openpond-sdk 0.9.7 → 0.9.9

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 (79) hide show
  1. package/dist/{chunk-SR55WVPK.js → chunk-4HZORGK3.js} +2 -2
  2. package/dist/{chunk-CNGJF6AM.js → chunk-4JRSH2TT.js} +4 -4
  3. package/dist/chunk-4JRSH2TT.js.map +7 -0
  4. package/dist/{chunk-XR6CLSCS.js → chunk-BGXWHDYM.js} +3 -2
  5. package/dist/chunk-BGXWHDYM.js.map +7 -0
  6. package/dist/{chunk-A4ZDOZKP.js → chunk-K3JCP4FO.js} +11 -5
  7. package/dist/chunk-K3JCP4FO.js.map +7 -0
  8. package/dist/{chunk-TQ5ASK7D.js → chunk-XMKEFQ62.js} +2 -2
  9. package/dist/{chunk-62LWLACP.js → chunk-YYJ4T4U4.js} +2 -2
  10. package/dist/dataset-marketplace.js +11 -0
  11. package/dist/dataset-marketplace.js.map +2 -2
  12. package/dist/dataset-workspaces.js +78 -25
  13. package/dist/dataset-workspaces.js.map +4 -4
  14. package/dist/experiments.js +92 -5
  15. package/dist/experiments.js.map +4 -4
  16. package/dist/index.js +4 -4
  17. package/dist/learning.js +8 -0
  18. package/dist/learning.js.map +3 -3
  19. package/dist/model-starter-attempts.js +1 -1
  20. package/dist/model-taskset-runs.js +2 -2
  21. package/dist/taskset-catalog.js +1 -1
  22. package/dist/taskset-drafts.js +2 -2
  23. package/dist/taskset-packages.js +2 -2
  24. package/dist/types/packages/sdk/src/dataset-grader-selection.d.ts +24 -0
  25. package/dist/types/packages/sdk/src/dataset-grader-selection.d.ts.map +1 -0
  26. package/dist/types/packages/sdk/src/dataset-marketplace.d.ts +9 -0
  27. package/dist/types/packages/sdk/src/dataset-marketplace.d.ts.map +1 -1
  28. package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts +1099 -34
  29. package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts.map +1 -1
  30. package/dist/types/packages/sdk/src/dataset-workspace-contracts.d.ts +23 -12
  31. package/dist/types/packages/sdk/src/dataset-workspace-contracts.d.ts.map +1 -1
  32. package/dist/types/packages/sdk/src/dataset-workspace-operations.d.ts +1617 -0
  33. package/dist/types/packages/sdk/src/dataset-workspace-operations.d.ts.map +1 -0
  34. package/dist/types/packages/sdk/src/dataset-workspaces.d.ts +1 -0
  35. package/dist/types/packages/sdk/src/dataset-workspaces.d.ts.map +1 -1
  36. package/dist/types/packages/sdk/src/experiment-client.d.ts.map +1 -1
  37. package/dist/types/packages/sdk/src/experiment-history.d.ts +266 -0
  38. package/dist/types/packages/sdk/src/experiment-history.d.ts.map +1 -0
  39. package/dist/types/packages/sdk/src/experiments.d.ts +2 -0
  40. package/dist/types/packages/sdk/src/experiments.d.ts.map +1 -1
  41. package/dist/types/packages/sdk/src/grader-feedback-key.d.ts +3 -0
  42. package/dist/types/packages/sdk/src/grader-feedback-key.d.ts.map +1 -0
  43. package/dist/types/packages/sdk/src/learning-batch-assets.d.ts +2 -2
  44. package/dist/types/packages/sdk/src/learning-client.d.ts +1 -1
  45. package/dist/types/packages/sdk/src/learning.d.ts +1 -0
  46. package/dist/types/packages/sdk/src/learning.d.ts.map +1 -1
  47. package/dist/types/packages/sdk/src/model-batch-review-preparation.d.ts +3 -3
  48. package/dist/types/packages/sdk/src/model-batch-review.d.ts +7 -7
  49. package/dist/types/packages/sdk/src/model-learning-policy.d.ts +2 -2
  50. package/dist/types/packages/sdk/src/model-starter-attempts.d.ts +2 -1
  51. package/dist/types/packages/sdk/src/model-starter-attempts.d.ts.map +1 -1
  52. package/dist/types/packages/sdk/src/model-starter-catalog.d.ts +1 -1
  53. package/dist/types/packages/sdk/src/model-starters.d.ts +1 -1
  54. package/dist/types/packages/sdk/src/task-inventory.d.ts +1 -1
  55. package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts +6 -6
  56. package/dist/types/packages/sdk/src/taskset-authored-files.d.ts +6 -6
  57. package/dist/types/packages/sdk/src/taskset-authored-portable-release.d.ts +2 -2
  58. package/dist/types/packages/sdk/src/taskset-catalog.d.ts +7 -0
  59. package/dist/types/packages/sdk/src/taskset-catalog.d.ts.map +1 -1
  60. package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts +12 -12
  61. package/dist/types/packages/sdk/src/taskset-draft-client.d.ts +36 -36
  62. package/dist/types/packages/sdk/src/taskset-draft-core.d.ts +6 -6
  63. package/dist/types/packages/sdk/src/taskset-draft-dataset-artifacts.d.ts +11 -11
  64. package/dist/types/packages/sdk/src/taskset-draft-document.d.ts +8 -8
  65. package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts +10 -10
  66. package/dist/types/packages/sdk/src/taskset-draft-source.d.ts +12 -12
  67. package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts +12 -12
  68. package/dist/types/packages/sdk/src/taskset-package-learning.d.ts +4 -4
  69. package/dist/types/packages/sdk/src/taskset-package-projection.d.ts +11 -11
  70. package/dist/types/packages/sdk/src/training-learning-batch.d.ts +14 -14
  71. package/dist/types/packages/sdk/src/training-project-client.d.ts +8 -8
  72. package/dist/types/packages/sdk/src/training-project-contracts.d.ts +11 -11
  73. package/package.json +1 -1
  74. package/dist/chunk-A4ZDOZKP.js.map +0 -7
  75. package/dist/chunk-CNGJF6AM.js.map +0 -7
  76. package/dist/chunk-XR6CLSCS.js.map +0 -7
  77. /package/dist/{chunk-SR55WVPK.js.map → chunk-4HZORGK3.js.map} +0 -0
  78. /package/dist/{chunk-TQ5ASK7D.js.map → chunk-XMKEFQ62.js.map} +0 -0
  79. /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
- import { DatasetWorkspaceBeginVersionSchema, DatasetWorkspacePublishSchema, DatasetWorkspaceWriteSchema } from "./dataset-workspace-contracts.js";
1
+ import { DatasetWorkspaceListQuerySchema, 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 {
@@ -11,9 +12,7 @@ export declare class OpenPondDatasetWorkspaceClient {
11
12
  teamId: string;
12
13
  fetch?: typeof fetch;
13
14
  });
14
- list(options?: {
15
- cursor?: string;
16
- projectId?: string;
15
+ list(options?: z.input<typeof DatasetWorkspaceListQuerySchema> & {
17
16
  signal?: AbortSignal;
18
17
  }): Promise<{
19
18
  teamId: string;
@@ -93,8 +92,8 @@ export declare class OpenPondDatasetWorkspaceClient {
93
92
  releaseId: string;
94
93
  releaseHash: string;
95
94
  managedReleasePath: string;
96
- adaptationSplit: "train" | "validation" | "frozen_eval" | "test";
97
- evaluationSplit: "train" | "validation" | "frozen_eval" | "test";
95
+ adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
96
+ evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
98
97
  primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
99
98
  qualityGate: "none" | "non_regression" | "all_pass";
100
99
  source: "builtin" | "imported";
@@ -312,7 +311,7 @@ export declare class OpenPondDatasetWorkspaceClient {
312
311
  schemaVersion: "openpond.taskData.v1";
313
312
  id: string;
314
313
  clusterKey: string;
315
- split: "train" | "validation" | "frozen_eval" | "test";
314
+ split: "train" | "validation" | "test" | "frozen_eval";
316
315
  input: Record<string, unknown>;
317
316
  expectedOutput: Record<string, unknown> | null;
318
317
  policyVisibleContext: Record<string, unknown>;
@@ -328,7 +327,7 @@ export declare class OpenPondDatasetWorkspaceClient {
328
327
  mediaType: string;
329
328
  sha256: string;
330
329
  sizeBytes: number;
331
- split: "train" | "validation" | "frozen_eval" | "test";
330
+ split: "train" | "validation" | "test" | "frozen_eval";
332
331
  metadata: Record<string, unknown>;
333
332
  }[] | undefined;
334
333
  resourceRefs?: string[] | undefined;
@@ -528,7 +527,7 @@ export declare class OpenPondDatasetWorkspaceClient {
528
527
  };
529
528
  shards: {
530
529
  id: string;
531
- split: "train" | "validation" | "frozen_eval" | "test";
530
+ split: "train" | "validation" | "test" | "frozen_eval";
532
531
  path: string;
533
532
  contentHash: string;
534
533
  schemaHash: string;
@@ -537,7 +536,7 @@ export declare class OpenPondDatasetWorkspaceClient {
537
536
  rowGroupCount: number;
538
537
  }[];
539
538
  rowCount: number;
540
- splitCounts: Record<"train" | "validation" | "frozen_eval" | "test", number>;
539
+ splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
541
540
  sourceReceiptRefs: string[];
542
541
  mappingHash: string;
543
542
  qualityReportHash: string;
@@ -564,6 +563,1072 @@ export declare class OpenPondDatasetWorkspaceClient {
564
563
  originProjectId?: string | undefined;
565
564
  ownerScope?: "workspace" | "personal" | undefined;
566
565
  }>;
566
+ /** Recover only this actor's retained operation; base bytes allow adapters to
567
+ * reconstruct and verify the identical original CAS request before replay. */
568
+ operationResult(operationId: string, options: {
569
+ datasetId?: string;
570
+ kind: z.infer<typeof DatasetWorkspaceOperationKindSchema>;
571
+ requestHash?: string;
572
+ signal?: AbortSignal;
573
+ }): Promise<{
574
+ operationId: string;
575
+ kind: "create" | "save" | "publish" | "begin_version";
576
+ requestHash: string;
577
+ receipt: {
578
+ schemaVersion: "openpond.datasetWorkspaceReceipt.v1";
579
+ teamId: string;
580
+ datasetId: string;
581
+ revision: number;
582
+ workspace: {
583
+ schemaVersion: "openpond.tasksetDraftWorkspace.v1";
584
+ draft: {
585
+ schemaVersion: "openpond.tasksetDraft.v1";
586
+ id: string;
587
+ revision: number;
588
+ profileId: string;
589
+ modelScope: {
590
+ modelId: string;
591
+ expectedModelRevision: number;
592
+ source?: {
593
+ schemaVersion: "openpond.modelTasksetDraftPreparation.v1";
594
+ draftId: string;
595
+ requestHash: string;
596
+ sourcePackageHash: string;
597
+ sourceTasksetRef: {
598
+ id: string;
599
+ contentHash: string;
600
+ revision: number;
601
+ };
602
+ tasksetId: string;
603
+ tasksetRevision: number;
604
+ lineage: {
605
+ schemaVersion: "openpond.modelTasksetAuthoring.v1";
606
+ owner: {
607
+ scopeId: string;
608
+ modelId: string;
609
+ };
610
+ root: {
611
+ id: string;
612
+ contentHash: string;
613
+ revision: number;
614
+ };
615
+ parent: {
616
+ id: string;
617
+ contentHash: string;
618
+ revision: number;
619
+ };
620
+ };
621
+ authoringGraph?: "bound" | undefined;
622
+ } | undefined;
623
+ } | null;
624
+ name: string;
625
+ objective: string;
626
+ purpose: "general" | "benchmark";
627
+ benchmark: {
628
+ schemaVersion: "openpond.tasksetBenchmark.v1";
629
+ definitionId: string;
630
+ releaseId: string;
631
+ releaseHash: string;
632
+ managedReleasePath: string;
633
+ adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
634
+ evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
635
+ primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
636
+ qualityGate: "none" | "non_regression" | "all_pass";
637
+ source: "builtin" | "imported";
638
+ metadata: Record<string, unknown>;
639
+ } | null;
640
+ preferenceComparison: {
641
+ schemaVersion: "openpond.tasksetPreferenceComparisonBinding.v1";
642
+ releaseId: string;
643
+ releaseHash: string;
644
+ publishedAt: string;
645
+ metadata: Record<string, unknown>;
646
+ } | null;
647
+ status: "draft" | "validating" | "needs_review" | "published";
648
+ sourceRefs: ({
649
+ id: string;
650
+ profileId: string;
651
+ title: string;
652
+ sourceHash: string;
653
+ occurredAt: string;
654
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
655
+ secretScanStatus: "passed" | "pending" | "blocked";
656
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
657
+ metadata: Record<string, unknown>;
658
+ schemaVersion: "openpond.huggingFaceDatasetSource.v1";
659
+ kind: "huggingface";
660
+ repositoryId: string;
661
+ repositoryUrl: string;
662
+ revision: string;
663
+ configuration: string;
664
+ upstreamSplits: string[];
665
+ gated: boolean;
666
+ private: boolean;
667
+ declaredLicense: string | null;
668
+ sourceFileHashes: string[];
669
+ } | {
670
+ id: string;
671
+ profileId: string;
672
+ title: string;
673
+ sourceHash: string;
674
+ occurredAt: string;
675
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
676
+ secretScanStatus: "passed" | "pending" | "blocked";
677
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
678
+ metadata: Record<string, unknown>;
679
+ schemaVersion: "openpond.uploadedFileDatasetSource.v1";
680
+ kind: "uploaded_file";
681
+ originalFileNames: string[];
682
+ mediaTypes: string[];
683
+ sourceFileHashes: string[];
684
+ totalBytes: number;
685
+ parserVersion: string;
686
+ } | {
687
+ id: string;
688
+ profileId: string;
689
+ title: string;
690
+ sourceHash: string;
691
+ occurredAt: string;
692
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
693
+ secretScanStatus: "passed" | "pending" | "blocked";
694
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
695
+ metadata: Record<string, unknown>;
696
+ schemaVersion: "openpond.generatedDatasetSource.v1";
697
+ kind: "generated";
698
+ generatorId: string;
699
