npm - openpond-sdk - Versions diffs - 0.6.7 → 0.7.0 - Mend

openpond-sdk 0.6.7 → 0.7.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (86) hide show

package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts CHANGED Viewed

@@ -4,7 +4,7 @@ import { type TasksetPackage } from "./taskset-package-contracts.js";
 /** Use the same validation, file pinning and release projection as Desktop.
  * Compilation neither executes models/graders nor selects the resulting release.
  * The host supplies its retained preparation and later commits Model/draft CAS. */
-export declare function compileModelTasksetDraftWorkspace(input: {
+export declare function compileTasksetDraftWorkspace(input: {
     workspace: TasksetDraftWorkspace;
     preparation: ModelTasksetDraftPreparation | null;
     adapterId: string;
@@ -24,7 +24,7 @@ export declare function compileModelTasksetDraftWorkspace(input: {
         };
         environment: {
             protocolVersion: "openpond.environment.v1";
-            kind: "text" | "work" | "agent" | "custom_program";
+            kind: "agent" | "work" | "custom_program" | "text";
             entrypoint: string;
             stateful: boolean;
             deterministicSeeds: boolean;
@@ -85,7 +85,7 @@ export declare function compileModelTasksetDraftWorkspace(input: {
             hardGate: boolean;
             rewardEligible: boolean;
             privileged: boolean;
-            kind: "content" | "state" | "schema" | "artifact" | "runtime_event";
+            kind: "content" | "schema" | "runtime_event" | "state" | "artifact";
             config: Record<string, unknown>;
         } | {
             id: string;
@@ -179,7 +179,7 @@ export declare function compileModelTasksetDraftWorkspace(input: {
         revision: number;
         contract: {
             protocolVersion: "openpond.environment.v1";
-            kind: "text" | "work" | "agent" | "custom_program";
+            kind: "agent" | "work" | "custom_program" | "text";
             entrypoint: string;
             stateful: boolean;
             deterministicSeeds: boolean;
@@ -230,7 +230,7 @@ export declare function compileModelTasksetDraftWorkspace(input: {
             hardGate: boolean;
             rewardEligible: boolean;
             privileged: boolean;
-            kind: "content" | "state" | "schema" | "artifact" | "runtime_event";
+            kind: "content" | "schema" | "runtime_event" | "state" | "artifact";
             config: Record<string, unknown>;
         } | {
             id: string;
@@ -347,7 +347,7 @@ export declare function compileModelTasksetDraftWorkspace(input: {
                 };
                 environment: {
                     protocolVersion: "openpond.environment.v1";
-                    kind: "text" | "work" | "agent" | "custom_program";
+                    kind: "agent" | "work" | "custom_program" | "text";
                     entrypoint: string;
                     stateful: boolean;
                     deterministicSeeds: boolean;
@@ -395,45 +395,1051 @@ export declare function compileModelTasksetDraftWorkspace(input: {
             }[] | undefined;
             contextRequirements?: string[] | undefined;
         };
+        assets: {
+            schemaVersion: "openpond.learningTextAsset.v1";
+            id: string;
+            revision: 1;
+            asset: {
+                id: string;
+                path: string;
+                contentHash: string;
+                sizeBytes: number;
+                mediaType: string;
+                visibility: "policy" | "verifier" | "host_private";
+            };
+            text: string;
+            contentHash: string;
+        }[];
+        rewards: {
+            schemaVersion: "openpond.rewardRelease.v1";
+            id: string;
+            revision: number;
+            name: string;
+            description: string;
+            implementation: {
+                kind: "content" | "schema" | "runtime_event" | "state" | "artifact";
+                config: Record<string, unknown>;
+            } | {
+                kind: "model_judge";
+                rubricRef: {
+                    id: string;
+                    path: string;
+                    contentHash: string;
+                    sizeBytes: number;
+                    mediaType: string;
+                    visibility: "policy" | "verifier" | "host_private";
+                };
+                calibrationStatus: "failed" | "pending" | "passed";
+                model?: {
+                    providerId: string;
+                    modelId: string;
+                    revision: string | null;
+                } | undefined;
+                temperature?: number | undefined;
+            } | {
+                timeoutMs: number;
+                networkPolicy: "none";
+                kind: "custom_verifier";
+                verifierRef: {
+                    id: string;
+                    path: string;
+                    contentHash: string;
+                    sizeBytes: number;
+                    mediaType: string;
+                    visibility: "policy" | "verifier" | "host_private";
+                };
+                exportName?: string | undefined;
+                runtime?: "isolated_javascript" | "sandbox_process" | undefined;
+            } | {
+                kind: "human";
+                rubricRef: {
+                    id: string;
+                    path: string;
+                    contentHash: string;
+                    sizeBytes: number;
+                    mediaType: string;
+                    visibility: "policy" | "verifier" | "host_private";
+                };
+                reviewerRole: string;
+            } | {
+                kind: "learned_model";
+                modelVersion: {
+                    id: string;
+                    contentHash: string;
+                };
+                inputContract: {
+                    id: string;
+                    path: string;
+                    contentHash: string;
+                    sizeBytes: number;
+                    mediaType: string;
+                    visibility: "policy" | "verifier" | "host_private";
+                };
