openpond-sdk 0.9.2 → 0.9.4

package/dist/types/packages/sdk/src/experiment-client.d.ts

@@ -0,0 +1,2441 @@
+import { z } from "zod";
+import { ExperimentListQuerySchema, SaveExperimentSchema, StartExperimentSchema, type ExperimentDefinitionRef } from "./experiment-contracts.js";
+import { PrepareHarnessExperimentSchema } from "./experiment-contracts.js";
+import { type ExperimentScoringRequest } from "./experiment-scoring-contracts.js";
+export declare class OpenPondExperimentError extends Error {
+    readonly status: number;
+    readonly code: string;
+    constructor(status: number, code: string, message: string);
+}
+/** Saved configuration and execution identities share the host's service. */
+export declare class OpenPondExperimentsClient {
+    private readonly options;
+    private readonly baseUrl;
+    constructor(options: {
+        baseUrl: string;
+        apiKey: string;
+        teamId: string;
+        fetch?: typeof fetch;
+    });
+    /** Resolves a released target without saving a definition or dispatching it. */
+    prepareHarness(value: z.input<typeof PrepareHarnessExperimentSchema>, signal?: AbortSignal): Promise<{
+        request: {
+            schemaVersion: "openpond.modelTasksetRunRequest.v1";
+            operationId: string;
+            teamId: string;
+            modelProjectId: string | null;
+            taskset: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            policy: {
+                kind: "hosted_chat";
+                modelId: string;
+                maxOutputTokens: number;
+                temperature: number;
+                topP: number;
+                messages?: {
+                    role: "user" | "system" | "assistant";
+                    content: string;
+                }[] | undefined;
+            } | {
+                kind: "hosted_harness";
+                modelId: string;
+                profileRepositoryId: string;
+                source: {
+                    profileId: string;
+                    sourceRevision: string;
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    catalogHash: string;
+                    definitionId: string;
+                    definitionHash: string;
+                    target: {
+                        kind: "profile";
+                    } | {
+                        kind: "workflow";
+                        workflowId: string;
+                    } | {
+                        kind: "skill";
+                        skillPath: string;
+                    } | {
+                        kind: "agent_action";
+                        actionId: string;
+                    };
+                    environmentHash: string;
+                };
+                modelConfigurationHash: string;
+                packageHash: string;
+                reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
+            } | {
+                kind: "fixture";
+            };
+            population: {
+                receiptId: string;
+                taskId: string;
+                seed: string;
+                fixtureId: string | null;
+            }[];
+            name?: string | undefined;
+            project?: {
+                id: string;
+                revision: number;
+                contentHash: string;
+                targetId: string | null;
+            } | undefined;
+        };
+        manifest: {
+            schemaVersion: "openpond.tasksetRunManifest.v1";
+            id: string;
+            tasksetRelease: {
+                id: string;
+                contentHash: string;
+            };
+            packageHash: string;
+            execution: {
+                kind: "taskset";
+                environmentRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                verifierSetRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                policyHash: string;
+            } | {
+                kind: "harness";
+                harnessRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+            };
+            policy: {
+                kind: "model";
+                model: {
+                    provider: string;
+                    model: string;
+                    revision: string | null;
+                    artifactHash: string | null;
+                    tokenizerRevision: string | null;
+                    chatTemplateHash: string | null;
+                };
+                configurationHash: string;
+            } | {
+                kind: "fixture";
+            };
+            gradingRole: "evaluation";
+            metricPolicy: {
+                schemaVersion: "openpond.tasksetMetricPolicy.v1";
+                primaryMetric: string;
+                aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
+                missingReward: "zero" | "exclude";
+                customAggregator: {
+                    module: string;
+                    exportName: string;
+                    contentHash: string;
+                    timeoutMs: number;
+                    networkPolicy: "none";
+                } | null;
+                taskWeights?: Record<string, number> | undefined;
+            };
+            population: {
+                receiptId: string;
+                taskId: string;
+                seed: string;
+                fixtureId: string | null;
+            }[];
+            runtimeTarget: {
+                adapterId: string;
+                placement: "local" | "remote" | "colocated" | "provider_native";
+                runtimeVersion: string;
+                capabilityReceipt: string;
+            };
+            limits: {
+                maxTurns: number;
+                timeoutMs: number;
+                maxOutputBytes: number;
+                maximumSpendUsd: number | null;
+            };
+            createdAt: string;
+            metadata: Record<string, unknown>;
+            contentHash: string;
+            profileEvaluation?: {
+                profileId: string;
+                sourceRevision: string;
+                harnessRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                catalogHash: string;
+                definitionId: string;
+                definitionHash: string;
+                target: {
+                    kind: "profile";
+                } | {
+                    kind: "workflow";
+                    workflowId: string;
+                } | {
+                    kind: "skill";
+                    skillPath: string;
+                } | {
+                    kind: "agent_action";
+                    actionId: string;
+                };
+                environmentHash: string;
+            } | undefined;
+        };
+        maximumCostUsd: number;
+        graders: {
+            id: string;
+            version: string;
+            contentHash: string;
+            feedbackKey: string;
+            release: {
+                id: string;
+                revision: number;
+                contentHash: string;
+            } | null;
+            mappings?: {
+                destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                path: string[];
+                source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                sourcePath: string[];
+            }[] | undefined;
+        }[];
+    }>;
+    save(value: z.input<typeof SaveExperimentSchema>, signal?: AbortSignal): Promise<{
+        schemaVersion: "openpond.experimentDefinition.v1";
+        id: string;
+        teamId: string;
+        revision: number;
+        request: {
+            schemaVersion: "openpond.modelTasksetRunRequest.v1";
+            operationId: string;
+            teamId: string;
+            modelProjectId: string | null;
+            taskset: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            policy: {
+                kind: "hosted_chat";
+                modelId: string;
+                maxOutputTokens: number;
+                temperature: number;
+                topP: number;
+                messages?: {
+                    role: "user" | "system" | "assistant";
+                    content: string;
+                }[] | undefined;
+            } | {
+                kind: "hosted_harness";
+                modelId: string;
+                profileRepositoryId: string;
+                source: {
+                    profileId: string;
+                    sourceRevision: string;
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    catalogHash: string;
+                    definitionId: string;
+                    definitionHash: string;
+                    target: {
+                        kind: "profile";
+                    } | {
+                        kind: "workflow";
+                        workflowId: string;
+                    } | {
+                        kind: "skill";
+                        skillPath: string;
+                    } | {
+                        kind: "agent_action";
+                        actionId: string;
+                    };
+                    environmentHash: string;
+                };
+                modelConfigurationHash: string;
+                packageHash: string;
+                reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
+            } | {
+                kind: "fixture";
+            };
+            population: {
+                receiptId: string;
+                taskId: string;
+                seed: string;
+                fixtureId: string | null;
+            }[];
+            name?: string | undefined;
+            project?: {
+                id: string;
+                revision: number;
+                contentHash: string;
+                targetId: string | null;
+            } | undefined;
+        };
+        maximumCostUsd: number;
+        graders: {
+            id: string;
+            version: string;
+            contentHash: string;
+            feedbackKey: string;
+            release: {
+                id: string;
+                revision: number;
+                contentHash: string;
+            } | null;
+            mappings?: {
+                destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                path: string[];
+                source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                sourcePath: string[];
+            }[] | undefined;
+        }[];
+        createdAt: string;