+ generatorVersion: string;
700
+ seed: number;
701
+ generatorHash: string;
702
+ } | {
703
+ id: string;
704
+ profileId: string;
705
+ title: string;
706
+ sourceHash: string;
707
+ occurredAt: string;
708
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
709
+ secretScanStatus: "passed" | "pending" | "blocked";
710
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
711
+ metadata: Record<string, unknown>;
712
+ schemaVersion: "openpond.learningBatchDatasetSource.v1";
713
+ kind: "learning_batch";
714
+ batch: {
715
+ id: string;
716
+ revision: number;
717
+ contentHash: string;
718
+ };
719
+ taskDefinition: {
720
+ id: string;
721
+ revision: number;
722
+ contentHash: string;
723
+ };
724
+ admittedBy: string;
725
+ } | {
726
+ schemaVersion: "openpond.trainingSource.v1";
727
+ id: string;
728
+ profileId: string;
729
+ sessionId: string;
730
+ turnIds: string[];
731
+ workspaceId: string | null;
732
+ sourceHash: string;
733
+ clusterKey: string;
734
+ title: string;
735
+ occurredAt: string;
736
+ consent: {
737
+ status: "pending" | "granted" | "denied" | "revoked";
738
+ scope: "metadata_only" | "selected_turns" | "full_session";
739
+ grantedBy: string | null;
740
+ grantedAt: string | null;
741
+ purpose: "task_authoring_and_evaluation";
742
+ };
743
+ connectedAppIds: string[];
744
+ secretScanStatus: "passed" | "pending" | "blocked";
745
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
746
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
747
+ metadata: Record<string, unknown>;
748
+ })[];
749
+ policy: {
750
+ policyVisibleFields: string[];
751
+ privilegedFields: string[];
752
+ hiddenGraderRefs: string[];
753
+ connectedAppScopes: string[];
754
+ };
755
+ environment: {
756
+ protocolVersion: "openpond.taskEnvironment.v1";
757
+ kind: "chat" | "agent" | "program" | "stateful_harness" | "work";
758
+ entrypoint: string;
759
+ stateful: boolean;
760
+ deterministicSeeds: boolean;
761
+ toolNames: string[];
762
+ lifecycle: ("grade" | "create" | "reset" | "step" | "cleanup")[];
763
+ defaultTimeoutMs: number;
764
+ networkPolicy: "none" | "declared_read_only" | "declared_scoped";
765
+ metadata: Record<string, unknown>;
766
+ actionBindings?: {
767
+ schemaVersion: "openpond.harnessActionBinding.v1";
768
+ actionId: string;
769
+ modelToolName: string;
770
+ description: string;
771
+ inputSchema: Record<string, unknown>;
772
+ actionSchemaHash: string;
773
+ agentRelease: {
774
+ id: string;
775
+ contentHash: string;
776
+ };
777
+ implementationHash: string;
778
+ runtimeBindingId: string;
779
+ capabilityReceiptHash: string;
780
+ sideEffect: "read" | "write";
781
+ studentVisible: boolean;
782
+ timeoutMs: number;
783
+ episodeArgumentBindings: {
784
+ argument: string;
785
+ source: "case_id";
786
+ }[];
787
+ }[] | undefined;
788
+ resources?: {
789
+ id: string;
790
+ kind: "file" | "configuration" | "catalog" | "code_module";
791
+ path: string;
792
+ visibility: "privileged" | "policy_visible" | "policy_hidden";
793
+ required: boolean;
794
+ metadata: Record<string, unknown>;
795
+ mediaType?: string | null | undefined;
796
+ }[] | undefined;
797
+ };
798
+ output: {
799
+ mode: "artifacts" | "text" | "structured_json";
800
+ jsonSchema: Record<string, unknown> | null;
801
+ renderer: {
802
+ module: string;
803
+ exportName: string;
804
+ configRef: string | null;
805
+ } | null;
806
+ };
807
+ capabilities: {
808
+ schemaVersion: "openpond.tasksetCapabilities.v1";
809
+ taskKind: "chat" | "single_agent" | "multi_agent" | "custom_program";
810
+ supportedSignals: ("label" | "reward" | "preference" | "feedback" | "demonstration" | "correction")[];
811
+ compatibleMethods: ("none" | "sft" | "dpo" | "grpo" | "ppo" | "sdft" | "opd" | "opsd" | "sdpo" | "retrieval")[];
812
+ rewardKinds: ("none" | "exact" | "deterministic" | "model_judge" | "human")[];
813
+ requiresTools: boolean;
814
+ requiresState: boolean;
815
+ requiresPrivilegedGrading: boolean;
816
+ environmentPlacements: ("local" | "remote" | "colocated" | "provider_native")[];
817
+ exportable: boolean;
818
+ portabilityBlockers: string[];
819
+ };
820
+ metrics: {
821
+ schemaVersion: "openpond.tasksetMetricPolicy.v1";
822
+ primaryMetric: string;
823
+ aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
824
+ missingReward: "zero" | "exclude";
825
+ customAggregator: {
826
+ module: string;
827
+ exportName: string;
828
+ contentHash: string;
829
+ timeoutMs: number;
830
+ networkPolicy: "none";
831
+ } | null;
832
+ taskWeights?: Record<string, number> | undefined;
833
+ };
834
+ review: {
835
+ enabled: boolean;
836
+ candidateCount: number;
837
+ minimumSamples: number;
838
+ allowTies: boolean;
839
+ allowRejectAll: boolean;
840
+ rubric: string;
841
+ criteria: {
842
+ id: string;
843
+ label: string;
844
+ description: string;
845
+ weight: number;
846
+ }[];
847
+ };
848
+ tasks: {
849
+ schemaVersion: "openpond.taskData.v1";
850
+ id: string;
851
+ clusterKey: string;
852
+ split: "train" | "validation" | "test" | "frozen_eval";
853
+ input: Record<string, unknown>;
854
+ expectedOutput: Record<string, unknown> | null;
855
+ policyVisibleContext: Record<string, unknown>;
856
+ privilegedContextRef: string | null;
857
+ tags: string[];
858
+ metadata: Record<string, unknown>;
859
+ sourceRefs: string[];
860
+ assets?: {
861
+ id: string;
862
+ sourceRefId: string;
863
+ artifactRef: string;
864
+ fileName: string;
865
+ mediaType: string;
866
+ sha256: string;
867
+ sizeBytes: number;
868
+ split: "train" | "validation" | "test" | "frozen_eval";
869
+ metadata: Record<string, unknown>;
870
+ }[] | undefined;
871
+ resourceRefs?: string[] | undefined;
872
+ requiredOutputs?: {
873
+ path: string;
874
+ mediaType: string;
875
+ metadata: Record<string, unknown>;
876
+ schemaRef?: string | null | undefined;
877
+ maxBytes?: number | undefined;
878
+ }[] | undefined;
879
+ }[];
880
+ graders: ({
881
+ id: string;
882
+ version: string;
883
+ label: string;
884
+ weight: number;
885
+ hardGate: boolean;
886
+ rewardEligible: boolean;
887
+ privileged: boolean;
888
+ metadata: Record<string, unknown>;
889
+ kind: "file" | "content" | "state" | "test" | "schema" | "diff" | "runtime_event";
890
+ config: Record<string, unknown>;
891
+ } | {
892
+ id: string;
893
+ version: string;
894
+ label: string;
895
+ weight: number;
896
+ hardGate: boolean;
897
