;
1077
+ }[];
1078
+ rowCount: number;
1079
+ splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
1080
+ sourceReceiptRefs: string[];
1081
+ mappingHash: string;
1082
+ qualityReportHash: string;
1083
+ createdAt: string;
1084
+ } | null | undefined;
1085
+ };
1086
+ files: {
1087
+ path: string;
1088
+ contentHash: string;
1089
+ sizeBytes: number;
1090
+ base64: string;
1091
+ }[];
1092
+ contentHash: string;
1093
+ };
1094
+ publication: {
1095
+ tasksetId: string;
1096
+ release: {
1097
+ id: string;
1098
+ revision: number;
1099
+ contentHash: string;
1100
+ };
1101
+ packageHash: string;
1102
+ } | null;
1103
+ originProjectId?: string | undefined;
1104
+ ownerScope?: "workspace" | "personal" | undefined;
1105
+ };
1106
+ base: {
1107
+ schemaVersion: "openpond.datasetWorkspaceReceipt.v1";
1108
+ teamId: string;
1109
+ datasetId: string;
1110
+ revision: number;
1111
+ workspace: {
1112
+ schemaVersion: "openpond.tasksetDraftWorkspace.v1";
1113
+ draft: {
1114
+ schemaVersion: "openpond.tasksetDraft.v1";
1115
+ id: string;
1116
+ revision: number;
1117
+ profileId: string;
1118
+ modelScope: {
1119
+ modelId: string;
1120
+ expectedModelRevision: number;
1121
+ source?: {
1122
+ schemaVersion: "openpond.modelTasksetDraftPreparation.v1";
1123
+ draftId: string;
1124
+ requestHash: string;
1125
+ sourcePackageHash: string;
1126
+ sourceTasksetRef: {
1127
+ id: string;
1128
+ contentHash: string;
1129
+ revision: number;
1130
+ };
1131
+ tasksetId: string;
1132
+ tasksetRevision: number;
1133
+ lineage: {
1134
+ schemaVersion: "openpond.modelTasksetAuthoring.v1";
1135
+ owner: {
1136
+ scopeId: string;
1137
+ modelId: string;
1138
+ };
1139
+ root: {
1140
+ id: string;
1141
+ contentHash: string;
1142
+ revision: number;
1143
+ };
1144
+ parent: {
1145
+ id: string;
1146
+ contentHash: string;
1147
+ revision: number;
1148
+ };
1149
+ };
1150
+ authoringGraph?: "bound" | undefined;
1151
+ } | undefined;
1152
+ } | null;
1153
+ name: string;
1154
+ objective: string;
1155
+ purpose: "general" | "benchmark";
1156
+ benchmark: {
1157
+ schemaVersion: "openpond.tasksetBenchmark.v1";
1158
+ definitionId: string;
1159
+ releaseId: string;
1160
+ releaseHash: string;
1161
+ managedReleasePath: string;
1162
+ adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
1163
+ evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
1164
+ primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
1165
+ qualityGate: "none" | "non_regression" | "all_pass";
1166
+ source: "builtin" | "imported";
1167
+ metadata: Record<string, unknown>;
1168
+ } | null;
1169
+ preferenceComparison: {
1170
+ schemaVersion: "openpond.tasksetPreferenceComparisonBinding.v1";
1171
+ releaseId: string;
1172
+ releaseHash: string;
1173
+ publishedAt: string;
1174
+ metadata: Record<string, unknown>;
1175
+ } | null;
1176
+ status: "draft" | "validating" | "needs_review" | "published";
1177
+ sourceRefs: ({
1178
+ id: string;
1179
+ profileId: string;
1180
+ title: string;
1181
+ sourceHash: string;
1182
+ occurredAt: string;
1183
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
1184
+ secretScanStatus: "passed" | "pending" | "blocked";
1185
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
1186
+ metadata: Record<string, unknown>;
1187
+ schemaVersion: "openpond.huggingFaceDatasetSource.v1";
