openpond-sdk 0.9.1 → 0.9.3

package/dist/{chunk-IKTZ2CD4.js → chunk-XR6CLSCS.js}

@@ -3,9 +3,13 @@ import {
   canonicalJson,
   canonicalSha256
 } from "./chunk-TCHAWJMB.js";
+import {
+  FailureClassSchema,
+  contentHash
+} from "./chunk-QUOW6URQ.js";
 
 // src/model-starter-attempts.ts
-import { z as z2 } from "zod";
+import { z as z3 } from "zod";
 
 // src/evaluation-project-context.ts
 import { z } from "zod";
@@ -25,71 +29,108 @@ var EvaluationMessagesSchema = z.array(z.object({
 
 // src/model-starter-attempts.ts
 import { RewardCompositionSchema } from "@openpond/evals/rewards";
+import { TaskGradeSchema as TaskGradeSchema2 } from "@openpond/evals/graders";
+
+// src/experiment-grading-contracts.ts
+import { z as z2 } from "zod";
 import { TaskGradeSchema } from "@openpond/evals/graders";
-var IdSchema = z2.string().trim().min(1).max(200);
-var HashSchema = z2.string().regex(/^[a-f0-9]{64}$/);
-var ModelStarterAttemptPolicySchema = z2.discriminatedUnion("kind", [
-  z2.object({ kind: z2.literal("hosted_chat"), modelId: IdSchema, messages: EvaluationMessagesSchema.optional(), maxOutputTokens: z2.number().int().min(1).max(4096).default(1024), temperature: z2.number().min(0).max(2).default(0), topP: z2.number().gt(0).max(1).default(1) }).strict(),
-  z2.object({ kind: z2.literal("fixture"), fixtureId: IdSchema }).strict()
+var ExperimentAttemptGradeSchema = z2.object({
+  schemaVersion: z2.literal("openpond.experimentAttemptGrade.v1"),
+  selectionHash: z2.string().regex(/^[a-f0-9]{64}$/),
+  grades: z2.array(TaskGradeSchema).min(1).max(100),
+  score: z2.number().min(0).max(1).nullable(),
+  passed: z2.boolean(),
+  rewardEligible: z2.boolean(),
+  gradingStatus: z2.enum(["scored", "unscorable", "not_configured"]),
+  failureClass: FailureClassSchema.nullable(),
+  contentHash: z2.string().regex(/^[a-f0-9]{64}$/)
+}).strict().superRefine((grade, context) => {
+  const { contentHash: actual, ...body } = grade;
+  if (contentHash(body) !== actual) context.addIssue({ code: "custom", message: "Experiment grading integrity failed." });
+  const components = grade.grades.flatMap((value) => value.components);
+  if (components.length !== grade.grades.length || new Set(components.map((component) => component.graderId)).size !== components.length)
+    context.addIssue({ code: "custom", message: "Each selected criterion requires exactly one distinct grader result." });
+  for (const item of grade.grades) {
+    const { contentHash: itemHash, ...itemBody } = item;
+    if (contentHash(itemBody) !== itemHash) context.addIssue({ code: "custom", message: "A retained criterion changed." });
+    for (const component of item.components) {
+      const { contentHash: componentHash, ...componentBody } = component;
+      if (contentHash(componentBody) !== componentHash) context.addIssue({ code: "custom", message: "A retained grader result changed." });
+    }
+  }
+});
+
+// src/model-starter-attempts.ts
+var IdSchema = z3.string().trim().min(1).max(200);
+var HashSchema = z3.string().regex(/^[a-f0-9]{64}$/);
+var ModelStarterAttemptPolicySchema = z3.discriminatedUnion("kind", [
+  z3.object({ kind: z3.literal("hosted_chat"), modelId: IdSchema, messages: EvaluationMessagesSchema.optional(), maxOutputTokens: z3.number().int().min(1).max(4096).default(1024), temperature: z3.number().min(0).max(2).default(0), topP: z3.number().gt(0).max(1).default(1) }).strict(),
+  z3.object({ kind: z3.literal("fixture"), fixtureId: IdSchema }).strict()
 ]);
-var ModelStarterAttemptRequestSchema = z2.object({
-  schemaVersion: z2.literal("openpond.modelStarterAttemptRequest.v1"),
+var ModelStarterAttemptRequestSchema = z3.object({
+  schemaVersion: z3.literal("openpond.modelStarterAttemptRequest.v1"),
   operationId: IdSchema,
   teamId: IdSchema,
   modelProjectId: IdSchema.nullable(),
   project: EvaluationProjectContextSchema.optional(),
   taskset: ModelProjectVersionedRefSchema,
   taskId: IdSchema,
-  environmentSeed: z2.number().int().min(0).max(2147483647).default(0),
+  environmentSeed: z3.number().int().min(0).max(2147483647).default(0),
   policy: ModelStarterAttemptPolicySchema
 }).strict();
-var ModelStarterAttemptChoicesQuerySchema = z2.object({ modelProjectId: IdSchema.nullable(), taskset: ModelProjectVersionedRefSchema, afterTaskId: IdSchema.optional(), limit: z2.number().int().min(1).max(100).default(25) }).strict();
-var ModelStarterAttemptChoicesSchema = z2.object({
+var ModelStarterAttemptChoicesQuerySchema = z3.object({ modelProjectId: IdSchema.nullable(), taskset: ModelProjectVersionedRefSchema, afterTaskId: IdSchema.optional(), limit: z3.number().int().min(1).max(100).default(25) }).strict();
+var ModelStarterAttemptChoicesSchema = z3.object({
   modelProjectId: IdSchema.nullable(),
   taskset: ModelProjectVersionedRefSchema,
-  available: z2.boolean(),
-  unavailableReason: z2.string().max(1e3).nullable(),
-  models: z2.array(z2.object({ id: IdSchema, name: z2.string().max(500) }).strict()).max(200),
-  tasks: z2.array(z2.object({ id: IdSchema, split: z2.enum(["train", "validation", "frozen_eval"]), inputPreview: z2.string().max(500), fixtures: z2.array(z2.object({ id: IdSchema, label: z2.string().max(200) }).strict()).max(1e3) }).strict()).max(100),
+  available: z3.boolean(),
+  unavailableReason: z3.string().max(1e3).nullable(),
+  models: z3.array(z3.object({ id: IdSchema, name: z3.string().max(500) }).strict()).max(200),
+  tasks: z3.array(z3.object({ id: IdSchema, split: z3.enum(["train", "validation", "frozen_eval"]), inputPreview: z3.string().max(500), fixtures: z3.array(z3.object({ id: IdSchema, label: z3.string().max(200) }).strict()).max(1e3) }).strict()).max(100),
   nextCursor: IdSchema.nullable()
 }).strict();
-var ModelStarterAttemptSummarySchema = z2.object({
-  schemaVersion: z2.literal("openpond.modelStarterAttemptSummary.v1"),
+var ModelStarterAttemptSummarySchema = z3.object({
+  schemaVersion: z3.literal("openpond.modelStarterAttemptSummary.v1"),
   id: IdSchema,
-  revision: z2.number().int().positive(),
+  revision: z3.number().int().positive(),
   request: ModelStarterAttemptRequestSchema,
-  status: z2.enum(["queued", "running", "cancelling", "completed", "failed", "cancelled"]),
-  policySnapshot: z2.object({ modelId: IdSchema, provider: IdSchema, upstreamModelId: z2.string().min(1).max(500), configurationHash: HashSchema }).strict().nullable(),
-  createdAt: z2.iso.datetime(),
-  startedAt: z2.iso.datetime().nullable(),
-  completedAt: z2.iso.datetime().nullable(),
-  cleanupComplete: z2.boolean(),
-  resultAvailable: z2.boolean(),
-  score: z2.number().min(0).max(1).nullable(),
-  gradingStatus: z2.enum(["not_started", "scored", "unscorable", "not_configured"]).default("not_started"),
-  passed: z2.boolean().nullable(),
-  outputPreview: z2.string().max(500).nullable(),
-  error: z2.object({ code: IdSchema, message: z2.string().max(2e3) }).strict().nullable()
+  status: z3.enum(["queued", "running", "cancelling", "completed", "failed", "cancelled"]),
+  policySnapshot: z3.object({ modelId: IdSchema, provider: IdSchema, upstreamModelId: z3.string().min(1).max(500), configurationHash: HashSchema }).strict().nullable(),
+  createdAt: z3.iso.datetime(),
+  startedAt: z3.iso.datetime().nullable(),
+  completedAt: z3.iso.datetime().nullable(),
+  cleanupComplete: z3.boolean(),
+  resultAvailable: z3.boolean(),
+  score: z3.number().min(0).max(1).nullable(),
+  gradingStatus: z3.enum(["not_started", "scored", "unscorable", "not_configured"]).default("not_started"),
+  passed: z3.boolean().nullable(),
+  outputPreview: z3.string().max(500).nullable(),
+  error: z3.object({ code: IdSchema, message: z3.string().max(2e3) }).strict().nullable()
 }).strict().superRefine((value, context) => {
   if (value.request.policy.kind === "fixture" !== (value.policySnapshot === null)) context.addIssue({ code: "custom", path: ["policySnapshot"], message: "Fixture execution must not claim a model identity." });
   if (value.status === "completed" && (!value.cleanupComplete || !value.resultAvailable || !value.completedAt)) context.addIssue({ code: "custom", message: "Completed attempts require retained results and confirmed cleanup." });
 });
-var ModelStarterAttemptListQuerySchema = z2.object({ modelProjectId: IdSchema, afterId: IdSchema.optional(), limit: z2.number().int().min(1).max(100).default(25) }).strict();
-var ModelStarterAttemptPageSchema = z2.object({ items: z2.array(ModelStarterAttemptSummarySchema).max(100), nextCursor: IdSchema.nullable() }).strict();
-var ModelStarterAttemptResultSchema = z2.object({
-  schemaVersion: z2.literal("openpond.modelStarterAttemptResult.v1"),
+var ModelStarterAttemptListQuerySchema = z3.object({ modelProjectId: IdSchema, afterId: IdSchema.optional(), limit: z3.number().int().min(1).max(100).default(25) }).strict();
+var ModelStarterAttemptPageSchema = z3.object({ items: z3.array(ModelStarterAttemptSummarySchema).max(100), nextCursor: IdSchema.nullable() }).strict();
+var ModelStarterAttemptResultSchema = z3.object({
+  schemaVersion: z3.literal("openpond.modelStarterAttemptResult.v1"),
   attempt: ModelStarterAttemptSummarySchema,
-  output: z2.string().max(1048576).nullable(),
+  output: z3.string().max(1048576).nullable(),
   // Only model-facing messages, never the private environment snapshot.
-  messages: z2.array(z2.record(z2.string(), z2.unknown())).max(4002),
+  messages: z3.array(z3.record(z3.string(), z3.unknown())).max(4002),
   composition: RewardCompositionSchema.nullable(),
   // Ordinary Tasksets have no Reward binding. Absence preserves the hash of
   // already retained v1 results; new producers explicitly return null or a grade.
-  grade: TaskGradeSchema.nullable().optional(),
-  environment: z2.object({ status: z2.enum(["completed", "budget_exhausted", "cancelled", "timed_out", "policy_failure", "environment_failure"]), collected: z2.boolean(), definition: ModelProjectVersionedRefSchema, initialStateHash: HashSchema.nullable(), finalStateHash: HashSchema.nullable(), attemptHash: HashSchema }).strict(),
-  providerRequestIds: z2.array(IdSchema).max(1001),
+  grade: TaskGradeSchema2.nullable().optional(),
+  experimentGrade: ExperimentAttemptGradeSchema.nullable().optional(),
+  environment: z3.object({ status: z3.enum(["completed", "budget_exhausted", "cancelled", "timed_out", "policy_failure", "environment_failure"]), collected: z3.boolean(), definition: ModelProjectVersionedRefSchema, initialStateHash: HashSchema.nullable(), finalStateHash: HashSchema.nullable(), attemptHash: HashSchema }).strict(),
+  providerRequestIds: z3.array(IdSchema).max(1001),
   contentHash: HashSchema
 }).strict().superRefine((value, context) => {
+  if (value.experimentGrade) {
+    if (value.grade || value.composition) context.addIssue({ code: "custom", message: "Selected Experiment grading cannot claim another grade population." });
+    if (!["completed", "failed"].includes(value.attempt.status) || !value.attempt.resultAvailable || value.attempt.score !== value.experimentGrade.score || value.attempt.gradingStatus !== value.experimentGrade.gradingStatus || value.attempt.passed !== (value.experimentGrade.score === null ? null : value.experimentGrade.passed))
+      context.addIssue({ code: "custom", message: "Experiment grading differs from its retained attempt summary." });
+  }
   if (!value.grade) return;
   if (value.composition) context.addIssue({ code: "custom", path: ["grade"], message: "An ordinary grade cannot also claim a Reward composition." });
   if (!["completed", "failed"].includes(value.attempt.status) || !value.attempt.resultAvailable || value.attempt.score !== value.grade.score || value.attempt.gradingStatus !== value.grade.gradingStatus || value.attempt.passed !== (value.grade.score === null ? null : value.grade.passed)) context.addIssue({ code: "custom", path: ["grade"], message: "Ordinary grade differs from its terminal attempt summary." });
@@ -135,8 +176,8 @@ var OpenPondModelStarterAttemptsClient = class {
   async result(id, options = {}) {
     const value = ModelStarterAttemptResultSchema.parse(await this.#request(`/${encodeURIComponent(IdSchema.parse(id))}/result`, "GET", void 0, options.signal));
     this.#summary(value.attempt, id);
-    const { contentHash, ...content } = value;
-    if (!value.attempt.resultAvailable || await canonicalSha256(content) !== contentHash) throw new OpenPondModelStarterAttemptError(502, "attempt_result_integrity", "Attempt result integrity failed.");
+    const { contentHash: contentHash2, ...content } = value;
+    if (!value.attempt.resultAvailable || await canonicalSha256(content) !== contentHash2) throw new OpenPondModelStarterAttemptError(502, "attempt_result_integrity", "Attempt result integrity failed.");
     return value;
   }
   async list(value, options = {}) {
@@ -174,7 +215,7 @@ var OpenPondModelStarterAttemptsClient = class {
     }
     const value = JSON.parse(text);
     if (!response.ok) {
-      const error = z2.object({ code: IdSchema, message: z2.string().max(2e3) }).safeParse(value);
+      const error = z3.object({ code: IdSchema, message: z3.string().max(2e3) }).safeParse(value);
       throw new OpenPondModelStarterAttemptError(response.status, error.success ? error.data.code : "attempt_request_failed", error.success ? error.data.message : "Attempt request failed.");
     }
     return value;
@@ -184,6 +225,7 @@ var OpenPondModelStarterAttemptsClient = class {
 export {
   EvaluationProjectContextSchema,
   EvaluationMessagesSchema,
+  ExperimentAttemptGradeSchema,
   ModelStarterAttemptPolicySchema,
   ModelStarterAttemptRequestSchema,
   ModelStarterAttemptChoicesQuerySchema,
@@ -195,4 +237,4 @@ export {
   OpenPondModelStarterAttemptError,
   OpenPondModelStarterAttemptsClient
 };
-//# sourceMappingURL=chunk-IKTZ2CD4.js.map
+//# sourceMappingURL=chunk-XR6CLSCS.js.map