+ rewardEligible: boolean;
898
+ privileged: boolean;
899
+ metadata: Record<string, unknown>;
900
+ kind: "model_judge";
901
+ rubric: string;
902
+ judge: {
903
+ providerId: "openpond" | "codex" | "anthropic" | "openai" | "xai" | "google" | "openrouter" | "deepseek" | "zai" | "moonshot" | "together" | "groq" | "custom-openai-compatible";
904
+ modelId: string;
905
+ };
906
+ calibrationFixtureRefs: string[];
907
+ calibrationStatus: "passed" | "failed" | "pending";
908
+ temperature: number;
909
+ } | {
910
+ id: string;
911
+ version: string;
912
+ label: string;
913
+ weight: number;
914
+ hardGate: boolean;
915
+ rewardEligible: boolean;
916
+ privileged: boolean;
917
+ metadata: Record<string, unknown>;
918
+ kind: "human";
919
+ rubric: string;
920
+ reviewerRole: string;
921
+ } | {
922
+ id: string;
923
+ version: string;
924
+ label: string;
925
+ weight: number;
926
+ hardGate: boolean;
927
+ rewardEligible: boolean;
928
+ privileged: boolean;
929
+ metadata: Record<string, unknown>;
930
+ kind: "custom_verifier";
931
+ module: string;
932
+ exportName: string;
933
+ timeoutMs: number;
934
+ networkPolicy: "none";
935
+ runtime?: "isolated_javascript" | "sandbox_process" | undefined;
936
+ })[];
937
+ graderFixtures: {
938
+ id: string;
939
+ taskId: string;
940
+ label: "infrastructure_failure" | "positive" | "negative" | "boundary" | "adversarial" | "prompt_injection";
941
+ output: Record<string, unknown>;
942
+ infrastructureError: string | null;
943
+ expectedPassed: boolean;
944
+ expectedRewardEligible: boolean;
945
+ metadata: Record<string, unknown>;
946
+ }[];
947
+ learningSignals: {
948
+ demonstrations: {
949
+ id: string;
950
+ taskId: string | null;
951
+ sourceRefs: string[];
952
+ artifactRef: string;
953
+ approved: boolean;
954
+ confidence: number;
955
+ metadata: Record<string, unknown>;
956
+ kind: "demonstration";
957
+ prompt: string | null;
958
+ response: string | null;
959
+ }[];
960
+ preferences: {
961
+ id: string;
962
+ taskId: string | null;
963
+ sourceRefs: string[];
964
+ artifactRef: string;
965
+ approved: boolean;
966
+ confidence: number;
967
+ metadata: Record<string, unknown>;
968
+ kind: "preference";
969
+ prompt: string;
970
+ chosen: string;
971
+ rejected: string;
972
+ rationale: string | null;
973
+ }[];
974
+ corrections: {
975
+ id: string;
976
+ taskId: string | null;
977
+ sourceRefs: string[];
978
+ artifactRef: string;
979
+ approved: boolean;
980
+ confidence: number;
981
+ metadata: Record<string, unknown>;
982
+ kind: "correction";
983
+ original: string;
984
+ corrected: string;
985
+ rationale: string | null;
986
+ }[];
987
+ feedback: {
988
+ id: string;
989
+ taskId: string | null;
990
+ sourceRefs: string[];
991
+ artifactRef: string;
992
+ approved: boolean;
993
+ confidence: number;
994
+ metadata: Record<string, unknown>;
995
+ kind: "feedback";
996
+ feedback: string;
997
+ polarity: "neutral" | "positive" | "negative" | "mixed";
998
+ }[];
999
+ rewards: {
1000
+ id: string;
1001
+ taskId: string | null;
1002
+ sourceRefs: string[];
1003
+ artifactRef: string;
1004
+ approved: boolean;
1005
+ confidence: number;
1006
+ metadata: Record<string, unknown>;
1007
+ kind: "reward";
1008
+ task: string;
1009
+ rules: {
1010
+ id: string;
1011
+ points: number;
1012
+ condition: string;
1013
+ }[];
1014
+ otherwisePoints: number;
1015
+ executable: boolean;
1016
+ }[];
1017
+ labels: {
1018
+ id: string;
1019
+ taskId: string | null;
1020
+ sourceRefs: string[];
1021
+ artifactRef: string;
1022
+ approved: boolean;
1023
+ confidence: number;
1024
+ metadata: Record<string, unknown>;
1025
+ kind: "label";
1026
+ labelKind: "rubric";
1027
+ task: string;
1028
+ criteria: {
1029
+ id: string;
1030
+ label: string;
1031
+ description: string;
1032
+ }[];
1033
+ calibrationExamples: {
1034
+ positive: string;
1035
+ negative: string;
1036
+ boundary: string;
1037
+ };
1038
+ }[];
1039
+ };
1040
+ publishedTasksetRef: {
1041
+ id: string;
1042
+ revision: number;
1043
+ contentHash: string;
1044
+ } | null;
1045
+ createdAt: string;
1046
+ updatedAt: string;
1047
+ metadata: Record<string, unknown>;
1048
+ datasetArtifact?: {
1049
+ schemaVersion: "openpond.datasetArtifact.v1";
1050
+ id: string;
1051
+ tasksetId: string;
1052
+ tasksetRevision: number;
1053
+ contentHash: string;
1054
+ format: "parquet";
1055
+ schema: {
1056
+ schemaVersion: "openpond.datasetSemanticSchema.v1";
1057
+ fields: {
1058
+ name: string;
1059
+ semanticRole: "metadata" | "reward" | "feedback" | "messages" | "row_id" | "cluster_id" | "prompt" | "demonstration" | "chosen" | "rejected" | "expected_output" | "privileged_context" | "tag";
1060
+ logicalType: "string" | "boolean" | "messages" | "integer" | "float" | "json";
1061
+ nullable: boolean;
1062
+ policy: "metadata" | "visible" | "privileged";
1063
+ }[];
1064
+ schemaHash: string;
1065
+ };
1066
+ shards: {
1067
+ id: string;
1068
+ split: "train" | "validation" | "test" | "frozen_eval";
1069
+ path: string;
1070
+ contentHash: string;
1071
+ schemaHash: string;
1072
+ sizeBytes: number;
1073
+ rowCount: number;
1074
+ rowGroupCount: number;
1075
+ }[];
1076
+ rowCount: number;
1077
+ splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
1078
+ sourceReceiptRefs: string[];
1079
+ mappingHash: string;
1080
+ qualityReportHash: string;
1081
+ createdAt: string;
1082
+ } | null | undefined;
1083
+ };
1084
+ files: {
1085
+ path: string;
1086
+ contentHash: string;
1087
+ sizeBytes: number;
1088
+ base64: string;
1089
+ }[];
1090
+ contentHash: string;
1091
+ };
1092
+ publication: {
1093
+ tasksetId: string;
1094
+ release: {
1095
+ id: string;
1096
+ revision: number;
1097
+ contentHash: string;
1098
+ };
1099
+ packageHash: string;
1100
+ } | null;
1101
+ originProjectId?: string | undefined;
1102
+ ownerScope?: "workspace" | "personal" | undefined;
1103
+ };
1104
+ base: {
1105
+ schemaVersion: "openpond.datasetWorkspaceReceipt.v1";
1106
+ teamId: string;
1107
+ datasetId: string;
1108
+ revision: number;
1109
+ workspace: {
1110
+ schemaVersion: "openpond.tasksetDraftWorkspace.v1";
1111
+ draft: {
1112
+ schemaVersion: "openpond.tasksetDraft.v1";
1113
+ id: string;
1114
+ revision: number;
1115
+ profileId: string;
1116
+ modelScope: {
1117
+ modelId: string;
1118
+ expectedModelRevision: number;
1119
+ source?: {
1120