+            };
+            rawScore: {
+                minimum: number;
+                maximum: number;
+            };
+            assets: {
+                id: string;
+                path: string;
+                contentHash: string;
+                sizeBytes: number;
+                mediaType: string;
+                visibility: "policy" | "verifier" | "host_private";
+            }[];
+            contentHash: string;
+            fixtureSetRef?: {
+                id: string;
+                path: string;
+                contentHash: string;
+                sizeBytes: number;
+                mediaType: string;
+                visibility: "policy" | "verifier" | "host_private";
+            } | undefined;
+            calibrationCheckRef?: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            } | undefined;
+        }[];
         rewardBinding: {
             schemaVersion: "openpond.rewardBinding.v1";
             id: string;
             revision: number;
-            sources: {
-                graderId: string;
-                reward: {
-                    id: string;
-                    contentHash: string;
-                    revision: number;
+            sources: {
+                graderId: string;
+                reward: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                role: "evaluation" | "training";
+                normalization: {
+                    kind: "identity";
+                } | {
+                    kind: "linear";
+                    minimum: number;
+                    maximum: number;
+                    direction: "higher" | "lower";
+                };
+                weight: number;
+                required: boolean;
+                hardGate: boolean;
+                privileged: boolean;
+                fixtureRefs: {
+                    id: string;
+                    contentHash: string;
+                }[];
+            }[];
+            aggregation: "weighted_mean";
+            unscorable: "exclude_optional_require_all_required";
+            contentHash: string;
+            name?: string | undefined;
+            description?: string | undefined;
+            recipeRef?: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            } | undefined;
+        };
+        execution?: {
+            environment: {
+                schemaVersion: "openpond.environmentRelease.v1";
+                id: string;
+                revision: number;
+                contract: {
+                    protocolVersion: "openpond.environment.v1";
+                    kind: "agent" | "work" | "custom_program" | "text";
+                    entrypoint: string;
+                    stateful: boolean;
+                    deterministicSeeds: boolean;
+                    lifecycle: ("create" | "reset" | "step" | "collect" | "destroy")[];
+                    networkPolicy: "none" | "declared_read_only" | "declared_scoped";
+                    defaultTimeoutMs: number;
+                };
+                actionSchemaRef: {
+                    id: string;
+                    path: string;
+                    contentHash: string;
+                    sizeBytes: number;
+                    mediaType: string;
+                    visibility: "policy" | "verifier" | "host_private";
+                } | null;
+                observationSchemaRef: {
+                    id: string;
+                    path: string;
+                    contentHash: string;
+                    sizeBytes: number;
+                    mediaType: string;
+                    visibility: "policy" | "verifier" | "host_private";
+                } | null;
+                stateSchemaRef: {
+                    id: string;
+                    path: string;
+                    contentHash: string;
+                    sizeBytes: number;
+                    mediaType: string;
+                    visibility: "policy" | "verifier" | "host_private";
+                } | null;
+                artifactCollection: {
+                    maxArtifacts: number;
+                    maxTotalBytes: number;
+                };
+                adapterConformanceHashes: Record<string, string>;
+                metadata: Record<string, unknown>;
+                contentHash: string;
+            };
+            verifierSet: {
+                schemaVersion: "openpond.verifierSetRelease.v1";
+                id: string;
+                revision: number;
+                graders: ({
+                    id: string;
+                    version: string;
+                    weight: number;
+                    hardGate: boolean;
+                    rewardEligible: boolean;
+                    privileged: boolean;
+                    kind: "content" | "schema" | "runtime_event" | "state" | "artifact";
+                    config: Record<string, unknown>;
+                } | {
+                    id: string;
+                    version: string;
+                    weight: number;
+                    hardGate: boolean;
+                    rewardEligible: boolean;
+                    privileged: boolean;
+                    kind: "model_judge";
+                    rubricRef: {
+                        id: string;
+                        path: string;
+                        contentHash: string;
+                        sizeBytes: number;
+                        mediaType: string;
+                        visibility: "policy" | "verifier" | "host_private";
+                    };
+                    calibrationStatus: "failed" | "pending" | "passed";
+                    model?: {
+                        providerId: string;
+                        modelId: string;
+                        revision: string | null;
+                    } | undefined;
+                    temperature?: number | undefined;
+                } | {
+                    id: string;
+                    version: string;
+                    weight: number;
+                    hardGate: boolean;
+                    rewardEligible: boolean;
+                    privileged: boolean;
+                    kind: "custom_verifier";
+                    verifierRef: {
