@tangle-network/agent-eval 0.221.7 → 0.222.0

package/dist/{client-hwo3cJjR.d.ts → client-BIFwUg7I.d.ts}

@@ -1,14 +1,13 @@
-import { A as SearchLedgerHash, O as SearchLedgerEntry } from "./search-ledger-types-p2xKfiDJ.js";
+import { A as SearchLedgerHash, O as SearchLedgerEntry } from "./search-ledger-types-qiYRUo_N.js";
 import { z } from "zod";
-//#region src/hosted/search-ledger-wire.d.ts
 /** At most this many UTF-8 bytes across one request's lines. A single line
  * larger than this cannot be shipped. */
 declare const SEARCH_LEDGER_BATCH_MAX_BYTES = 1048576;
 declare const SEARCH_LEDGER_INGEST_PATH = "/v1/ingest/search-ledger";
 /** What the run is for: an optimization, or an eval (a one-node search). */
 declare const SearchRunKindSchema: z.ZodEnum<{
-  eval: "eval";
   optimization: "optimization";
+  eval: "eval";
 }>;
 type SearchRunKind = z.infer<typeof SearchRunKindSchema>;
 /** Where a store's copy of one search's chain ends. `headHash` is null exactly
@@ -23,8 +22,8 @@ declare const IngestSearchLedgerRequestSchema: z.ZodObject<{
   wireVersion: z.ZodLiteral<"2026-07-24.v1">;
   searchId: z.ZodString;
   runKind: z.ZodEnum<{
-    eval: "eval";
     optimization: "optimization";
+    eval: "eval";
   }>;
   fromSequence: z.ZodNumber;
   lines: z.ZodArray<z.ZodString>;
@@ -50,8 +49,8 @@ declare const SearchBlobPutResponseSchema: z.ZodObject<{
   sha256: z.ZodCustom<`sha256:${string}`, `sha256:${string}`>;
   byteLength: z.ZodNumber;
   state: z.ZodEnum<{
-    masked: "masked";
     stored: "stored";
+    masked: "masked";
     withheld: "withheld";
   }>;
 }, z.core.$strict>;
@@ -81,8 +80,6 @@ interface AdmitSearchLedgerBatchInput {
  * state machine (`SearchState`) runs after admission, on `entries`.
  */
 declare function admitSearchLedgerBatch(input: AdmitSearchLedgerBatchInput): SearchLedgerBatchAdmission;
-//#endregion
-//#region src/hosted/types.d.ts
 /**
  * # Hosted-tier wire format: the shapes every orchestrator (ours, a partner's
  * self-hosted one, a future open implementation) must accept.
@@ -105,7 +102,7 @@ declare function admitSearchLedgerBatch(input: AdmitSearchLedgerBatchInput): Sea
  * header. Tenants isolate everything downstream of ingest; no tenant
  * ever sees another tenant's data.
  */
-declare const HOSTED_WIRE_VERSION: '2026-07-24.v1';
+declare const HOSTED_WIRE_VERSION: "2026-07-24.v1";
 type HostedWireVersion = typeof HOSTED_WIRE_VERSION;
 /** Every ingest request carries these. */
 interface HostedIngestHeaders {
@@ -167,8 +164,6 @@ interface IngestResponse {
     reason: string;
   }>;
 }
-//#endregion
-//#region src/hosted/client.d.ts
 interface HostedTenant {
   /** Orchestrator endpoint base URL (no trailing slash). Required. */
   endpoint: string;
@@ -218,6 +213,5 @@ declare function hostedTenantFromEnv(overrides?: Partial<HostedTenant> & {
 declare function hostedClientFromEnv(overrides?: Partial<HostedTenant> & {
   env?: Record<string, string | undefined>;
 }): HostedClient | undefined;
-//#endregion