1188
+ kind: "huggingface";
1189
+ repositoryId: string;
1190
+ repositoryUrl: string;
1191
+ revision: string;
1192
+ configuration: string;
1193
+ upstreamSplits: string[];
1194
+ gated: boolean;
1195
+ private: boolean;
1196
+ declaredLicense: string | null;
1197
+ sourceFileHashes: string[];
1198
+ } | {
1199
+ id: string;
1200
+ profileId: string;
1201
+ title: string;
1202
+ sourceHash: string;
1203
+ occurredAt: string;
1204
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
1205
+ secretScanStatus: "passed" | "pending" | "blocked";
1206
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
1207
+ metadata: Record<string, unknown>;
1208
+ schemaVersion: "openpond.uploadedFileDatasetSource.v1";
1209
+ kind: "uploaded_file";
1210
+ originalFileNames: string[];
1211
+ mediaTypes: string[];
1212
+ sourceFileHashes: string[];
1213
+ totalBytes: number;
1214
+ parserVersion: string;
1215
+ } | {
1216
+ id: string;
1217
+ profileId: string;
1218
+ title: string;
1219
+ sourceHash: string;
1220
+ occurredAt: string;
1221
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
1222
+ secretScanStatus: "passed" | "pending" | "blocked";
1223
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
1224
+ metadata: Record<string, unknown>;
1225
+ schemaVersion: "openpond.generatedDatasetSource.v1";
1226
+ kind: "generated";
1227
+ generatorId: string;
1228
+ generatorVersion: string;
1229
+ seed: number;
1230
+ generatorHash: string;
1231
+ } | {
1232
+ id: string;
1233
+ profileId: string;
1234
+ title: string;
1235
+ sourceHash: string;
1236
+ occurredAt: string;
1237
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
1238
+ secretScanStatus: "passed" | "pending" | "blocked";
1239
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
1240
+ metadata: Record<string, unknown>;
1241
+ schemaVersion: "openpond.learningBatchDatasetSource.v1";
1242
+ kind: "learning_batch";
1243
+ batch: {
1244
+ id: string;
1245
+ revision: number;
1246
+ contentHash: string;
1247
+ };
1248
+ taskDefinition: {
1249
+ id: string;
1250
+ revision: number;
1251
+ contentHash: string;
1252
+ };
1253
+ admittedBy: string;
1254
+ } | {
1255
+ schemaVersion: "openpond.trainingSource.v1";
1256
+ id: string;
1257
+ profileId: string;
1258
+ sessionId: string;
1259
+ turnIds: string[];
1260
+ workspaceId: string | null;
1261
+ sourceHash: string;
1262
+ clusterKey: string;
1263
+ title: string;
1264
+ occurredAt: string;
1265
+ consent: {
1266
+ status: "pending" | "granted" | "denied" | "revoked";
1267
+ scope: "metadata_only" | "selected_turns" | "full_session";
1268
+ grantedBy: string | null;
1269
+ grantedAt: string | null;
1270
+ purpose: "task_authoring_and_evaluation";
1271
+ };
1272
+ connectedAppIds: string[];
1273
+ secretScanStatus: "passed" | "pending" | "blocked";
1274
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
1275
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
1276
+ metadata: Record<string, unknown>;
1277
+ })[];
1278
+ policy: {
1279
+ policyVisibleFields: string[];
1280
+ privilegedFields: string[];
1281
+ hiddenGraderRefs: string[];
1282
+ connectedAppScopes: string[];
1283
+ };
1284
+ environment: {
1285
+ protocolVersion: "openpond.taskEnvironment.v1";
1286
+ kind: "chat" | "agent" | "program" | "stateful_harness" | "work";
1287
+ entrypoint: string;
