cayu.evals
All 819 declared imports from cayu.evals. Optional imports are included even when wildcard imports omit them. Select a name for its signature, fields, methods and release source.
from cayu.evals import ASSERTION_EVIDENCE_MAX_BYTESASSERTION_EVIDENCE_MAX_BYTESASSERTION_EVIDENCE_SCHEMA_VERSIONArtifactAssertionSpecArtifactCreatedArtifactScopeEvidenceV1ArtifactStructuralEvidenceV1AssertionCostEvidenceV1AssertionEvidenceViewAssertionSpecBROWSER_ACCEPTANCE_FIXTURE_REVISIONBROWSER_ACCEPTANCE_HTML_MAX_BYTESBROWSER_ACCEPTANCE_MANIFEST_MAX_BYTESBROWSER_ACCEPTANCE_REPORT_MAX_BYTESBROWSER_ACCEPTANCE_SCHEMA_VERSIONBenchmarkFileV1BenchmarkPackageV1BenchmarkRequirementsV1BrowserAcceptanceAccessStateBrowserAcceptanceAgentReportStateBrowserAcceptanceAggregateV1BrowserAcceptanceArtifactEvidenceV1BrowserAcceptanceAuthenticatedConfigV1BrowserAcceptanceAuthenticationCollectorBrowserAcceptanceAuthenticationPhaseV1BrowserAcceptanceCaseAggregateV1BrowserAcceptanceCaseCategoryBrowserAcceptanceCaseV1BrowserAcceptanceCompletionStateBrowserAcceptanceConformanceBrowserAcceptanceDiagnosticStateBrowserAcceptanceDiagnosticV1BrowserAcceptanceFaultEvidenceV1BrowserAcceptanceFaultScenarioBrowserAcceptanceFixtureV1BrowserAcceptanceInfrastructureStateBrowserAcceptanceLimitsV1BrowserAcceptanceManifestV1BrowserAcceptanceModeBrowserAcceptanceOperationEvidenceV1BrowserAcceptanceOperationStateBrowserAcceptanceOperatorEvidenceV1BrowserAcceptancePlanV1BrowserAcceptanceProfileEvidenceV1BrowserAcceptanceReportV1BrowserAcceptanceRequestSummaryV1BrowserAcceptanceRuntimeIdentityV1BrowserAcceptanceScenarioExecutionV1BrowserAcceptanceSemanticOracleBrowserAcceptanceSemanticStateBrowserAcceptanceStateBrowserAcceptanceTrialReceiptV1BrowserAcceptanceUsageV1BrowserAcceptanceVariabilityStateBrowserAllocationDispositionCAPTURED_EVALUATION_CANDIDATE_MAX_BYTESCAPTURED_EVALUATION_CANDIDATE_SCHEMA_VERSIONCAPTURED_EVALUATION_RESULT_MAX_BYTESCAPTURED_EVALUATION_RESULT_SCHEMA_VERSIONCAPTURED_RUN_SCORE_MAX_BYTESCAPTURED_RUN_SCORE_SCHEMA_VERSIONCAUSAL_MEMORY_CAMPAIGN_EXPERIMENT_IDCAUSAL_MEMORY_CAMPAIGN_REPETITIONSCAUSAL_MEMORY_CAMPAIGN_SUITE_IDCAUSAL_MEMORY_CAMPAIGN_TARGET_KEYCAUSAL_MEMORY_CAMPAIGN_VARIANTSCORPUS_EXECUTION_COMPARISON_MAX_BYTESCORPUS_EXECUTION_COMPARISON_MAX_HTML_BYTESCORPUS_EXECUTION_COMPARISON_MAX_JSON_BYTESCORPUS_EXECUTION_DEFAULT_MAX_CONCURRENCYCORPUS_EXECUTION_MAX_APP_MANIFEST_BYTESCORPUS_EXECUTION_MAX_BOOTSTRAP_MESSAGESCORPUS_EXECUTION_MAX_COMPILED_INPUT_CHARSCORPUS_EXECUTION_MAX_CONCURRENCYCORPUS_EXECUTION_MAX_MODEL_JUDGESCORPUS_EXECUTION_MAX_REQUEST_BASE_BYTESCORPUS_EXECUTION_MAX_TOTAL_INPUT_CHARSCORPUS_EXECUTION_RESULT_MAX_BYTESCORPUS_EXECUTION_RESULT_MAX_HTML_BYTESCORPUS_EXECUTION_RESULT_MAX_JSON_BYTESCORPUS_EXECUTION_RESULT_SCHEMA_VERSIONCapturedEvaluationCandidateV1CapturedEvaluationResultV1CapturedEvaluationSourceV1CapturedEvaluationWarningCodeCapturedRunScoreV1ChildSessionCompletedChildStatusAssertionSpecCompiledCorpusSuiteCompiledEvalScenarioV2CorpusCaseComparisonCorpusComparisonCompatibilityCorpusComparisonReasonCorpusComparisonResultSummaryCorpusExecutionComparisonCorpusExecutionLimitsCorpusExecutionRegressionCorpusExecutionResultCorpusRegressionKindCorpusRegressionScopeCorpusReliabilityDistributionV1CorpusTargetCorpusUserMessageSpecDEFAULT_EVAL_MAX_ACTIVE_TRIALSDETERMINISTIC_BROWSER_ACCEPTANCE_SUITE_IDEVAL_CORPUS_MAX_ASSERTIONS_PER_CASEEVAL_CORPUS_MAX_BYTESEVAL_CORPUS_MAX_CASESEVAL_CORPUS_MAX_MERGE_INPUTSEVAL_CORPUS_MAX_MESSAGES_PER_CASEEVAL_CORPUS_MAX_MESSAGE_CHARSEVAL_CORPUS_MAX_PUBLISHED_ASSERTION_RESULTSEVAL_CORPUS_MAX_SUITESEVAL_CORPUS_MAX_TIMEOUT_SECONDSEVAL_CORPUS_MAX_TOTAL_MESSAGE_CHARSEVAL_CORPUS_MAX_TRIALSEVAL_CORPUS_SCHEMA_VERSIONEVAL_JUDGE_CALIBRATION_MAX_BYTESEVAL_JUDGE_CALIBRATION_MAX_TRIALSEVAL_JUDGE_CALIBRATION_SCHEMA_VERSIONEVAL_MEMORY_ATTRIBUTION_MAX_BYTESEVAL_MEMORY_ATTRIBUTION_PROJECTION_BUDGET_BYTESEVAL_MEMORY_ATTRIBUTION_RESULT_BUDGET_BYTESEVAL_MEMORY_ATTRIBUTION_SCHEMA_VERSIONEVAL_MEMORY_ATTRIBUTION_SOURCE_BUDGET_BYTESEVAL_RESULT_PRESENTATION_MAX_BYTESEVAL_RESULT_PRESENTATION_SCHEMA_VERSIONEVAL_RESULT_PROJECTION_MAX_BYTESEVAL_RESULT_PROJECTION_SCHEMA_VERSIONEVAL_RESULT_REPORT_MAX_BYTESEVAL_RESULT_REPORT_SCHEMA_VERSIONEVAL_RUN_INVOCATION_MAX_BYTESEVAL_RUN_MAX_OBSERVATION_INTERVAL_SECONDSEVAL_RUN_MAX_TERMINAL_WAIT_SECONDSEVAL_RUN_MIN_OBSERVATION_INTERVAL_SECONDSEVAL_SCENARIO_MAX_ARTIFACT_REQUIREMENTSEVAL_SCENARIO_MAX_BYTESEVAL_SCENARIO_MAX_EVENTSEVAL_SCENARIO_MAX_JSON_PART_BYTESEVAL_SCENARIO_MAX_MESSAGES_PER_EVENTEVAL_SCENARIO_MAX_PARTS_PER_MESSAGEEVAL_SCENARIO_MAX_SECRET_REQUIREMENTSEVAL_SCENARIO_MAX_TEXT_CHARSEVAL_SCENARIO_MAX_TOTAL_ARTIFACT_BYTESEVAL_SCENARIO_MAX_TOTAL_TEXT_CHARSEVAL_SCENARIO_SCHEMA_VERSIONEVAL_SCHEMA_VERSIONEVAL_STORE_DEFAULT_PAGE_BYTESEVAL_STORE_DEFAULT_PAGE_SIZEEVAL_STORE_MAX_CURSOR_BYTESEVAL_STORE_MAX_LEASE_SECONDSEVAL_STORE_MAX_PAGE_BYTESEVAL_STORE_MAX_PAGE_SIZEEVAL_SUITE_AUTHORING_MAX_BYTESEVAL_SUITE_AUTHORING_SCHEMA_VERSIONEVAL_SUITE_AUTHORING_V2_SCHEMA_VERSIONEVAL_SUITE_AUTHORING_V3_SCHEMA_VERSIONEVAL_SUITE_RUN_EXPOSURE_SCHEMA_VERSIONEVAL_SUITE_SELECTION_SCHEMA_VERSIONEVAL_SUITE_TRIAL_POLICY_SCHEMA_VERSIONEVAL_TRIAL_OUTPUT_MAX_PREVIEW_BYTESEVAL_TRIAL_OUTPUT_MAX_RETAINED_BYTESEVAL_TRIAL_OUTPUT_MAX_RETAINED_CHARSEXTERNAL_BODY_MAX_BYTESEXTERNAL_BODY_MAX_FILESEXTERNAL_CONTAINER_MAX_INPUT_BYTESEXTERNAL_CONTAINER_MAX_OUTPUT_BYTESEXTERNAL_CONTAINER_RESET_CONTRACT_REVISIONEXTERNAL_CONTAINER_RUNNER_REVISIONEXTERNAL_CONTAINER_STREAM_PROTOCOLEXTERNAL_PRIVATE_MEMORY_ABLATION_COMPLETION_FILENAMEEXTERNAL_PRIVATE_MEMORY_ABLATION_MAX_ARTIFACT_BYTESEXTERNAL_PRIVATE_MEMORY_ABLATION_MAX_REPORT_EVIDENCE_BYTES_PER_TRIALEXTERNAL_PRIVATE_MEMORY_ABLATION_MAX_SUPPLEMENTAL_BYTES_PER_TRIALEXTERNAL_PRIVATE_MEMORY_ABLATION_METHODOLOGY_FILENAMEEXTERNAL_PRIVATE_MEMORY_ABLATION_MIN_REPORT_EVIDENCE_BYTES_PER_TRIALEXTERNAL_PRIVATE_MEMORY_ABLATION_REPORT_FILENAMEEXTERNAL_PRIVATE_MEMORY_ABLATION_SCHEMA_VERSIONEXTERNAL_PRIVATE_MEMORY_ABLATION_TRIAL_BUDGET_KEYEXTERNAL_PROCESS_PROTOCOL_VERSIONEXTERNAL_TRIAL_ENVELOPE_PREFIXEvalAssertionEvalAssertionPresentationV1EvalAssertionResultEvalAuthoredSuiteCatalogEntryEvalAuthoredSuiteCatalogPageEvalAuthoredSuiteCatalogQueryEvalAuthoredSuiteConflictEvalAuthoredSuiteReferenceErrorEvalBaselineConflictEvalBaselineKeyEvalBaselineMutationRecordEvalBaselineRecordEvalBaselineUpdateEvalCandidateCostBudgetV1EvalCandidateLaunchExposureEvalCaseEvalCaseCatalogEntryEvalCaseCatalogPageEvalCaseCatalogQueryEvalCaseComparisonEvalCaseContractV1EvalCaseDefinitionV1EvalCaseDefinitionV2EvalCaseDraftV1EvalCaseDraftV2EvalCasePresentationV1EvalCasePresentationV2EvalCaseReliabilityV1EvalCaseResultEvalCaseSpecEvalCaseStimulusV1EvalCatalogQueryEvalContextEvalCorpusCatalogEntryEvalCorpusCatalogPageEvalCorpusConflictEvalCorpusDocumentEvalCorpusInspectionV1EvalCorpusSuiteInspectionV1EvalDiagnosticV1EvalExecutionCandidateIdentityV1EvalExecutionCapacityEvalExecutionProfileBindingV1EvalExecutionProfileExposureV1EvalExecutionProfilePolicyV1EvalExecutionProfileV1EvalExecutionResourceCeilingsV1EvalExecutionTargetMaterialIdentityV1EvalJudgeCalibrationConflictEvalJudgeCalibrationCriterionLabelV1EvalJudgeCalibrationDefinitionV1EvalJudgeCalibrationDraftV1EvalJudgeCalibrationEvidenceProvenanceV1EvalJudgeCalibrationEvidenceV1EvalJudgeCalibrationHumanLabelV1EvalJudgeCalibrationReportV1EvalJudgeCalibrationTrialV1EvalJudgeEvidenceSelectionV1EvalJudgeProfileExposureV1EvalMaximumCostExposureV1EvalMaximumCostTotalV1EvalMaximumCostUnavailableReasonEvalMemoryAttributionCapturePolicyV1EvalMemoryAttributionEvidenceV1EvalMemoryAttributionSourceV1EvalMemoryEvidenceCompletenessEvalMemoryEvidenceLimitationEvalMemorySourceAliasV1EvalMemorySourceReferenceV1EvalOutcomeEvalPlanEvalProcessCaseInspectionV1EvalProcessEventKindEvalProcessInspectionV1EvalProcessWorkerInspectionV1EvalResultAssertionIdentityV1EvalResultCaseProjectionV1EvalResultConflictEvalResultOriginEvalResultOutcomeDimensionsV1EvalResultPageEvalResultPresentationV1EvalResultPresentationV2EvalResultProjectionV1EvalResultProjectionV2EvalResultQueryEvalResultRecordEvalResultReportV1EvalResultReportV2EvalResultTargetIdentityV1EvalRunEvalRunAdmissionConflictEvalRunClaimEvalRunClaimLostEvalRunComparisonEvalRunContractV1EvalRunContractV2EvalRunCostBudgetEvalRunFailureCodeEvalRunFailureDiagnosticEvalRunFailureReasonEvalRunInvocationEvalRunLeaseEvalRunObservationEvalRunOwnershipEvalRunPageEvalRunQueryEvalRunRecordEvalRunRequestEvalRunResultSummaryEvalRunSpecEvalRunStateConflictEvalRunStatusEvalScenarioApprovalDecisionRecordEvalScenarioApprovalSubmissionEvalScenarioArtifactReferenceEvalScenarioCatalogEntryEvalScenarioCatalogPageEvalScenarioCatalogQueryEvalScenarioConflictEvalScenarioDocumentV2EvalScenarioDraftV2EvalScenarioInspectionV2EvalScenarioRunInvocationEvalScenarioRunProgressEvalScenarioStimulusV1EvalScenarioTrialFailureCodeEvalScenarioTrialPhaseEvalScenarioTrialProgressEvalSelectedCaseV1EvalSessionInspectionV1EvalSessionObservationV1EvalSessionReferenceV1EvalSimpleInputStimulusV1EvalStatusEvalStoreEvalStorePublicationRejectedEvalStoreResultTooLargeEvalStoreTransientContentionEvalStructuredJudgeComparisonV1EvalStructuredJudgeCriterionComparisonV1EvalStructuredJudgeCriterionPresentationV1EvalStructuredJudgeObservationMismatchV1EvalStructuredJudgePresentationV1EvalSuiteEvalSuiteCatalogEntryEvalSuiteCatalogPageEvalSuiteCatalogQueryEvalSuiteDocumentV1EvalSuiteDocumentV2EvalSuiteDocumentV3EvalSuiteDraftV1EvalSuiteDraftV2EvalSuiteDraftV3EvalSuiteRunExposureV1EvalSuiteSelectionV1EvalSuiteSpecEvalSuiteTrialPolicyV1EvalSuiteTrialRequestDraftV3EvalToolJsonAssertionComparisonV1EvalToolJsonObservationMismatchV1EvalTrialDiagnosticCodeEvalTrialOutputPreviewV1EvalTrialPresentationV1EvalTrialResultEvaluationEvidencePolicySpecEvaluationSourceIdentityV1EvaluationTargetIdentityEventNotOccurredEventOccurredEventPayloadContainsEvidenceSpoolExternalBodyReleaseV1ExternalContainerLaunchRequestV1ExternalContainerOperationAdapterExternalContainerOutputV1ExternalContainerUsageV1ExternalPrivateEvalCorpusExternalPrivateMemoryAblationArtifactPathsExternalPrivateMemoryAblationAuthorizationExternalPrivateMemoryAblationCacheEvidencePolicyExternalPrivateMemoryAblationCacheEvidenceStateExternalPrivateMemoryAblationDestinationExternalPrivateMemoryAblationEvidenceCollectorExternalPrivateMemoryAblationExecutionModeExternalPrivateMemoryAblationLimitationExternalPrivateMemoryAblationMethodologyExternalPrivateMemoryAblationResultExternalPrivateMemoryAblationRunFailureCodeExternalPrivateMemoryAblationRunStatusExternalPrivateMemoryAblationScheduleEntryExternalPrivateMemoryAblationSchedulePolicyExternalPrivateMemoryAblationScheduleStrategyExternalPrivateMemoryAblationSupplementalEvidenceExternalPrivateMemoryAblationTrialExternalPrivateMemoryAblationTrialMethodologyExternalProcessModelProviderExternalProcessTargetIdentityV1ExternalTrialEnvelopeV1ExternalTrialIdentityV1FinalOutputContainsFinalOutputContainsAssertionSpecFinalOutputEqualsAssertionSpecFinalOutputMatchesInMemoryEvalStoreIncrementalCaptureProgressIncrementalEvidenceAdmissionIncrementalEvidenceErrorIncrementalEvidenceLimitsIncrementalSessionSealIncrementalWorkflowCaptureErrorJudgePrivacyPolicyV1JudgeProfileIdentityV1JudgeReferenceV1KNOWLEDGE_MAINTENANCE_EVALUATION_CORPUS_SCHEMA_VERSIONKNOWLEDGE_MAINTENANCE_EVALUATION_RESULT_SCHEMA_VERSIONKnowledgeMaintenanceEvaluationCaseKnowledgeMaintenanceEvaluationCaseResultKnowledgeMaintenanceEvaluationClaimKnowledgeMaintenanceEvaluationCorpusKnowledgeMaintenanceEvaluationDispositionKnowledgeMaintenanceEvaluationEntryKnowledgeMaintenanceEvaluationMetricsKnowledgeMaintenanceEvaluationResultKnowledgeMaintenanceEvaluationScenarioLIVE_AUTHENTICATED_BROWSER_ACCEPTANCE_SUITE_IDLIVE_PUBLIC_BROWSER_ACCEPTANCE_SUITE_IDLLMJudgeLoadedBenchmarkPackageMEMORY_EXPERIMENT_REPORT_MAX_BYTESMEMORY_EXPERIMENT_REPORT_SCHEMA_VERSIONMEMORY_RETRIEVAL_BASELINE_SCHEMA_VERSIONMEMORY_RETRIEVAL_CORPUS_SCHEMA_VERSIONMaxEstimatedCostMaxEstimatedCostAssertionSpecMaxModelStepsMaxModelStepsAssertionSpecMaxToolCallsMaxToolCallsAssertionSpecMaxTotalTokensMaxTotalTokensAssertionSpecMemoryAttributionAssertionSpecMemoryCaseComparisonMemoryExperimentCaseMemoryExperimentGatePolicyMemoryExperimentReportMemoryExperimentReportRequestMemoryExperimentTrialEvidenceMemoryExperimentVariantMemoryMetricAvailabilityMemoryMetricBindingMemoryMetricDeltaMemoryMetricDirectionMemoryMetricDistributionMemoryMetricGateMemoryMetricObservationMemoryMetricRoleMemoryOperationalDeltaMemoryOperationalDimensionMemoryOperationalDistributionMemoryPairStatusMemoryPreparationOverheadEvidenceMemoryPublishedResultEvidenceMemoryRankingTermMemoryRetrievalAccessSpecMemoryRetrievalBaselineMetricsMemoryRetrievalBaselineResultMemoryRetrievalCaseMemoryRetrievalCaseResultMemoryRetrievalCorpusMemoryRetrievalCorpusEntryMemoryRetrievalIdProbeMemoryRetrievalLanguageSliceMemoryTrialAvailabilityMemoryTrialPairComparisonMemoryTrialReportRowMemoryVariantCostQualityReportMemoryVariantDispositionMemoryVariantDispositionStatusMemoryVariantOperationalReportModelJudgeAssertionSpecModelJudgeTargetOpaqueExternalCaseRefV1PRICING_PROFILE_SEMANTICS_VERSIONPROMOTABLE_RUN_INPUT_SCHEMA_VERSIONPROMOTION_CANDIDATE_MAX_BYTESPROMOTION_CANDIDATE_SCHEMA_VERSIONPROMOTION_SOURCE_SCHEMA_VERSIONPUBLISHED_EVAL_MAX_BYTESPUBLISHED_EVAL_OUTPUT_PREVIEW_BUDGET_BYTESPUBLISHED_EVAL_SCHEMA_VERSIONPreparedEvalJudgeCalibrationPreparedExternalPrivateMemoryAblationPricingProfileIdentityV1PrivateJudgeReferenceTargetPrivateJudgeReferenceV1ProbeRequirementsProcessEventAssertionSpecProcessEventsInOrderAssertionSpecPromotableRunInputV1PromotionCandidateV1PromotionCaseV1PromotionSourceV1PromotionWarningCodePublicJudgeReferenceDraftV1PublicJudgeReferenceV1PublishedArtifactDetailPublishedAssertionDetailPublishedAssertionResultPublishedChildStatusDetailPublishedEvalCaseResultPublishedEvalRunPublishedEvalTrialResultPublishedFinalOutputContainsDetailPublishedFinalOutputEqualsDetailPublishedJudgeReferenceIdentityV1PublishedMaxEstimatedCostDetailPublishedMaxModelStepsDetailPublishedMaxToolCallsDetailPublishedMaxTotalTokensDetailPublishedMemoryAttributionDetailPublishedModelJudgeCostV1PublishedModelJudgeDetailPublishedModelJudgeUsageV1PublishedProcessEventDetailPublishedProcessEventsInOrderDetailPublishedRootStatusDetailPublishedStructuredJudgeCostV1PublishedStructuredJudgeCriterionV1PublishedStructuredJudgeUsageV1PublishedStructuredModelJudgeDetailPublishedToolArgumentsContainDetailPublishedToolCalledDetailPublishedToolResultContainsDetailPublishedToolsCalledInOrderDetailPublishedUsageRecordedDetailPublishedUsageSummaryV1PublishedWorkspaceFileDetailRECALL_BASELINE_CORPUS_SCHEMA_VERSIONRECALL_BASELINE_RESULT_SCHEMA_VERSIONRUNTIME_REPLAY_DEFAULT_MAX_EVENTSRUNTIME_REPLAY_DEFAULT_MAX_MODEL_STEPSRUNTIME_REPLAY_DEFAULT_MAX_TOOL_CALLSRUNTIME_REPLAY_DEFAULT_MAX_TRANSCRIPT_MESSAGESRUNTIME_REPLAY_DEFAULT_TIMEOUT_SECONDSRUNTIME_REPLAY_HARD_MAX_EVENTSRUNTIME_REPLAY_HARD_MAX_MODEL_STEPSRUNTIME_REPLAY_HARD_MAX_TOOL_CALLSRUNTIME_REPLAY_HARD_MAX_TRANSCRIPT_MESSAGESRUNTIME_REPLAY_HARD_TIMEOUT_SECONDSRUNTIME_REPLAY_SCHEMA_VERSIONRecallBaselineCaseRecallBaselineCaseResultRecallBaselineCorpusRecallBaselineExpectedIdentityRecallBaselineKnowledgeEntryRecallBaselineMetricsRecallBaselineResultRecallBaselineTranscriptRecallBaselineTranscriptMessageRootStatusAssertionSpecRunInputSpecRuntimeReplayAttemptComparisonRuntimeReplayBoundaryKindRuntimeReplayBoundsRuntimeReplayDispositionRuntimeReplayDivergenceRuntimeReplayDivergenceKindRuntimeReplayFingerprintIdentityRuntimeReplayReasonRuntimeReplayReportRuntimeReplayRequestRuntimeReplayWarningSCENARIO_CAPTURE_ARTIFACT_READ_CONCURRENCYSCENARIO_CAPTURE_MAX_DIAGNOSTICSSCENARIO_PREFLIGHT_ARTIFACT_READ_CONCURRENCYSCENARIO_PREFLIGHT_MAX_DIAGNOSTICSSavedIncrementalWorkflowCaptureSavedIncrementalWorkflowScoreSavedWorkflowEvalCaptureSavedWorkflowEvalScoreScenarioApprovalCheckpointEventV2ScenarioArtifactLaunchBindingV2ScenarioArtifactMaterializationErrorScenarioArtifactMaterializationV2ScenarioArtifactRequirementV2ScenarioCaptureDiagnosticCodeScenarioCaptureDiagnosticV2ScenarioCaptureResultV2ScenarioEventV2ScenarioExecutionErrorScenarioFilePartV2ScenarioInitialInputEventV2ScenarioInputPartV2ScenarioInputV2ScenarioJsonPartV2ScenarioLaunchBindingV2ScenarioLaunchDiagnosticCodeScenarioLaunchDiagnosticV2ScenarioLaunchPreflightResultV2ScenarioLaunchSettingsV2ScenarioQueuedInputEventV2ScenarioResumedInputEventV2ScenarioSecretLaunchBindingV2ScenarioSecretRequirementV2ScenarioTextPartV2ScenarioUserMessageV2ScriptedModelProviderSessionCompletedSessionFailedSessionInterruptedSessionPromotionErrorSessionPromotionErrorCodeSessionStatusIsSessionTrajectoryBoundsSessionTrajectoryErrorSessionTrajectoryErrorCodeStructuredModelJudgeAssertionDraftV1StructuredModelJudgeAssertionSpecStructuredRubricCriterionV1StructuredRubricDraftV1StructuredRubricV1TERMINAL_EVAL_RUN_STATUSESTRAJECTORY_SCHEMA_VERSIONToolArgsContainToolArgumentsContainAssertionSpecToolCallEvidenceV1ToolCallValueEvidenceV1ToolCalledToolCalledAssertionSpecToolNotCalledToolResultContainsToolResultContainsAssertionSpecToolsCalledInOrderToolsCalledInOrderAssertionSpecTrajectoryTrajectoryProbesTranscriptContainsTrialRequestSpecUsageRecordedUsageRecordedAssertionSpecWORKFLOW_EVAL_DEFAULT_CLOSE_TIMEOUT_SECONDSWORKFLOW_EVAL_MAX_APPLICATION_CONTEXT_BYTESWORKFLOW_EVAL_MAX_CLOSE_TIMEOUT_SECONDSWORKFLOW_EVAL_MAX_FINAL_OUTPUT_CHARSWORKFLOW_EVAL_MAX_INPUT_BYTESWORKFLOW_EVAL_MAX_INPUT_MESSAGESWORKFLOW_EVAL_MAX_STRUCTURED_OUTPUT_BYTESWORKSPACE_PROBE_MAX_BYTESWorkflowEvalExecutionWorkflowEvalFactoryWorkflowEvalInstanceScopeWorkflowEvalInvocationWorkflowEvalOutputEvidenceV1WorkflowEvalResultWorkflowEvalResultProjectorWorkflowEvalTargetWorkflowEvalTargetIdentityV1WorkflowEvalTerminalEvidenceWorkspaceFileAssertionSpecWorkspaceFileContainsWorkspaceFileExistsWorkspaceFileProbeWorkspaceStructuralEvidenceV1add_eval_caseassertion_spec_revisionauthored_suite_launch_settingsbenchmark_package_from_jsonbenchmark_package_scenariosbenchmark_package_to_jsonbenchmark_suite_selectionbrowser_acceptance_report_from_jsonbrowser_acceptance_report_to_jsonbuild_browser_acceptance_reportbuild_browser_acceptance_retry_reportbuild_captured_evaluation_candidatebuild_causal_memory_reference_corpusbuild_memory_experiment_reportbuild_promotion_candidatecapture_eval_scenario_from_sessioncapture_incremental_workflow_eval_attemptcapture_workflow_eval_attemptcaptured_evaluation_result_from_jsoncaptured_evaluation_result_to_jsoncompare_corpus_execution_resultscompare_eval_resultscompare_eval_runscomparison_to_jsoncompile_assertion_speccompile_authored_suite_run_exposurecompile_corpus_suitecompile_eval_judge_calibration_draftcompile_eval_scenariocompile_eval_scenario_draftcompile_eval_suite_authoring_draftcompile_eval_suite_draftcompile_eval_suite_draft_v2compile_eval_suite_draft_v3corpus_execution_comparison_to_jsoncorpus_execution_compatibilitycorpus_execution_result_from_jsoncorpus_execution_result_to_jsoncorpus_for_authored_scenario_casecorpus_for_authored_simple_selectioncorpus_for_eval_scenariocorpus_from_captured_evaluation_candidatecorpus_from_promotion_candidatedeterministic_browser_acceptance_manifestduplicate_eval_caseeval_corpus_from_jsoneval_corpus_inspection_to_jsoneval_corpus_to_jsoneval_judge_calibration_report_from_jsoneval_judge_calibration_report_to_jsoneval_memory_attribution_bounds_for_trial_counteval_memory_attribution_fingerprinteval_memory_attribution_max_bytes_for_trial_counteval_memory_attribution_source_limit_for_trial_counteval_memory_source_aliaseval_result_compatibilityeval_result_projectioneval_result_report_from_jsoneval_result_report_to_jsoneval_result_to_jsoneval_run_contract_for_corpuseval_run_observationeval_run_to_jsoneval_scenario_from_jsoneval_scenario_to_jsoneval_suite_document_from_jsoneval_suite_document_to_jsoneval_suite_selectionevaluate_assertion_specevaluate_assertion_specsevaluate_assertionsevaluation_target_identityexport_captured_evaluation_corpusexport_process_eval_runexport_promotion_corpusexternal