+                        id: string;
+                        path: string;
+                        contentHash: string;
+                        sizeBytes: number;
+                        mediaType: string;
+                        visibility: "policy" | "verifier" | "host_private";
+                    };
+                    timeoutMs: number;
+                    networkPolicy: "none";
+                    runtime?: "isolated_javascript" | "sandbox_process" | undefined;
+                    exportName?: string | undefined;
+                } | {
+                    id: string;
+                    version: string;
+                    weight: number;
+                    hardGate: boolean;
+                    rewardEligible: boolean;
+                    privileged: boolean;
+                    kind: "human";
+                    rubricRef: {
+                        id: string;
+                        path: string;
+                        contentHash: string;
+                        sizeBytes: number;
+                        mediaType: string;
+                        visibility: "policy" | "verifier" | "host_private";
+                    };
+                    reviewerRole: string;
+                })[];
+                isolation: {
+                    processBoundary: "same_process" | "isolated_process" | "container";
+                    networkPolicy: "none";
+                    defaultTimeoutMs: number;
+                };
+                calibrationReceiptRefs: {
+                    id: string;
+                    contentHash: string;
+                }[];
+                metadata: Record<string, unknown>;
+                contentHash: string;
+            };
+            javascript: {
+                schemaVersion: "openpond.javascriptEnvironment.v1";
+                id: string;
+                revision: number;
+                module: {
+                    id: string;
+                    path: string;
+                    contentHash: string;
+                    sizeBytes: number;
+                    mediaType: string;
+                    visibility: "policy" | "verifier" | "host_private";
+                };
+                tools: {
+                    name: string;
+                    description: string;
+                    inputSchema: Record<string, unknown>;
+                    inputSchemaHash: string;
+                    sideEffect: "read" | "write";
+                    timeoutMs: number;
+                }[];
+                maxSteps: number;
+                maxStateBytes: number;
+                maxObservationBytes: number;
+                operationTimeoutMs: number;
+                contentHash: string;
+                executionServices?: ({
+                    kind: "sqlite.v1";
+                    sql: {
+                        scope: "input" | "state" | "arguments" | "initialState";
+                        path: string[];
+                    };
+                    snapshot: {
+                        scope: "input" | "state" | "arguments" | "initialState";
+                        path: string[];
+                    };
+                    maxRows: number;
+                    maxResultBytes: number;
+                    id: string;
+                    operation: "step" | "collect";
+                    timeoutMs: number;
+                    toolName?: string | undefined;
+                } | {
+                    kind: "javascript.v1";
+                    source: {
+                        scope: "input" | "state" | "arguments" | "initialState";
+                        path: string[];
+                    };
+                    cases: {
+                        scope: "input" | "state" | "arguments" | "initialState";
+                        path: string[];
+                    };
+                    exportName: string;
+                    maxCases: number;
+                    maxResultBytes: number;
+                    id: string;
+                    operation: "step" | "collect";
+                    timeoutMs: number;
+                    toolName?: string | undefined;
+                })[] | undefined;
+            };
+        } | undefined;
+        instructionMode?: "per_task" | undefined;
+    } | undefined;
+    learningResources?: {
+        batch: {
+            schemaVersion: "openpond.taskBatch.v1";
+            id: string;
+            revision: 1;
+            taskDefinition: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            rewardBinding: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            purpose: "evaluation" | "supervised_training" | "reward_training";
+            examples: {
+                evidence: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                decision: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                familyKey: string;
+                inputHash: string;
+                split: "train" | "validation" | "test" | "frozen_eval";
+            }[];
+            sealedAt: string;
+            sealedBy: string;
+            contentHash: string;
+        };
+        evidence: {
+            schemaVersion: "openpond.taskEvidence.v1";
+            id: string;
+            revision: number;
+            source: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            submission: {
+                schemaVersion: "openpond.taskExample.v1";
+                sourceId: string;
+                idempotencyKey: string;
+                taskDefinition: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                exampleId: string;
+                attemptId: string;
+                occurredAt: string;
+                familyKey: string | null;
+                split: "train" | "validation" | "test" | "frozen_eval";
+                input: Record<string, import("zod").JSONType>;
+                observedOutput: Record<string, import("zod").JSONType> | null;
+                expected: Record<string, import("zod").JSONType> | null;
+                evaluatorContext: Record<string, import("zod").JSONType> | null;
+                assets: {
+                    id: string;
+                    path: string;
+                    contentHash: string;