1288
+ stateful: boolean;
1289
+ deterministicSeeds: boolean;
1290
+ toolNames: string[];
1291
+ lifecycle: ("grade" | "create" | "reset" | "step" | "cleanup")[];
1292
+ defaultTimeoutMs: number;
1293
+ networkPolicy: "none" | "declared_read_only" | "declared_scoped";
1294
+ metadata: Record<string, unknown>;
1295
+ actionBindings?: {
1296
+ schemaVersion: "openpond.harnessActionBinding.v1";
1297
+ actionId: string;
1298
+ modelToolName: string;
1299
+ description: string;
1300
+ inputSchema: Record<string, unknown>;
1301
+ actionSchemaHash: string;
1302
+ agentRelease: {
1303
+ id: string;
1304
+ contentHash: string;
1305
+ };
1306
+ implementationHash: string;
1307
+ runtimeBindingId: string;
1308
+ capabilityReceiptHash: string;
1309
+ sideEffect: "read" | "write";
1310
+ studentVisible: boolean;
1311
+ timeoutMs: number;
1312
+ episodeArgumentBindings: {
1313
+ argument: string;
1314
+ source: "case_id";
1315
+ }[];
1316
+ }[] | undefined;
1317
+ resources?: {
1318
+ id: string;
1319
+ kind: "file" | "configuration" | "catalog" | "code_module";
1320
+ path: string;
1321
+ visibility: "privileged" | "policy_visible" | "policy_hidden";
1322
+ required: boolean;
1323
+ metadata: Record<string, unknown>;
1324
+ mediaType?: string | null | undefined;
1325
+ }[] | undefined;
1326
+ };
1327
+ output: {
1328
+ mode: "artifacts" | "text" | "structured_json";
1329
+ jsonSchema: Record<string, unknown> | null;
1330
+ renderer: {
1331
+ module: string;
1332
+ exportName: string;
1333
+ configRef: string | null;
1334
+ } | null;
1335
+ };
1336
+ capabilities: {
1337
+ schemaVersion: "openpond.tasksetCapabilities.v1";
1338
+ taskKind: "chat" | "single_agent" | "multi_agent" | "custom_program";
1339
+ supportedSignals: ("label" | "reward" | "preference" | "feedback" | "demonstration" | "correction")[];
1340
+ compatibleMethods: ("none" | "sft" | "dpo" | "grpo" | "ppo" | "sdft" | "opd" | "opsd" | "sdpo" | "retrieval")[];
1341
+ rewardKinds: ("none" | "exact" | "deterministic" | "model_judge" | "human")[];
1342
+ requiresTools: boolean;
1343
+ requiresState: boolean;
1344
+ requiresPrivilegedGrading: boolean;
1345
+ environmentPlacements: ("local" | "remote" | "colocated" | "provider_native")[];
1346
+ exportable: boolean;
1347
+ portabilityBlockers: string[];
1348
+ };
1349
+ metrics: {
1350
+ schemaVersion: "openpond.tasksetMetricPolicy.v1";
1351
+ primaryMetric: string;
1352
+ aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
1353
+ missingReward: "zero" | "exclude";
1354
+ customAggregator: {
1355
+ module: string;
1356
+ exportName: string;
1357
+ contentHash: string;
1358
+ timeoutMs: number;
1359
+ networkPolicy: "none";
1360
+ } | null;
1361
+ taskWeights?: Record<string, number> | undefined;
1362
+ };
1363
+ review: {
1364
+ enabled: boolean;
1365
+ candidateCount: number;
1366
+ minimumSamples: number;
1367
+ allowTies: boolean;
1368
+ allowRejectAll: boolean;
1369
+ rubric: string;
1370
+ criteria: {
1371
+ id: string;
1372
+ label: string;
1373
+ description: string;
1374
+ weight: number;
1375
+ }[];
1376
+ };
1377
+ tasks: {
1378
+ schemaVersion: "openpond.taskData.v1";
1379
+ id: string;
1380
+ clusterKey: string;
1381
+ split: "train" | "validation" | "test" | "frozen_eval";
1382
+ input: Record<string, unknown>;
1383
+ expectedOutput: Record<string, unknown> | null;