+ schemaVersion: "openpond.modelTasksetDraftPreparation.v1";
1121
+ draftId: string;
1122
+ requestHash: string;
1123
+ sourcePackageHash: string;
1124
+ sourceTasksetRef: {
1125
+ id: string;
1126
+ contentHash: string;
1127
+ revision: number;
1128
+ };
1129
+ tasksetId: string;
1130
+ tasksetRevision: number;
1131
+ lineage: {
1132
+ schemaVersion: "openpond.modelTasksetAuthoring.v1";
1133
+ owner: {
1134
+ scopeId: string;
1135
+ modelId: string;
1136
+ };
1137
+ root: {
1138
+ id: string;
1139
+ contentHash: string;
1140
+ revision: number;
1141
+ };
1142
+ parent: {
1143
+ id: string;
1144
+ contentHash: string;
1145
+ revision: number;
1146
+ };
1147
+ };
1148
+ authoringGraph?: "bound" | undefined;
1149
+ } | undefined;
1150
+ } | null;
1151
+ name: string;
1152
+ objective: string;
1153
+ purpose: "general" | "benchmark";
1154
+ benchmark: {
1155
+ schemaVersion: "openpond.tasksetBenchmark.v1";
1156
+ definitionId: string;
1157
+ releaseId: string;
1158
+ releaseHash: string;
1159
+ managedReleasePath: string;
1160
+ adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
1161
+ evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
1162
+ primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
1163
+ qualityGate: "none" | "non_regression" | "all_pass";
1164
+ source: "builtin" | "imported";
1165
+ metadata: Record<string, unknown>;
1166
+ } | null;
1167
+ preferenceComparison: {
1168
+ schemaVersion: "openpond.tasksetPreferenceComparisonBinding.v1";
1169
+ releaseId: string;
1170
+ releaseHash: string;
1171
+ publishedAt: string;
1172
+ metadata: Record<string, unknown>;
1173
+ } | null;
1174
+ status: "draft" | "validating" | "needs_review" | "published";
1175
+ sourceRefs: ({
1176
+ id: string;
1177
+ profileId: string;
1178
+ title: string;
1179
+ sourceHash: string;
1180
+ occurredAt: string;
1181
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
1182
+ secretScanStatus: "passed" | "pending" | "blocked";
1183
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
1184
+ metadata: Record<string, unknown>;
1185
+ schemaVersion: "openpond.huggingFaceDatasetSource.v1";
1186
+ kind: "huggingface";
1187
+ repositoryId: string;
1188
+ repositoryUrl: string;
1189
+ revision: string;
1190
+ configuration: string;
1191
+ upstreamSplits: string[];
1192
+ gated: boolean;
1193
+ private: boolean;
1194
+ declaredLicense: string | null;
1195
+ sourceFileHashes: string[];
1196
+ } | {
1197
+ id: string;
1198
+ profileId: string;
1199
+ title: string;
1200
+ sourceHash: string;
1201
+ occurredAt: string;
1202
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
1203
+ secretScanStatus: "passed" | "pending" | "blocked";
1204
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
1205
+ metadata: Record<string, unknown>;
1206
+ schemaVersion: "openpond.uploadedFileDatasetSource.v1";
1207
+ kind: "uploaded_file";
1208
+ originalFileNames: string[];
1209
+ mediaTypes: string[];
1210
+ sourceFileHashes: string[];
1211
+ totalBytes: number;
1212
+ parserVersion: string;
1213
+ } | {
1214
+ id: string;
1215
+ profileId: string;
1216
+ title: string;
1217
+ sourceHash: string;
1218
+ occurredAt: string;
1219
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
1220
+ secretScanStatus: "passed" | "pending" | "blocked";
1221
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
1222
+ metadata: Record<string, unknown>;
1223
+ schemaVersion: "openpond.generatedDatasetSource.v1";
1224
+ kind: "generated";
1225
+ generatorId: string;
1226
+ generatorVersion: string;
1227
+ seed: number;
1228
+ generatorHash: string;
1229
+ } | {
1230
+ id: string;
1231
+ profileId: string;
1232
+ title: string;
1233
+ sourceHash: string;
1234
+ occurredAt: string;
1235
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
1236
+ secretScanStatus: "passed" | "pending" | "blocked";
1237
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
1238
+ metadata: Record<string, unknown>;
1239
+ schemaVersion: "openpond.learningBatchDatasetSource.v1";
1240
+ kind: "learning_batch";
1241
+ batch: {
1242
+ id: string;
1243
+ revision: number;
1244
+ contentHash: string;
1245
+ };
1246
+ taskDefinition: {
1247
+ id: string;
1248
+ revision: number;
1249
+ contentHash: string;
1250
+ };
1251
+ admittedBy: string;
1252
+ } | {
1253
+ schemaVersion: "openpond.trainingSource.v1";
1254
+ id: string;
1255
+ profileId: string;
1256
+ sessionId: string;
1257
+ turnIds: string[];
1258
+ workspaceId: string | null;
1259
+ sourceHash: string;
1260
+ clusterKey: string;
1261
+ title: string;
1262
+ occurredAt: string;
1263
+ consent: {
1264
+ status: "pending" | "granted" | "denied" | "revoked";
1265
+ scope: "metadata_only" | "selected_turns" | "full_session";
1266
+ grantedBy: string | null;
1267
+ grantedAt: string | null;
1268
+ purpose: "task_authoring_and_evaluation";
1269
+ };
1270
+ connectedAppIds: string[];
1271
+ secretScanStatus: "passed" | "pending" | "blocked";
1272
+ piiScanStatus: "passed" | "pending" | "blocked" | "review";
1273
+ licensingStatus: "pending" | "blocked" | "approved" | "review";
1274
+ metadata: Record<string, unknown>;
1275
+ })[];
1276
+ policy: {
1277
+ policyVisibleFields: string[];
1278
+ privilegedFields: string[];
1279
+ hiddenGraderRefs: string[];
1280
+ connectedAppScopes: string[];
1281
+ };
1282
+ environment: {
1283
+ protocolVersion: "openpond.taskEnvironment.v1";
1284
+ kind: "chat" | "agent" | "program" | "stateful_harness" | "work";
1285
+ entrypoint: string;
1286
+ stateful: boolean;
1287
+ deterministicSeeds: boolean;
1288
+ toolNames: string[];
1289
+ lifecycle: ("grade" | "create" | "reset" | "step" | "cleanup")[];
1290
+ defaultTimeoutMs: number;
1291
+ networkPolicy: "none" | "declared_read_only" | "declared_scoped";
1292
+ metadata: Record<string, unknown>;
1293
+ actionBindings?: {
1294
+ schemaVersion: "openpond.harnessActionBinding.v1";
1295
+ actionId: string;
1296
+ modelToolName: string;
1297
+ description: string;
1298
+ inputSchema: Record<string, unknown>;
1299
+ actionSchemaHash: string;
1300
+ agentRelease: {
1301
+ id: string;
1302
+ contentHash: string;
1303
+ };
1304
+ implementationHash: string;
1305
+ runtimeBindingId: string;
1306
+ capabilityReceiptHash: string;
1307
+ sideEffect: "read" | "write";
1308
+ studentVisible: boolean;
1309
+ timeoutMs: number;
1310
+ episodeArgumentBindings: {
1311
+ argument: string;
1312
+ source: "case_id";
1313