+                    sizeBytes: number;
+                    mediaType: string;
+                    visibility: "policy" | "verifier" | "host_private";
+                }[];
+                provenance: {
+                    sourceRecordRef: string | null;
+                    mappingHash: string | null;
+                };
+                requiredOutputs?: {
+                    path: string;
+                    mediaType: string;
+                    schemaRef: {
+                        id: string;
+                        path: string;
+                        contentHash: string;
+                        sizeBytes: number;
+                        mediaType: string;
+                        visibility: "policy" | "verifier" | "host_private";
+                    } | null;
+                    maxBytes: number | null;
+                    metadata: Record<string, unknown>;
+                }[] | undefined;
+            };
+            supersedes: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            } | null;
+            correctionFeedbackId: string | null;
+            receivedAt: string;
+            contentHash: string;
+        }[];
+        decisions: {
+            schemaVersion: "openpond.taskAdmissionDecision.v1";
+            id: string;
+            revision: number;
+            evidence: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            supersedes: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            } | null;
+            actor: {
+                kind: "human" | "qualified_policy";
+                id: string;
+                policy: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                } | null;
+            };
+            evidenceValidity: "pending" | "valid" | "invalid";
+            taskAdmissibility: "rejected" | "pending" | "approved";
+            observedQuality: "failed" | "passed" | "unavailable" | "unscored";
+            targetApproval: "rejected" | "pending" | "approved" | "not_required";
+            approvedTarget: Record<string, import("zod").JSONType> | null;
+            grade: {
+                schemaVersion: "openpond.rewardComposition.v1";
+                binding: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                taskHash: string;
+                outputHash: string;
+                results: {
+                    graderId: string;
+                    reward: {
+                        id: string;
+                        contentHash: string;
+                        revision: number;
+                    };
+                    role: "evaluation" | "training";
+                    status: "failed" | "pending" | "unavailable" | "scored";
+                    rawScore: number | null;
+                    normalizedScore: number | null;
+                    passed: boolean | null;
+                    evidenceHashes: string[];
+                    message: string | null;
+                    graderEvidence?: {
+                        schemaVersion: "openpond.graderEvidence.v1";
+                        graderId: string;
+                        graderVersion: string;
+                        score: number | null;
+                        passed: boolean;
+                        rewardEligible: boolean;
+                        failureClass: "infrastructure_failure" | "cancelled" | "policy_failure" | "environment_failure" | "grader_failure" | "timeout" | null;
+                        feedback: string[];
+                        visibleEvidenceRefs: string[];
+                        privilegedEvidenceRefs: string[];
+                        contentHash: string;
+                        modelJudgeReceipt?: {
+                            schemaVersion: "openpond.modelJudgeReceipt.v1";
+                            providerId: string;
+                            modelId: string;
+                            modelRevision: string | null;
+                            responseId: string | null;
+                            requestHash: string;
+                            responseHash: string;
+                            inputTokens: number | null;
+                            outputTokens: number | null;
+                            costUsd: number | null;
+                        } | undefined;
+                    } | undefined;
+                }[];
+                training: {
+                    status: "unscorable" | "scored" | "not_configured";
+                    score: number | null;
+                    passed: boolean | null;
+                };
+                evaluation: {
+                    status: "unscorable" | "scored" | "not_configured";
+                    score: number | null;
+                    passed: boolean | null;
+                };
+                contentHash: string;
+            } | null;
+            targetGrade: {
+                schemaVersion: "openpond.rewardComposition.v1";
+                binding: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                taskHash: string;
+                outputHash: string;
+                results: {
+                    graderId: string;
+                    reward: {
+                        id: string;
+                        contentHash: string;
+                        revision: number;
+                    };
+                    role: "evaluation" | "training";
+                    status: "failed" | "pending" | "unavailable" | "scored";
+                    rawScore: number | null;
+                    normalizedScore: number | null;
+                    passed: boolean | null;
+                    evidenceHashes: string[];
+                    message: string | null;
+                    graderEvidence?: {
+                        schemaVersion: "openpond.graderEvidence.v1";
+                        graderId: string;
+                        graderVersion: string;
+                        score: number | null;
+                        passed: boolean;
+                        rewardEligible: boolean;
+                        failureClass: "infrastructure_failure" | "cancelled" | "policy_failure" | "environment_failure" | "grader_failure" | "timeout" | null;
+                        feedback: string[];