1384
+ policyVisibleContext: Record<string, unknown>;
1385
+ privilegedContextRef: string | null;
1386
+ tags: string[];
1387
+ metadata: Record<string, unknown>;
1388
+ sourceRefs: string[];
1389
+ assets?: {
1390
+ id: string;
1391
+ sourceRefId: string;
1392
+ artifactRef: string;
1393
+ fileName: string;
1394
+ mediaType: string;
1395
+ sha256: string;
1396
+ sizeBytes: number;
1397
+ split: "train" | "validation" | "test" | "frozen_eval";
1398
+ metadata: Record<string, unknown>;
1399
+ }[] | undefined;
1400
+ resourceRefs?: string[] | undefined;
1401
+ requiredOutputs?: {
1402
+ path: string;
1403
+ mediaType: string;
1404
+ metadata: Record<string, unknown>;
1405
+ schemaRef?: string | null | undefined;
1406
+ maxBytes?: number | undefined;
1407
+ }[] | undefined;
1408
+ }[];
1409
+ graders: ({
1410
+ id: string;
1411
+ version: string;
1412
+ label: string;
1413
+ weight: number;
1414
+ hardGate: boolean;
1415
+ rewardEligible: boolean;
1416
+ privileged: boolean;
1417
+ metadata: Record<string, unknown>;
1418
+ kind: "file" | "content" | "state" | "test" | "schema" | "diff" | "runtime_event";
1419
+ config: Record<string, unknown>;
1420
+ } | {
1421
+ id: string;
1422
+ version: string;
1423
+ label: string;
1424
+ weight: number;
1425
+ hardGate: boolean;
1426
+ rewardEligible: boolean;
1427
+ privileged: boolean;
1428
+ metadata: Record<string, unknown>;
1429
+ kind: "model_judge";
1430
+ rubric: string;
1431
+ judge: {
1432
+ providerId: "openpond" | "codex" | "anthropic" | "openai" | "xai" | "google" | "openrouter" | "deepseek" | "zai" | "moonshot" | "together" | "groq" | "custom-openai-compatible";
1433
+ modelId: string;
1434
+ };
1435
+ calibrationFixtureRefs: string[];
1436
+ calibrationStatus: "passed" | "failed" | "pending";
1437
+ temperature: number;
1438
+ } | {
1439
+ id: string;
1440
+ version: string;
1441
+ label: string;
1442
+ weight: number;
1443
+ hardGate: boolean;
1444
+ rewardEligible: boolean;
1445
+ privileged: boolean;
1446
+ metadata: Record<string, unknown>;
1447
+ kind: "human";
1448
+ rubric: string;
1449
+ reviewerRole: string;
1450
+ } | {
1451
+ id: string;
1452
+ version: string;
1453
+ label: string;
1454
+ weight: number;
1455
+ hardGate: boolean;
1456
+ rewardEligible: boolean;
1457
+ privileged: boolean;
1458
+ metadata: Record<string, unknown>;
1459
+ kind: "custom_verifier";
1460
+ module: string;
1461
+ exportName: string;
1462
+ timeoutMs: number;
1463
+ networkPolicy: "none";
1464
+ runtime?: "isolated_javascript" | "sandbox_process" | undefined;
1465
+ })[];
1466
+ graderFixtures: {
1467
+ id: string;
1468
+ taskId: string;
1469
+ label: "infrastructure_failure" | "positive" | "negative" | "boundary" | "adversarial" | "prompt_injection";
1470
+ output: Record<string, unknown>;
1471
+ infrastructureError: string | null;
1472
+ expectedPassed: boolean;
1473
+ expectedRewardEligible: boolean;
1474
+ metadata: Record<string, unknown>;
1475
+ }[];
1476
+ learningSignals: {
1477
+ demonstrations: {
1478
+ id: string;
1479
+ taskId: string | null;
1480
+ sourceRefs: string[];
1481
+ artifactRef: string;
1482
+ approved: boolean;
1483
+ confidence: number;
1484
+ metadata: Record<string, unknown>;
1485
+ kind: "demonstration";
1486
+ prompt: string | null;
1487
+ response: string | null;
1488
+ }[];
1489