+ }[];
1314
+ }[] | undefined;
1315
+ resources?: {
1316
+ id: string;
1317
+ kind: "file" | "configuration" | "catalog" | "code_module";
1318
+ path: string;
1319
+ visibility: "privileged" | "policy_visible" | "policy_hidden";
1320
+ required: boolean;
1321
+ metadata: Record<string, unknown>;
1322
+ mediaType?: string | null | undefined;
1323
+ }[] | undefined;
1324
+ };
1325
+ output: {
1326
+ mode: "artifacts" | "text" | "structured_json";
1327
+ jsonSchema: Record<string, unknown> | null;
1328
+ renderer: {
1329
+ module: string;
1330
+ exportName: string;
1331
+ configRef: string | null;
1332
+ } | null;
1333
+ };
1334
+ capabilities: {
1335
+ schemaVersion: "openpond.tasksetCapabilities.v1";
1336
+ taskKind: "chat" | "single_agent" | "multi_agent" | "custom_program";
1337
+ supportedSignals: ("label" | "reward" | "preference" | "feedback" | "demonstration" | "correction")[];
1338
+ compatibleMethods: ("none" | "sft" | "dpo" | "grpo" | "ppo" | "sdft" | "opd" | "opsd" | "sdpo" | "retrieval")[];
1339
+ rewardKinds: ("none" | "exact" | "deterministic" | "model_judge" | "human")[];
1340
+ requiresTools: boolean;
1341
+ requiresState: boolean;
1342
+ requiresPrivilegedGrading: boolean;
1343
+ environmentPlacements: ("local" | "remote" | "colocated" | "provider_native")[];
1344
+ exportable: boolean;
1345
+ portabilityBlockers: string[];
1346
+ };
1347
+ metrics: {
1348
+ schemaVersion: "openpond.tasksetMetricPolicy.v1";
1349
+ primaryMetric: string;
1350
+ aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
1351
+ missingReward: "zero" | "exclude";
1352
+ customAggregator: {
1353
+ module: string;
1354
+ exportName: string;
1355
+ contentHash: string;
1356
+ timeoutMs: number;
1357
+ networkPolicy: "none";
1358
+ } | null;
1359
+ taskWeights?: Record<string, number> | undefined;
1360
+ };
1361
+ review: {
1362
+ enabled: boolean;
1363
+ candidateCount: number;
1364
+ minimumSamples: number;
1365
+ allowTies: boolean;
1366
+ allowRejectAll: boolean;
1367
+ rubric: string;
1368
+ criteria: {
1369
+ id: string;
1370
+ label: string;
1371
+ description: string;
1372
+ weight: number;
1373
+ }[];
1374
+ };
1375
+ tasks: {
1376
+ schemaVersion: "openpond.taskData.v1";
1377
+ id: string;
1378
+ clusterKey: string;
1379
+ split: "train" | "validation" | "test" | "frozen_eval";
1380
+ input: Record<string, unknown>;
1381
+ expectedOutput: Record<string, unknown> | null;
1382
+ policyVisibleContext: Record<string, unknown>;
1383
+ privilegedContextRef: string | null;
1384
+ tags: string[];
1385
+ metadata: Record<string, unknown>;
1386
+ sourceRefs: string[];
1387
+ assets?: {
1388
+ id: string;
1389
+ sourceRefId: string;
1390
+ artifactRef: string;
1391
+ fileName: string;
1392
+ mediaType: string;
1393
+ sha256: string;
1394
+ sizeBytes: number;
1395
+ split: "train" | "validation" | "test" | "frozen_eval";
1396
+ metadata: Record<string, unknown>;
1397
+ }[] | undefined;
1398
+ resourceRefs?: string[] | undefined;
1399
+ requiredOutputs?: {
1400
+ path: string;
1401
+ mediaType: string;
1402
+ metadata: Record<string, unknown>;
1403
+ schemaRef?: string | null | undefined;
1404
+ maxBytes?: number | undefined;
1405
+ }[] | undefined;
1406
+ }[];
1407
+ graders: ({
1408
+ id: string;
1409
+ version: string;
1410
+ label: string;
1411
+ weight: number;
1412
+ hardGate: boolean;
1413
+ rewardEligible: boolean;
1414
+ privileged: boolean;
1415
+ metadata: Record<string, unknown>;
1416
+ kind: "file" | "content" | "state" | "test" | "schema" | "diff" | "runtime_event";
1417
+ config: Record<string, unknown>;
1418
+ } | {
1419
+ id: string;
1420
+ version: string;
1421
+ label: string;
1422
+ weight: number;
1423
+ hardGate: boolean;
1424
+ rewardEligible: boolean;
1425
+ privileged: boolean;
1426
+ metadata: Record<string, unknown>;
1427
+ kind: "model_judge";
1428
+ rubric: string;
1429
+ judge: {
1430
+ providerId: "openpond" | "codex" | "anthropic" | "openai" | "xai" | "google" | "openrouter" | "deepseek" | "zai" | "moonshot" | "together" | "groq" | "custom-openai-compatible";
1431
+ modelId: string;
1432
+ };
1433
+ calibrationFixtureRefs: string[];
1434
+ calibrationStatus: "passed" | "failed" | "pending";
1435
+ temperature: number;
1436
+ } | {
1437
+ id: string;
1438
+ version: string;
1439
+ label: string;
1440
+ weight: number;
1441
+ hardGate: boolean;
1442
+ rewardEligible: boolean;
1443
+ privileged: boolean;
1444
+ metadata: Record<string, unknown>;
1445
+ kind: "human";
1446
+ rubric: string;
1447
+ reviewerRole: string;
1448
+ } | {
1449
+ id: string;
1450
+ version: string;
1451
+ label: string;
1452
+ weight: number;
1453
+ hardGate: boolean;
1454
+ rewardEligible: boolean;
1455
+ privileged: boolean;
1456
+ metadata: Record<string, unknown>;
1457
+ kind: "custom_verifier";
1458
+ module: string;
1459
+ exportName: string;
1460
+ timeoutMs: number;
1461
+ networkPolicy: "none";
1462
+ runtime?: "isolated_javascript" | "sandbox_process" | undefined;
1463
+ })[];
1464
+ graderFixtures: {
1465
+ id: string;
1466
+ taskId: string;
1467
+ label: "infrastructure_failure" | "positive" | "negative" | "boundary" | "adversarial" | "prompt_injection";
1468
+ output: Record<string, unknown>;
1469
+ infrastructureError: string | null;
1470
+ expectedPassed: boolean;
1471
+ expectedRewardEligible: boolean;
1472
+ metadata: Record<string, unknown>;
1473
+ }[];
1474
+ learningSignals: {
1475
+ demonstrations: {
1476
+ id: string;
1477
+ taskId: string | null;
1478
+ sourceRefs: string[];
1479
+ artifactRef: string;
1480
+ approved: boolean;
1481
+ confidence: number;
1482
+ metadata: Record<string, unknown>;
1483
+ kind: "demonstration";
1484
+ prompt: string | null;
1485
+ response: string | null;
1486
+ }[];
1487
+ preferences: {
1488
+ id: string;
1489
+ taskId: string | null;
1490
+ sourceRefs: string[];
1491
+ artifactRef: string;
1492
+ approved: boolean;
1493
+ confidence: number;
1494
+ metadata: Record<string, unknown>;
1495
+ kind: "preference";
1496
+ prompt: string;
1497
+ chosen: string;
1498
+ rejected: string;
1499
+ rationale: string | null;
1500
+ }[];
1501
+ corrections: {
1502
+ id: string;
1503
+ taskId: string | null;
1504
+ sourceRefs: string[];
1505
+ artifactRef: string;
1506
+ approved: boolean;
1507
+ confidence: number;
1508
+ metadata: Record<string, unknown>;
1509
+ kind: "correction";
1510
+ original: string;
1511