+                        visibleEvidenceRefs: string[];
+                        privilegedEvidenceRefs: string[];
+                        contentHash: string;
+                        modelJudgeReceipt?: {
+                            schemaVersion: "openpond.modelJudgeReceipt.v1";
+                            providerId: string;
+                            modelId: string;
+                            modelRevision: string | null;
+                            responseId: string | null;
+                            requestHash: string;
+                            responseHash: string;
+                            inputTokens: number | null;
+                            outputTokens: number | null;
+                            costUsd: number | null;
+                        } | undefined;
+                    } | undefined;
+                }[];
+                training: {
+                    status: "unscorable" | "scored" | "not_configured";
+                    score: number | null;
+                    passed: boolean | null;
+                };
+                evaluation: {
+                    status: "unscorable" | "scored" | "not_configured";
+                    score: number | null;
+                    passed: boolean | null;
+                };
+                contentHash: string;
+            } | null;
+            note: string;
+            decidedAt: string;
+            contentHash: string;
+        }[];
+        sources: {
+            schemaVersion: "openpond.learningSource.v1";
+            id: string;
+            revision: number;
+            name: string;
+            kind: "benchmark" | "work" | "provider" | "direct" | "opentelemetry";
+            taskDefinition: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            enabled: boolean;
+            allowedSplits: ("train" | "validation" | "test" | "frozen_eval")[];
+            mapping: {
+                schemaVersion: "openpond.taskSourceMapping.v1";
+                exampleId: string;
+                attemptId: string;
+                occurredAt: string;
+                input: string;
+                observedOutput: string | null;
+                expected: string | null;
+                evaluatorContext: string | null;
+                familyKey: string;
+                split: "train" | "validation" | "test" | "frozen_eval";
+            } | null;
+            adapterVersion: string | null;
+            contentHash: string;
+            reviewOrigin?: {
+                modelId: string;
+                packageHash: string;
+                taskset: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                batch: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                rewardBinding: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+            } | undefined;
+        }[];
+        assets: {
+            schemaVersion: "openpond.learningTextAsset.v1";
+            id: string;
+            revision: 1;
+            asset: {
+                id: string;
+                path: string;
+                contentHash: string;
+                sizeBytes: number;
+                mediaType: string;
+                visibility: "policy" | "verifier" | "host_private";
+            };
+            text: string;
+            contentHash: string;
+        }[];
+    } | undefined;
+};
+/** Model-derived authoring retains its stronger owner/source contract. */
+export declare function compileModelTasksetDraftWorkspace(input: Parameters<typeof compileTasksetDraftWorkspace>[0]): {
+    schemaVersion: "openpond.tasksetPackage.v1";
+    taskset: {
+        schemaVersion: "openpond.tasksetRelease.v2";
+        id: string;
+        revision: number;
+        policy: {
+            policyVisibleFields: string[];
+            privilegedFields: string[];
+            hiddenGraderRefs: string[];
+            connectedAppScopes: string[];
+        };
+        environment: {
+            protocolVersion: "openpond.environment.v1";
+            kind: "agent" | "work" | "custom_program" | "text";
+            entrypoint: string;
+            stateful: boolean;
+            deterministicSeeds: boolean;
+            lifecycle: ("create" | "reset" | "step" | "collect" | "destroy")[];
+            networkPolicy: "none" | "declared_read_only" | "declared_scoped";
+            defaultTimeoutMs: number;
+        };
+        tools: {
+            name: string;
+            description: string;
+            inputSchema: Record<string, unknown>;
+            inputSchemaHash: string;
+            sideEffect: "read" | "write";
+            timeoutMs: number;
+        }[];
+        capabilities: {
+            id: string;
+            required: boolean;
+            scopes: string[];
+            portability: "portable" | "host_adapter" | "local_only" | "hosted_only";
+        }[];
+        tasks: {
+            id: string;
+            clusterKey: string;
+            split: "train" | "validation" | "test" | "frozen_eval";
+            input: Record<string, unknown>;
+            expectedOutput: Record<string, unknown> | null;
+            policyVisibleContext: Record<string, unknown>;
+            privilegedContextRef: string | null;
+            artifactRefs: {
+                id: string;
+                path: string;
+                contentHash: string;
+                sizeBytes: number;
+                mediaType: string;
+                visibility: "policy" | "verifier" | "host_private";
+            }[];
+            tags: string[];
+            requiredOutputs?: {
+                path: string;
+                mediaType: string;
+                schemaRef: {
+                    id: string;
+                    path: string;
+                    contentHash: string;
+                    sizeBytes: number;
+                    mediaType: string;
+                    visibility: "policy" | "verifier" | "host_private";
+                } | null;
+                maxBytes: number | null;
+                metadata: Record<string, unknown>;
+            }[] | undefined;
+        }[];
+        graders: ({
+            id: string;
+            version: string;
+            weight: number;
+            hardGate: boolean;