+ preferences: {
1490
+ id: string;
1491
+ taskId: string | null;
1492
+ sourceRefs: string[];
1493
+ artifactRef: string;
1494
+ approved: boolean;
1495
+ confidence: number;
1496
+ metadata: Record<string, unknown>;
1497
+ kind: "preference";
1498
+ prompt: string;
1499
+ chosen: string;
1500
+ rejected: string;
1501
+ rationale: string | null;
1502
+ }[];
1503
+ corrections: {
1504
+ id: string;
1505
+ taskId: string | null;
1506
+ sourceRefs: string[];
1507
+ artifactRef: string;
1508
+ approved: boolean;
1509
+ confidence: number;
1510
+ metadata: Record<string, unknown>;
1511
+ kind: "correction";
1512
+ original: string;
1513
+ corrected: string;
1514
+ rationale: string | null;
1515
+ }[];
1516
+ feedback: {
1517
+ id: string;
1518
+ taskId: string | null;
1519
+ sourceRefs: string[];
1520
+ artifactRef: string;
1521
+ approved: boolean;
1522
+ confidence: number;
1523
+ metadata: Record<string, unknown>;
1524
+ kind: "feedback";
1525
+ feedback: string;
1526
+ polarity: "neutral" | "positive" | "negative" | "mixed";
1527
+ }[];
1528
+ rewards: {
1529
+ id: string;
1530
+ taskId: string | null;
1531
+ sourceRefs: string[];
1532
+ artifactRef: string;
1533
+ approved: boolean;
1534
+ confidence: number;
1535
+ metadata: Record<string, unknown>;
1536
+ kind: "reward";
1537
+ task: string;
1538
+ rules: {
1539
+ id: string;
1540
+ points: number;
1541
+ condition: string;
1542
+ }[];
1543
+ otherwisePoints: number;
1544
+ executable: boolean;
1545
+ }[];
1546
+ labels: {
1547
+ id: string;
1548
+ taskId: string | null;
1549
+ sourceRefs: string[];
1550
+ artifactRef: string;
1551
+ approved: boolean;
1552
+ confidence: number;
1553
+ metadata: Record<string, unknown>;
1554
+ kind: "label";
1555
+ labelKind: "rubric";
1556
+ task: string;
1557
+ criteria: {
1558
+ id: string;
1559
+ label: string;
1560
+ description: string;
1561
+ }[];
1562
+ calibrationExamples: {
1563
+ positive: string;
1564
+ negative: string;
1565
+ boundary: string;
1566
+ };
1567
+ }[];
1568
+ };
1569
+ publishedTasksetRef: {
1570
+ id: string;
1571
+ revision: number;
1572
+ contentHash: string;
1573
+ } | null;
1574
+ createdAt: string;
1575
+ updatedAt: string;
1576
+ metadata: Record<string, unknown>;
1577
+ datasetArtifact?: {
1578
+ schemaVersion: "openpond.datasetArtifact.v1";
1579
+ id: string;
1580
+ tasksetId: string;
1581
+ tasksetRevision: number;
1582
+ contentHash: string;
1583
+ format: "parquet";
1584
+ schema: {
1585
+ schemaVersion: "openpond.datasetSemanticSchema.v1";
1586
+ fields: {
1587
+ name: string;
1588
+ semanticRole: "metadata" | "reward" | "feedback" | "messages" | "row_id" | "cluster_id" | "prompt" | "demonstration" | "chosen" | "rejected" | "expected_output" | "privileged_context" | "tag";
1589
+ logicalType: "string" | "boolean" | "messages" | "integer" | "float" | "json";
1590
+ nullable: boolean;
1591
+ policy: "metadata" | "visible" | "privileged";
1592
+ }[];
1593
+ schemaHash: string;
1594
+ };
1595
+ shards: {
1596
+ id: string;
1597
+ split: "train" | "validation" | "test" | "frozen_eval";
1598
+ path: string;
1599
+ contentHash: string;
1600
+ schemaHash: string;
1601
+ sizeBytes: number;
1602
+ rowCount: number;
1603
+ rowGroupCount: number;
1604
+ }[];
1605
+ rowCount: number;
1606
+ splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
1607