package/dist/chunk-XR6CLSCS.js.map

@@ -0,0 +1,7 @@
+{
+  "version": 3,
+  "sources": ["../src/model-starter-attempts.ts", "../src/evaluation-project-context.ts", "../src/experiment-grading-contracts.ts"],
+  "sourcesContent": ["import { z } from \"zod\";\nimport { EvaluationProjectContextSchema, EvaluationMessagesSchema } from \"./evaluation-project-context.js\";\nexport { EvaluationProjectContextSchema, EvaluationMessagesSchema, type EvaluationProjectContext } from \"./evaluation-project-context.js\";\nimport { RewardCompositionSchema } from \"@openpond/evals/rewards\";\nimport { TaskGradeSchema } from \"@openpond/evals/graders\";\nimport { ExperimentAttemptGradeSchema } from \"./experiment-grading-contracts.js\";\nimport { ModelProjectVersionedRefSchema } from \"./model-projects.js\";\nimport { canonicalJson, canonicalSha256 } from \"./protocol.js\";\n\nconst IdSchema = z.string().trim().min(1).max(200);\nconst HashSchema = z.string().regex(/^[a-f0-9]{64}$/);\nexport const ModelStarterAttemptPolicySchema = z.discriminatedUnion(\"kind\", [\n  z.object({ kind: z.literal(\"hosted_chat\"), modelId: IdSchema, messages: EvaluationMessagesSchema.optional(), maxOutputTokens: z.number().int().min(1).max(4_096).default(1_024), temperature: z.number().min(0).max(2).default(0), topP: z.number().gt(0).max(1).default(1) }).strict(),\n  z.object({ kind: z.literal(\"fixture\"), fixtureId: IdSchema }).strict(),\n]);\n/** The server resolves task input, private state, verifier and fixture script.\n * A caller can select releases, but cannot submit replacement execution bytes. */\nexport const ModelStarterAttemptRequestSchema = z.object({\n  schemaVersion: z.literal(\"openpond.modelStarterAttemptRequest.v1\"),\n  operationId: IdSchema, teamId: IdSchema, modelProjectId: IdSchema.nullable(),\n  project: EvaluationProjectContextSchema.optional(),\n  taskset: ModelProjectVersionedRefSchema, taskId: IdSchema,\n  environmentSeed: z.number().int().min(0).max(2_147_483_647).default(0),\n  policy: ModelStarterAttemptPolicySchema,\n}).strict();\nexport type ModelStarterAttemptRequest = z.infer<typeof ModelStarterAttemptRequestSchema>;\nexport const ModelStarterAttemptChoicesQuerySchema = z.object({ modelProjectId: IdSchema.nullable(), taskset: ModelProjectVersionedRefSchema, afterTaskId: IdSchema.optional(), limit: z.number().int().min(1).max(100).default(25) }).strict();\nexport const ModelStarterAttemptChoicesSchema = z.object({\n  modelProjectId: IdSchema.nullable(), taskset: ModelProjectVersionedRefSchema, available: z.boolean(), unavailableReason: z.string().max(1_000).nullable(),\n  models: z.array(z.object({ id: IdSchema, name: z.string().max(500) }).strict()).max(200),\n  tasks: z.array(z.object({ id: IdSchema, split: z.enum([\"train\", \"validation\", \"frozen_eval\"]), inputPreview: z.string().max(500), fixtures: z.array(z.object({ id: IdSchema, label: z.string().max(200) }).strict()).max(1_000) }).strict()).max(100),\n  nextCursor: IdSchema.nullable(),\n}).strict();\nexport const ModelStarterAttemptSummarySchema = z.object({\n  schemaVersion: z.literal(\"openpond.modelStarterAttemptSummary.v1\"),\n  id: IdSchema, revision: z.number().int().positive(), request: ModelStarterAttemptRequestSchema,\n  status: z.enum([\"queued\", \"running\", \"cancelling\", \"completed\", \"failed\", \"cancelled\"]),\n  policySnapshot: z.object({ modelId: IdSchema, provider: IdSchema, upstreamModelId: z.string().min(1).max(500), configurationHash: HashSchema }).strict().nullable(),\n  createdAt: z.iso.datetime(), startedAt: z.iso.datetime().nullable(), completedAt: z.iso.datetime().nullable(),\n  cleanupComplete: z.boolean(), resultAvailable: z.boolean(), score: z.number().min(0).max(1).nullable(),\n  gradingStatus: z.enum([\"not_started\", \"scored\", \"unscorable\", \"not_configured\"]).default(\"not_started\"),\n  passed: z.boolean().nullable(), outputPreview: z.string().max(500).nullable(),\n  error: z.object({ code: IdSchema, message: z.string().max(2_000) }).strict().nullable(),\n}).strict().superRefine((value, context) => {\n  if ((value.request.policy.kind === \"fixture\") !== (value.policySnapshot === null)) context.addIssue({ code: \"custom\", path: [\"policySnapshot\"], message: \"Fixture execution must not claim a model identity.\" });\n  if (value.status === \"completed\" && (!value.cleanupComplete || !value.resultAvailable || !value.completedAt)) context.addIssue({ code: \"custom\", message: \"Completed attempts require retained results and confirmed cleanup.\" });\n});\nexport type ModelStarterAttemptSummary = z.infer<typeof ModelStarterAttemptSummarySchema>;\nexport const ModelStarterAttemptListQuerySchema = z.object({ modelProjectId: IdSchema, afterId: IdSchema.optional(), limit: z.number().int().min(1).max(100).default(25) }).strict();\nexport const ModelStarterAttemptPageSchema = z.object({ items: z.array(ModelStarterAttemptSummarySchema).max(100), nextCursor: IdSchema.nullable() }).strict();\nexport const ModelStarterAttemptResultSchema = z.object({\n  schemaVersion: z.literal(\"openpond.modelStarterAttemptResult.v1\"),\n  attempt: ModelStarterAttemptSummarySchema,\n  output: z.string().max(1_048_576).nullable(),\n  // Only model-facing messages, never the private environment snapshot.\n  messages: z.array(z.record(z.string(), z.unknown())).max(4_002),\n  composition: RewardCompositionSchema.nullable(),\n  // Ordinary Tasksets have no Reward binding. Absence preserves the hash of\n  // already retained v1 results; new producers explicitly return null or a grade.\n  grade: TaskGradeSchema.nullable().optional(),\n  experimentGrade: ExperimentAttemptGradeSchema.nullable().optional(),\n  environment: z.object({ status: z.enum([\"completed\", \"budget_exhausted\", \"cancelled\", \"timed_out\", \"policy_failure\", \"environment_failure\"]), collected: z.boolean(), definition: ModelProjectVersionedRefSchema, initialStateHash: HashSchema.nullable(), finalStateHash: HashSchema.nullable(), attemptHash: HashSchema }).strict(),\n  providerRequestIds: z.array(IdSchema).max(1_001),\n  contentHash: HashSchema,\n}).strict().superRefine((value, context) => {\n  if (value.experimentGrade) {\n    if (value.grade || value.composition) context.addIssue({ code: \"custom\", message: \"Selected Experiment grading cannot claim another grade population.\" });\n    if (![\"completed\", \"failed\"].includes(value.attempt.status) || !value.attempt.resultAvailable || value.attempt.score !== value.experimentGrade.score\n      || value.attempt.gradingStatus !== value.experimentGrade.gradingStatus || value.attempt.passed !== (value.experimentGrade.score === null ? null : value.experimentGrade.passed))\n      context.addIssue({ code: \"custom\", message: \"Experiment grading differs from its retained attempt summary.\" });\n  }\n  if (!value.grade) return;\n  if (value.composition) context.addIssue({ code: \"custom\", path: [\"grade\"], message: \"An ordinary grade cannot also claim a Reward composition.\" });\n  if (![\"completed\", \"failed\"].includes(value.attempt.status) || !value.attempt.resultAvailable || value.attempt.score !== value.grade.score || value.attempt.gradingStatus !== value.grade.gradingStatus || value.attempt.passed !== (value.grade.score === null ? null : value.grade.passed)) context.addIssue({ code: \"custom\", path: [\"grade\"], message: \"Ordinary grade differs from its terminal attempt summary.\" });\n});\nexport type ModelStarterAttemptResult = z.infer<typeof ModelStarterAttemptResultSchema>;\n\nexport class OpenPondModelStarterAttemptError extends Error {\n  constructor(readonly status: number, readonly code: string, message: string) { super(message); this.name = \"OpenPondModelStarterAttemptError\"; }\n}\n\nexport class OpenPondModelStarterAttemptsClient {\n  readonly #options: { baseUrl: string; apiKey: string; teamId: string; fetch?: typeof globalThis.fetch };\n  constructor(options: { baseUrl: string; apiKey: string; teamId: string; fetch?: typeof globalThis.fetch }) {\n    const url = new URL(options.baseUrl);\n    if (!options.apiKey.trim() || !options.teamId.trim() || ![\"https:\", \"http:\"].includes(url.protocol) || url.username || url.password || url.search || url.hash) throw new Error(\"Starter attempts require an HTTP(S) endpoint and workspace credentials.\");\n    this.#options = { ...options, baseUrl: url.toString().replace(/\\/+$/, \"\") };\n  }\n  async create(value: z.input<typeof ModelStarterAttemptRequestSchema>, options: { signal?: AbortSignal } = {}) {\n    const request = ModelStarterAttemptRequestSchema.parse(value);\n    if (request.teamId !== this.#options.teamId) throw new Error(\"Attempt belongs to another workspace.\");\n    const result = this.#summary(await this.#request(\"\", \"POST\", request, options.signal));\n    if (canonicalJson(result.request) !== canonicalJson(request)) throw new OpenPondModelStarterAttemptError(502, \"attempt_request_mismatch\", \"Attempt receipt differs from the submitted request.