@openpond/evals 0.4.2 → 0.6.0

package/dist/telemetry.js

@@ -0,0 +1,112 @@
+import { z } from "zod";
+import { MetadataSchema, ReleaseHashSchema, ReleaseIdSchema, ReleaseTimestampSchema, } from "@openpond/harness";
+export const RUN_TELEMETRY_SCHEMA_VERSION = "openpond.runTelemetryEvent.v1";
+export const METRIC_DEFINITION_SCHEMA_VERSION = "openpond.metricDefinition.v1";
+export const METRIC_OBSERVATION_SCHEMA_VERSION = "openpond.metricObservation.v1";
+export const TelemetryVisibilitySchema = z.enum([
+    "policy_visible",
+    "team_visible",
+    "host_private",
+]);
+export const TelemetrySourceSchema = z.enum([
+    "runtime",
+    "environment",
+    "grader",
+    "optimizer",
+    "control_plane",
+    "evaluation",
+]);
+export const TelemetryEventTypeSchema = z.enum([
+    "run_started",
+    "run_state_changed",
+    "rollout_group_started",
+    "attempt_completed",
+    "grader_completed",
+    "reward_composed",
+    "optimizer_step_completed",
+    "checkpoint_committed",
+    "evaluation_completed",
+    "run_completed",
+    "run_failed",
+    "cleanup_completed",
+]);
+export const RunTelemetryLineageSchema = z.object({
+    modelProjectId: ReleaseIdSchema,
+    runId: ReleaseIdSchema,
+    modelVersionId: ReleaseIdSchema.nullable(),
+    harnessReleaseHash: ReleaseHashSchema,
+    tasksetReleaseHash: ReleaseHashSchema,
+    environmentReleaseHash: ReleaseHashSchema.nullable(),
+    checkpointId: ReleaseIdSchema.nullable(),
+    step: z.number().int().nonnegative().nullable(),
+    rolloutGroupId: ReleaseIdSchema.nullable(),
+    attemptId: ReleaseIdSchema.nullable(),
+    scenarioId: ReleaseIdSchema.nullable(),
+}).strict();
+export const RunTelemetryEventSchema = z.object({
+    schemaVersion: z.literal(RUN_TELEMETRY_SCHEMA_VERSION),
+    eventId: ReleaseIdSchema,
+    sequence: z.number().int().nonnegative(),
+    occurredAt: ReleaseTimestampSchema,
+    source: TelemetrySourceSchema,
+    type: TelemetryEventTypeSchema,
+    visibility: TelemetryVisibilitySchema,
+    lineage: RunTelemetryLineageSchema,
+    attributes: MetadataSchema,
+}).strict();
+export const MetricValueTypeSchema = z.enum(["gauge", "counter", "distribution"]);
+export const MetricUnitSchema = z.enum([
+    "ratio",
+    "count",
+    "seconds",
+    "milliseconds",
+    "tokens",
+    "bytes",
+    "usd",
+    "scalar",
+]);
+export const MetricDirectionSchema = z.enum(["higher", "lower", "neutral"]);
+export const MetricAggregationSchema = z.enum(["last", "sum", "mean", "min", "max", "p50", "p95"]);
+export const MetricDefinitionSchema = z.object({
+    schemaVersion: z.literal(METRIC_DEFINITION_SCHEMA_VERSION),
+    id: ReleaseIdSchema,
+    displayName: z.string().trim().min(1).max(200),
+    description: z.string().trim().min(1).max(2_000),
+    valueType: MetricValueTypeSchema,
+    unit: MetricUnitSchema,
+    direction: MetricDirectionSchema,
+    aggregation: MetricAggregationSchema,
+    visibility: TelemetryVisibilitySchema,
+    boundedDimensions: z.array(z.string().trim().min(1).max(100)).max(32),
+}).strict();
+export const MetricObservationSchema = z.object({
+    schemaVersion: z.literal(METRIC_OBSERVATION_SCHEMA_VERSION),
+    observationId: ReleaseIdSchema,
+    metricId: ReleaseIdSchema,
+    eventId: ReleaseIdSchema,
+    sequence: z.number().int().nonnegative(),
+    observedAt: ReleaseTimestampSchema,
+    value: z.number().finite(),
+    lineage: RunTelemetryLineageSchema,
+    dimensions: z.record(z.string().trim().min(1).max(100), z.string().max(500)),
+}).strict();
+export const RunTelemetryBatchSchema = z.object({
+    schemaVersion: z.literal("openpond.runTelemetryBatch.v1"),
+    events: z.array(RunTelemetryEventSchema).max(1_000),
+    observations: z.array(MetricObservationSchema).max(10_000),
+}).strict().superRefine((batch, context) => {
+    if (batch.events.length + batch.observations.length === 0) {
+        context.addIssue({ code: "custom", message: "Telemetry batch cannot be empty." });
+    }
+    const keys = new Set();
+    for (const item of [...batch.events, ...batch.observations]) {
+        const id = item.schemaVersion === METRIC_OBSERVATION_SCHEMA_VERSION
+            ? item.observationId
+            : item.eventId;
+        const key = `${item.lineage.runId}:${item.sequence}:${id}`;
+        if (keys.has(key)) {
+            context.addIssue({ code: "custom", message: "Telemetry batch contains a duplicate idempotency key." });
+        }
+        keys.add(key);
+    }
+});