+ sourceReceiptRefs: string[];
1608
+ mappingHash: string;
1609
+ qualityReportHash: string;
1610
+ createdAt: string;
1611
+ } | null | undefined;
1612
+ };
1613
+ files: {
1614
+ path: string;
1615
+ contentHash: string;
1616
+ sizeBytes: number;
1617
+ base64: string;
1618
+ }[];
1619
+ contentHash: string;
1620
+ };
1621
+ publication: {
1622
+ tasksetId: string;
1623
+ release: {
1624
+ id: string;
1625
+ revision: number;
1626
+ contentHash: string;
1627
+ };
1628
+ packageHash: string;
1629
+ } | null;
1630
+ originProjectId?: string | undefined;
1631
+ ownerScope?: "workspace" | "personal" | undefined;
1632
+ } | null;
1633
+ } | null>;
567
1634
  beginVersion(id: string, input: z.input<typeof DatasetWorkspaceBeginVersionSchema>, signal?: AbortSignal): Promise<{
568
1635
  schemaVersion: "openpond.datasetWorkspaceReceipt.v1";
569
1636
  teamId: string;
@@ -620,8 +1687,8 @@ export declare class OpenPondDatasetWorkspaceClient {
620
1687
  releaseId: string;
621
1688
  releaseHash: string;
622
1689
  managedReleasePath: string;
623
- adaptationSplit: "train" | "validation" | "frozen_eval" | "test";
624
- evaluationSplit: "train" | "validation" | "frozen_eval" | "test";
1690
+ adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
1691
+ evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
625
1692
  primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
626
1693
  qualityGate: "none" | "non_regression" | "all_pass";
627
1694
  source: "builtin" | "imported";
@@ -839,7 +1906,7 @@ export declare class OpenPondDatasetWorkspaceClient {
839
1906
  schemaVersion: "openpond.taskData.v1";
840
1907
  id: string;
841
1908
  clusterKey: string;
842
- split: "train" | "validation" | "frozen_eval" | "test";
1909
+ split: "train" | "validation" | "test" | "frozen_eval";
843
1910
  input: Record<string, unknown>;
844
1911
  expectedOutput: Record<string, unknown> | null;
845
1912
  policyVisibleContext: Record<string, unknown>;
@@ -855,7 +1922,7 @@ export declare class OpenPondDatasetWorkspaceClient {
855
1922
  mediaType: string;
856
1923
  sha256: string;
857
1924
  sizeBytes: number;
858
- split: "train" | "validation" | "frozen_eval" | "test";
1925
+ split: "train" | "validation" | "test" | "frozen_eval";
859
1926
  metadata: Record<string, unknown>;
860
1927
  }[] | undefined;
861
1928
  resourceRefs?: string[] | undefined;
@@ -1055,7 +2122,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1055
2122
  };
1056
2123
  shards: {
1057
2124
  id: string;
1058
- split: "train" | "validation" | "frozen_eval" | "test";
2125
+ split: "train" | "validation" | "test" | "frozen_eval";
1059
2126
  path: string;
1060
2127
  contentHash: string;
1061
2128
  schemaHash: string;
@@ -1064,7 +2131,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1064
2131
  rowGroupCount: number;
1065
2132
  }[];
1066
2133
  rowCount: number;
1067
- splitCounts: Record<"train" | "validation" | "frozen_eval" | "test", number>;
2134
+ splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
1068
2135
  sourceReceiptRefs: string[];
1069
2136
  mappingHash: string;
1070
2137
  qualityReportHash: string;
@@ -1168,8 +2235,8 @@ export declare class OpenPondDatasetWorkspaceClient {
1168
2235
  releaseId: string;
1169
2236
  releaseHash: string;
1170
2237
  managedReleasePath: string;
1171