\");\n    return result;\n  }\n  async get(id: string, options: { signal?: AbortSignal } = {}) { return this.#summary(await this.#request(`/${encodeURIComponent(IdSchema.parse(id))}`, \"GET\", undefined, options.signal), id); }\n  async choices(value: z.input<typeof ModelStarterAttemptChoicesQuerySchema>, options: { signal?: AbortSignal } = {}) {\n    const query = ModelStarterAttemptChoicesQuerySchema.parse(value);\n    const params = new URLSearchParams({ ...(query.modelProjectId ? { modelProjectId: query.modelProjectId } : {}), tasksetId: query.taskset.id, revision: String(query.taskset.revision), contentHash: query.taskset.contentHash, limit: String(query.limit) });\n    if (query.afterTaskId) params.set(\"afterTaskId\", query.afterTaskId);\n    const result = ModelStarterAttemptChoicesSchema.parse(await this.#request(`/choices?${params}`, \"GET\", undefined, options.signal));\n    if (result.modelProjectId !== query.modelProjectId || canonicalJson(result.taskset) !== canonicalJson(query.taskset) || result.tasks.length > query.limit || new Set(result.tasks.map(task => task.id)).size !== result.tasks.length) throw new OpenPondModelStarterAttemptError(502, \"attempt_choices_mismatch\", \"Attempt choices differ from the selected model or Taskset.\");\n    return result;\n  }\n  async cancel(id: string, options: { signal?: AbortSignal } = {}) { return this.#summary(await this.#request(`/${encodeURIComponent(IdSchema.parse(id))}/cancel`, \"POST\", undefined, options.signal), id); }\n  async result(id: string, options: { signal?: AbortSignal } = {}) {\n    const value = ModelStarterAttemptResultSchema.parse(await this.#request(`/${encodeURIComponent(IdSchema.parse(id))}/result`, \"GET\", undefined, options.signal));\n    this.#summary(value.attempt, id);\n    const { contentHash, ...content } = value;\n    if (!value.attempt.resultAvailable || await canonicalSha256(content) !== contentHash) throw new OpenPondModelStarterAttemptError(502, \"attempt_result_integrity\", \"Attempt result integrity failed.\");\n    return value;\n  }\n  async list(value: z.input<typeof ModelStarterAttemptListQuerySchema>, options: { signal?: AbortSignal } = {}) {\n    const query = ModelStarterAttemptListQuerySchema.parse(value);\n    const params = new URLSearchParams({ modelProjectId: query.modelProjectId, limit: String(query.limit) });\n    if (query.afterId) params.set(\"afterId\", query.afterId);\n    const page = ModelStarterAttemptPageSchema.parse(await this.#request(`?${params}`, \"GET\", undefined, options.signal));\n    if (page.items.length > query.limit || new Set(page.items.map(item => item.id)).size !== page.items.length || page.items.some(item => this.#summary(item).request.modelProjectId !== query.modelProjectId)) throw new OpenPondModelStarterAttemptError(502, \"attempt_page_mismatch\", \"Attempt page differs from the selected model.\");\n    return page;\n  }\n  #summary(value: unknown, id?: string): ModelStarterAttemptSummary {\n    const result = ModelStarterAttemptSummarySchema.parse(value);\n    if (result.request.teamId !== this.#options.teamId || (id && result.id !== id)) throw new OpenPondModelStarterAttemptError(502, \"attempt_identity_mismatch\", \"Attempt receipt belongs to another identity or workspace.\");\n    return result;\n  }\n  async #request(path: string, method: string, body: unknown, signal?: AbortSignal): Promise<unknown> {\n    const response = await (this.#options.fetch ?? globalThis.fetch)(`${this.#options.baseUrl}/v1/model-starter-attempts${path}`, { method, headers: { Authorization: `Bearer ${this.#options.apiKey}`, \"X-OpenPond-Team-Id\": this.#options.teamId, \"Content-Type\": \"application/json\", Accept: \"application/json\" }, ...(body === undefined ? {} : { body: JSON.stringify(body) }), signal, redirect: \"error\" });\n    const reader = response.body?.getReader();\n    if (!reader) throw new OpenPondModelStarterAttemptError(response.status, \"attempt_empty_response\", \"Attempt response was empty.\");\n    let bytes = 0; let text = \"\"; const decoder = new TextDecoder();\n    try {\n      while (true) {\n        const chunk = await reader.read(); if (chunk.done) break;\n        bytes += chunk.value.byteLength;\n        if (bytes > 16_777_216) throw new OpenPondModelStarterAttemptError(502, \"attempt_response_too_large\", \"Attempt response exceeded its byte limit.\");\n        text += decoder.decode(chunk.value, { stream: true });\n      }\n      text += decoder.decode();\n    } finally { await reader.cancel(); reader.releaseLock(); }\n    const value: unknown = JSON.parse(text);\n    if (!response.ok) {\n      const error = z.object({ code: IdSchema, message: z.string().max(2_000) }).safeParse(value);\n      throw new OpenPondModelStarterAttemptError(response.status, error.success ? error.data.code : \"attempt_request_failed\", error.success ? error.data.message : \"Attempt request failed.\");\n    }\n    return value;\n  }\n}\n", "import { z } from \"zod\";\n\n/** The selected Project revision is retained independently of model artifacts.\n * Hosts resolve and authorize this reference before admitting execution. */\nexport const EvaluationProjectContextSchema = z.object({\n  id: z.string().trim().min(1).max(240),\n  revision: z.number().int().positive(),\n  contentHash: z.string().regex(/^[a-f0-9]{64}$/),\n  targetId: z.string().trim().min(1).max(240).nullable(),\n}).strict();\nexport type EvaluationProjectContext = z.infer<typeof EvaluationProjectContextSchema>;\n\nexport const EvaluationMessagesSchema = z.array(z.object({\n  role: z.enum([\"system\", \"user\", \"assistant\"]),\n  content: z.string().max(100_000),\n}).strict()).max(100).superRefine((messages, context) => {\n  if (messages.reduce((total, message) => total + message.content.length, 0) > 100_000)\n    context.addIssue({ code: \"custom\", message: \"Evaluation prompt exceeds 100,000 characters.\" });\n});\n", "import { z } from \"zod\";\nimport { FailureClassSchema, contentHash } from \"@openpond/harness\";\nimport { TaskGradeSchema } from \"@openpond/evals/graders\";\n\nexport const ExperimentAttemptGradeSchema = z.object({\n  schemaVersion: z.literal(\"openpond.experimentAttemptGrade.v1\"), selectionHash: z.string().regex(/^[a-f0-9]{64}$/),\n  grades: z.array(TaskGradeSchema).min(1).max(100), score: z.number().min(0).max(1).nullable(),\n  passed: z.boolean(), rewardEligible: z.boolean(), gradingStatus: z.enum([\"scored\", \"unscorable\", \"not_configured\"]),\n  failureClass: FailureClassSchema.nullable(), contentHash: z.string().regex(/^[a-f0-9]{64}$/),\n}).strict().superRefine((grade, context) => {\n  const { contentHash: actual, ...body } = grade;\n  if (contentHash(body) !== actual) context.addIssue({ code: \"custom\", message: \"Experiment grading integrity failed.\" });\n  const components = grade.grades.flatMap(value => value.components);\n  if (components.length !== grade.grades.length || new Set(components.map(component => component.graderId)).size !== components.length)\n    context.addIssue({ code: \"custom\", message: \"Each selected criterion requires exactly one distinct grader result.\" });\n  for (const item of grade.grades) {\n    const { contentHash: itemHash, ...itemBody } = item;\n    if (contentHash(itemBody) !== itemHash) context.addIssue({ code: \"custom\", message: \"A retained criterion changed.\" });\n    for (const component of item.components) {\n      const { contentHash: componentHash, ...componentBody } = component;\n      if (contentHash(componentBody) !== componentHash) context.addIssue({ code: \"custom\", message: \"A retained grader result changed.\" });\n    }\n  }\n});\nexport type ExperimentAttemptGrade = z.infer<typeof ExperimentAttemptGradeSchema>;\n"],