+        updatedAt: string;
+        contentHash: string;
+    }>;
+    get(id: string, options?: {
+        reference?: ExperimentDefinitionRef;
+        signal?: AbortSignal;
+    }): Promise<{
+        schemaVersion: "openpond.experimentDefinition.v1";
+        id: string;
+        teamId: string;
+        revision: number;
+        request: {
+            schemaVersion: "openpond.modelTasksetRunRequest.v1";
+            operationId: string;
+            teamId: string;
+            modelProjectId: string | null;
+            taskset: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            policy: {
+                kind: "hosted_chat";
+                modelId: string;
+                maxOutputTokens: number;
+                temperature: number;
+                topP: number;
+                messages?: {
+                    role: "user" | "system" | "assistant";
+                    content: string;
+                }[] | undefined;
+            } | {
+                kind: "hosted_harness";
+                modelId: string;
+                profileRepositoryId: string;
+                source: {
+                    profileId: string;
+                    sourceRevision: string;
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    catalogHash: string;
+                    definitionId: string;
+                    definitionHash: string;
+                    target: {
+                        kind: "profile";
+                    } | {
+                        kind: "workflow";
+                        workflowId: string;
+                    } | {
+                        kind: "skill";
+                        skillPath: string;
+                    } | {
+                        kind: "agent_action";
+                        actionId: string;
+                    };
+                    environmentHash: string;
+                };
+                modelConfigurationHash: string;
+                packageHash: string;
+                reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
+            } | {
+                kind: "fixture";
+            };
+            population: {
+                receiptId: string;
+                taskId: string;
+                seed: string;
+                fixtureId: string | null;
+            }[];
+            name?: string | undefined;
+            project?: {
+                id: string;
+                revision: number;
+                contentHash: string;
+                targetId: string | null;
+            } | undefined;
+        };
+        maximumCostUsd: number;
+        graders: {
+            id: string;
+            version: string;
+            contentHash: string;
+            feedbackKey: string;
+            release: {
+                id: string;
+                revision: number;
+                contentHash: string;
+            } | null;
+            mappings?: {
+                destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                path: string[];
+                source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                sourcePath: string[];
+            }[] | undefined;
+        }[];
+        createdAt: string;
+        updatedAt: string;
+        contentHash: string;
+    }>;
+    list(value?: z.input<typeof ExperimentListQuerySchema>, signal?: AbortSignal): Promise<{
+        items: {
+            schemaVersion: "openpond.experimentDefinition.v1";
+            id: string;
+            teamId: string;
+            revision: number;
+            request: {
+                schemaVersion: "openpond.modelTasksetRunRequest.v1";
+                operationId: string;
+                teamId: string;
+                modelProjectId: string | null;
+                taskset: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                policy: {
+                    kind: "hosted_chat";
+                    modelId: string;
+                    maxOutputTokens: number;
+                    temperature: number;
+                    topP: number;
+                    messages?: {
+                        role: "user" | "system" | "assistant";
+                        content: string;
+                    }[] | undefined;
+                } | {
+                    kind: "hosted_harness";
+                    modelId: string;
+                    profileRepositoryId: string;
+                    source: {
+                        profileId: string;
+                        sourceRevision: string;
+                        harnessRelease: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        catalogHash: string;
+                        definitionId: string;
+                        definitionHash: string;
+                        target: {
+                            kind: "profile";
+                        } | {
+                            kind: "workflow";
+                            workflowId: string;
+                        } | {
+                            kind: "skill";
+                            skillPath: string;
+                        } | {
+                            kind: "agent_action";
+                            actionId: string;
+                        };
+                        environmentHash: string;
+                    };
+                    modelConfigurationHash: string;
+                    packageHash: string;
+                    reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
+                } | {
+                    kind: "fixture";
+                };
+                population: {
+                    receiptId: string;
+                    taskId: string;
+                    seed: string;
+                    fixtureId: string | null;
+                }[];
+                name?: string | undefined;
+                project?: {
+                    id: string;
+                    revision: number;
+                    contentHash: string;
+                    targetId: string | null;
+                } | undefined;
+            };
+            maximumCostUsd: number;
+            graders: {
+                id: string;
+                version: string;
+                contentHash: string;
+                feedbackKey: string;
+                release: {
+                    id: string;
+                    revision: number;
+                    contentHash: string;
+                } | null;
+                mappings?: {
+                    destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    path: string[];
+                    source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    sourcePath: string[];
+                }[] | undefined;
+            }[];
+            createdAt: string;
+            updatedAt: string;
+            contentHash: string;
+        }[];
+        nextCursor: string | null;
+    }>;
+    start(value: z.input<typeof StartExperimentSchema>, signal?: AbortSignal): Promise<{
+        summary: {
+            schemaVersion: "openpond.modelTasksetRunSummary.v1";
+            id: string;
+            revision: number;
+            teamId: string;
+            modelProjectId: string | null;
+            operationId: string;
+            taskset: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            policyKind: "fixture" | "hosted_chat" | "hosted_harness";
+            manifestHash: string;
+            status: "failed" | "completed" | "cancelled" | "running" | "cancelling" | "queued";
+            totalCount: number;
+            counts: {
+                pending: number;
+                running: number;
+                completed: number;
+                failed: number;
+                cancelled: number;
+            };
+            createdAt: string;
+            startedAt: string | null;
+            completedAt: string | null;
+            cleanupComplete: boolean;
+            resultAvailable: boolean;
+            score: number | null;
+            metricName: string;
+            error: {
+                code: string;
+                message: string;
+            } | null;
+            name?: string | undefined;
+            project?: {
+                id: string;
+                revision: number;
+                contentHash: string;
+                targetId: string | null;
+            } | undefined;
+        };
+        request: {
+            schemaVersion: "openpond.modelTasksetRunRequest.v1";
+            operationId: string;
+            teamId: string;
+            modelProjectId: string | null;
+            taskset: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            policy: {
+                kind: "hosted_chat";
+                modelId: string;
+                maxOutputTokens: number;
+                temperature: number;
+                topP: number;
+                messages?: {
+                    role: "user" | "system" | "assistant";
+                    content: string;
+                }[] | undefined;
+            } | {
+                kind: "hosted_harness";
+                modelId: string;
+                profileRepositoryId: string;
+                source: {
+                    profileId: string;
+                    sourceRevision: string;
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    catalogHash: string;
+                    definitionId: string;
+                    definitionHash: string;
+                    target: {
+                        kind: "profile";
+                    } | {
+                        kind: "workflow";
+                        workflowId: string;
+                    } | {
+                        kind: "skill";
+                        skillPath: string;
+                    } | {
+                        kind: "agent_action";
+                        actionId: string;
+                    };
+                    environmentHash: string;
+                };
+                modelConfigurationHash: string;
+                packageHash: string;
+                reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
+            } | {
+                kind: "fixture";
+            };
+            population: {
+                receiptId: string;
+                taskId: string;
+                seed: string;
+                fixtureId: string | null;
+            }[];
+            name?: string | undefined;
+            project?: {
+                id: string;
+                revision: number;