+ corrected: string;
1512
+ rationale: string | null;
1513
+ }[];
1514
+ feedback: {
1515
+ id: string;
1516
+ taskId: string | null;
1517
+ sourceRefs: string[];
1518
+ artifactRef: string;
1519
+ approved: boolean;
1520
+ confidence: number;
1521
+ metadata: Record<string, unknown>;
1522
+ kind: "feedback";
1523
+ feedback: string;
1524
+ polarity: "neutral" | "positive" | "negative" | "mixed";
1525
+ }[];
1526
+ rewards: {
1527
+ id: string;
1528
+ taskId: string | null;
1529
+ sourceRefs: string[];
1530
+ artifactRef: string;
1531
+ approved: boolean;
1532
+ confidence: number;
1533
+ metadata: Record<string, unknown>;
1534
+ kind: "reward";
1535
+ task: string;
1536
+ rules: {
1537
+ id: string;
1538
+ points: number;
1539
+ condition: string;
1540
+ }[];
1541
+ otherwisePoints: number;
1542
+ executable: boolean;
1543
+ }[];
1544
+ labels: {
1545
+ id: string;
1546
+ taskId: string | null;
1547
+ sourceRefs: string[];
1548
+ artifactRef: string;
1549
+ approved: boolean;
1550
+ confidence: number;
1551
+ metadata: Record<string, unknown>;
1552
+ kind: "label";
1553
+ labelKind: "rubric";
1554
+ task: string;
1555
+ criteria: {
1556
+ id: string;
1557
+ label: string;
1558
+ description: string;
1559
+ }[];
1560
+ calibrationExamples: {
1561
+ positive: string;
1562
+ negative: string;
1563
+ boundary: string;
1564
+ };
1565
+ }[];
1566
+ };
1567
+ publishedTasksetRef: {
1568
+ id: string;
1569
+ revision: number;
1570
+ contentHash: string;
1571
+ } | null;
1572
+ createdAt: string;
1573
+ updatedAt: string;
1574
+ metadata: Record<string, unknown>;
1575
+ datasetArtifact?: {
1576
+ schemaVersion: "openpond.datasetArtifact.v1";
1577
+ id: string;
1578
+ tasksetId: string;
1579
+ tasksetRevision: number;
1580
+ contentHash: string;
1581
+ format: "parquet";
1582
+ schema: {
1583
+ schemaVersion: "openpond.datasetSemanticSchema.v1";
1584
+ fields: {
1585
+ name: string;
1586
+ semanticRole: "metadata" | "reward" | "feedback" | "messages" | "row_id" | "cluster_id" | "prompt" | "demonstration" | "chosen" | "rejected" | "expected_output" | "privileged_context" | "tag";
1587
+ logicalType: "string" | "boolean" | "messages" | "integer" | "float" | "json";
1588
+ nullable: boolean;
1589
+ policy: "metadata" | "visible" | "privileged";
1590
+ }[];
1591
+ schemaHash: string;
1592
+ };
1593
+ shards: {
1594
+ id: string;
1595
+ split: "train" | "validation" | "test" | "frozen_eval";
1596
+ path: string;
1597
+ contentHash: string;
1598
+ schemaHash: string;
1599
+ sizeBytes: number;
1600
+ rowCount: number;
1601
+ rowGroupCount: number;
1602
+ }[];
1603
+ rowCount: number;
1604
+ splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
1605
+ sourceReceiptRefs: string[];
1606
+ mappingHash: string;
1607
+ qualityReportHash: string;
1608
+ createdAt: string;
1609
+ } | null | undefined;
1610
+ };
1611
+ files: {
1612
+ path: string;
1613
+ contentHash: string;
1614
+ sizeBytes: number;
1615
+ base64: string;
1616
+ }[];
1617
+ contentHash: string;
1618
+ };
1619
+ publication: {
1620
+ tasksetId: string;
1621
+ release: {
1622
+ id: string;
1623
+ revision: number;
1624
+ contentHash: string;
1625
+ };
1626
+ packageHash: string;
1627
+ } | null;
1628
+ originProjectId?: string | undefined;
1629
+ ownerScope?: "workspace" | "personal" | undefined;
1630
+ } | null;
1631
+ } | null>;
567
1632
  beginVersion(id: string, input: z.input<typeof DatasetWorkspaceBeginVersionSchema>, signal?: AbortSignal): Promise<{
568
1633
  schemaVersion: "openpond.datasetWorkspaceReceipt.v1";
569
1634
  teamId: string;
@@ -620,8 +1685,8 @@ export declare class OpenPondDatasetWorkspaceClient {
620
1685
  releaseId: string;
621
1686
  releaseHash: string;
622
1687
  managedReleasePath: string;
623
- adaptationSplit: "train" | "validation" | "frozen_eval" | "test";
624
- evaluationSplit: "train" | "validation" | "frozen_eval" | "test";
1688
+ adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
1689
+ evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
625
1690
  primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
626
1691
  qualityGate: "none" | "non_regression" | "all_pass";
627
1692
  source: "builtin" | "imported";
@@ -839,7 +1904,7 @@ export declare class OpenPondDatasetWorkspaceClient {
839
1904
  schemaVersion: "openpond.taskData.v1";
840
1905
  id: string;
841
1906
  clusterKey: string;
842
- split: "train" | "validation" | "frozen_eval" | "test";
1907
+ split: "train" | "validation" | "test" | "frozen_eval";
843
1908
  input: Record<string, unknown>;
844
1909
  expectedOutput: Record<string, unknown> | null;
845
1910
  policyVisibleContext: Record<string, unknown>;
@@ -855,7 +1920,7 @@ export declare class OpenPondDatasetWorkspaceClient {
855
1920
  mediaType: string;
856
1921
  sha256: string;
857
1922
  sizeBytes: number;
858
- split: "train" | "validation" | "frozen_eval" | "test";
1923
+ split: "train" | "validation" | "test" | "frozen_eval";
859
1924
  metadata: Record<string, unknown>;
860
1925
  }[] | undefined;
861
1926
  resourceRefs?: string[] | undefined;
@@ -1055,7 +2120,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1055
2120
  };
1056
2121
  shards: {
1057
2122
  id: string;
1058
- split: "train" | "validation" | "frozen_eval" | "test";
2123
+ split: "train" | "validation" | "test" | "frozen_eval";
1059
2124
  path: string;
1060
2125
  contentHash: string;
1061
2126
  schemaHash: string;
@@ -1064,7 +2129,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1064
2129
  rowGroupCount: number;
1065
2130
  }[];
1066
2131
  rowCount: number;
1067
- splitCounts: Record<"train" | "validation" | "frozen_eval" | "test", number>;
2132
+ splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
1068
2133
  sourceReceiptRefs: string[];
1069
2134
  mappingHash: string;
1070
2135
  qualityReportHash: string;
@@ -1168,8 +2233,8 @@ export declare class OpenPondDatasetWorkspaceClient {
1168
2233
  releaseId: string;
1169
2234
  releaseHash: string;
1170
2235
  managedReleasePath: string;
1171
- adaptationSplit: "train" | "validation" | "frozen_eval" | "test";
1172
- evaluationSplit: "train" | "validation" | "frozen_eval" | "test";
2236
+ adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
2237
+ evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
1173
2238
  primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
1174
2239