+  "mappings": ";;;;;;;;;;;AAAA,SAAS,KAAAA,UAAS;;;ACAlB,SAAS,SAAS;AAIX,IAAM,iCAAiC,EAAE,OAAO;AAAA,EACrD,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG;AAAA,EACpC,UAAU,EAAE,OAAO,EAAE,IAAI,EAAE,SAAS;AAAA,EACpC,aAAa,EAAE,OAAO,EAAE,MAAM,gBAAgB;AAAA,EAC9C,UAAU,EAAE,OAAO,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,EAAE,SAAS;AACvD,CAAC,EAAE,OAAO;AAGH,IAAM,2BAA2B,EAAE,MAAM,EAAE,OAAO;AAAA,EACvD,MAAM,EAAE,KAAK,CAAC,UAAU,QAAQ,WAAW,CAAC;AAAA,EAC5C,SAAS,EAAE,OAAO,EAAE,IAAI,GAAO;AACjC,CAAC,EAAE,OAAO,CAAC,EAAE,IAAI,GAAG,EAAE,YAAY,CAAC,UAAU,YAAY;AACvD,MAAI,SAAS,OAAO,CAAC,OAAO,YAAY,QAAQ,QAAQ,QAAQ,QAAQ,CAAC,IAAI;AAC3E,YAAQ,SAAS,EAAE,MAAM,UAAU,SAAS,gDAAgD,CAAC;AACjG,CAAC;;;ADfD,SAAS,+BAA+B;AACxC,SAAS,mBAAAC,wBAAuB;;;AEJhC,SAAS,KAAAC,UAAS;AAElB,SAAS,uBAAuB;AAEzB,IAAM,+BAA+BC,GAAE,OAAO;AAAA,EACnD,eAAeA,GAAE,QAAQ,oCAAoC;AAAA,EAAG,eAAeA,GAAE,OAAO,EAAE,MAAM,gBAAgB;AAAA,EAChH,QAAQA,GAAE,MAAM,eAAe,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG;AAAA,EAAG,OAAOA,GAAE,OAAO,EAAE,IAAI,CAAC,EAAE,IAAI,CAAC,EAAE,SAAS;AAAA,EAC3F,QAAQA,GAAE,QAAQ;AAAA,EAAG,gBAAgBA,GAAE,QAAQ;AAAA,EAAG,eAAeA,GAAE,KAAK,CAAC,UAAU,cAAc,gBAAgB,CAAC;AAAA,EAClH,cAAc,mBAAmB,SAAS;AAAA,EAAG,aAAaA,GAAE,OAAO,EAAE,MAAM,gBAAgB;AAC7F,CAAC,EAAE,OAAO,EAAE,YAAY,CAAC,OAAO,YAAY;AAC1C,QAAM,EAAE,aAAa,QAAQ,GAAG,KAAK,IAAI;AACzC,MAAI,YAAY,IAAI,MAAM,OAAQ,SAAQ,SAAS,EAAE,MAAM,UAAU,SAAS,uCAAuC,CAAC;AACtH,QAAM,aAAa,MAAM,OAAO,QAAQ,WAAS,MAAM,UAAU;AACjE,MAAI,WAAW,WAAW,MAAM,OAAO,UAAU,IAAI,IAAI,WAAW,IAAI,eAAa,UAAU,QAAQ,CAAC,EAAE,SAAS,WAAW;AAC5H,YAAQ,SAAS,EAAE,MAAM,UAAU,SAAS,uEAAuE,CAAC;AACtH,aAAW,QAAQ,MAAM,QAAQ;AAC/B,UAAM,EAAE,aAAa,UAAU,GAAG,SAAS,IAAI;AAC/C,QAAI,YAAY,QAAQ,MAAM,SAAU,SAAQ,SAAS,EAAE,MAAM,UAAU,SAAS,gCAAgC,CAAC;AACrH,eAAW,aAAa,KAAK,YAAY;AACvC,YAAM,EAAE,aAAa,eAAe,GAAG,cAAc,IAAI;AACzD,UAAI,YAAY,aAAa,MAAM,cAAe,SAAQ,SAAS,EAAE,MAAM,UAAU,SAAS,oCAAoC,CAAC;AAAA,IACrI;AAAA,EACF;AACF,CAAC;;;AFdD,IAAM,WAAWC,GAAE,OAAO,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG;AACjD,IAAM,aAAaA,GAAE,OAAO,EAAE,MAAM,gBAAgB;AAC7C,IAAM,kCAAkCA,GAAE,mBAAmB,QAAQ;AAAA,EAC1EA,GAAE,OAAO,EAAE,MAAMA,GAAE,QAAQ,aAAa,GAAG,SAAS,UAAU,UAAU,yBAAyB,SAAS,GAAG,iBAAiBA,GAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,IAAI,IAAK,EAAE,QAAQ,IAAK,GAAG,aAAaA,GAAE,OAAO,EAAE,IAAI,CAAC,EAAE,IAAI,CAAC,EAAE,QAAQ,CAAC,GAAG,MAAMA,GAAE,OAAO,EAAE,GAAG,CAAC,EAAE,IAAI,CAAC,EAAE,QAAQ,CAAC,EAAE,CAAC,EAAE,OAAO;AAAA,EACtRA,GAAE,OAAO,EAAE,MAAMA,GAAE,QAAQ,SAAS,GAAG,WAAW,SAAS,CAAC,EAAE,OAAO;AACvE,CAAC;AAGM,IAAM,mCAAmCA,GAAE,OAAO;AAAA,EACvD,eAAeA,GAAE,QAAQ,wCAAwC;AAAA,EACjE,aAAa;AAAA,EAAU,QAAQ;AAAA,EAAU,gBAAgB,SAAS,SAAS;AAAA,EAC3E,SAAS,+BAA+B,SAAS;AAAA,EACjD,SAAS;AAAA,EAAgC,QAAQ;AAAA,EACjD,iBAAiBA,GAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,IAAI,UAAa,EAAE,QAAQ,CAAC;AAAA,EACrE,QAAQ;AACV,CAAC,EAAE,OAAO;AAEH,IAAM,wCAAwCA,GAAE,OAAO,EAAE,gBAAgB,SAAS,SAAS,GAAG,SAAS,gCAAgC,aAAa,SAAS,SAAS,GAAG,OAAOA,GAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,EAAE,QAAQ,EAAE,EAAE,CAAC,EAAE,OAAO;AACvO,IAAM,mCAAmCA,GAAE,OAAO;AAAA,EACvD,gBAAgB,SAAS,SAAS;AAAA,EAAG,SAAS;AAAA,EAAgC,WAAWA,GAAE,QAAQ;AAAA,EAAG,mBAAmBA,GAAE,OAAO,EAAE,IAAI,GAAK,EAAE,SAAS;AAAA,EACxJ,QAAQA,GAAE,MAAMA,GAAE,OAAO,EAAE,IAAI,UAAU,MAAMA,GAAE,OAAO,EAAE,IAAI,GAAG,EAAE,CAAC,EAAE,OAAO,CAAC,EAAE,IAAI,GAAG;AAAA,EACvF,OAAOA,GAAE,MAAMA,GAAE,OAAO,EAAE,IAAI,UAAU,OAAOA,GAAE,KAAK,CAAC,SAAS,cAAc,aAAa,CAAC,GAAG,cAAcA,GAAE,OAAO,EAAE,IAAI,GAAG,GAAG,UAAUA,GAAE,MAAMA,GAAE,OAAO,EAAE,IAAI,UAAU,OAAOA,GAAE,OAAO,EAAE,IAAI,GAAG,EAAE,CAAC,EAAE,OAAO,CAAC,EAAE,IAAI,GAAK,EAAE,CAAC,EAAE,OAAO,CAAC,EAAE,IAAI,GAAG;AAAA,EACpP,YAAY,SAAS,SAAS;AAChC,CAAC,EAAE,OAAO;AACH,IAAM,mCAAmCA,GAAE,OAAO;AAAA,EACvD,eAAeA,GAAE,QAAQ,wCAAwC;AAAA,EACjE,IAAI;AAAA,EAAU,UAAUA,GAAE,OAAO,EAAE,IAAI,EAAE,SAAS;AAAA,EAAG,SAAS;AAAA,EAC9D,QAAQA,GAAE,KAAK,CAAC,UAAU,WAAW,cAAc,aAAa,UAAU,WAAW,CAAC;AAAA,EACtF,gBAAgBA,GAAE,OAAO,EAAE,SAAS,UAAU,UAAU,UAAU,iBAAiBA,GAAE,OAAO,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,GAAG,mBAAmB,WAAW,CAAC,EAAE,OAAO,EAAE,SAAS;AAAA,EAClK,WAAWA,GAAE,IAAI,SAAS;AAAA,EAAG,WAAWA,GAAE,IAAI,SAAS,EAAE,SAAS;AAAA,EAAG,aAAaA,GAAE,IAAI,SAAS,EAAE,SAAS;AAAA,EAC5G,iBAAiBA,GAAE,QAAQ;AAAA,EAAG,iBAAiBA,GAAE,QAAQ;AAAA,EAAG,OAAOA,GAAE,OAAO,EAAE,IAAI,CAAC,EAAE,IAAI,CAAC,EAAE,SAAS;AAAA,EACrG,eAAeA,GAAE,KAAK,CAAC,eAAe,UAAU,cAAc,gBAAgB,CAAC,EAAE,QAAQ,aAAa;AAAA,EACtG,QAAQA,GAAE,QAAQ,EAAE,SAAS;AAAA,EAAG,eAAeA,GAAE,OAAO,EAAE,IAAI,GAAG,EAAE,SAAS;AAAA,EAC5E,OAAOA,GAAE,OAAO,EAAE,MAAM,UAAU,SAASA,GAAE,OAAO,EAAE,IAAI,GAAK,EAAE,CAAC,EAAE,OAAO,EAAE,SAAS;AACxF,CAAC,EAAE,OAAO,EAAE,YAAY,CAAC,OAAO,YAAY;AAC1C,MAAK,MAAM,QAAQ,OAAO,SAAS,eAAgB,MAAM,mBAAmB,MAAO,SAAQ,SAAS,EAAE,MAAM,UAAU,MAAM,CAAC,gBAAgB,GAAG,SAAS,qDAAqD,CAAC;AAC/M,MAAI,MAAM,WAAW,gBAAgB,CAAC,MAAM,mBAAmB,CAAC,MAAM,mBAAmB,CAAC,MAAM,aAAc,SAAQ,SAAS,EAAE,MAAM,UAAU,SAAS,qEAAqE,CAAC;AAClO,CAAC;AAEM,IAAM,qCAAqCA,GAAE,OAAO,EAAE,gBAAgB,UAAU,SAAS,SAAS,SAAS,GAAG,OAAOA,GAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,EAAE,QAAQ,EAAE,EAAE,CAAC,EAAE,OAAO;AAC5K,IAAM,gCAAgCA,GAAE,OAAO,EAAE,OAAOA,GAAE,MAAM,gCAAgC,EAAE,IAAI,GAAG,GAAG,YAAY,SAAS,SAAS,EAAE,CAAC,EAAE,OAAO;AACtJ,IAAM,kCAAkCA,GAAE,OAAO;AAAA,EACtD,eAAeA,GAAE,QAAQ,uCAAuC;AAAA,EAChE,SAAS;AAAA,EACT,QAAQA,GAAE,OAAO,EAAE,IAAI,OAAS,EAAE,SAAS;AAAA;AAAA,EAE3C,UAAUA,GAAE,MAAMA,GAAE,OAAOA,GAAE,OAAO,GAAGA,GAAE,QAAQ,CAAC,CAAC,EAAE,IAAI,IAAK;AAAA,EAC9D,aAAa,wBAAwB,SAAS;AAAA;AAAA;AAAA,EAG9C,OAAOC,iBAAgB,SAAS,EAAE,SAAS;AAAA,EAC3C,iBAAiB,6BAA6B,SAAS,EAAE,SAAS;AAAA,EAClE,aAAaD,GAAE,OAAO,EAAE,QAAQA,GAAE,KAAK,CAAC,aAAa,oBAAoB,aAAa,aAAa,kBAAkB,qBAAqB,CAAC,GAAG,WAAWA,GAAE,QAAQ,GAAG,YAAY,gCAAgC,kBAAkB,WAAW,SAAS,GAAG,gBAAgB,WAAW,SAAS,GAAG,aAAa,WAAW,CAAC,EAAE,OAAO;AAAA,EACpU,oBAAoBA,GAAE,MAAM,QAAQ,EAAE,IAAI,IAAK;AAAA,EAC/C,aAAa;AACf,CAAC,EAAE,OAAO,EAAE,YAAY,CAAC,OAAO,YAAY;AAC1C,MAAI,MAAM,iBAAiB;AACzB,QAAI,MAAM,SAAS,MAAM,YAAa,SAAQ,SAAS,EAAE,MAAM,UAAU,SAAS,qEAAqE,CAAC;AACxJ,QAAI,CAAC,CAAC,aAAa,QAAQ,EAAE,SAAS,MAAM,QAAQ,MAAM,KAAK,CAAC,MAAM,QAAQ,mBAAmB,MAAM,QAAQ,UAAU,MAAM,gBAAgB,SAC1I,MAAM,QAAQ,kBAAkB,MAAM,gBAAgB,iBAAiB,MAAM,QAAQ,YAAY,MAAM,gBAAgB,UAAU,OAAO,OAAO,MAAM,gBAAgB;AACxK,cAAQ,SAAS,EAAE,MAAM,UAAU,SAAS,gEAAgE,CAAC;AAAA,EACjH;AACA,MAAI,CAAC,MAAM,MAAO;AAClB,MAAI,MAAM,YAAa,SAAQ,SAAS,EAAE,MAAM,UAAU,MAAM,CAAC,OAAO,GAAG,SAAS,4DAA4D,CAAC;AACjJ,MAAI,CAAC,CAAC,aAAa,QAAQ,EAAE,SAAS,MAAM,QAAQ,MAAM,KAAK,CAAC,MAAM,QAAQ,mBAAmB,MAAM,QAAQ,UAAU,MAAM,MAAM,SAAS,MAAM,QAAQ,kBAAkB,MAAM,MAAM,iBAAiB,MAAM,QAAQ,YAAY,MAAM,MAAM,UAAU,OAAO,OAAO,MAAM,MAAM,QAAS,SAAQ,SAAS,EAAE,MAAM,UAAU,MAAM,CAAC,OAAO,GAAG,SAAS,4DAA4D,CAAC;AAC1Z,CAAC;AAGM,IAAM,mCAAN,cAA+C,MAAM;AAAA,EAC1D,YAAqB,QAAyB,MAAc,SAAiB;AAAE,UAAM,OAAO;AAAvE;AAAyB;AAAiD,SAAK,OAAO;AAAA,EAAoC;AAAA,EAA1H;AAAA,EAAyB;AAChD;AAEO,IAAM,qCAAN,MAAyC;AAAA,EACrC;AAAA,EACT,YAAY,SAA+F;AACzG,UAAM,MAAM,IAAI,IAAI,QAAQ,OAAO;AACnC,QAAI,CAAC,QAAQ,OAAO,KAAK,KAAK,CAAC,QAAQ,OAAO,KAAK,KAAK,CAAC,CAAC,UAAU,OAAO,EAAE,SAAS,IAAI,QAAQ,KAAK,IAAI,YAAY,IAAI,YAAY,IAAI,UAAU,IAAI,KAAM,OAAM,IAAI,MAAM,yEAAyE;AACxP,SAAK,WAAW,EAAE,GAAG,SAAS,SAAS,IAAI,SAAS,EAAE,QAAQ,QAAQ,EAAE,EAAE;AAAA,EAC5E;AAAA,EACA,MAAM,OAAO,OAAyD,UAAoC,CAAC,GAAG;AAC5G,UAAM,UAAU,iCAAiC,MAAM,KAAK;AAC5D,QAAI,QAAQ,WAAW,KAAK,SAAS,OAAQ,OAAM,IAAI,MAAM,uCAAuC;AACpG,UAAM,SAAS,KAAK,SAAS,MAAM,KAAK,SAAS,IAAI,QAAQ,SAAS,QAAQ,MAAM,CAAC;AACrF,QAAI,cAAc,OAAO,OAAO,MAAM,cAAc,OAAO,EAAG,OAAM,IAAI,iCAAiC,KAAK,4BAA4B,qDAAqD;AAC/L,WAAO;AAAA,EACT;AAAA,EACA,MAAM,IAAI,IAAY,UAAoC,CAAC,GAAG;AAAE,WAAO,KAAK,SAAS,MAAM,KAAK,SAAS,IAAI,mBAAmB,SAAS,MAAM,EAAE,CAAC,CAAC,IAAI,OAAO,QAAW,QAAQ,MAAM,GAAG,EAAE;AAAA,EAAG;AAAA,EAC/L,MAAM,QAAQ,OAA8D,UAAoC,CAAC,GAAG;AAClH,UAAM,QAAQ,sCAAsC,MAAM,KAAK;AAC/D,UAAM,SAAS,IAAI,gBAAgB,EAAE,GAAI,MAAM,iBAAiB,EAAE,gBAAgB,MAAM,eAAe,IAAI,CAAC,GAAI,WAAW,MAAM,QAAQ,IAAI,UAAU,OAAO,MAAM,QAAQ,QAAQ,GAAG,aAAa,MAAM,QAAQ,aAAa,OAAO,OAAO,MAAM,KAAK,EAAE,CAAC;AAC3P,QAAI,MAAM,YAAa,QAAO,IAAI,eAAe,MAAM,WAAW;AAClE,UAAM,SAAS,iCAAiC,MAAM,MAAM,KAAK,SAAS,YAAY,MAAM,IAAI,OAAO,QAAW,QAAQ,MAAM,CAAC;AACjI,QAAI,OAAO,mBAAmB,MAAM,kBAAkB,cAAc,OAAO,OAAO,MAAM,cAAc,MAAM,OAAO,KAAK,OAAO,MAAM,SAAS,MAAM,SAAS,IAAI,IAAI,OAAO,MAAM,IAAI,UAAQ,KAAK,EAAE,CAAC,EAAE,SAAS,OAAO,MAAM,OAAQ,OAAM,IAAI,iCAAiC,KAAK,4BAA4B,4DAA4D;AAC9W,WAAO;AAAA,EACT;AAAA,EACA,MAAM,OAAO,IAAY,UAAoC,CAAC,GAAG;AAAE,WAAO,KAAK,SAAS,MAAM,KAAK,SAAS,IAAI,mBAAmB,SAAS,MAAM,EAAE,CAAC,CAAC,WAAW,QAAQ,QAAW,QAAQ,MAAM,GAAG,EAAE;AAAA,EAAG;AAAA,EAC1M,MAAM,OAAO,IAAY,UAAoC,CAAC,GAAG;AAC/D,UAAM,QAAQ,gCAAgC,MAAM,MAAM,KAAK,SAAS,IAAI,mBAAmB,SAAS,MAAM,EAAE,CAAC,CAAC,WAAW,OAAO,QAAW,QAAQ,MAAM,CAAC;AAC9J,SAAK,SAAS,MAAM,SAAS,EAAE;AAC/B,UAAM,EAAE,aAAAE,cAAa,GAAG,QAAQ,IAAI;AACpC,QAAI,CAAC,MAAM,QAAQ,mBAAmB,MAAM,gBAAgB,OAAO,MAAMA,aAAa,OAAM,IAAI,iCAAiC,KAAK,4BAA4B,kCAAkC;AACpM,WAAO;AAAA,EACT;AAAA,EACA,MAAM,KAAK,OAA2D,UAAoC,CAAC,GAAG;AAC5G,UAAM,QAAQ,mCAAmC,MAAM,KAAK;AAC5D,UAAM,SAAS,IAAI,gBAAgB,EAAE,gBAAgB,MAAM,gBAAgB,OAAO,OAAO,MAAM,KAAK,EAAE,CAAC;AACvG,QAAI,MAAM,QAAS,QAAO,IAAI,WAAW,MAAM,OAAO;AACtD,UAAM,OAAO,8BAA8B,MAAM,MAAM,KAAK,SAAS,IAAI,MAAM,IAAI,OAAO,QAAW,QAAQ,MAAM,CAAC;AACpH,QAAI,KAAK,MAAM,SAAS,MAAM,SAAS,IAAI,IAAI,KAAK,MAAM,IAAI,UAAQ,KAAK,EAAE,CAAC,EAAE,SAAS,KAAK,MAAM,UAAU,KAAK,MAAM,KAAK,UAAQ,KAAK,SAAS,IAAI,EAAE,QAAQ,mBAAmB,MAAM,cAAc,EAAG,OAAM,IAAI,iCAAiC,KAAK,yBAAyB,+CAA+C;AACpU,WAAO;AAAA,EACT;AAAA,EACA,SAAS,OAAgB,IAAyC;AAChE,UAAM,SAAS,iCAAiC,MAAM,KAAK;AAC3D,QAAI,OAAO,QAAQ,WAAW,KAAK,SAAS,UAAW,MAAM,OAAO,OAAO,GAAK,OAAM,IAAI,iCAAiC,KAAK,6BAA6B,2DAA2D;AACxN,WAAO;AAAA,EACT;AAAA,EACA,MAAM,SAAS,MAAc,QAAgB,MAAe,QAAwC;AAClG,UAAM,WAAW,OAAO,KAAK,SAAS,SAAS,WAAW,OAAO,GAAG,KAAK,SAAS,OAAO,6BAA6B,IAAI,IAAI,EAAE,QAAQ,SAAS,EAAE,eAAe,UAAU,KAAK,SAAS,MAAM,IAAI,sBAAsB,KAAK,SAAS,QAAQ,gBAAgB,oBAAoB,QAAQ,mBAAmB,GAAG,GAAI,SAAS,SAAY,CAAC,IAAI,EAAE,MAAM,KAAK,UAAU,IAAI,EAAE,GAAI,QAAQ,UAAU,QAAQ,CAAC;AAC5Y,UAAM,SAAS,SAAS,MAAM,UAAU;AACxC,QAAI,CAAC,OAAQ,OAAM,IAAI,iCAAiC,SAAS,QAAQ,0BAA0B,6BAA6B;AAChI,QAAI,QAAQ;AAAG,QAAI,OAAO;AAAI,UAAM,UAAU,IAAI,YAAY;AAC9D,QAAI;AACF,aAAO,MAAM;AACX,cAAM,QAAQ,MAAM,OAAO,KAAK;AAAG,YAAI,MAAM,KAAM;AACnD,iBAAS,MAAM,MAAM;AACrB,YAAI,QAAQ,SAAY,OAAM,IAAI,iCAAiC,KAAK,8BAA8B,2CAA2C;AACjJ,gBAAQ,QAAQ,OAAO,MAAM,OAAO,EAAE,QAAQ,KAAK,CAAC;AAAA,MACtD;AACA,cAAQ,QAAQ,OAAO;AAAA,IACzB,UAAE;AAAU,YAAM,OAAO,OAAO;AAAG,aAAO,YAAY;AAAA,IAAG;AACzD,UAAM,QAAiB,KAAK,MAAM,IAAI;AACtC,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,QAAQF,GAAE,OAAO,EAAE,MAAM,UAAU,SAASA,GAAE,OAAO,EAAE,IAAI,GAAK,EAAE,CAAC,EAAE,UAAU,KAAK;AAC1F,YAAM,IAAI,iCAAiC,SAAS,QAAQ,MAAM,UAAU,MAAM,KAAK,OAAO,0BAA0B,MAAM,UAAU,MAAM,KAAK,UAAU,yBAAyB;AAAA,IACxL;AACA,WAAO;AAAA,EACT;AACF;",
+  "names": ["z", "TaskGradeSchema", "z", "z", "z", "TaskGradeSchema", "contentHash"]
+}