+                contentHash: string;
+                targetId: string | null;
+            } | undefined;
+        };
+        manifest: {
+            schemaVersion: "openpond.tasksetRunManifest.v1";
+            id: string;
+            tasksetRelease: {
+                id: string;
+                contentHash: string;
+            };
+            packageHash: string;
+            execution: {
+                kind: "taskset";
+                environmentRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                verifierSetRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                policyHash: string;
+            } | {
+                kind: "harness";
+                harnessRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+            };
+            policy: {
+                kind: "model";
+                model: {
+                    provider: string;
+                    model: string;
+                    revision: string | null;
+                    artifactHash: string | null;
+                    tokenizerRevision: string | null;
+                    chatTemplateHash: string | null;
+                };
+                configurationHash: string;
+            } | {
+                kind: "fixture";
+            };
+            gradingRole: "evaluation";
+            metricPolicy: {
+                schemaVersion: "openpond.tasksetMetricPolicy.v1";
+                primaryMetric: string;
+                aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
+                missingReward: "zero" | "exclude";
+                customAggregator: {
+                    module: string;
+                    exportName: string;
+                    contentHash: string;
+                    timeoutMs: number;
+                    networkPolicy: "none";
+                } | null;
+                taskWeights?: Record<string, number> | undefined;
+            };
+            population: {
+                receiptId: string;
+                taskId: string;
+                seed: string;
+                fixtureId: string | null;
+            }[];
+            runtimeTarget: {
+                adapterId: string;
+                placement: "local" | "remote" | "colocated" | "provider_native";
+                runtimeVersion: string;
+                capabilityReceipt: string;
+            };
+            limits: {
+                maxTurns: number;
+                timeoutMs: number;
+                maxOutputBytes: number;
+                maximumSpendUsd: number | null;
+            };
+            createdAt: string;
+            metadata: Record<string, unknown>;
+            contentHash: string;
+            profileEvaluation?: {
+                profileId: string;
+                sourceRevision: string;
+                harnessRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                catalogHash: string;
+                definitionId: string;
+                definitionHash: string;
+                target: {
+                    kind: "profile";
+                } | {
+                    kind: "workflow";
+                    workflowId: string;
+                } | {
+                    kind: "skill";
+                    skillPath: string;
+                } | {
+                    kind: "agent_action";
+                    actionId: string;
+                };
+                environmentHash: string;
+            } | undefined;
+        };
+        policySnapshot: {
+            modelId: string;
+            provider: string;
+            upstreamModelId: string;
+            configurationHash: string;
+        } | null;
+    }>;
+    result(executionId: string, signal?: AbortSignal): Promise<{
+        executionId: string;
+        passId: string | null;
+        manifest: {
+            schemaVersion: "openpond.experimentManifest.v1";
+            id: string;
+            name: string;
+            teamId: string;
+            operationId: string;
+            maximumCostUsd: number | null;
+            dataset: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            target: {
+                kind: "model";
+                modelId: string;
+                configurationHash: string;
+            } | {
+                kind: "harness";
+                source: {
+                    profileId: string;
+                    sourceRevision: string;
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    catalogHash: string;
+                    definitionId: string;
+                    definitionHash: string;
+                    target: {
+                        kind: "profile";
+                    } | {
+                        kind: "workflow";
+                        workflowId: string;
+                    } | {
+                        kind: "skill";
+                        skillPath: string;
+                    } | {
+                        kind: "agent_action";
+                        actionId: string;
+                    };
+                    environmentHash: string;
+                };
+            } | {
+                kind: "fixture";
+                configurationHash: string;
+            };
+            evaluators: {
+                release: {
+                    id: string;
+                    contentHash: string;
+                    revision: string | number;
+                };
+                feedbackKey: string;
+                output: "boolean" | "score" | "category";
+                categories: string[];
+                configurationHash?: string | undefined;
+            }[];
+            population: {
+                caseId: string;
+                seed: string;
+                fixtureId: string | null;
+            }[];
+            createdAt: string;
+            contentHash: string;
+            execution?: {
+                packageHash: string;
+                runtimeTargetHash: string;
+                metricPolicyHash: string;
+            } | undefined;
+            lineage?: {
+                definition: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                } | null;
+                execution: {
+                    id: string;
+                    contentHash: string;
+                };
+                scoringPassId: string | null;
+            } | undefined;
+        };
+        result: {
+            schemaVersion: "openpond.experimentResult.v1";
+            manifest: {
+                id: string;
+                contentHash: string;
+            };
+            status: "failed" | "completed" | "cancelled";
+            cases: {
+                identity: {
+                    caseId: string;
+                    seed: string;
+                    fixtureId: string | null;
+                };
+                status: "failed" | "completed" | "cancelled" | "unavailable";
+                output: unknown;
+                error: {
+                    code: string;
+                    message: string;
+                } | null;
+                feedback: {
+                    feedbackKey: string;
+                    evaluator: {
+                        id: string;
+                        contentHash: string;
+                        revision: string | number;
+                    };
+                    status: "pending" | "failed" | "scored" | "unavailable";
+                    value: string | number | boolean | null;
+                    reasoning: string | null;
+                    evidenceRefs: {
+                        id: string;
+                        contentHash: string;
+                        mediaType: string | null;
+                        sizeBytes: number | null;
+                    }[];
+                    passed?: boolean | null | undefined;
+                }[];
+                usage: {
+                    inputTokens: number | null;
+                    outputTokens: number | null;
+                    totalTokens: number | null;
+                    costUsd: number | null;
+                    latencyMs: number | null;
+                };
+                traceRef: {
+                    id: string;
+                    contentHash: string;
+                    mediaType: string | null;
+                    sizeBytes: number | null;
+                } | null;
+                startedAt: string | null;
+                completedAt: string | null;
+            }[];
+            completedAt: string;
+            contentHash: string;
+        };
+    }>;
+    executions(definitionId: string, options?: {
+        afterId?: string;
+        signal?: AbortSignal;
+    }): Promise<{
+        items: {
+            summary: {
+                schemaVersion: "openpond.modelTasksetRunSummary.v1";
+                id: string;
+                revision: number;
+                teamId: string;
+                modelProjectId: string | null;
+                operationId: string;
+                taskset: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                policyKind: "fixture" | "hosted_chat" | "hosted_harness";
+                manifestHash: string;
+                status: "failed" | "completed" | "cancelled" | "running" | "cancelling" | "queued";
+                totalCount: number;
+                counts: {
+                    pending: number;
+                    running: number;
+                    completed: number;
+                    failed: number;
+                    cancelled: number;
+                };
+                createdAt: string;
+                startedAt: string | null;
+                completedAt: string | null;
+                cleanupComplete: boolean;
+                resultAvailable: boolean;
+                score: number | null;
+                metricName: string;
+                error: {
+                    code: string;
+                    message: string;
+                } | null;
+                name?: string | undefined;
+                project?: {
+                    id: string;
+                    revision: number;
+                    contentHash: string;
+                    targetId: string | null;
+                } | undefined;
+            };
+            request: {
+                schemaVersion: "openpond.modelTasksetRunRequest.v1";
+                operationId: string;
+                teamId: string;
+                modelProjectId: string | null;