+            rewardEligible: boolean;
+            privileged: boolean;
+            kind: "content" | "schema" | "runtime_event" | "state" | "artifact";
+            config: Record<string, unknown>;
+        } | {
+            id: string;
+            version: string;
+            weight: number;
+            hardGate: boolean;
+            rewardEligible: boolean;
+            privileged: boolean;
+            kind: "model_judge";
+            rubricRef: {
+                id: string;
+                path: string;
+                contentHash: string;
+                sizeBytes: number;
+                mediaType: string;
+                visibility: "policy" | "verifier" | "host_private";
+            };
+            calibrationStatus: "failed" | "pending" | "passed";
+            model?: {
+                providerId: string;
+                modelId: string;
+                revision: string | null;
+            } | undefined;
+            temperature?: number | undefined;
+        } | {
+            id: string;
+            version: string;
+            weight: number;
+            hardGate: boolean;
+            rewardEligible: boolean;
+            privileged: boolean;
+            kind: "custom_verifier";
+            verifierRef: {
+                id: string;
+                path: string;
+                contentHash: string;
+                sizeBytes: number;
+                mediaType: string;
+                visibility: "policy" | "verifier" | "host_private";
+            };
+            timeoutMs: number;
+            networkPolicy: "none";
+            runtime?: "isolated_javascript" | "sandbox_process" | undefined;
+            exportName?: string | undefined;
+        } | {
+            id: string;
+            version: string;
+            weight: number;
+            hardGate: boolean;
+            rewardEligible: boolean;
+            privileged: boolean;
+            kind: "human";
+            rubricRef: {
+                id: string;
+                path: string;
+                contentHash: string;
+                sizeBytes: number;
+                mediaType: string;
+                visibility: "policy" | "verifier" | "host_private";
+            };
+            reviewerRole: string;
+        })[];
+        metadata: Record<string, unknown>;
+        contentHash: string;
+        environmentRelease?: {
+            id: string;
+            contentHash: string;
+        } | undefined;
+        verifierSetRelease?: {
+            id: string;
+            contentHash: string;
+        } | undefined;
+        metrics?: {
+            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;
+        } | undefined;
+    };
+    environment: {
+        schemaVersion: "openpond.environmentRelease.v1";
+        id: string;
+        revision: number;
+        contract: {
+            protocolVersion: "openpond.environment.v1";
+            kind: "agent" | "work" | "custom_program" | "text";
+            entrypoint: string;
+            stateful: boolean;
+            deterministicSeeds: boolean;
+            lifecycle: ("create" | "reset" | "step" | "collect" | "destroy")[];
+            networkPolicy: "none" | "declared_read_only" | "declared_scoped";
+            defaultTimeoutMs: number;
+        };
+        actionSchemaRef: {
+            id: string;
+            path: string;
+            contentHash: string;
+            sizeBytes: number;
+            mediaType: string;
+            visibility: "policy" | "verifier" | "host_private";
+        } | null;
+        observationSchemaRef: {
+            id: string;
+            path: string;
+            contentHash: string;
+            sizeBytes: number;
+            mediaType: string;
+            visibility: "policy" | "verifier" | "host_private";
+        } | null;
+        stateSchemaRef: {
+            id: string;
+            path: string;
+            contentHash: string;
+            sizeBytes: number;
+            mediaType: string;
+            visibility: "policy" | "verifier" | "host_private";
+        } | null;
+        artifactCollection: {
+            maxArtifacts: number;
+            maxTotalBytes: number;
+        };
+        adapterConformanceHashes: Record<string, string>;
+        metadata: Record<string, unknown>;
+        contentHash: string;
+    };
+    verifierSet: {
+        schemaVersion: "openpond.verifierSetRelease.v1";
+        id: string;
+        revision: number;
+        graders: ({
+            id: string;
+            version: string;
+            weight: number;
+            hardGate: boolean;
+            rewardEligible: boolean;
+            privileged: boolean;
+            kind: "content" | "schema" | "runtime_event" | "state" | "artifact";
+            config: Record<string, unknown>;
+        } | {
+            id: string;
+            version: string;
+            weight: number;
+            hardGate: boolean;
+            rewardEligible: boolean;
+            privileged: boolean;
+            kind: "model_judge";
+            rubricRef: {
+                id: string;
+                path: string;
+                contentHash: string;
+                sizeBytes: number;
+                mediaType: string;
+                visibility: "policy" | "verifier" | "host_private";
+            };
+            calibrationStatus: "failed" | "pending" | "passed";
+            model?: {
+                providerId: string;
+                modelId: string;
+                revision: string | null;
+            } | undefined;
+            temperature?: number | undefined;
+        } | {
+            id: string;
+            version: string;
+            weight: number;
+            hardGate: boolean;
+            rewardEligible: boolean;
+            privileged: boolean;
+            kind: "custom_verifier";
+            verifierRef: {
+                id: string;
+                path: string;
+                contentHash: string;
+                sizeBytes: number;
+                mediaType: string;
+                visibility: "policy" | "verifier" | "host_private";
+            };
+            timeoutMs: number;
+            networkPolicy: "none";
+            runtime?: "isolated_javascript" | "sandbox_process" | undefined;
+            exportName?: string | undefined;
+        } | {
+            id: string;
+            version: string;
+            weight: number;
+            hardGate: boolean;
+            rewardEligible: boolean;
+            privileged: boolean;
+            kind: "human";
+            rubricRef: {
+                id: string;
+                path: string;
+                contentHash: string;
+                sizeBytes: number;
+                mediaType: string;
+                visibility: "policy" | "verifier" | "host_private";
+            };
+            reviewerRole: string;
+        })[];
+        isolation: {
+            processBoundary: "same_process" | "isolated_process" | "container";
+            networkPolicy: "none";
+            defaultTimeoutMs: number;
+        };
+        calibrationReceiptRefs: {
+            id: string;
+            contentHash: string;
+        }[];
+        metadata: Record<string, unknown>;
+        contentHash: string;
+    };
+    files: {
+        asset: {
+            id: string;
+            path: string;
+            contentHash: string;
+            sizeBytes: number;
+            mediaType: string;
+            visibility: "policy" | "verifier" | "host_private";
+        };
+        base64: string;
+    }[];
+    contentHash: string;
+    modelResources?: {
+        taskDefinition: {
+            schemaVersion: "openpond.taskDefinition.v1";
+            id: string;
+            revision: number;
+            name: string;
+            description: string;
+            instructions: string;
+            category: "custom" | "structured" | "question_answering" | "coding" | "tool_workflow";
+            familyNamespace: string;
+            inputSchema: Record<string, unknown>;
+            outputSchema: Record<string, unknown>;
+            rewardBinding: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            };
+            harness: {
+                id: string;
+                contentHash: string;
+            } | null;
+            execution: {
+                policy: {
+                    policyVisibleFields: string[];
+                    privilegedFields: string[];
+                    hiddenGraderRefs: string[];
+                    connectedAppScopes: string[];
                 };
-                role: "evaluation" | "training";
-                normalization: {
-                    kind: "identity";
-                } | {
-                    kind: "linear";
-                    minimum: number;
-                    maximum: number;
-                    direction: "higher" | "lower";
+                environment: {
+                    protocolVersion: "openpond.environment.v1";
+                    kind: "agent" | "work" | "custom_program" | "text";
+                    entrypoint: string;
+                    stateful: boolean;
+                    deterministicSeeds: boolean;
+                    lifecycle: ("create" | "reset" | "step" | "collect" | "destroy")[];
+                    networkPolicy: "none" | "declared_read_only" | "declared_scoped";
+                    defaultTimeoutMs: number;
                 };
-                weight: number;
-                required: boolean;
-                hardGate: boolean;
-                privileged: boolean;
-                fixtureRefs: {
+                tools: {
+                    name: string;
+                    description: string;
+                    inputSchema: Record<string, unknown>;
+                    inputSchemaHash: string;
+                    sideEffect: "read" | "write";
+                    timeoutMs: number;
+                }[];
+                capabilities: {
                     id: string;
-                    contentHash: string;
+                    required: boolean;
+                    scopes: string[];
+                    portability: "portable" | "host_adapter" | "local_only" | "hosted_only";
                 }[];
-            }[];
-            aggregation: "weighted_mean";
-            unscorable: "exclude_optional_require_all_required";
+                environmentRelease?: {
+                    id: string;
+                    contentHash: string;
+                } | undefined;
+                verifierSetRelease?: {
+                    id: string;
+                    contentHash: string;
+                } | undefined;
+            };
             contentHash: string;
-            name?: string | undefined;
-            description?: string | undefined;
-            recipeRef?: {
-                id: string;
-                contentHash: string;
-                revision: number;
-            } | undefined;
+            requiredOutputs?: {
+                path: string;
+                mediaType: string;
+                schemaRef: {
+                    id: string;
+                    path: string;
+                    contentHash: string;
+                    sizeBytes: number;
+                    mediaType: string;
+                    visibility: "policy" | "verifier" | "host_private";
+                } | null;
+                maxBytes: number | null;
+                metadata: Record<string, unknown>;
+            }[] | undefined;
+            contextRequirements?: string[] | undefined;
         };
         assets: {
             schemaVersion: "openpond.learningTextAsset.v1";
@@ -457,7 +1463,7 @@ export declare function compileModelTasksetDraftWorkspace(input: {
             name: string;
             description: string;
             implementation: {
-                kind: "content" | "state" | "schema" | "artifact" | "runtime_event";
+                kind: "content" | "schema" | "runtime_event" | "state" | "artifact";
                 config: Record<string, unknown>;
             } | {
                 kind: "model_judge";
@@ -543,6 +1549,46 @@ export declare function compileModelTasksetDraftWorkspace(input: {
                 revision: number;
             } | undefined;
         }[];
+        rewardBinding: {
+            schemaVersion: "openpond.rewardBinding.v1";
+            id: string;
+            revision: number;
+            sources: {
+                graderId: string;