package/dist/chunk-ZV2BLO3D.js

@@ -0,0 +1,123 @@
+import {
+  EvaluationProjectContextSchema,
+  ModelStarterAttemptPolicySchema,
+  ModelStarterAttemptSummarySchema
+} from "./chunk-XR6CLSCS.js";
+import {
+  ModelProjectVersionedRefSchema,
+  canonicalJson,
+  canonicalSha256
+} from "./chunk-TCHAWJMB.js";
+import {
+  assertContentHash
+} from "./chunk-QUOW6URQ.js";
+
+// src/model-taskset-runs-contracts.ts
+import { z } from "zod";
+import { TasksetRunManifestSchema, TasksetRunMemberSchema, TasksetMetricResultSchema, assertTasksetMetricResult, orderTasksetRunReceipts } from "@openpond/evals/metrics";
+import { AttemptReceiptSchema } from "@openpond/evals/runs";
+var IdSchema = z.string().trim().min(1).max(200);
+var HashSchema = z.string().regex(/^[a-f0-9]{64}$/);
+var ModelTasksetRunPolicySchema = z.discriminatedUnion("kind", [
+  ModelStarterAttemptPolicySchema.options[0],
+  z.object({ kind: z.literal("fixture") }).strict()
+]);
+var ModelTasksetRunRequestSchema = z.object({
+  schemaVersion: z.literal("openpond.modelTasksetRunRequest.v1"),
+  operationId: IdSchema,
+  teamId: IdSchema,
+  modelProjectId: IdSchema.nullable(),
+  name: z.string().trim().min(1).max(200).optional(),
+  project: EvaluationProjectContextSchema.optional(),
+  taskset: ModelProjectVersionedRefSchema,
+  policy: ModelTasksetRunPolicySchema,
+  population: z.array(TasksetRunMemberSchema).min(1).max(1e4)
+}).strict().superRefine((value, context) => {
+  if (new Set(value.population.map((member) => member.receiptId)).size !== value.population.length) context.addIssue({ code: "custom", path: ["population"], message: "Run receipt identities must be unique." });
+  if (value.population.some((member) => member.fixtureId !== null !== (value.policy.kind === "fixture"))) context.addIssue({ code: "custom", path: ["population"], message: "Only fixture runs require a fixture identity for every member." });
+  if (value.population.some((member) => !/^(0|[1-9][0-9]*)$/.test(member.seed) || Number(member.seed) > 2147483647)) context.addIssue({ code: "custom", path: ["population"], message: "Hosted environment seeds must be integers from 0 to 2147483647." });
+});
+var ModelTasksetRunSummarySchema = z.object({
+  schemaVersion: z.literal("openpond.modelTasksetRunSummary.v1"),
+  id: IdSchema,
+  revision: z.number().int().positive(),
+  teamId: IdSchema,
+  modelProjectId: IdSchema.nullable(),
+  name: z.string().trim().min(1).max(200).optional(),
+  project: EvaluationProjectContextSchema.optional(),
+  operationId: IdSchema,
+  taskset: ModelProjectVersionedRefSchema,
+  policyKind: z.enum(["hosted_chat", "fixture"]),
+  manifestHash: HashSchema,
+  status: z.enum(["queued", "running", "cancelling", "completed", "failed", "cancelled"]),
+  totalCount: z.number().int().min(1).max(1e4),
+  counts: z.object({ pending: z.number().int().nonnegative(), running: z.number().int().nonnegative(), completed: z.number().int().nonnegative(), failed: z.number().int().nonnegative(), cancelled: z.number().int().nonnegative() }).strict(),
+  createdAt: z.iso.datetime(),
+  startedAt: z.iso.datetime().nullable(),
+  completedAt: z.iso.datetime().nullable(),
+  cleanupComplete: z.boolean(),
+  resultAvailable: z.boolean(),
+  score: z.number().min(0).max(1).nullable(),
+  metricName: z.string().min(1).max(240),
+  error: z.object({ code: IdSchema, message: z.string().max(2e3) }).strict().nullable()
+}).strict().superRefine((value, context) => {
+  if (Object.values(value.counts).reduce((sum, count) => sum + count, 0) !== value.totalCount) context.addIssue({ code: "custom", path: ["counts"], message: "Run counts must account for the complete population." });
+  if (value.status !== "completed" && value.score !== null) context.addIssue({ code: "custom", path: ["score"], message: "Only completed evaluations publish a metric value." });
+  if (["completed", "failed", "cancelled"].includes(value.status) && (!value.completedAt || value.counts.pending || value.counts.running)) context.addIssue({ code: "custom", message: "Terminal runs require complete population accounting." });
+  if (value.status === "cancelled" && !value.cleanupComplete) context.addIssue({ code: "custom", message: "Cancelled runs require confirmed cleanup." });
+  if (value.resultAvailable && !["completed", "failed", "cancelled"].includes(value.status)) context.addIssue({ code: "custom", message: "Run results require terminal population accounting." });
+  if (value.status === "completed" && (!value.cleanupComplete || !value.resultAvailable)) context.addIssue({ code: "custom", message: "Completed evaluations require retained results and confirmed cleanup." });
+});
+var ModelTasksetRunDetailsSchema = z.object({
+  summary: ModelTasksetRunSummarySchema,
+  request: ModelTasksetRunRequestSchema,
+  manifest: TasksetRunManifestSchema,
+  policySnapshot: ModelStarterAttemptSummarySchema.shape.policySnapshot
+}).strict();
+var ModelTasksetRunListQuerySchema = z.object({ modelProjectId: IdSchema.optional(), projectId: EvaluationProjectContextSchema.shape.id.optional(), afterId: IdSchema.optional(), limit: z.number().int().min(1).max(100).default(25) }).strict();
+var ModelTasksetRunPageSchema = z.object({ items: z.array(ModelTasksetRunSummarySchema).max(100), nextCursor: IdSchema.nullable() }).strict();
+var ModelTasksetRunResultSchema = z.object({
+  schemaVersion: z.literal("openpond.modelTasksetRunResult.v1"),
+  run: ModelTasksetRunDetailsSchema,
+  receipts: z.array(AttemptReceiptSchema).min(1).max(1e4),
+  metric: TasksetMetricResultSchema.nullable(),
+  contentHash: HashSchema
+}).strict();
+async function verifyModelTasksetRunDetails(value) {
+  const result = ModelTasksetRunDetailsSchema.parse(value);
+  const { summary, request, manifest, policySnapshot } = result;
+  assertContentHash(manifest, "Taskset run manifest");
+  if (manifest.contentHash !== summary.manifestHash || manifest.id !== summary.id || manifest.createdAt !== summary.createdAt || summary.teamId !== request.teamId || summary.modelProjectId !== request.modelProjectId || canonicalJson(summary.project ?? null) !== canonicalJson(request.project ?? null) || canonicalJson(manifest.metadata.project ?? null) !== canonicalJson(request.project ?? null) || (summary.name ?? null) !== (request.name ?? null) || summary.operationId !== request.operationId || summary.policyKind !== request.policy.kind || summary.totalCount !== request.population.length || summary.metricName !== manifest.metricPolicy.primaryMetric || canonicalJson(summary.taskset) !== canonicalJson(request.taskset) || manifest.tasksetRelease.id !== request.taskset.id || manifest.tasksetRelease.contentHash !== request.taskset.contentHash || canonicalJson(manifest.population) !== canonicalJson(request.population)) throw new Error("Evaluation run differs from its admitted request or manifest.");
+  if (request.policy.kind === "fixture") {
+    if (manifest.policy.kind !== "fixture" || policySnapshot !== null) throw new Error("Fixture checks cannot claim a model identity.");
+  } else if (manifest.policy.kind !== "model" || !policySnapshot || policySnapshot.modelId !== request.policy.modelId || manifest.policy.model.provider !== policySnapshot.provider || manifest.policy.model.model !== policySnapshot.upstreamModelId || manifest.policy.model.revision !== null || manifest.policy.model.artifactHash !== null || manifest.policy.model.tokenizerRevision !== null || manifest.policy.model.chatTemplateHash !== null || manifest.policy.configurationHash !== await canonicalSha256({ policy: request.policy, snapshot: policySnapshot })) throw new Error("Evaluation model differs from its admitted policy.");
+  return result;
+}
+async function verifyModelTasksetRunResult(value) {
+  const result = ModelTasksetRunResultSchema.parse(value);
+  const { contentHash: actual, ...content } = result;
+  if (await canonicalSha256(content) !== actual) throw new Error("Evaluation result integrity failed.");
+  const { summary, manifest } = await verifyModelTasksetRunDetails(result.run);
+  if (!summary.resultAvailable || !["completed", "failed", "cancelled"].includes(summary.status)) throw new Error("Evaluation result is not terminal.");
+  const ordered = orderTasksetRunReceipts(manifest, result.receipts);
+  if (canonicalJson(ordered) !== canonicalJson(result.receipts)) throw new Error("Evaluation receipts differ from the admitted order.");
+  if (summary.status === "completed" !== (result.metric !== null)) throw new Error("Only completed evaluations retain a metric.");
+  if (result.metric) {
+    assertTasksetMetricResult(result.metric);
+    if (result.metric.runManifest.id !== manifest.id || result.metric.runManifest.contentHash !== manifest.contentHash || canonicalJson(result.metric.tasksetRelease) !== canonicalJson(manifest.tasksetRelease) || canonicalJson(result.metric.policy) !== canonicalJson(manifest.metricPolicy) || canonicalJson(result.metric.receiptRefs) !== canonicalJson(result.receipts.map(({ id, contentHash }) => ({ id, contentHash }))) || result.metric.value !== summary.score) throw new Error("Evaluation metric differs from its admitted run or receipt population.");
+  }
+  return result;
+}
+
+export {
+  ModelTasksetRunPolicySchema,
+  ModelTasksetRunRequestSchema,
+  ModelTasksetRunSummarySchema,
+  ModelTasksetRunDetailsSchema,
+  ModelTasksetRunListQuerySchema,
+  ModelTasksetRunPageSchema,
+  ModelTasksetRunResultSchema,
+  verifyModelTasksetRunDetails,
+  verifyModelTasksetRunResult
+};
+//# sourceMappingURL=chunk-ZV2BLO3D.js.map