+                taskset: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                policy: {
+                    kind: "hosted_chat";
+                    modelId: string;
+                    maxOutputTokens: number;
+                    temperature: number;
+                    topP: number;
+                    messages?: {
+                        role: "user" | "system" | "assistant";
+                        content: string;
+                    }[] | undefined;
+                } | {
+                    kind: "hosted_harness";
+                    modelId: string;
+                    profileRepositoryId: string;
+                    source: {
+                        profileId: string;
+                        sourceRevision: string;
+                        harnessRelease: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        catalogHash: string;
+                        definitionId: string;
+                        definitionHash: string;
+                        target: {
+                            kind: "profile";
+                        } | {
+                            kind: "workflow";
+                            workflowId: string;
+                        } | {
+                            kind: "skill";
+                            skillPath: string;
+                        } | {
+                            kind: "agent_action";
+                            actionId: string;
+                        };
+                        environmentHash: string;
+                    };
+                    modelConfigurationHash: string;
+                    packageHash: string;
+                    reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
+                } | {
+                    kind: "fixture";
+                };
+                population: {
+                    receiptId: string;
+                    taskId: string;
+                    seed: string;
+                    fixtureId: string | null;
+                }[];
+                name?: string | undefined;
+                project?: {
+                    id: string;
+                    revision: number;
+                    contentHash: string;
+                    targetId: string | null;
+                } | undefined;
+            };
+            manifest: {
+                schemaVersion: "openpond.tasksetRunManifest.v1";
+                id: string;
+                tasksetRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                packageHash: string;
+                execution: {
+                    kind: "taskset";
+                    environmentRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    verifierSetRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    policyHash: string;
+                } | {
+                    kind: "harness";
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                };
+                policy: {
+                    kind: "model";
+                    model: {
+                        provider: string;
+                        model: string;
+                        revision: string | null;
+                        artifactHash: string | null;
+                        tokenizerRevision: string | null;
+                        chatTemplateHash: string | null;
+                    };
+                    configurationHash: string;
+                } | {
+                    kind: "fixture";
+                };
+                gradingRole: "evaluation";
+                metricPolicy: {
+                    schemaVersion: "openpond.tasksetMetricPolicy.v1";
+                    primaryMetric: string;
+                    aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
+                    missingReward: "zero" | "exclude";
+                    customAggregator: {
+                        module: string;
+                        exportName: string;
+                        contentHash: string;
+                        timeoutMs: number;
+                        networkPolicy: "none";
+                    } | null;
+                    taskWeights?: Record<string, number> | undefined;
+                };
+                population: {
+                    receiptId: string;
+                    taskId: string;
+                    seed: string;
+                    fixtureId: string | null;
+                }[];
+                runtimeTarget: {
+                    adapterId: string;
+                    placement: "local" | "remote" | "colocated" | "provider_native";
+                    runtimeVersion: string;
+                    capabilityReceipt: string;
+                };
+                limits: {
+                    maxTurns: number;
+                    timeoutMs: number;
+                    maxOutputBytes: number;
+                    maximumSpendUsd: number | null;
+                };
+                createdAt: string;
+                metadata: Record<string, unknown>;
+                contentHash: string;
+                profileEvaluation?: {
+                    profileId: string;
+                    sourceRevision: string;
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    catalogHash: string;
+                    definitionId: string;
+                    definitionHash: string;
+                    target: {
+                        kind: "profile";
+                    } | {
+                        kind: "workflow";
+                        workflowId: string;
+                    } | {
+                        kind: "skill";
+                        skillPath: string;
+                    } | {
+                        kind: "agent_action";
+                        actionId: string;
+                    };
+                    environmentHash: string;
+                } | undefined;
+            };
+            policySnapshot: {
+                modelId: string;
+                provider: string;
+                upstreamModelId: string;
+                configurationHash: string;
+            } | null;
+        }[];
+        nextCursor: string | null;
+    }>;
+    execution(id: string, signal?: AbortSignal): Promise<{
+        summary: {
+            schemaVersion: "openpond.modelTasksetRunSummary.v1";
+            id: string;
+            revision: number;
+            teamId: string;
+            modelProjectId: string | null;
+            operationId: string;
+            taskset: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            policyKind: "fixture" | "hosted_chat" | "hosted_harness";
+            manifestHash: string;
+            status: "failed" | "completed" | "cancelled" | "running" | "cancelling" | "queued";
+            totalCount: number;
+            counts: {
+                pending: number;
+                running: number;
+                completed: number;
+                failed: number;
+                cancelled: number;
+            };
+            createdAt: string;
+            startedAt: string | null;
+            completedAt: string | null;
+            cleanupComplete: boolean;
+            resultAvailable: boolean;
+            score: number | null;
+            metricName: string;
+            error: {
+                code: string;
+                message: string;
+            } | null;
+            name?: string | undefined;
+            project?: {
+                id: string;
+                revision: number;
+                contentHash: string;
+                targetId: string | null;
+            } | undefined;
+        };
+        request: {
+            schemaVersion: "openpond.modelTasksetRunRequest.v1";
+            operationId: string;
+            teamId: string;
+            modelProjectId: string | null;
+            taskset: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            policy: {
+                kind: "hosted_chat";
+                modelId: string;
+                maxOutputTokens: number;
+                temperature: number;
+                topP: number;
+                messages?: {
+                    role: "user" | "system" | "assistant";
+                    content: string;
+                }[] | undefined;
+            } | {
+                kind: "hosted_harness";
+                modelId: string;
+                profileRepositoryId: string;
+                source: {
+                    profileId: string;
+                    sourceRevision: string;
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    catalogHash: string;
+                    definitionId: string;
+                    definitionHash: string;
+                    target: {
+                        kind: "profile";
+                    } | {
+                        kind: "workflow";
+                        workflowId: string;
+                    } | {
+                        kind: "skill";
+                        skillPath: string;
+                    } | {
+                        kind: "agent_action";
+                        actionId: string;
+                    };
+                    environmentHash: string;
+                };
+                modelConfigurationHash: string;
+                packageHash: string;
+                reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
+            } | {
+                kind: "fixture";
+            };
+            population: {
+                receiptId: string;
+                taskId: string;
+                seed: string;
+                fixtureId: string | null;
+            }[];
+            name?: string | undefined;
+            project?: {
+                id: string;
+                revision: number;
+                contentHash: string;
+                targetId: string | null;
+            } | undefined;
+        };
+        manifest: {
+            schemaVersion: "openpond.tasksetRunManifest.v1";
+            id: string;
+            tasksetRelease: {
+                id: string;
+                contentHash: string;
+            };
+            packageHash: string;
+            execution: {
+                kind: "taskset";
+                environmentRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                verifierSetRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                policyHash: string;
+            } | {
+                kind: "harness";
+                harnessRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+            };
+            policy: {
+                kind: "model";