package/dist/types/artifact-verification.d.ts

@@ -0,0 +1,31 @@
+import type { ImmutableArtifactRef } from "@openpond/harness";
+import { type ArtifactManifest, type RewardComponentReceipt } from "./execution-contracts.js";
+import type { RequiredOutputContract } from "./tasksets.js";
+export type CollectedArtifact = {
+    path: string;
+    artifact: ImmutableArtifactRef | null;
+    detectedMediaType: string | null;
+    status: "collected" | "failed";
+    parseStatus?: "not_requested" | "passed" | "failed";
+    schemaStatus?: "not_requested" | "passed" | "failed";
+    errorCode?: string | null;
+    failureOwner?: "policy" | "collector" | "environment" | null;
+    evidenceRefs?: ImmutableArtifactRef[];
+    metadata?: Record<string, unknown>;
+};
+export declare function buildArtifactManifest(input: {
+    id: string;
+    attemptRef: {
+        id: string;
+        contentHash: string;
+    };
+    requiredOutputs: RequiredOutputContract[];
+    collectedArtifacts: CollectedArtifact[];
+    createdAt: string;
+    metadata?: Record<string, unknown>;
+}): ArtifactManifest;
+export declare function verifyRequiredOutputs(input: {
+    requiredOutputs: RequiredOutputContract[];
+    manifest: ArtifactManifest;
+}): RewardComponentReceipt[];
+//# sourceMappingURL=artifact-verification.d.ts.map