package/dist/chunk-ZV2BLO3D.js.map

@@ -0,0 +1,7 @@
+{
+  "version": 3,
+  "sources": ["../src/model-taskset-runs-contracts.ts"],
+  "sourcesContent": ["import { z } from \"zod\";\nimport { EvaluationProjectContextSchema } from \"./evaluation-project-context.js\";\nimport { assertContentHash } from \"@openpond/harness\";\nimport { TasksetRunManifestSchema, TasksetRunMemberSchema, TasksetMetricResultSchema, assertTasksetMetricResult, orderTasksetRunReceipts } from \"@openpond/evals/metrics\";\nimport { AttemptReceiptSchema } from \"@openpond/evals/runs\";\nimport { ModelProjectVersionedRefSchema } from \"./model-projects.js\";\nimport { ModelStarterAttemptPolicySchema, ModelStarterAttemptSummarySchema } from \"./model-starter-attempts.js\";\nimport { canonicalJson, canonicalSha256 } from \"./protocol.js\";\n\nconst IdSchema = z.string().trim().min(1).max(200);\nconst HashSchema = z.string().regex(/^[a-f0-9]{64}$/);\nexport const ModelTasksetRunPolicySchema = z.discriminatedUnion(\"kind\", [\n  ModelStarterAttemptPolicySchema.options[0],\n  z.object({ kind: z.literal(\"fixture\") }).strict(),\n]);\nexport const ModelTasksetRunRequestSchema = z.object({\n  schemaVersion: z.literal(\"openpond.modelTasksetRunRequest.v1\"),\n  operationId: IdSchema, teamId: IdSchema, modelProjectId: IdSchema.nullable(),\n  name: z.string().trim().min(1).max(200).optional(),\n  project: EvaluationProjectContextSchema.optional(),\n  taskset: ModelProjectVersionedRefSchema, policy: ModelTasksetRunPolicySchema,\n  population: z.array(TasksetRunMemberSchema).min(1).max(10_000),\n}).strict().superRefine((value, context) => {\n  if (new Set(value.population.map(member => member.receiptId)).size !== value.population.length) context.addIssue({ code: \"custom\", path: [\"population\"], message: \"Run receipt identities must be unique.\" });\n  if (value.population.some(member => (member.fixtureId !== null) !== (value.policy.kind === \"fixture\"))) context.addIssue({ code: \"custom\", path: [\"population\"], message: \"Only fixture runs require a fixture identity for every member.\" });\n  if (value.population.some(member => !/^(0|[1-9][0-9]*)$/.test(member.seed) || Number(member.seed) > 2_147_483_647)) context.addIssue({ code: \"custom\", path: [\"population\"], message: \"Hosted environment seeds must be integers from 0 to 2147483647.\" });\n});\nexport type ModelTasksetRunRequest = z.infer<typeof ModelTasksetRunRequestSchema>;\nexport const ModelTasksetRunSummarySchema = z.object({\n  schemaVersion: z.literal(\"openpond.modelTasksetRunSummary.v1\"),\n  id: IdSchema, revision: z.number().int().positive(), teamId: IdSchema, modelProjectId: IdSchema.nullable(),\n  name: z.string().trim().min(1).max(200).optional(),\n  project: EvaluationProjectContextSchema.optional(),\n  operationId: IdSchema, taskset: ModelProjectVersionedRefSchema, policyKind: z.enum([\"hosted_chat\", \"fixture\"]),\n  manifestHash: HashSchema,\n  status: z.enum([\"queued\", \"running\", \"cancelling\", \"completed\", \"failed\", \"cancelled\"]),\n  totalCount: z.number().int().min(1).max(10_000),\n  counts: z.object({ pending: z.number().int().nonnegative(), running: z.number().int().nonnegative(), completed: z.number().int().nonnegative(), failed: z.number().int().nonnegative(), cancelled: z.number().int().nonnegative() }).strict(),\n  createdAt: z.iso.datetime(), startedAt: z.iso.datetime().nullable(), completedAt: z.iso.datetime().nullable(),\n  cleanupComplete: z.boolean(), resultAvailable: z.boolean(), score: z.number().min(0).max(1).nullable(),\n  metricName: z.string().min(1).max(240),\n  error: z.object({ code: IdSchema, message: z.string().max(2_000) }).strict().nullable(),\n}).strict().superRefine((value, context) => {\n  if (Object.values(value.counts).reduce((sum, count) => sum + count, 0) !== value.totalCount) context.addIssue({ code: \"custom\", path: [\"counts\"], message: \"Run counts must account for the complete population.\" });\n  if (value.status !== \"completed\" && value.score !== null) context.addIssue({ code: \"custom\", path: [\"score\"], message: \"Only completed evaluations publish a metric value.\" });\n  if ([\"completed\", \"failed\", \"cancelled\"].includes(value.status) && (!value.completedAt || value.counts.pending || value.counts.running)) context.addIssue({ code: \"custom\", message: \"Terminal runs require complete population accounting.\" });\n  if (value.status === \"cancelled\" && !value.cleanupComplete) context.addIssue({ code: \"custom\", message: \"Cancelled runs require confirmed cleanup.\" });\n  if (value.resultAvailable && ![\"completed\", \"failed\", \"cancelled\"].includes(value.status)) context.addIssue({ code: \"custom\", message: \"Run results require terminal population accounting.\" });\n  if (value.status === \"completed\" && (!value.cleanupComplete || !value.resultAvailable)) context.addIssue({ code: \"custom\", message: \"Completed evaluations require retained results and confirmed cleanup.\" });\n});\nexport type ModelTasksetRunSummary = z.infer<typeof ModelTasksetRunSummarySchema>;\nexport const ModelTasksetRunDetailsSchema = z.object({\n  summary: ModelTasksetRunSummarySchema,\n  request: ModelTasksetRunRequestSchema,\n  manifest: TasksetRunManifestSchema,\n  policySnapshot: ModelStarterAttemptSummarySchema.shape.policySnapshot,\n}).strict();\nexport type ModelTasksetRunDetails = z.infer<typeof ModelTasksetRunDetailsSchema>;\nexport const ModelTasksetRunListQuerySchema = z.object({ modelProjectId: IdSchema.optional(), projectId: EvaluationProjectContextSchema.shape.id.optional(), afterId: IdSchema.optional(), limit: z.number().int().min(1).max(100).default(25) }).strict();\nexport const ModelTasksetRunPageSchema = z.object({ items: z.array(ModelTasksetRunSummarySchema).max(100), nextCursor: IdSchema.nullable() }).strict();\nexport const ModelTasksetRunResultSchema = z.object({\n  schemaVersion: z.literal(\"openpond.modelTasksetRunResult.v1\"),\n  run: ModelTasksetRunDetailsSchema,\n  receipts: z.array(AttemptReceiptSchema).min(1).max(10_000),\n  metric: TasksetMetricResultSchema.nullable(),\n  contentHash: HashSchema,\n}).strict();\nexport type ModelTasksetRunResult = z.infer<typeof ModelTasksetRunResultSchema>;\n\n/** Server artifact ownership authenticates the producer. These checks bind all\n * portable projections to the admitted request, population and policy. */\nexport async function verifyModelTasksetRunDetails(value: unknown): Promise<ModelTasksetRunDetails> {\n  const result = ModelTasksetRunDetailsSchema.parse(value);\n  const { summary, request, manifest, policySnapshot } = result;\n  assertContentHash(manifest, \"Taskset run manifest\");\n  if (manifest.contentHash !== summary.manifestHash\n    || manifest.id !== summary.id || manifest.createdAt !== summary.createdAt || summary.teamId !== request.teamId || summary.modelProjectId !== request.modelProjectId\n    || canonicalJson(summary.project ?? null) !== canonicalJson(request.project ?? null)\n    || canonicalJson(manifest.metadata.project ?? null) !== canonicalJson(request.project ?? null)\n    || (summary.name ?? null) !== (request.name ?? null)\n    || summary.operationId !== request.operationId || summary.policyKind !== request.policy.kind\n    || summary.totalCount !== request.population.length || summary.metricName !== manifest.metricPolicy.primaryMetric\n    || canonicalJson(summary.taskset) !== canonicalJson(request.taskset)\n    || manifest.tasksetRelease.id !== request.taskset.id || manifest.tasksetRelease.contentHash !== request.taskset.contentHash\n    || canonicalJson(manifest.population) !== canonicalJson(request.population)) throw new Error(\"Evaluation run differs from its admitted request or manifest.\");\n  if (request.policy.kind === \"fixture\") {\n    if (manifest.policy.kind !== \"fixture\" || policySnapshot !== null) throw new Error(\"Fixture checks cannot claim a model identity.\");\n  } else if (manifest.policy.kind !== \"model\" || !policySnapshot || policySnapshot.modelId !