 export { SearchLedgerConflictSchema as C, admitSearchLedgerBatch as D, SearchRunKind as E, SearchLedgerConflict as S, SearchLedgerHeadSchema as T, SEARCH_LEDGER_BATCH_MAX_BYTES as _, hostedClientFromEnv as a, SearchBlobPutResponseSchema as b, HostedIngestHeaders as c, IngestTracesRequest as d, TraceSpanEvent as f, IngestSearchLedgerRequestSchema as g, IngestSearchLedgerRequest as h, createHostedClient as i, HostedWireVersion as l, AdmitSearchLedgerBatchInput as m, HostedTenant as n, hostedTenantFromEnv as o, UnixNanoTimestamp as p, SearchLedgerIngestOutcome as r, HOSTED_WIRE_VERSION as s, HostedClient as t, IngestResponse as u, SEARCH_LEDGER_INGEST_PATH as v, SearchLedgerHead as w, SearchLedgerBatchAdmission as x, SearchBlobPutResponse as y };
-//# sourceMappingURL=client-hwo3cJjR.d.ts.map
+//# sourceMappingURL=client-BIFwUg7I.d.ts.map

package/dist/client-BIFwUg7I.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"client-BIFwUg7I.d.ts","names":[],"sources":["../src/hosted/search-ledger-wire.ts","../src/hosted/types.ts","../src/hosted/client.ts"],"mappings":";;;;cAqCa;cAEA;;cAsBA,qBAAmB,EAAA;;;;KACpB,gBAAgB,EAAE,aAAa;;;cAI9B,wBAAsB,EAAA;;;;GAM/B,EAAA,KAAA;KACQ,mBAAmB,EAAE,aAAa;cAEjC,iCAA+B,EAAA;;;;;;;;;GAkBxC,EAAA,KAAA;KACQ,4BAA4B,EAAE,aAAa;;cAG1C,4BAA0B,EAAA;;;;;;;;;;;GAM5B,EAAA,KAAA;KACC,uBAAuB,EAAE,aAAa;;;cAIrC,6BAA2B,EAAA;;;;;;;;GAM7B,EAAA,KAAA;KACC,wBAAwB,EAAE,aAAa;KA+BvC;EAEN;;EAEA,SAAS;EACT;;EAEA,MAAM;;EAEN;EAAoB,UAAU;;UAEnB;EACf,SAAS;;EAET,MAAM;;EAEN,gBAAgB,mBAAmB;;;;;;;;iBASrB,uBACd,OAAO,8BACN;;;;;;;;;;;;;;;;;;;;;;;cCtJU;KACD,2BAA2B;;UAKtB;;EAEf;;EAEA;;EAEA,yBAAyB;;EAEzB;;;;;;KASU;;;;;;UAOK;EACf;EACA;EACA;EACA;EACA,mBAAmB;EACnB,iBAAiB;EACjB,YAAY;EACZ,SAAS;IACP,cAAc;IACd;IACA,aAAa;;EAEf;IAAW;IAAgC;;;EAE3C;;EAEA;;EAEA;;EAEA;;UAKe;EACf,aAAa;EACb,OAAO;;UAGQ;;EAEf;;EAEA,UAAU;IAAQ;IAAe;;;UC7ClB;;EAEf;;EAEA;;EAEA;;EAEA,mBAAmB;;EAEnB;;EAEA;;;KAIU;EACN;EAAoB,MAAM;;EAC1B;EAAoB,UAAU;;UAEnB;EACf,aAAa,OAAO,kBAAkB,0BAA0B,QAAQ;;EAExE,iBAAiB,kBAAkB,SAAS,cAAc,QAAQ;EAClE,mBACE,SAAS,2BACT,SAAS,cACR,QAAQ;EACX,cACE,4BACA,OAAO,WAAW,cAClB,qBACA,SAAS,cACR,QAAQ;WACF,QAAQ;WACR,aAAa;;iBA8IR,mBAAmB,QAAQ,eAAe;;;;;;;;;;iBAqG1C,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C;;;iBAmBa,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C"}

package/dist/{cluster-bootstrap-DEwwhl0C.js → cluster-bootstrap-SbaeM44s.js}

@@ -1,6 +1,6 @@
-import { s as ValidationError } from "./errors-BcOs5MKW.js";
-import { i as makeRng } from "./internal-BTK91Bqt.js";
-//#region src/meta-eval/cluster-bootstrap.ts
+import { s as ValidationError } from "./errors-DeEEV9Vx.js";
+import { i as makeRng } from "./internal-v6XfOeJW.js";
+// src/meta-eval/cluster-bootstrap.ts
 /**
 * Percentile cluster bootstrap of an observation mean. Whole clusters are resampled with
 * replacement and each resample's statistic is Σ values / Σ observations over the drawn
@@ -60,7 +60,6 @@ function clusterBootstrapMean(observations, options = {}) {
 		resamples
 	};
 }
-//#endregion
 export { clusterBootstrapMean as t };
 
-//# sourceMappingURL=cluster-bootstrap-DEwwhl0C.js.map
+//# sourceMappingURL=cluster-bootstrap-SbaeM44s.js.map

package/dist/{cluster-bootstrap-DEwwhl0C.js.map → cluster-bootstrap-SbaeM44s.js.map}