package/dist/types/artifact-verification.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"artifact-verification.d.ts","sourceRoot":"","sources":["../../../../src/artifact-verification.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,mBAAmB,CAAC;AAE9D,OAAO,EAGL,KAAK,gBAAgB,EAGrB,KAAK,sBAAsB,EAC5B,MAAM,0BAA0B,CAAC;AAElC,OAAO,KAAK,EAAE,sBAAsB,EAAE,MAAM,eAAe,CAAC;AAE5D,MAAM,MAAM,iBAAiB,GAAG;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,QAAQ,EAAE,oBAAoB,GAAG,IAAI,CAAC;IACtC,iBAAiB,EAAE,MAAM,GAAG,IAAI,CAAC;IACjC,MAAM,EAAE,WAAW,GAAG,QAAQ,CAAC;IAC/B,WAAW,CAAC,EAAE,eAAe,GAAG,QAAQ,GAAG,QAAQ,CAAC;IACpD,YAAY,CAAC,EAAE,eAAe,GAAG,QAAQ,GAAG,QAAQ,CAAC;IACrD,SAAS,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAC1B,YAAY,CAAC,EAAE,QAAQ,GAAG,WAAW,GAAG,aAAa,GAAG,IAAI,CAAC;IAC7D,YAAY,CAAC,EAAE,oBAAoB,EAAE,CAAC;IACtC,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CACpC,CAAC;AAEF,wBAAgB,qBAAqB,CAAC,KAAK,EAAE;IAC3C,EAAE,EAAE,MAAM,CAAC;IACX,UAAU,EAAE;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,WAAW,EAAE,MAAM,CAAA;KAAE,CAAC;IAChD,eAAe,EAAE,sBAAsB,EAAE,CAAC;IAC1C,kBAAkB,EAAE,iBAAiB,EAAE,CAAC;IACxC,SAAS,EAAE,MAAM,CAAC;IAClB,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CACpC,GAAG,gBAAgB,CAwEnB;AAED,wBAAgB,qBAAqB,CAAC,KAAK,EAAE;IAC3C,eAAe,EAAE,sBAAsB,EAAE,CAAC;IAC1C,QAAQ,EAAE,gBAAgB,CAAC;CAC5B,GAAG,sBAAsB,EAAE,CAqC3B"}

package/dist/types/builtin-benchmarks/harness-refiner.d.ts

@@ -49,6 +49,20 @@ export declare const harnessRefinerBenchmarkRelease: {
             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;
@@ -114,6 +128,14 @@ export declare const harnessRefinerBenchmarkRelease: {
     })[];
     metadata: Record<string, unknown>;
     contentHash: string;
+    environmentRelease?: {
+        id: string;
+        contentHash: string;
+    } | undefined;
+    verifierSetRelease?: {
+        id: string;
+        contentHash: string;
+    } | undefined;
 };
 export declare const harnessRefinerBenchmarkAssets: Readonly<Record<string, string>>;
 //# sourceMappingURL=harness-refiner.d.ts.map

package/dist/types/builtin-benchmarks/harness-refiner.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"harness-refiner.d.ts","sourceRoot":"","sources":["../../../../../src/builtin-benchmarks/harness-refiner.ts"],"names":[],"mappings":"AAGA,eAAO,MAAM,8BAA8B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAwuC1C,CAAC;AAEF,eAAO,MAAM,6BAA6B,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CASzE,CAAC"}
+{"version":3,"file":"harness-refiner.d.ts","sourceRoot":"","sources":["../../../../../src/builtin-benchmarks/harness-refiner.ts"],"names":[],"mappings":"AAGA,eAAO,MAAM,8BAA8B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAqqD1C,CAAC;AAEF,eAAO,MAAM,6BAA6B,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CASzE,CAAC"}

package/dist/types/conformance.d.ts

@@ -159,6 +159,20 @@ export declare const genericToolConformance: {
                 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;
@@ -224,6 +238,14 @@ export declare const genericToolConformance: {
         })[];
         metadata: Record<string, unknown>;
         contentHash: string;
+        environmentRelease?: {
+            id: string;
+            contentHash: string;
+        } | undefined;
+        verifierSetRelease?: {
+            id: string;
+            contentHash: string;
+        } | undefined;
     };
     manifest: {
         schemaVersion: "openpond.runManifest.v1";
@@ -426,6 +448,20 @@ export declare const marketingPortfolioConformance: {
                 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;
@@ -491,6 +527,14 @@ export declare const marketingPortfolioConformance: {
         })[];
         metadata: Record<string, unknown>;
         contentHash: string;
+        environmentRelease?: {
+            id: string;
+            contentHash: string;
+        } | undefined;
+        verifierSetRelease?: {
+            id: string;
+            contentHash: string;
+        } | undefined;
     };
     manifest: {
         schemaVersion: "openpond.runManifest.v1";

package/dist/types/conformance.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"conformance.d.ts","sourceRoot":"","sources":["../../../../src/conformance.ts"],"names":[],"mappings":"AAaA,eAAO,MAAM,sBAAsB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAEjC,CAAC;AAEH,eAAO,MAAM,6BAA6B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAGxC,CAAC"}
+{"version":3,"file":"conformance.d.ts","sourceRoot":"","sources":["../../../../src/conformance.ts"],"names":[],"mappings":"AAaA,eAAO,MAAM,sBAAsB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAEjC,CAAC;AAEH,eAAO,MAAM,6BAA6B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAGxC,CAAC"}

package/dist/types/evidence/conformance.d.ts

@@ -9,9 +9,9 @@ export declare const completeWorkProcessTraceFixture: {
         sequence: number;
         timestamp: string;
         layer: "agent" | "environment";
-        kind: "approval" | "validation" | "artifact" | "tool" | "state_transition" | "question" | "cleanup";
+        kind: "validation" | "artifact" | "approval" | "tool" | "state_transition" | "question" | "cleanup";
         action: "turn_started" | "turn_completed" | "turn_failed" | "turn_cancelled" | "turn_timed_out" | "tool_invoked" | "tool_completed" | "tool_failed" | "validation_completed" | "validation_failed" | "approval_requested" | "approval_resolved" | "question_asked" | "question_answered" | "question_dismissed" | "artifact_created" | "workspace_changed" | "environment_created" | "environment_reset" | "environment_destroyed" | "cleanup_completed" | "cleanup_failed";
-        status: "cancelled" | "failed" | "started" | "completed";
+        status: "failed" | "completed" | "cancelled" | "started";
         inputHash: string | null;
         outputHash: string | null;
         receiptHash: string | null;
@@ -25,14 +25,14 @@ export declare const completeWorkProcessTraceFixture: {
         attributes: {
             toolCategory: "agent" | "command" | "filesystem" | "source_control" | "browser" | "connected_app" | "sandbox" | "other" | null;
             validationKind: "test" | "structural" | "visual" | "user_review" | "other" | null;
-            transitionState: "timeout" | "cancelled" | "failed" | "completed" | "running" | null;
+            transitionState: "failed" | "completed" | "cancelled" | "timeout" | "running" | null;
             interventionOutcome: "requested" | "approved" | "denied" | "answered" | "dismissed" | null;
             artifactCount: number;
             exitCode: number | null;
             durationMs: number | null;
             cpuTimeMs: number | null;
             memoryPeakBytes: number | null;
-            errorClass: "unknown" | "timeout" | "cancelled" | "validation" | "policy" | "environment" | "infrastructure" | null;
+            errorClass: "validation" | "policy" | "unknown" | "environment" | "cancelled" | "timeout" | "infrastructure" | null;
         };
     }[];
     contentHash: string;
@@ -70,8 +70,8 @@ export declare const completeWorkEvidenceFixture: {
     };
     inputHash: string;
     terminal: {
-        status: "timeout" | "cancelled" | "failed" | "completed";
-        failureClass: "unknown" | "policy_failure" | "environment_failure" | "infrastructure_failure" | "timeout" | "cancelled" | "validation_failure" | "model_failure" | null;
+        status: "failed" | "completed" | "cancelled" | "timeout";
+        failureClass: "unknown" | "policy_failure" | "environment_failure" | "cancelled" | "infrastructure_failure" | "timeout" | "validation_failure" | "model_failure" | null;
     };
     trace: {
         sanitizedRef: {
@@ -186,9 +186,9 @@ export declare const incompleteWorkProcessTraceFixture: {
         sequence: number;
         timestamp: string;
         layer: "agent" | "environment";
-        kind: "approval" | "validation" | "artifact" | "tool" | "state_transition" | "question" | "cleanup";
+        kind: "validation" | "artifact" | "approval" | "tool" | "state_transition" | "question" | "cleanup";