- adaptationSplit: "train" | "validation" | "frozen_eval" | "test";
1172
- evaluationSplit: "train" | "validation" | "frozen_eval" | "test";
2238
+ adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
2239
+ evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
1173
2240
  primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
1174
2241
  qualityGate: "none" | "non_regression" | "all_pass";
1175
2242
  source: "builtin" | "imported";
@@ -1387,7 +2454,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1387
2454
  schemaVersion: "openpond.taskData.v1";
1388
2455
  id: string;
1389
2456
  clusterKey: string;
1390
- split: "train" | "validation" | "frozen_eval" | "test";
2457
+ split: "train" | "validation" | "test" | "frozen_eval";
1391
2458
  input: Record<string, unknown>;
1392
2459
  expectedOutput: Record<string, unknown> | null;
1393
2460
  policyVisibleContext: Record<string, unknown>;
@@ -1403,7 +2470,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1403
2470
  mediaType: string;
1404
2471
  sha256: string;
1405
2472
  sizeBytes: number;
1406
- split: "train" | "validation" | "frozen_eval" | "test";
2473
+ split: "train" | "validation" | "test" | "frozen_eval";
1407
2474
  metadata: Record<string, unknown>;
1408
2475
  }[] | undefined;
1409
2476
  resourceRefs?: string[] | undefined;
@@ -1603,7 +2670,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1603
2670
  };
1604
2671
  shards: {
1605
2672
  id: string;
1606
- split: "train" | "validation" | "frozen_eval" | "test";
2673
+ split: "train" | "validation" | "test" | "frozen_eval";
1607
2674
  path: string;
1608
2675
  contentHash: string;
1609
2676
  schemaHash: string;
@@ -1612,7 +2679,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1612
2679
  rowGroupCount: number;
1613
2680
  }[];
1614
2681
  rowCount: number;
1615
- splitCounts: Record<"train" | "validation" | "frozen_eval" | "test", number>;
2682
+ splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
1616
2683
  sourceReceiptRefs: string[];
1617
2684
  mappingHash: string;
1618
2685
  qualityReportHash: string;
@@ -1695,8 +2762,8 @@ export declare class OpenPondDatasetWorkspaceClient {
1695
2762
  releaseId: string;
1696
2763
  releaseHash: string;
1697
2764
  managedReleasePath: string;
1698
- adaptationSplit: "train" | "validation" | "frozen_eval" | "test";
1699
- evaluationSplit: "train" | "validation" | "frozen_eval" | "test";
2765
+ adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
2766
+ evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
1700
2767
  primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
1701
2768
  qualityGate: "none" | "non_regression" | "all_pass";
1702
2769
  source: "builtin" | "imported";
@@ -1914,7 +2981,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1914
2981
  schemaVersion: "openpond.taskData.v1";
1915
2982
  id: string;
1916
2983
  clusterKey: string;
1917
- split: "train" | "validation" | "frozen_eval" | "test";
2984
+ split: "train" | "validation" | "test" | "frozen_eval";
1918
2985
  input: Record<string, unknown>;
1919
2986
  expectedOutput: Record<string, unknown> | null;
1920
2987
  policyVisibleContext: Record<string, unknown>;
@@ -1930,7 +2997,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1930
2997
  mediaType: string;
1931
2998
  sha256: string;
1932
2999
  sizeBytes: number;