_body_content_revisionexternal_body_file_revisionexternal_container_environment_revisionexternal_private_memory_ablation_destinationexternal_private_memory_ablation_experiment_revisionexternal_private_memory_ablation_methodology_to_jsonexternal_trial_envelope_from_requestfinal_output_textimport_workflow_eval_attemptinspect_browser_acceptance_runtime_identityinspect_eval_corpusinspect_eval_scenarioinspect_eval_sessionsinspect_process_eval_runlive_authenticated_browser_acceptance_manifestlive_public_browser_acceptance_manifestload_benchmark_packageload_causal_memory_reference_corpusload_corpus_execution_resultload_eval_corpusload_eval_runload_eval_scenarioload_external_private_memory_ablation_corpusload_knowledge_maintenance_evaluation_corpusload_memory_retrieval_corpusload_recall_baseline_corpusload_trajectorymaterialize_eval_scenario_artifact_fixturememory_experiment_accounting_source_idmemory_experiment_accounting_task_idmemory_experiment_report_from_jsonmemory_experiment_report_to_jsonmemory_experiment_request_from_jsonmerge_eval_corporamerge_eval_corpus_filesmodel_facing_textmodel_facing_tool_resultmodel_judge_implementation_revisionmodel_judge_profilepreflight_eval_scenarioprepare_eval_judge_calibrationprepare_external_private_memory_ablationpresent_eval_resultpricing_profile_identityproject_assertion_evidence_viewproject_browser_acceptance_diagnosticproject_browser_acceptance_trialpromotable_run_inputpublish_eval_runrender_browser_acceptance_htmlrender_captured_evaluation_htmlrender_comparison_htmlrender_corpus_execution_comparison_htmlrender_corpus_execution_htmlrender_eval_result_htmlrender_html_reportrender_memory_experiment_report_htmlreplace_eval_scenario_artifact_requirementreplay_sessionrevise_eval_caserun_browser_acceptancerun_causal_memory_reference_campaignrun_compiled_eval_scenariorun_corpus_suiterun_eval_caserun_eval_judge_calibration_trialrun_eval_planrun_eval_suiterun_external_private_memory_ablationrun_knowledge_maintenance_evaluationrun_memory_retrieval_baselinerun_recall_baselinerun_workflow_eval_suiterunnable_promotion_candidatescenario_from_corpus_casescenario_launch_settings_from_invocationscore_captured_evaluation_candidatescore_incremental_workflow_eval_capturescore_promotion_candidatescore_workflow_eval_capturescripted_structured_outputstandard_eval_memory_attribution_boundstrajectory_from_sessiontrajectory_to_jsonvalidate_captured_result_for_corpusvalidate_eval_suite_selectionvalidate_expected_eval_suite_revisionvalidate_expected_scenario_revisionwith_external_trial_envelopeworkflow_eval_input_messages_sha256workflow_eval_output_sha256workflow_eval_trial_session_idworkflow_spec_revisionwrite_browser_acceptance_reportwrite_corpus_execution_htmlwrite_corpus_execution_resultwrite_eval_run_jsonwrite_external_private_memory_ablation_artifactswrite_html_reportwrite_trajectory_json