         action: "turn_started" | "turn_completed" | "turn_failed" | "turn_cancelled" | "turn_timed_out" | "tool_invoked" | "tool_completed" | "tool_failed" | "validation_completed" | "validation_failed" | "approval_requested" | "approval_resolved" | "question_asked" | "question_answered" | "question_dismissed" | "artifact_created" | "workspace_changed" | "environment_created" | "environment_reset" | "environment_destroyed" | "cleanup_completed" | "cleanup_failed";
-        status: "cancelled" | "failed" | "started" | "completed";
+        status: "failed" | "completed" | "cancelled" | "started";
         inputHash: string | null;
         outputHash: string | null;
         receiptHash: string | null;
@@ -202,14 +202,14 @@ export declare const incompleteWorkProcessTraceFixture: {
         attributes: {
             toolCategory: "agent" | "command" | "filesystem" | "source_control" | "browser" | "connected_app" | "sandbox" | "other" | null;
             validationKind: "test" | "structural" | "visual" | "user_review" | "other" | null;
-            transitionState: "timeout" | "cancelled" | "failed" | "completed" | "running" | null;
+            transitionState: "failed" | "completed" | "cancelled" | "timeout" | "running" | null;
             interventionOutcome: "requested" | "approved" | "denied" | "answered" | "dismissed" | null;
             artifactCount: number;
             exitCode: number | null;
             durationMs: number | null;
             cpuTimeMs: number | null;
             memoryPeakBytes: number | null;
-            errorClass: "unknown" | "timeout" | "cancelled" | "validation" | "policy" | "environment" | "infrastructure" | null;
+            errorClass: "validation" | "policy" | "unknown" | "environment" | "cancelled" | "timeout" | "infrastructure" | null;
         };
     }[];
     contentHash: string;
@@ -308,8 +308,8 @@ export declare const invalidRawEvidenceFixture: {
     };
     inputHash: string;
     terminal: {
-        status: "timeout" | "cancelled" | "failed" | "completed";
-        failureClass: "unknown" | "policy_failure" | "environment_failure" | "infrastructure_failure" | "timeout" | "cancelled" | "validation_failure" | "model_failure" | null;
+        status: "failed" | "completed" | "cancelled" | "timeout";
+        failureClass: "unknown" | "policy_failure" | "environment_failure" | "cancelled" | "infrastructure_failure" | "timeout" | "validation_failure" | "model_failure" | null;
     };
     trace: {
         sanitizedRef: {
@@ -392,9 +392,9 @@ export declare const workEvidenceConformance: {
             sequence: number;
             timestamp: string;
             layer: "agent" | "environment";
-            kind: "approval" | "validation" | "artifact" | "tool" | "state_transition" | "question" | "cleanup";
+            kind: "validation" | "artifact" | "approval" | "tool" | "state_transition" | "question" | "cleanup";
             action: "turn_started" | "turn_completed" | "turn_failed" | "turn_cancelled" | "turn_timed_out" | "tool_invoked" | "tool_completed" | "tool_failed" | "validation_completed" | "validation_failed" | "approval_requested" | "approval_resolved" | "question_asked" | "question_answered" | "question_dismissed" | "artifact_created" | "workspace_changed" | "environment_created" | "environment_reset" | "environment_destroyed" | "cleanup_completed" | "cleanup_failed";
-            status: "cancelled" | "failed" | "started" | "completed";
+            status: "failed" | "completed" | "cancelled" | "started";
             inputHash: string | null;
             outputHash: string | null;
             receiptHash: string | null;
@@ -408,14 +408,14 @@ export declare const workEvidenceConformance: {
             attributes: {
                 toolCategory: "agent" | "command" | "filesystem" | "source_control" | "browser" | "connected_app" | "sandbox" | "other" | null;
                 validationKind: "test" | "structural" | "visual" | "user_review" | "other" | null;
-                transitionState: "timeout" | "cancelled" | "failed" | "completed" | "running" | null;
+                transitionState: "failed" | "completed" | "cancelled" | "timeout" | "running" | null;
                 interventionOutcome: "requested" | "approved" | "denied" | "answered" | "dismissed" | null;
                 artifactCount: number;
                 exitCode: number | null;