+                model: {
+                    provider: string;
+                    model: string;
+                    revision: string | null;
+                    artifactHash: string | null;
+                    tokenizerRevision: string | null;
+                    chatTemplateHash: string | null;
+                };
+                configurationHash: string;
+            } | {
+                kind: "fixture";
+            };
+            gradingRole: "evaluation";
+            metricPolicy: {
+                schemaVersion: "openpond.tasksetMetricPolicy.v1";
+                primaryMetric: string;
+                aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
+                missingReward: "zero" | "exclude";
+                customAggregator: {
+                    module: string;
+                    exportName: string;
+                    contentHash: string;
+                    timeoutMs: number;
+                    networkPolicy: "none";
+                } | null;
+                taskWeights?: Record<string, number> | undefined;
+            };
+            population: {
+                receiptId: string;
+                taskId: string;
+                seed: string;
+                fixtureId: string | null;
+            }[];
+            runtimeTarget: {
+                adapterId: string;
+                placement: "local" | "remote" | "colocated" | "provider_native";
+                runtimeVersion: string;
+                capabilityReceipt: string;
+            };
+            limits: {
+                maxTurns: number;
+                timeoutMs: number;
+                maxOutputBytes: number;
+                maximumSpendUsd: number | null;
+            };
+            createdAt: string;
+            metadata: Record<string, unknown>;
+            contentHash: string;
+            profileEvaluation?: {
+                profileId: string;
+                sourceRevision: string;
+                harnessRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                catalogHash: string;
+                definitionId: string;
+                definitionHash: string;
+                target: {
+                    kind: "profile";
+                } | {
+                    kind: "workflow";
+                    workflowId: string;
+                } | {
+                    kind: "skill";
+                    skillPath: string;
+                } | {
+                    kind: "agent_action";
+                    actionId: string;
+                };
+                environmentHash: string;
+            } | undefined;
+        };
+        policySnapshot: {
+            modelId: string;
+            provider: string;
+            upstreamModelId: string;
+            configurationHash: string;
+        } | null;
+    }>;
+    cancel(id: string, signal?: AbortSignal): Promise<{
+        summary: {
+            schemaVersion: "openpond.modelTasksetRunSummary.v1";
+            id: string;
+            revision: number;
+            teamId: string;
+            modelProjectId: string | null;
+            operationId: string;
+            taskset: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            policyKind: "fixture" | "hosted_chat" | "hosted_harness";
+            manifestHash: string;
+            status: "failed" | "completed" | "cancelled" | "running" | "cancelling" | "queued";
+            totalCount: number;
+            counts: {
+                pending: number;
+                running: number;
+                completed: number;
+                failed: number;
+                cancelled: number;
+            };
+            createdAt: string;
+            startedAt: string | null;
+            completedAt: string | null;
+            cleanupComplete: boolean;
+            resultAvailable: boolean;
+            score: number | null;
+            metricName: string;
+            error: {
+                code: string;
+                message: string;
+            } | null;
+            name?: string | undefined;
+            project?: {
+                id: string;
+                revision: number;
+                contentHash: string;
+                targetId: string | null;
+            } | undefined;
+        };
+        request: {
+            schemaVersion: "openpond.modelTasksetRunRequest.v1";
+            operationId: string;
+            teamId: string;
+            modelProjectId: string | null;
+            taskset: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            policy: {
+                kind: "hosted_chat";
+                modelId: string;
+                maxOutputTokens: number;
+                temperature: number;
+                topP: number;
+                messages?: {
+                    role: "user" | "system" | "assistant";
+                    content: string;
+                }[] | undefined;
+            } | {
+                kind: "hosted_harness";
+                modelId: string;
+                profileRepositoryId: string;
+                source: {
+                    profileId: string;
+                    sourceRevision: string;
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    catalogHash: string;
+                    definitionId: string;
+                    definitionHash: string;
+                    target: {
+                        kind: "profile";
+                    } | {
+                        kind: "workflow";
+                        workflowId: string;
+                    } | {
+                        kind: "skill";
+                        skillPath: string;
+                    } | {
+                        kind: "agent_action";
+                        actionId: string;
+                    };
+                    environmentHash: string;
+                };
+                modelConfigurationHash: string;
+                packageHash: string;
+                reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
+            } | {
+                kind: "fixture";
+            };
+            population: {
+                receiptId: string;
+                taskId: string;
+                seed: string;
+                fixtureId: string | null;
+            }[];
+            name?: string | undefined;
+            project?: {
+                id: string;
+                revision: number;
+                contentHash: string;
+                targetId: string | null;
+            } | undefined;
+        };
+        manifest: {
+            schemaVersion: "openpond.tasksetRunManifest.v1";
+            id: string;
+            tasksetRelease: {
+                id: string;
+                contentHash: string;
+            };
+            packageHash: string;
+            execution: {
+                kind: "taskset";
+                environmentRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                verifierSetRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                policyHash: string;
+            } | {
+                kind: "harness";
+                harnessRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+            };
+            policy: {
+                kind: "model";
+                model: {
+                    provider: string;
+                    model: string;
+                    revision: string | null;
+                    artifactHash: string | null;
+                    tokenizerRevision: string | null;
+                    chatTemplateHash: string | null;
+                };
+                configurationHash: string;
+            } | {
+                kind: "fixture";
+            };
+            gradingRole: "evaluation";
+            metricPolicy: {
+                schemaVersion: "openpond.tasksetMetricPolicy.v1";
+                primaryMetric: string;
+                aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
+                missingReward: "zero" | "exclude";
+                customAggregator: {
+                    module: string;
+                    exportName: string;
+                    contentHash: string;
+                    timeoutMs: number;
+                    networkPolicy: "none";
+                } | null;
+                taskWeights?: Record<string, number> | undefined;
+            };
+            population: {
+                receiptId: string;
+                taskId: string;
+                seed: string;
+                fixtureId: string | null;
+            }[];
+            runtimeTarget: {
+                adapterId: string;
+                placement: "local" | "remote" | "colocated" | "provider_native";
+                runtimeVersion: string;
+                capabilityReceipt: string;
+            };
+            limits: {
+                maxTurns: number;
+                timeoutMs: number;
+                maxOutputBytes: number;
+                maximumSpendUsd: number | null;
+            };
+            createdAt: string;
+            metadata: Record<string, unknown>;
+            contentHash: string;
+            profileEvaluation?: {
+                profileId: string;
+                sourceRevision: string;
+                harnessRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                catalogHash: string;
+                definitionId: string;
+                definitionHash: string;
+                target: {
+                    kind: "profile";
+                } | {
+                    kind: "workflow";
+                    workflowId: string;
+                } | {
+                    kind: "skill";
+                    skillPath: string;
+                } | {
+                    kind: "agent_action";
+                    actionId: string;
+                };
+                environmentHash: string;
+            } | undefined;
+        };
+        policySnapshot: {
+            modelId: string;
+            provider: string;
+            upstreamModelId: string;
+            configurationHash: string;
+        } | null;
+    }>;
+    /** Retry intentionally creates another execution; a transport retry reuses the same operation id. */
+    retry(id: string, operationId: string, signal?: AbortSignal): Promise<{
+        summary: {
+            schemaVersion: "openpond.modelTasksetRunSummary.v1";
+            id: string;
+            revision: number;
+            teamId: string;
+            modelProjectId: string | null;