@@ -1 +1 @@
-{"version":3,"file":"cluster-bootstrap-DEwwhl0C.js","names":[],"sources":["../src/meta-eval/cluster-bootstrap.ts"],"sourcesContent":["import { ValidationError } from '../errors'\nimport { makeRng } from '../statistics/internal'\n\n/** One observation and the source cluster it belongs to. */\nexport interface ClusteredObservation {\n  cluster: string\n  value: number\n}\n\nexport interface ClusterBootstrapOptions {\n  /** Confidence level. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resample count. Default 2000. */\n  resamples?: number\n  /** Deterministic seed. If omitted, derived from the values so the interval is reproducible. */\n  seed?: number\n}\n\nexport interface ClusterBootstrapResult {\n  n: number\n  clusters: number\n  /** Mean over observations, each observation weighted equally. */\n  mean: number\n  low: number\n  high: number\n  confidence: number\n  resamples: number\n}\n\n/**\n * Percentile cluster bootstrap of an observation mean. Whole clusters are resampled with\n * replacement and each resample's statistic is Σ values / Σ observations over the drawn\n * clusters, so observations that share a source move together and a large cluster cannot\n * pose as many independent results. For a paired comparison pass the per-observation\n * difference (candidate − baseline) under the observation's cluster.\n *\n * With one cluster the interval collapses to the point estimate: there is no between-cluster\n * spread to measure, which callers must treat as no interval, not as certainty.\n */\nexport function clusterBootstrapMean(\n  observations: readonly ClusteredObservation[],\n  options: ClusterBootstrapOptions = {},\n): ClusterBootstrapResult {\n  const confidence = options.confidence ?? 0.95\n  const resamples = options.resamples ?? 2000\n  if (!(confidence > 0 && confidence < 1)) {\n    throw new ValidationError(\n      `clusterBootstrapMean: confidence must be in (0,1), got ${confidence}`,\n    )\n  }\n  if (!Number.isInteger(resamples) || resamples < 1) {\n    throw new ValidationError(\n      `clusterBootstrapMean: resamples must be a positive integer, got ${resamples}`,\n    )\n  }\n  if (!observations.length) {\n    throw new ValidationError('clusterBootstrapMean: requires at least one observation')\n  }\n\n  const sums = new Map<string, { sum: number; count: number }>()\n  const values: number[] = []\n  for (const [index, observation] of observations.entries()) {\n    if (typeof observation?.cluster !== 'string' || observation.cluster === '') {\n      throw new ValidationError(`clusterBootstrapMean: observation ${index} needs a cluster id`)\n    }\n    if (!Number.isFinite(observation.value)) {\n      throw new ValidationError(`clusterBootstrapMean: observation ${index} value must be finite`)\n    }\n    const cell = sums.get(observation.cluster) ?? { sum: 0, count: 0 }\n    cell.sum += observation.value\n    cell.count += 1\n    sums.set(observation.cluster, cell)\n    values.push(observation.value)\n  }\n\n  const cells = [...sums.values()]\n  const total = cells.reduce((acc, cell) => acc + cell.sum, 0)\n  const mean = total / observations.length\n  const k = cells.length\n  const rng = makeRng(options.seed, values)\n  const samples = new Array<number>(resamples)\n  for (let b = 0; b < resamples; b++) {\n    let sum = 0\n    let count = 0\n    for (let draw = 0; draw < k; draw++) {\n      const cell = cells[Math.floor(rng() * k)]!