== request.policy.modelId\n    || manifest.policy.model.provider !== policySnapshot.provider || manifest.policy.model.model !== policySnapshot.upstreamModelId\n    || manifest.policy.model.revision !== null || manifest.policy.model.artifactHash !== null || manifest.policy.model.tokenizerRevision !== null || manifest.policy.model.chatTemplateHash !== null\n    || manifest.policy.configurationHash !== await canonicalSha256({ policy: request.policy, snapshot: policySnapshot })) throw new Error(\"Evaluation model differs from its admitted policy.\");\n  return result;\n}\n\nexport async function verifyModelTasksetRunResult(value: unknown): Promise<ModelTasksetRunResult> {\n  const result = ModelTasksetRunResultSchema.parse(value);\n  const { contentHash: actual, ...content } = result;\n  if (await canonicalSha256(content) !== actual) throw new Error(\"Evaluation result integrity failed.\");\n  const { summary, manifest } = await verifyModelTasksetRunDetails(result.run);\n  if (!summary.resultAvailable || ![\"completed\", \"failed\", \"cancelled\"].includes(summary.status)) throw new Error(\"Evaluation result is not terminal.\");\n  const ordered = orderTasksetRunReceipts(manifest, result.receipts);\n  if (canonicalJson(ordered) !== canonicalJson(result.receipts)) throw new Error(\"Evaluation receipts differ from the admitted order.\");\n  if ((summary.status === \"completed\") !== (result.metric !== null)) throw new Error(\"Only completed evaluations retain a metric.\");\n  if (result.metric) {\n    assertTasksetMetricResult(result.metric);\n    if (result.metric.runManifest.id !== manifest.id || result.metric.runManifest.contentHash !== manifest.contentHash\n      || canonicalJson(result.metric.tasksetRelease) !== canonicalJson(manifest.tasksetRelease)\n      || canonicalJson(result.metric.policy) !== canonicalJson(manifest.metricPolicy)\n      || canonicalJson(result.metric.receiptRefs) !== canonicalJson(result.receipts.map(({ id, contentHash }) => ({ id, contentHash })))\n      || result.metric.value !== summary.score) throw new Error(\"Evaluation metric differs from its admitted run or receipt population.\");\n  }\n  return result;\n}\n"],
+  "mappings": ";;;;;;;;;;;;;;;AAAA,SAAS,SAAS;AAGlB,SAAS,0BAA0B,wBAAwB,2BAA2B,2BAA2B,+BAA+B;AAChJ,SAAS,4BAA4B;AAKrC,IAAM,WAAW,EAAE,OAAO,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG;AACjD,IAAM,aAAa,EAAE,OAAO,EAAE,MAAM,gBAAgB;AAC7C,IAAM,8BAA8B,EAAE,mBAAmB,QAAQ;AAAA,EACtE,gCAAgC,QAAQ,CAAC;AAAA,EACzC,EAAE,OAAO,EAAE,MAAM,EAAE,QAAQ,SAAS,EAAE,CAAC,EAAE,OAAO;AAClD,CAAC;AACM,IAAM,+BAA+B,EAAE,OAAO;AAAA,EACnD,eAAe,EAAE,QAAQ,oCAAoC;AAAA,EAC7D,aAAa;AAAA,EAAU,QAAQ;AAAA,EAAU,gBAAgB,SAAS,SAAS;AAAA,EAC3E,MAAM,EAAE,OAAO,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,EAAE,SAAS;AAAA,EACjD,SAAS,+BAA+B,SAAS;AAAA,EACjD,SAAS;AAAA,EAAgC,QAAQ;AAAA,EACjD,YAAY,EAAE,MAAM,sBAAsB,EAAE,IAAI,CAAC,EAAE,IAAI,GAAM;AAC/D,CAAC,EAAE,OAAO,EAAE,YAAY,CAAC,OAAO,YAAY;AAC1C,MAAI,IAAI,IAAI,MAAM,WAAW,IAAI,YAAU,OAAO,SAAS,CAAC,EAAE,SAAS,MAAM,WAAW,OAAQ,SAAQ,SAAS,EAAE,MAAM,UAAU,MAAM,CAAC,YAAY,GAAG,SAAS,yCAAyC,CAAC;AAC5M,MAAI,MAAM,WAAW,KAAK,YAAW,OAAO,cAAc,UAAW,MAAM,OAAO,SAAS,UAAU,EAAG,SAAQ,SAAS,EAAE,MAAM,UAAU,MAAM,CAAC,YAAY,GAAG,SAAS,iEAAiE,CAAC;AAC5O,MAAI,MAAM,WAAW,KAAK,YAAU,CAAC,oBAAoB,KAAK,OAAO,IAAI,KAAK,OAAO,OAAO,IAAI,IAAI,UAAa,EAAG,SAAQ,SAAS,EAAE,MAAM,UAAU,MAAM,CAAC,YAAY,GAAG,SAAS,kEAAkE,CAAC;AAC3P,CAAC;AAEM,IAAM,+BAA+B,EAAE,OAAO;AAAA,EACnD,eAAe,EAAE,QAAQ,oCAAoC;AAAA,EAC7D,IAAI;AAAA,EAAU,UAAU,EAAE,OAAO,EAAE,IAAI,EAAE,SAAS;AAAA,EAAG,QAAQ;AAAA,EAAU,gBAAgB,SAAS,SAAS;AAAA,EACzG,MAAM,EAAE,OAAO,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,EAAE,SAAS;AAAA,EACjD,SAAS,+BAA+B,SAAS;AAAA,EACjD,aAAa;AAAA,EAAU,SAAS;AAAA,EAAgC,YAAY,EAAE,KAAK,CAAC,eAAe,SAAS,CAAC;AAAA,EAC7G,cAAc;AAAA,EACd,QAAQ,EAAE,KAAK,CAAC,UAAU,WAAW,cAAc,aAAa,UAAU,WAAW,CAAC;AAAA,EACtF,YAAY,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,IAAI,GAAM;AAAA,EAC9C,QAAQ,EAAE,OAAO,EAAE,SAAS,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,GAAG,SAAS,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,GAAG,WAAW,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,GAAG,QAAQ,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,GAAG,WAAW,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,CAAC,EAAE,OAAO;AAAA,EAC5O,WAAW,EAAE,IAAI,SAAS;AAAA,EAAG,WAAW,EAAE,IAAI,SAAS,EAAE,SAAS;AAAA,EAAG,aAAa,EAAE,IAAI,SAAS,EAAE,SAAS;AAAA,EAC5G,iBAAiB,EAAE,QAAQ;AAAA,EAAG,iBAAiB,EAAE,QAAQ;AAAA,EAAG,OAAO,EAAE,OAAO,EAAE,IAAI,CAAC,EAAE,IAAI,CAAC,EAAE,SAAS;AAAA,EACrG,YAAY,EAAE,OAAO,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG;AAAA,EACrC,OAAO,EAAE,OAAO,EAAE,MAAM,UAAU,SAAS,EAAE,OAAO,EAAE,IAAI,GAAK,EAAE,CAAC,EAAE,OAAO,EAAE,SAAS;AACxF,CAAC,EAAE,OAAO,EAAE,YAAY,CAAC,OAAO,YAAY;AAC1C,MAAI,OAAO,OAAO,MAAM,MAAM,EAAE,OAAO,CAAC,KAAK,UAAU,MAAM,OAAO,CAAC,MAAM,MAAM,WAAY,SAAQ,SAAS,EAAE,MAAM,UAAU,MAAM,CAAC,QAAQ,GAAG,SAAS,uDAAuD,CAAC;AACnN,MAAI,MAAM,WAAW,eAAe,MAAM,UAAU,KAAM,SAAQ,SAAS,EAAE,MAAM,UAAU,MAAM,CAAC,OAAO,GAAG,SAAS,qDAAqD,CAAC;AAC7K,MAAI,CAAC,aAAa,UAAU,WAAW,EAAE,SAAS,MAAM,MAAM,MAAM,CAAC,MAAM,eAAe,MAAM,OAAO,WAAW,MAAM,OAAO,SAAU,SAAQ,SAAS,EAAE,MAAM,UAAU,SAAS,wDAAwD,CAAC;AAC9O,MAAI,MAAM,WAAW,eAAe,CAAC,MAAM,gBAAiB,SAAQ,SAAS,EAAE,MAAM,UAAU,SAAS,4CAA4C,CAAC;AACrJ,MAAI,MAAM,mBAAmB,CAAC,CAAC,aAAa,UAAU,WAAW,EAAE,SAAS,MAAM,MAAM,EAAG,SAAQ,SAAS,EAAE,MAAM,UAAU,SAAS,sDAAsD,CAAC;AAC9L,MAAI,MAAM,WAAW,gBAAgB,CAAC,MAAM,mBAAmB,CAAC,MAAM,iBAAkB,SAAQ,SAAS,EAAE,MAAM,UAAU,SAAS,wEAAwE,CAAC;AAC/M,CAAC;AAEM,IAAM,+BAA+B,EAAE,OAAO;AAAA,EACnD,SAAS;AAAA,EACT,SAAS;AAAA,EACT,UAAU;AAAA,EACV,gBAAgB,iCAAiC,MAAM;AACzD,CAAC,EAAE,OAAO;AAEH,IAAM,iCAAiC,EAAE,OAAO,EAAE,gBAAgB,SAAS,SAAS,GAAG,WAAW,+BAA+B,MAAM,GAAG,SAAS,GAAG,SAAS,SAAS,SAAS,GAAG,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,EAAE,QAAQ,EAAE,EAAE,CAAC,EAAE,OAAO;AAClP,IAAM,4BAA4B,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,4BAA4B,EAAE,IAAI,GAAG,GAAG,YAAY,SAAS,SAAS,EAAE,CAAC,EAAE,OAAO;AAC9I,IAAM,8BAA8B,EAAE,OAAO;AAAA,EAClD,eAAe,EAAE,QAAQ,mCAAmC;AAAA,EAC5D,KAAK;AAAA,EACL,UAAU,EAAE,MAAM,oBAAoB,EAAE,IAAI,CAAC,EAAE,IAAI,GAAM;AAAA,EACzD,QAAQ,0BAA0B,SAAS;AAAA,EAC3C,aAAa;AACf,CAAC,EAAE,OAAO;AAKV,eAAsB,6BAA6B,OAAiD;AAClG,QAAM,SAAS,6BAA6B,MAAM,KAAK;AACvD,QAAM,EAAE,SAAS,SAAS,UAAU,eAAe,IAAI;AACvD,oBAAkB,UAAU,sBAAsB;AAClD,MAAI,SAAS,gBAAgB,QAAQ,gBAChC,SAAS,OAAO,QAAQ,MAAM,SAAS,cAAc,QAAQ,aAAa,QAAQ,WAAW,QAAQ,UAAU,QAAQ,mBAAmB,QAAQ,kBAClJ,cAAc,QAAQ,WAAW,IAAI,MAAM,cAAc,QAAQ,WAAW,IAAI,KAChF,cAAc,SAAS,SAAS,WAAW,IAAI,MAAM,cAAc,QAAQ,WAAW,IAAI,MACzF,QAAQ,QAAQ,WAAW,QAAQ,QAAQ,SAC5C,QAAQ,gBAAgB,QAAQ,eAAe,QAAQ,eAAe,QAAQ,OAAO,QACrF,QAAQ,eAAe,QAAQ,WAAW,UAAU,QAAQ,eAAe,SAAS,aAAa,iBACjG,cAAc,QAAQ,OAAO,MAAM,cAAc,QAAQ,OAAO,KAChE,SAAS,eAAe,OAAO,QAAQ,QAAQ,MAAM,SAAS,eAAe,gBAAgB,QAAQ,QAAQ,eAC7G,cAAc,SAAS,UAAU,MAAM,cAAc,QAAQ,UAAU,EAAG,OAAM,IAAI,MAAM,+DAA+D;AAC9J,MAAI,QAAQ,OAAO,SAAS,WAAW;AACrC,QAAI,SAAS,OAAO,SAAS,aAAa,mBAAmB,KAAM,OAAM,IAAI,MAAM,+CAA+C;AAAA,EACpI,WAAW,SAAS,OAAO,SAAS,WAAW,CAAC,kBAAkB,eAAe,YAAY,QAAQ,OAAO,WACvG,SAAS,OAAO,MAAM,aAAa,eAAe,YAAY,SAAS,OAAO,MAAM,UAAU,eAAe,mBAC7G,SAAS,OAAO,MAAM,aAAa,QAAQ,SAAS,OAAO,MAAM,iBAAiB,QAAQ,SAAS,OAAO,MAAM,sBAAsB,QAAQ,SAAS,OAAO,MAAM,qBAAqB,QACzL,SAAS,OAAO,sBAAsB,MAAM,gBAAgB,EAAE,QAAQ,QAAQ,QAAQ,UAAU,eAAe,CAAC,EAAG,OAAM,IAAI,MAAM,oDAAoD;AAC5L,SAAO;AACT;AAEA,eAAsB,4BAA4B,OAAgD;AAChG,QAAM,SAAS,4BAA4B,MAAM,KAAK;AACtD,QAAM,EAAE,aAAa,QAAQ,GAAG,QAAQ,IAAI;AAC5C,MAAI,MAAM,gBAAgB,OAAO,MAAM,OAAQ,OAAM,IAAI,MAAM,qCAAqC;AACpG,QAAM,EAAE,SAAS,SAAS,IAAI,MAAM,6BAA6B,OAAO,GAAG;AAC3E,MAAI,CAAC,QAAQ,mBAAmB,CAAC,CAAC,aAAa,UAAU,WAAW,EAAE,SAAS,QAAQ,MAAM,EAAG,OAAM,IAAI,MAAM,oCAAoC;AACpJ,QAAM,UAAU,wBAAwB,UAAU,OAAO,QAAQ;AACjE,MAAI,cAAc,OAAO,MAAM,cAAc,OAAO,QAAQ,EAAG,OAAM,IAAI,MAAM,qDAAqD;AACpI,MAAK,QAAQ,WAAW,iBAAkB,OAAO,WAAW,MAAO,OAAM,IAAI,MAAM,6CAA6C;AAChI,MAAI,OAAO,QAAQ;AACjB,8BAA0B,OAAO,MAAM;AACvC,QAAI,OAAO,OAAO,YAAY,OAAO,SAAS,MAAM,OAAO,OAAO,YAAY,gBAAgB,SAAS,eAClG,cAAc,OAAO,OAAO,cAAc,MAAM,cAAc,SAAS,cAAc,KACrF,cAAc,OAAO,OAAO,MAAM,MAAM,cAAc,SAAS,YAAY,KAC3E,cAAc,OAAO,OAAO,WAAW,MAAM,cAAc,OAAO,SAAS,IAAI,CAAC,EAAE,IAAI,YAAY,OAAO,EAAE,IAAI,YAAY,EAAE,CAAC,KAC9H,OAAO,OAAO,UAAU,QAAQ,MAAO,OAAM,IAAI,MAAM,wEAAwE;AAAA,EACtI;AACA,SAAO;AACT;",
+  "names": []
+}