+            operationId: string;
+            taskset: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            policyKind: "fixture" | "hosted_chat" | "hosted_harness";
+            manifestHash: string;
+            status: "failed" | "completed" | "cancelled" | "running" | "cancelling" | "queued";
+            totalCount: number;
+            counts: {
+                pending: number;
+                running: number;
+                completed: number;
+                failed: number;
+                cancelled: number;
+            };
+            createdAt: string;
+            startedAt: string | null;
+            completedAt: string | null;
+            cleanupComplete: boolean;
+            resultAvailable: boolean;
+            score: number | null;
+            metricName: string;
+            error: {
+                code: string;
+                message: string;
+            } | null;
+            name?: string | undefined;
+            project?: {
+                id: string;
+                revision: number;
+                contentHash: string;
+                targetId: string | null;
+            } | undefined;
+        };
+        request: {
+            schemaVersion: "openpond.modelTasksetRunRequest.v1";
+            operationId: string;
+            teamId: string;
+            modelProjectId: string | null;
+            taskset: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            policy: {
+                kind: "hosted_chat";
+                modelId: string;
+                maxOutputTokens: number;
+                temperature: number;
+                topP: number;
+                messages?: {
+                    role: "user" | "system" | "assistant";
+                    content: string;
+                }[] | undefined;
+            } | {
+                kind: "hosted_harness";
+                modelId: string;
+                profileRepositoryId: string;
+                source: {
+                    profileId: string;
+                    sourceRevision: string;
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    catalogHash: string;
+                    definitionId: string;
+                    definitionHash: string;
+                    target: {
+                        kind: "profile";
+                    } | {
+                        kind: "workflow";
+                        workflowId: string;
+                    } | {
+                        kind: "skill";
+                        skillPath: string;
+                    } | {
+                        kind: "agent_action";
+                        actionId: string;
+                    };
+                    environmentHash: string;
+                };
+                modelConfigurationHash: string;
+                packageHash: string;
+                reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
+            } | {
+                kind: "fixture";
+            };
+            population: {
+                receiptId: string;
+                taskId: string;
+                seed: string;
+                fixtureId: string | null;
+            }[];
+            name?: string | undefined;
+            project?: {
+                id: string;
+                revision: number;
+                contentHash: string;
+                targetId: string | null;
+            } | undefined;
+        };
+        manifest: {
+            schemaVersion: "openpond.tasksetRunManifest.v1";
+            id: string;
+            tasksetRelease: {
+                id: string;
+                contentHash: string;
+            };
+            packageHash: string;
+            execution: {
+                kind: "taskset";
+                environmentRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                verifierSetRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                policyHash: string;
+            } | {
+                kind: "harness";
+                harnessRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+            };
+            policy: {
+                kind: "model";
+                model: {
+                    provider: string;
+                    model: string;
+                    revision: string | null;
+                    artifactHash: string | null;
+                    tokenizerRevision: string | null;
+                    chatTemplateHash: string | null;
+                };
+                configurationHash: string;
+            } | {
+                kind: "fixture";
+            };
+            gradingRole: "evaluation";
+            metricPolicy: {
+                schemaVersion: "openpond.tasksetMetricPolicy.v1";
+                primaryMetric: string;
+                aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
+                missingReward: "zero" | "exclude";
+                customAggregator: {
+                    module: string;
+                    exportName: string;
+                    contentHash: string;
+                    timeoutMs: number;
+                    networkPolicy: "none";
+                } | null;
+                taskWeights?: Record<string, number> | undefined;
+            };
+            population: {
+                receiptId: string;
+                taskId: string;
+                seed: string;
+                fixtureId: string | null;
+            }[];
+            runtimeTarget: {
+                adapterId: string;
+                placement: "local" | "remote" | "colocated" | "provider_native";
+                runtimeVersion: string;
+                capabilityReceipt: string;
+            };
+            limits: {
+                maxTurns: number;
+                timeoutMs: number;
+                maxOutputBytes: number;
+                maximumSpendUsd: number | null;
+            };
+            createdAt: string;
+            metadata: Record<string, unknown>;
+            contentHash: string;
+            profileEvaluation?: {
+                profileId: string;
+                sourceRevision: string;
+                harnessRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                catalogHash: string;
+                definitionId: string;
+                definitionHash: string;
+                target: {
+                    kind: "profile";
+                } | {
+                    kind: "workflow";
+                    workflowId: string;
+                } | {
+                    kind: "skill";
+                    skillPath: string;
+                } | {
+                    kind: "agent_action";
+                    actionId: string;
+                };
+                environmentHash: string;
+            } | undefined;
+        };
+        policySnapshot: {
+            modelId: string;
+            provider: string;
+            upstreamModelId: string;
+            configurationHash: string;
+        } | null;
+    }>;
+    score(value: ExperimentScoringRequest, signal?: AbortSignal): Promise<{
+        schemaVersion: "openpond.experimentScoringPass.v1";
+        id: string;
+        teamId: string;
+        revision: number;
+        request: {
+            operationId: string;
+            execution: {
+                id: string;
+                contentHash: string;
+            };
+            graders: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            }[];
+            maximumCostUsd: number;
+            mappings?: {
+                graderId: string;
+                fields: {
+                    destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    path: string[];
+                    source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    sourcePath: string[];
+                }[];
+            }[] | undefined;
+        };
+        graders: {
+            id: string;
+            version: string;
+            contentHash: string;
+            feedbackKey: string;
+            release: {
+                id: string;
+                revision: number;
+                contentHash: string;
+            } | null;
+            mappings?: {
+                destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                path: string[];
+                source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                sourcePath: string[];
+            }[] | undefined;
+        }[];
+        status: "failed" | "completed" | "cancelled" | "running" | "cancelling" | "queued";
+        totalCount: number;
+        counts: {
+            pending: number;
+            running: number;
+            scored: number;
+            unavailable: number;
+            failed: number;
+            cancelled: number;
+        };
+        resultAvailable: boolean;
+        createdAt: string;
+        startedAt: string | null;
+        completedAt: string | null;
+        error: {
+            code: string;
+            message: string;
+        } | null;
+        contentHash: string;
+    }>;
+    scoringPass(id: string, signal?: AbortSignal): Promise<{
+        schemaVersion: "openpond.experimentScoringPass.v1";
+        id: string;
+        teamId: string;
+        revision: number;
+        request: {
+            operationId: string;
+            execution: {
+                id: string;
+                contentHash: string;
+            };
+            graders: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            }[];
+            maximumCostUsd: number;
+            mappings?: {
+                graderId: string;
+                fields: {
+                    destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    path: string[];
+                    source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    sourcePath: string[];
+                }[];
+            }[] | undefined;
+        };
+        graders: {
+            id: string;
+            version: string;
+            contentHash: string;
+            feedbackKey: string;
+            release: {
+                id: string;
+                revision: number;
+                contentHash: string;
+            } | null;
+            mappings?: {
+                destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                path: string[];
+                source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                sourcePath: string[];
+            }[] | undefined;
+        }[];