+                reward: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                role: "evaluation" | "training";
+                normalization: {
+                    kind: "identity";
+                } | {
+                    kind: "linear";
+                    minimum: number;
+                    maximum: number;
+                    direction: "higher" | "lower";
+                };
+                weight: number;
+                required: boolean;
+                hardGate: boolean;
+                privileged: boolean;
+                fixtureRefs: {
+                    id: string;
+                    contentHash: string;
+                }[];
+            }[];
+            aggregation: "weighted_mean";
+            unscorable: "exclude_optional_require_all_required";
+            contentHash: string;
+            name?: string | undefined;
+            description?: string | undefined;
+            recipeRef?: {
+                id: string;
+                contentHash: string;
+                revision: number;
+            } | undefined;
+        };
         execution?: {
             environment: {
                 schemaVersion: "openpond.environmentRelease.v1";
@@ -550,7 +1596,7 @@ export declare function compileModelTasksetDraftWorkspace(input: {
                 revision: number;
                 contract: {
                     protocolVersion: "openpond.environment.v1";
-                    kind: "text" | "work" | "agent" | "custom_program";
+                    kind: "agent" | "work" | "custom_program" | "text";
                     entrypoint: string;
                     stateful: boolean;
                     deterministicSeeds: boolean;
@@ -601,7 +1647,7 @@ export declare function compileModelTasksetDraftWorkspace(input: {
                     hardGate: boolean;
                     rewardEligible: boolean;
                     privileged: boolean;
-                    kind: "content" | "state" | "schema" | "artifact" | "runtime_event";
+                    kind: "content" | "schema" | "runtime_event" | "state" | "artifact";
                     config: Record<string, unknown>;
                 } | {
                     id: string;
@@ -704,11 +1750,11 @@ export declare function compileModelTasksetDraftWorkspace(input: {
                 executionServices?: ({
                     kind: "sqlite.v1";
                     sql: {
-                        scope: "input" | "arguments" | "state" | "initialState";
+                        scope: "input" | "state" | "arguments" | "initialState";
                         path: string[];
                     };
                     snapshot: {
-                        scope: "input" | "arguments" | "state" | "initialState";
+                        scope: "input" | "state" | "arguments" | "initialState";
                         path: string[];
                     };
                     maxRows: number;
@@ -720,11 +1766,11 @@ export declare function compileModelTasksetDraftWorkspace(input: {
                 } | {
                     kind: "javascript.v1";
                     source: {
-                        scope: "input" | "arguments" | "state" | "initialState";
+                        scope: "input" | "state" | "arguments" | "initialState";
                         path: string[];
                     };
                     cases: {
-                        scope: "input" | "arguments" | "state" | "initialState";
+                        scope: "input" | "state" | "arguments" | "initialState";
                         path: string[];
                     };
                     exportName: string;
@@ -861,9 +1907,9 @@ export declare function compileModelTasksetDraftWorkspace(input: {
                 } | null;
             };
             evidenceValidity: "pending" | "valid" | "invalid";
-            taskAdmissibility: "pending" | "rejected" | "approved";
+            taskAdmissibility: "rejected" | "pending" | "approved";
             observedQuality: "failed" | "passed" | "unavailable" | "unscored";
-            targetApproval: "pending" | "rejected" | "approved" | "not_required";
+            targetApproval: "rejected" | "pending" | "approved" | "not_required";
             approvedTarget: Record<string, import("zod").JSONType> | null;
             grade: {
                 schemaVersion: "openpond.rewardComposition.v1";
@@ -895,7 +1941,7 @@ export declare function compileModelTasksetDraftWorkspace(input: {
                         score: number | null;
                         passed: boolean;
                         rewardEligible: boolean;
-                        failureClass: "cancelled" | "policy_failure" | "grader_failure" | "environment_failure" | "infrastructure_failure" | "timeout" | null;
+                        failureClass: "infrastructure_failure" | "cancelled" | "policy_failure" | "environment_failure" | "grader_failure" | "timeout" | null;
                         feedback: string[];
                         visibleEvidenceRefs: string[];
                         privilegedEvidenceRefs: string[];
@@ -956,7 +2002,7 @@ export declare function compileModelTasksetDraftWorkspace(input: {
                         score: number | null;
                         passed: boolean;
                         rewardEligible: boolean;
-                        failureClass: "cancelled" | "policy_failure" | "grader_failure" | "environment_failure" | "infrastructure_failure" | "timeout" | null;
+                        failureClass: "infrastructure_failure" | "cancelled" | "policy_failure" | "environment_failure" | "grader_failure" | "timeout" | null;
                         feedback: string[];
                         visibleEvidenceRefs: string[];
                         privilegedEvidenceRefs: string[];
@@ -996,7 +2042,7 @@ export declare function compileModelTasksetDraftWorkspace(input: {
             id: string;
             revision: number;
             name: string;
-            kind: "work" | "provider" | "direct" | "benchmark" | "opentelemetry";
+            kind: "benchmark" | "work" | "provider" | "direct" | "opentelemetry";
             taskDefinition: {
                 id: string;
                 contentHash: string;