                 durationMs: number | null;
                 cpuTimeMs: number | null;
                 memoryPeakBytes: number | null;
-                errorClass: "unknown" | "timeout" | "cancelled" | "validation" | "policy" | "environment" | "infrastructure" | null;
+                errorClass: "validation" | "policy" | "unknown" | "environment" | "cancelled" | "timeout" | "infrastructure" | null;
             };
         }[];
         contentHash: string;
@@ -453,8 +453,8 @@ export declare const workEvidenceConformance: {
         };
         inputHash: string;
         terminal: {
-            status: "timeout" | "cancelled" | "failed" | "completed";
-            failureClass: "unknown" | "policy_failure" | "environment_failure" | "infrastructure_failure" | "timeout" | "cancelled" | "validation_failure" | "model_failure" | null;
+            status: "failed" | "completed" | "cancelled" | "timeout";
+            failureClass: "unknown" | "policy_failure" | "environment_failure" | "cancelled" | "infrastructure_failure" | "timeout" | "validation_failure" | "model_failure" | null;
         };
         trace: {
             sanitizedRef: {
@@ -625,9 +625,9 @@ export declare const workEvidenceConformance: {
             sequence: number;
             timestamp: string;
             layer: "agent" | "environment";
-            kind: "approval" | "validation" | "artifact" | "tool" | "state_transition" | "question" | "cleanup";
+            kind: "validation" | "artifact" | "approval" | "tool" | "state_transition" | "question" | "cleanup";
             action: "turn_started" | "turn_completed" | "turn_failed" | "turn_cancelled" | "turn_timed_out" | "tool_invoked" | "tool_completed" | "tool_failed" | "validation_completed" | "validation_failed" | "approval_requested" | "approval_resolved" | "question_asked" | "question_answered" | "question_dismissed" | "artifact_created" | "workspace_changed" | "environment_created" | "environment_reset" | "environment_destroyed" | "cleanup_completed" | "cleanup_failed";
-            status: "cancelled" | "failed" | "started" | "completed";
+            status: "failed" | "completed" | "cancelled" | "started";
             inputHash: string | null;
             outputHash: string | null;
             receiptHash: string | null;
@@ -641,14 +641,14 @@ export declare const workEvidenceConformance: {
             attributes: {
                 toolCategory: "agent" | "command" | "filesystem" | "source_control" | "browser" | "connected_app" | "sandbox" | "other" | null;
                 validationKind: "test" | "structural" | "visual" | "user_review" | "other" | null;
-                transitionState: "timeout" | "cancelled" | "failed" | "completed" | "running" | null;
+                transitionState: "failed" | "completed" | "cancelled" | "timeout" | "running" | null;
                 interventionOutcome: "requested" | "approved" | "denied" | "answered" | "dismissed" | null;
                 artifactCount: number;
                 exitCode: number | null;
                 durationMs: number | null;
                 cpuTimeMs: number | null;
                 memoryPeakBytes: number | null;
-                errorClass: "unknown" | "timeout" | "cancelled" | "validation" | "policy" | "environment" | "infrastructure" | null;
+                errorClass: "validation" | "policy" | "unknown" | "environment" | "cancelled" | "timeout" | "infrastructure" | null;
             };
         }[];
         contentHash: string;
@@ -691,8 +691,8 @@ export declare const workEvidenceConformance: {
         };
         inputHash: string;
         terminal: {
-            status: "timeout" | "cancelled" | "failed" | "completed";
-            failureClass: "unknown" | "policy_failure" | "environment_failure" | "infrastructure_failure" | "timeout" | "cancelled" | "validation_failure" | "model_failure" | null;
+            status: "failed" | "completed" | "cancelled" | "timeout";
+            failureClass: "unknown" | "policy_failure" | "environment_failure" | "cancelled" | "infrastructure_failure" | "timeout" | "validation_failure" | "model_failure" | null;
         };
         trace: {
             sanitizedRef: {