+        status: "failed" | "completed" | "cancelled" | "running" | "cancelling" | "queued";
+        totalCount: number;
+        counts: {
+            pending: number;
+            running: number;
+            scored: number;
+            unavailable: number;
+            failed: number;
+            cancelled: number;
+        };
+        resultAvailable: boolean;
+        createdAt: string;
+        startedAt: string | null;
+        completedAt: string | null;
+        error: {
+            code: string;
+            message: string;
+        } | null;
+        contentHash: string;
+    }>;
+    scoringPasses(executionId: string, options?: {
+        afterId?: string;
+        signal?: AbortSignal;
+    }): Promise<{
+        items: {
+            schemaVersion: "openpond.experimentScoringPass.v1";
+            id: string;
+            teamId: string;
+            revision: number;
+            request: {
+                operationId: string;
+                execution: {
+                    id: string;
+                    contentHash: string;
+                };
+                graders: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                }[];
+                maximumCostUsd: number;
+                mappings?: {
+                    graderId: string;
+                    fields: {
+                        destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                        path: string[];
+                        source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                        sourcePath: string[];
+                    }[];
+                }[] | undefined;
+            };
+            graders: {
+                id: string;
+                version: string;
+                contentHash: string;
+                feedbackKey: string;
+                release: {
+                    id: string;
+                    revision: number;
+                    contentHash: string;
+                } | null;
+                mappings?: {
+                    destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    path: string[];
+                    source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    sourcePath: string[];
+                }[] | undefined;
+            }[];
+            status: "failed" | "completed" | "cancelled" | "running" | "cancelling" | "queued";
+            totalCount: number;
+            counts: {
+                pending: number;
+                running: number;
+                scored: number;
+                unavailable: number;
+                failed: number;
+                cancelled: number;
+            };
+            resultAvailable: boolean;
+            createdAt: string;
+            startedAt: string | null;
+            completedAt: string | null;
+            error: {
+                code: string;
+                message: string;
+            } | null;
+            contentHash: string;
+        }[];
+        nextCursor: string | null;
+    }>;
+    cancelScoringPass(id: string, signal?: AbortSignal): Promise<{
+        schemaVersion: "openpond.experimentScoringPass.v1";
+        id: string;
+        teamId: string;
+        revision: number;
+        request: {
+            operationId: string;
+            execution: {
+                id: string;
+                contentHash: string;
+            };
+            graders: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            }[];
+            maximumCostUsd: number;
+            mappings?: {
+                graderId: string;
+                fields: {
+                    destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    path: string[];
+                    source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    sourcePath: string[];
+                }[];
+            }[] | undefined;
+        };
+        graders: {
+            id: string;
+            version: string;
+            contentHash: string;
+            feedbackKey: string;
+            release: {
+                id: string;
+                revision: number;
+                contentHash: string;
+            } | null;
+            mappings?: {
+                destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                path: string[];
+                source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                sourcePath: string[];
+            }[] | undefined;
+        }[];
+        status: "failed" | "completed" | "cancelled" | "running" | "cancelling" | "queued";
+        totalCount: number;
+        counts: {
+            pending: number;
+            running: number;
+            scored: number;
+            unavailable: number;
+            failed: number;
+            cancelled: number;
+        };
+        resultAvailable: boolean;
+        createdAt: string;
+        startedAt: string | null;
+        completedAt: string | null;
+        error: {
+            code: string;
+            message: string;
+        } | null;
+        contentHash: string;
+    }>;
+    scoringResult(id: string, signal?: AbortSignal): Promise<{
+        executionId: string;
+        passId: string;
+        manifest: {
+            schemaVersion: "openpond.experimentManifest.v1";
+            id: string;
+            name: string;
+            teamId: string;
+            operationId: string;
+            maximumCostUsd: number | null;
+            dataset: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            target: {
+                kind: "model";
+                modelId: string;
+                configurationHash: string;
+            } | {
+                kind: "harness";
+                source: {
+                    profileId: string;
+                    sourceRevision: string;
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    catalogHash: string;
+                    definitionId: string;
+                    definitionHash: string;
+                    target: {
+                        kind: "profile";
+                    } | {
+                        kind: "workflow";
+                        workflowId: string;
+                    } | {
+                        kind: "skill";
+                        skillPath: string;
+                    } | {
+                        kind: "agent_action";
+                        actionId: string;
+                    };
+                    environmentHash: string;
+                };
+            } | {
+                kind: "fixture";
+                configurationHash: string;
+            };
+            evaluators: {
+                release: {
+                    id: string;
+                    contentHash: string;
+                    revision: string | number;
+                };
+                feedbackKey: string;
+                output: "boolean" | "score" | "category";
+                categories: string[];
+                configurationHash?: string | undefined;
+            }[];
+            population: {
+                caseId: string;
+                seed: string;
+                fixtureId: string | null;
+            }[];
+            createdAt: string;
+            contentHash: string;
+            execution?: {
+                packageHash: string;
+                runtimeTargetHash: string;
+                metricPolicyHash: string;
+            } | undefined;
+            lineage?: {
+                definition: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                } | null;
+                execution: {
+                    id: string;
+                    contentHash: string;
+                };
+                scoringPassId: string | null;
+            } | undefined;
+        };
+        result: {
+            schemaVersion: "openpond.experimentResult.v1";
+            manifest: {
+                id: string;
+                contentHash: string;
+            };
+            status: "failed" | "completed" | "cancelled";
+            cases: {
+                identity: {
+                    caseId: string;
+                    seed: string;
+                    fixtureId: string | null;
+                };
+                status: "failed" | "completed" | "cancelled" | "unavailable";
+                output: unknown;
+                error: {
+                    code: string;
+                    message: string;
+                } | null;
+                feedback: {
+                    feedbackKey: string;
+                    evaluator: {
+                        id: string;
+                        contentHash: string;
+                        revision: string | number;
+                    };
+                    status: "pending" | "failed" | "scored" | "unavailable";
+                    value: string | number | boolean | null;
+                    reasoning: string | null;
+                    evidenceRefs: {
+                        id: string;
+                        contentHash: string;
+                        mediaType: string | null;
+                        sizeBytes: number | null;
+                    }[];
+                    passed?: boolean | null | undefined;
+                }[];
+                usage: {
+                    inputTokens: number | null;
+                    outputTokens: number | null;
+                    totalTokens: number | null;
+                    costUsd: number | null;
+                    latencyMs: number | null;
+                };
+                traceRef: {
+                    id: string;
+                    contentHash: string;
+                    mediaType: string | null;
+                    sizeBytes: number | null;
+                } | null;
+                startedAt: string | null;
+                completedAt: string | null;
+            }[];
+            completedAt: string;
+            contentHash: string;
+        };
+    }>;
+    compare(baselineId: string, candidateId: string, signal?: AbortSignal): Promise<{
+        baseline: {
+            executionId: string;
+            passId: string | null;
+            manifest: {
+                schemaVersion: "openpond.experimentManifest.v1";
+                id: string;
+                name: string;
+                teamId: string;
+                operationId: string;
+                maximumCostUsd: number | null;
+                dataset: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                target: {
+                    kind: "model";