  qualityGate: "none" | "non_regression" | "all_pass";
1175
2240
  source: "builtin" | "imported";
@@ -1387,7 +2452,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1387
2452
  schemaVersion: "openpond.taskData.v1";
1388
2453
  id: string;
1389
2454
  clusterKey: string;
1390
- split: "train" | "validation" | "frozen_eval" | "test";
2455
+ split: "train" | "validation" | "test" | "frozen_eval";
1391
2456
  input: Record<string, unknown>;
1392
2457
  expectedOutput: Record<string, unknown> | null;
1393
2458
  policyVisibleContext: Record<string, unknown>;
@@ -1403,7 +2468,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1403
2468
  mediaType: string;
1404
2469
  sha256: string;
1405
2470
  sizeBytes: number;
1406
- split: "train" | "validation" | "frozen_eval" | "test";
2471
+ split: "train" | "validation" | "test" | "frozen_eval";
1407
2472
  metadata: Record<string, unknown>;
1408
2473
  }[] | undefined;
1409
2474
  resourceRefs?: string[] | undefined;
@@ -1603,7 +2668,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1603
2668
  };
1604
2669
  shards: {
1605
2670
  id: string;
1606
- split: "train" | "validation" | "frozen_eval" | "test";
2671
+ split: "train" | "validation" | "test" | "frozen_eval";
1607
2672
  path: string;
1608
2673
  contentHash: string;
1609
2674
  schemaHash: string;
@@ -1612,7 +2677,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1612
2677
  rowGroupCount: number;
1613
2678
  }[];
1614
2679
  rowCount: number;
1615
- splitCounts: Record<"train" | "validation" | "frozen_eval" | "test", number>;
2680
+ splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
1616
2681
  sourceReceiptRefs: string[];
1617
2682
  mappingHash: string;
1618
2683
  qualityReportHash: string;
@@ -1695,8 +2760,8 @@ export declare class OpenPondDatasetWorkspaceClient {
1695
2760
  releaseId: string;
1696
2761
  releaseHash: string;
1697
2762
  managedReleasePath: string;
1698
- adaptationSplit: "train" | "validation" | "frozen_eval" | "test";
1699
- evaluationSplit: "train" | "validation" | "frozen_eval" | "test";
2763
+ adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
2764
+ evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
1700
2765
  primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
1701
2766
  qualityGate: "none" | "non_regression" | "all_pass";
1702
2767
  source: "builtin" | "imported";
@@ -1914,7 +2979,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1914
2979
  schemaVersion: "openpond.taskData.v1";
1915
2980
  id: string;
1916
2981
  clusterKey: string;
1917
- split: "train" | "validation" | "frozen_eval" | "test";
2982
+ split: "train" | "validation" | "test" | "frozen_eval";
1918
2983
  input: Record<string, unknown>;
1919
2984
  expectedOutput: Record<string, unknown> | null;
1920
2985
  policyVisibleContext: Record<string, unknown>;
@@ -1930,7 +2995,7 @@ export declare class OpenPondDatasetWorkspaceClient {
1930
2995
  mediaType: string;
1931
2996
  sha256: string;
1932
2997
  sizeBytes: number;
1933
- split: "train" | "validation" | "frozen_eval" | "test";
2998
+ split: "train" | "validation" | "test" | "frozen_eval";
1934
2999
  metadata: Record<string, unknown>;
1935
3000
  }[] | undefined;
1936
3001
  resourceRefs?: string[] | undefined;
@@ -2130,7 +3195,7 @@ export declare class OpenPondDatasetWorkspaceClient {
2130
3195
  };
2131
3196
  shards: {
2132
3197
  id: string;
2133
- split: "train" | "validation" | "frozen_eval" | "test";
3198
+ split: "train" | "validation" | "test" | "frozen_eval";
2134
3199
  path: string;
2135
3200
  contentHash: string;
2136
3201
  schemaHash: string;
@@ -2139,7 +3204,7 @@ export declare class OpenPondDatasetWorkspaceClient {
2139
3204
  rowGroupCount: number;
2140
3205
  }[];
2141
3206
  rowCount: number;
2142
- splitCounts: Record<"train" | "validation" | "frozen_eval" | "test", number>;
3207
+ splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
2143
3208
  sourceReceiptRefs: string[];
2144
3209
  mappingHash: string;
2145
3210
  qualityReportHash: string;
@@ -2234,8 +3299,8 @@ export declare class OpenPondDatasetWorkspaceClient {
2234
3299
  releaseId: string;
2235
3300
  releaseHash: string;
2236
3301
  managedReleasePath: string;
2237
- adaptationSplit: "train" | "validation" | "frozen_eval" | "test";
2238
- evaluationSplit: "train" | "validation" | "frozen_eval" | "test";
3302
+ adaptationSplit: "train" | "validation" | "test" | "frozen_eval";
3303
+ evaluationSplit: "train" | "validation" | "test" | "frozen_eval";
2239
3304
  primaryMetric: "foreground_tokens" | "success_rate" | "latency_ms" | "cost_usd";
2240
3305
  qualityGate: "none" | "non_regression" | "all_pass";
2241
3306
  source: "builtin" | "imported";
@@ -2453,7 +3518,7 @@ export declare class OpenPondDatasetWorkspaceClient {
2453
3518
  schemaVersion: "openpond.taskData.v1";
2454
3519
  id: string;
2455
3520
  clusterKey: string;
2456
- split: "train" | "validation" | "frozen_eval" | "test";
3521
+ split: "train" | "validation" | "test" | "frozen_eval";
2457
3522
  input: Record<string, unknown>;
2458
3523
  expectedOutput: Record<string, unknown> | null;
2459
3524
  policyVisibleContext: Record<string, unknown>;
@@ -2469,7 +3534,7 @@ export declare class OpenPondDatasetWorkspaceClient {
2469
3534
  mediaType: string;
2470
3535
  sha256: string;
2471
3536
  sizeBytes: number;
2472
- split: "train" | "validation" | "frozen_eval" | "test";
3537
+ split: "train" | "validation" | "test" | "frozen_eval";
2473
3538
  metadata: Record<string, unknown>;
2474
3539
  }[] | undefined;
2475
3540
  resourceRefs?: string[] | undefined;
@@ -2669,7 +3734,7 @@ export declare class OpenPondDatasetWorkspaceClient {
2669
3734
  };
2670
3735
  shards: {
2671
3736
  id: string;
2672
- split: "train" | "validation" | "frozen_eval" | "test";
3737
+ split: "train" | "validation" | "test" | "frozen_eval";
2673
3738
  path: string;
2674
3739
  contentHash: string;
2675
3740
  schemaHash: string;
@@ -2678,7 +3743,7 @@ export declare class OpenPondDatasetWorkspaceClient {
2678
3743
  rowGroupCount: number;
2679
3744
  }[];
2680
3745
  rowCount: number;
2681
- splitCounts: Record<"train" | "validation" | "frozen_eval" | "test", number>;
3746
+ splitCounts: Record<"train" | "validation" | "test" | "frozen_eval", number>;
2682
3747
  sourceReceiptRefs: string[];
2683
3748
  mappingHash: string;
2684
3749
  qualityReportHash: string;