package/dist/dataset-workspaces.js

@@ -5,6 +5,9 @@ import {
   TasksetDraftWorkspaceSchema,
   validateTasksetDraftWorkspace
 } from "./chunk-32H2EJRA.js";
+import {
+  TasksetCatalogReleaseRefSchema
+} from "./chunk-A4ZDOZKP.js";
 import "./chunk-DT7FRNIZ.js";
 import "./chunk-LGEF4S5I.js";
 import "./chunk-WOPM32SE.js";
@@ -12,9 +15,6 @@ import "./chunk-2U54EOXJ.js";
 import "./chunk-JCPLAIZL.js";
 import "./chunk-TCHAWJMB.js";
 import "./chunk-3KXDH6OB.js";
-import {
-  TasksetCatalogReleaseRefSchema
-} from "./chunk-A4ZDOZKP.js";
 import "./chunk-Y4NGWGTA.js";
 import "./chunk-QUOW6URQ.js";
 

package/dist/experiments.js

@@ -0,0 +1,50 @@
+import {
+  ExperimentDefinitionRefSchema,
+  ExperimentDefinitionSchema,
+  ExperimentExecutionContextSchema,
+  ExperimentFieldMappingSchema,
+  ExperimentFieldMappingsSchema,
+  ExperimentGraderPinSchema,
+  ExperimentGraderSelectionSchema,
+  ExperimentListQuerySchema,
+  ExperimentScoringPassSchema,
+  ExperimentScoringRequestSchema,
+  OpenPondExperimentError,
+  OpenPondExperimentsClient,
+  SaveExperimentSchema,
+  StartExperimentSchema,
+  experimentDefinitionRef,
+  mapExperimentGradingFields,
+  verifyExperimentDefinition,
+  verifyExperimentScoringPass
+} from "./chunk-KLX6OVHR.js";
+import "./chunk-ZV2BLO3D.js";
+import {
+  ExperimentAttemptGradeSchema
+} from "./chunk-XR6CLSCS.js";
+import "./chunk-TCHAWJMB.js";
+import "./chunk-3KXDH6OB.js";
+import "./chunk-Y4NGWGTA.js";
+import "./chunk-QUOW6URQ.js";
+export {
+  ExperimentAttemptGradeSchema,
+  ExperimentDefinitionRefSchema,
+  ExperimentDefinitionSchema,
+  ExperimentExecutionContextSchema,
+  ExperimentFieldMappingSchema,
+  ExperimentFieldMappingsSchema,
+  ExperimentGraderPinSchema,
+  ExperimentGraderSelectionSchema,
+  ExperimentListQuerySchema,
+  ExperimentScoringPassSchema,
+  ExperimentScoringRequestSchema,
+  OpenPondExperimentError,
+  OpenPondExperimentsClient,
+  SaveExperimentSchema,
+  StartExperimentSchema,
+  experimentDefinitionRef,
+  mapExperimentGradingFields,
+  verifyExperimentDefinition,
+  verifyExperimentScoringPass
+};
+//# sourceMappingURL=experiments.js.map

package/dist/experiments.js.map

@@ -0,0 +1,7 @@
+{
+  "version": 3,
+  "sources": [],
+  "sourcesContent": [],
+  "mappings": "",
+  "names": []
+}

package/dist/index.js

@@ -1,8 +1,18 @@
+import {
+  OpenPondWorkflowsClient
+} from "./chunk-MEKWESBF.js";
 import "./chunk-LXXQVGQY.js";
 import {
   OpenPondLearningClient,
   OpenPondLearningError
 } from "./chunk-K4C5QOSV.js";
+import {
+  OpenPondExperimentsClient
+} from "./chunk-KLX6OVHR.js";
+import "./chunk-ZV2BLO3D.js";
+import "./chunk-XR6CLSCS.js";
+import "./chunk-TCHAWJMB.js";
+import "./chunk-3KXDH6OB.js";
 import "./chunk-Y4NGWGTA.js";
 import {
   DEFAULT_REFINER_REVIEW_PROFILE,
@@ -23,9 +33,6 @@ import {
 import {
   OpenPondProfileActionsClient
 } from "./chunk-2DMYAJLS.js";
-import {
-  OpenPondWorkflowsClient
-} from "./chunk-MEKWESBF.js";
 import {
   OpenPondApiError,
   apiFetch,
@@ -2464,6 +2471,7 @@ var OpenPondClient = class {
   actions;
   profileActions;
   learning;
+  experiments;
   constructor(options) {
     const apiKey = options.apiKey?.trim() ?? "";
     if (!apiKey && (!options.sandbox || !options.model))
@@ -2480,6 +2488,7 @@ var OpenPondClient = class {
     } : void 0;
     const apiBaseUrl = options.baseUrl?.trim() || "https://api.openpond.ai";
     this.learning = (scope) => new OpenPondLearningClient({ apiKey, baseUrl: apiBaseUrl, scope });
+    this.experiments = (teamId) => new OpenPondExperimentsClient({ apiKey, baseUrl: apiBaseUrl, teamId });
     this.sandboxes = createOpenPondSandboxClient({
       apiKey: sandbox?.apiKey ?? apiKey,
       baseUrl: apiBaseUrl,