+                    modelId: string;
+                    configurationHash: string;
+                } | {
+                    kind: "harness";
+                    source: {
+                        profileId: string;
+                        sourceRevision: string;
+                        harnessRelease: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        catalogHash: string;
+                        definitionId: string;
+                        definitionHash: string;
+                        target: {
+                            kind: "profile";
+                        } | {
+                            kind: "workflow";
+                            workflowId: string;
+                        } | {
+                            kind: "skill";
+                            skillPath: string;
+                        } | {
+                            kind: "agent_action";
+                            actionId: string;
+                        };
+                        environmentHash: string;
+                    };
+                } | {
+                    kind: "fixture";
+                    configurationHash: string;
+                };
+                evaluators: {
+                    release: {
+                        id: string;
+                        contentHash: string;
+                        revision: string | number;
+                    };
+                    feedbackKey: string;
+                    output: "boolean" | "score" | "category";
+                    categories: string[];
+                    configurationHash?: string | undefined;
+                }[];
+                population: {
+                    caseId: string;
+                    seed: string;
+                    fixtureId: string | null;
+                }[];
+                createdAt: string;
+                contentHash: string;
+                execution?: {
+                    packageHash: string;
+                    runtimeTargetHash: string;
+                    metricPolicyHash: string;
+                } | undefined;
+                lineage?: {
+                    definition: {
+                        id: string;
+                        contentHash: string;
+                        revision: number;
+                    } | null;
+                    execution: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    scoringPassId: string | null;
+                } | undefined;
+            };
+            result: {
+                schemaVersion: "openpond.experimentResult.v1";
+                manifest: {
+                    id: string;
+                    contentHash: string;
+                };
+                status: "failed" | "completed" | "cancelled";
+                cases: {
+                    identity: {
+                        caseId: string;
+                        seed: string;
+                        fixtureId: string | null;
+                    };
+                    status: "failed" | "completed" | "cancelled" | "unavailable";
+                    output: unknown;
+                    error: {
+                        code: string;
+                        message: string;
+                    } | null;
+                    feedback: {
+                        feedbackKey: string;
+                        evaluator: {
+                            id: string;
+                            contentHash: string;
+                            revision: string | number;
+                        };
+                        status: "pending" | "failed" | "scored" | "unavailable";
+                        value: string | number | boolean | null;
+                        reasoning: string | null;
+                        evidenceRefs: {
+                            id: string;
+                            contentHash: string;
+                            mediaType: string | null;
+                            sizeBytes: number | null;
+                        }[];
+                        passed?: boolean | null | undefined;
+                    }[];
+                    usage: {
+                        inputTokens: number | null;
+                        outputTokens: number | null;
+                        totalTokens: number | null;
+                        costUsd: number | null;
+                        latencyMs: number | null;
+                    };
+                    traceRef: {
+                        id: string;
+                        contentHash: string;
+                        mediaType: string | null;
+                        sizeBytes: number | null;
+                    } | null;
+                    startedAt: string | null;
+                    completedAt: string | null;
+                }[];
+                completedAt: string;
+                contentHash: string;
+            };
+        };
+        candidate: {
+            executionId: string;
+            passId: string | null;
+            manifest: {
+                schemaVersion: "openpond.experimentManifest.v1";
+                id: string;
+                name: string;
+                teamId: string;
+                operationId: string;
+                maximumCostUsd: number | null;
+                dataset: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                target: {
+                    kind: "model";
+                    modelId: string;
+                    configurationHash: string;
+                } | {
+                    kind: "harness";
+                    source: {
+                        profileId: string;
+                        sourceRevision: string;
+                        harnessRelease: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        catalogHash: string;
+                        definitionId: string;
+                        definitionHash: string;
+                        target: {
+                            kind: "profile";
+                        } | {
+                            kind: "workflow";
+                            workflowId: string;
+                        } | {
+                            kind: "skill";
+                            skillPath: string;
+                        } | {
+                            kind: "agent_action";
+                            actionId: string;
+                        };
+                        environmentHash: string;
+                    };
+                } | {
+                    kind: "fixture";
+                    configurationHash: string;
+                };
+                evaluators: {
+                    release: {
+                        id: string;
+                        contentHash: string;
+                        revision: string | number;
+                    };
+                    feedbackKey: string;
+                    output: "boolean" | "score" | "category";
+                    categories: string[];
+                    configurationHash?: string | undefined;
+                }[];
+                population: {
+                    caseId: string;
+                    seed: string;
+                    fixtureId: string | null;
+                }[];
+                createdAt: string;
+                contentHash: string;
+                execution?: {
+                    packageHash: string;
+                    runtimeTargetHash: string;
+                    metricPolicyHash: string;
+                } | undefined;
+                lineage?: {
+                    definition: {
+                        id: string;
+                        contentHash: string;
+                        revision: number;
+                    } | null;
+                    execution: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    scoringPassId: string | null;
+                } | undefined;
+            };
+            result: {
+                schemaVersion: "openpond.experimentResult.v1";
+                manifest: {
+                    id: string;
+                    contentHash: string;
+                };
+                status: "failed" | "completed" | "cancelled";
+                cases: {
+                    identity: {
+                        caseId: string;
+                        seed: string;
+                        fixtureId: string | null;
+                    };
+                    status: "failed" | "completed" | "cancelled" | "unavailable";
+                    output: unknown;
+                    error: {
+                        code: string;
+                        message: string;
+                    } | null;
+                    feedback: {
+                        feedbackKey: string;
+                        evaluator: {
+                            id: string;
+                            contentHash: string;
+                            revision: string | number;
+                        };
+                        status: "pending" | "failed" | "scored" | "unavailable";
+                        value: string | number | boolean | null;
+                        reasoning: string | null;
+                        evidenceRefs: {
+                            id: string;
+                            contentHash: string;
+                            mediaType: string | null;
+                            sizeBytes: number | null;
+                        }[];
+                        passed?: boolean | null | undefined;
+                    }[];
+                    usage: {
+                        inputTokens: number | null;
+                        outputTokens: number | null;
+                        totalTokens: number | null;
+                        costUsd: number | null;
+                        latencyMs: number | null;
+                    };
+                    traceRef: {
+                        id: string;
+                        contentHash: string;
+                        mediaType: string | null;
+                        sizeBytes: number | null;
+                    } | null;
+                    startedAt: string | null;
+                    completedAt: string | null;
+                }[];
+                completedAt: string;
+                contentHash: string;
+            };
+        };
+        comparison: import("@openpond/evals/experiments").ExperimentComparison;
+    }>;
+    private definition;
+    private executionDetails;
+    private pass;
+    private request;
+}
+//# sourceMappingURL=experiment-client.d.ts.map