1933
- split: "train" | "validation" | "frozen_eval" | "test";
3000
+ split: "train" | "validation" | "test" | "frozen_eval";
1934
3001
  metadata: Record<string, unknown>;
1935
3002
  }[] | undefined;
1936
3003
  resourceRefs?: string[] | undefined;
@@ -2130,7 +3197,7 @@ export declare class OpenPondDatasetWorkspaceClient {
2130
3197
  };
2131
3198
  shards: {
2132
3199
  id: string;
2133
- split: "train" | "validation" | "frozen_eval" | "test";
3200
+ split: "train" | "validation" | "test" | "frozen_eval";
2134
3201
  path: string;
2135
3202
  contentHash: string;
2136
3203
  schemaHash: string;
@@ -2139,7 +3206,7 @@ export declare class OpenPondDatasetWorkspaceClient {
2139
3206
  rowGroupCount: number;
2140
3207
  }[];
2141
3208
  rowCount: number;
2142
- splitCounts: Record<"train" | "validation" | "frozen_eval" | "test", number>;
3209
+ splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
2143
3210
  sourceReceiptRefs: string[];
2144
3211
  mappingHash: string;
2145
3212
  qualityReportHash: string;
@@ -2234,8 +3301,8 @@ export declare class OpenPondDatasetWorkspaceClient {
2234
3301
  releaseId: string;
2235
3302
  releaseHash: string;
2236
3303
  managedReleasePath: string;
2237
- adaptationSplit: "train" | "validation" | "frozen_eval" | "test";
2238
- evaluationSplit: "train" | "validation" | "frozen_eval" | "test";
3304
+ adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
3305
+ evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
2239
3306
  primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
2240
3307
  qualityGate: "none" | "non_regression" | "all_pass";
2241
3308
  source: "builtin" | "imported";
@@ -2453,7 +3520,7 @@ export declare class OpenPondDatasetWorkspaceClient {
2453
3520
  schemaVersion: "openpond.taskData.v1";
2454
3521
  id: string;
2455
3522
  clusterKey: string;
2456
- split: "train" | "validation" | "frozen_eval" | "test";
3523
+ split: "train" | "validation" | "test" | "frozen_eval";
2457
3524
  input: Record<string, unknown>;
2458
3525
  expectedOutput: Record<string, unknown> | null;
2459
3526
  policyVisibleContext: Record<string, unknown>;
@@ -2469,7 +3536,7 @@ export declare class OpenPondDatasetWorkspaceClient {
2469
3536
  mediaType: string;
2470
3537
  sha256: string;
2471
3538
  sizeBytes: number;
2472
- split: "train" | "validation" | "frozen_eval" | "test";
3539
+ split: "train" | "validation" | "test" | "frozen_eval";
2473
3540
  metadata: Record<string, unknown>;
2474
3541
  }[] | undefined;
2475
3542
  resourceRefs?: string[] | undefined;
@@ -2669,7 +3736,7 @@ export declare class OpenPondDatasetWorkspaceClient {
2669
3736
  };
2670
3737
  shards: {
2671
3738
  id: string;
2672
- split: "train" | "validation" | "frozen_eval" | "test";
3739
+ split: "train" | "validation" | "test" | "frozen_eval";
2673
3740
  path: string;
2674
3741
  contentHash: string;
2675
3742
  schemaHash: string;
@@ -2678,7 +3745,7 @@ export declare class OpenPondDatasetWorkspaceClient {
2678
3745
  rowGroupCount: number;
2679
3746
  }[];
2680
3747
  rowCount: number;
2681
- splitCounts: Record<"train" | "validation" | "frozen_eval" | "test", number>;
3748
+ splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
2682
3749
  sourceReceiptRefs: string[];
2683
3750
  mappingHash: string;
2684
3751
  qualityReportHash: string;