\n      sum += cell.sum\n      count += cell.count\n    }\n    samples[b] = sum / count\n  }\n  samples.sort((a, b) => a - b)\n  const alpha = 1 - confidence\n  const lowIdx = Math.floor((alpha / 2) * resamples)\n  const highIdx = Math.min(resamples - 1, Math.ceil((1 - alpha / 2) * resamples) - 1)\n  return {\n    n: observations.length,\n    clusters: k,\n    mean,\n    low: samples[lowIdx]!,\n    high: samples[Math.max(highIdx, lowIdx)]!,\n    confidence,\n    resamples,\n  }\n}\n"],"mappings":";;;;;;;;;;;;;AAuCA,SAAgB,qBACd,cACA,UAAmC,CAAC,GACZ;CACxB,MAAM,aAAa,QAAQ,cAAc;CACzC,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,EAAE,aAAa,KAAK,aAAa,IACnC,MAAM,IAAI,gBACR,0DAA0D,YAC5D;CAEF,IAAI,CAAC,OAAO,UAAU,SAAS,KAAK,YAAY,GAC9C,MAAM,IAAI,gBACR,mEAAmE,WACrE;CAEF,IAAI,CAAC,aAAa,QAChB,MAAM,IAAI,gBAAgB,yDAAyD;CAGrF,MAAM,uBAAO,IAAI,IAA4C;CAC7D,MAAM,SAAmB,CAAC;CAC1B,KAAK,MAAM,CAAC,OAAO,gBAAgB,aAAa,QAAQ,GAAG;EACzD,IAAI,OAAO,aAAa,YAAY,YAAY,YAAY,YAAY,IACtE,MAAM,IAAI,gBAAgB,qCAAqC,MAAM,oBAAoB;EAE3F,IAAI,CAAC,OAAO,SAAS,YAAY,KAAK,GACpC,MAAM,IAAI,gBAAgB,qCAAqC,MAAM,sBAAsB;EAE7F,MAAM,OAAO,KAAK,IAAI,YAAY,OAAO,KAAK;GAAE,KAAK;GAAG,OAAO;EAAE;EACjE,KAAK,OAAO,YAAY;EACxB,KAAK,SAAS;EACd,KAAK,IAAI,YAAY,SAAS,IAAI;EAClC,OAAO,KAAK,YAAY,KAAK;CAC/B;CAEA,MAAM,QAAQ,CAAC,GAAG,KAAK,OAAO,CAAC;CAE/B,MAAM,OADQ,MAAM,QAAQ,KAAK,SAAS,MAAM,KAAK,KAAK,CACzC,IAAI,aAAa;CAClC,MAAM,IAAI,MAAM;CAChB,MAAM,MAAM,QAAQ,QAAQ,MAAM,MAAM;CACxC,MAAM,UAAU,IAAI,MAAc,SAAS;CAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;EAClC,IAAI,MAAM;EACV,IAAI,QAAQ;EACZ,KAAK,IAAI,OAAO,GAAG,OAAO,GAAG,QAAQ;GACnC,MAAM,OAAO,MAAM,KAAK,MAAM,IAAI,IAAI,CAAC;GACvC,OAAO,KAAK;GACZ,SAAS,KAAK;EAChB;EACA,QAAQ,KAAK,MAAM;CACrB;CACA,QAAQ,MAAM,GAAG,MAAM,IAAI,CAAC;CAC5B,MAAM,QAAQ,IAAI;CAClB,MAAM,SAAS,KAAK,MAAO,QAAQ,IAAK,SAAS;CACjD,MAAM,UAAU,KAAK,IAAI,YAAY,GAAG,KAAK,MAAM,IAAI,QAAQ,KAAK,SAAS,IAAI,CAAC;CAClF,OAAO;EACL,GAAG,aAAa;EAChB,UAAU;EACV;EACA,KAAK,QAAQ;EACb,MAAM,QAAQ,KAAK,IAAI,SAAS,MAAM;EACtC;EACA;CACF;AACF"}
+{"version":3,"file":"cluster-bootstrap-SbaeM44s.js","names":[],"sources":["../src/meta-eval/cluster-bootstrap.ts"],"sourcesContent":["import { ValidationError } from '../errors'\nimport { makeRng } from '../statistics/internal'\n\n/** One observation and the source cluster it belongs to. */\nexport interface ClusteredObservation {\n  cluster: string\n  value: number\n}\n\nexport interface ClusterBootstrapOptions {\n  /** Confidence level. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resample count. Default 2000. */\n  resamples?: number\n  /** Deterministic seed. If omitted, derived from the values so the interval is reproducible. */\n  seed?: number\n}\n\nexport interface ClusterBootstrapResult {\n  n: number\n  clusters: number\n  /** Mean over observations, each observation weighted equally. */\n  mean: number\n  low: number\n  high: number\n  confidence: number\n  resamples: number\n}\n\n/**\n * Percentile cluster bootstrap of an observation mean. Whole clusters are resampled with\n * replacement and each resample's statistic is Σ values / Σ observations over the drawn\n * clusters, so observations that share a source move together and a large cluster cannot\n * pose as many independent results. For a paired comparison pass the per-observation\n * difference (candidate − baseline) under the observation's cluster.\n *\n * With one cluster the interval collapses to the point estimate: there is no between-cluster\n * spread to measure, which callers must treat as no interval, not as certainty.\n */\nexport function clusterBootstrapMean(\n  observations: readonly ClusteredObservation[],\n  options: ClusterBootstrapOptions = {},\n): ClusterBootstrapResult {\n  const confidence = options.confidence ?? 0.95\n  const resamples = options.resamples ?? 2000\n  if (!(confidence > 0 && confidence < 1)) {\n    throw new ValidationError(\n      `clusterBootstrapMean: confidence must be in (0,1), got ${confidence}`,\n    )\n  }\n  if (!Number.isInteger(resamples) || resamples < 1) {\n    throw new ValidationError(\n      `clusterBootstrapMean: resamples must be a positive integer, got ${resamples}`,\n    )\n  }\n  if (!observations.length) {\n    throw new ValidationError('clusterBootstrapMean: requires at least one observation')\n  }\n\n  const sums = new Map<string, { sum: number; count: number }>()\n  const values: number[] = []\n  for (const [index, observation] of observations.entries()) {\n    if (typeof observation?.cluster !== 'string' || observation.cluster === '') {\n      throw new ValidationError(`clusterBootstrapMean: observation ${index} needs a cluster id`)\n    }\n    if (!Number.isFinite(observation.value)) {\n      throw new ValidationError(`clusterBootstrapMean: observation ${index} value must be finite`)\n    }\n    const cell = sums.get(observation.cluster) ?? { sum: 0, count: 0 }\n    cell.sum += observation.value\n    cell.count += 1\n    sums.set(observation.cluster, cell)\n    values.push(observation.value)\n  }\n\n  const cells = [...sums.values()]\n  const total = cells.reduce((acc, cell) => acc + cell.sum, 0)\n  const mean = total / observations.length\n  const k = cells.length\n  const rng = makeRng(options.seed, values)\n  const samples = new Array<number>(resamples)\n  for (let b = 0; b < resamples; b++) {\n    let sum = 0\n    let count = 0\n    for (let draw = 0; draw < k; draw++) {\n      const cell = cells[Math.floor(rng() * k)]!\n      sum += cell.sum\n      count += cell.count\n    }\n    samples[b] = sum / count\n  }\n  samples.sort((a, b) => a - b)\n  const alpha = 1 - confidence\n  const lowIdx = Math.floor((alpha / 2) * resamples)\n  const highIdx = Math.min(resamples - 1, Math.ceil((1 - alpha / 2) * resamples) - 1)\n  return {\n    n: observations.length,\n    clusters: k,\n    mean,\n    low: samples[lowIdx]!,\n    high: samples[Math.max(highIdx, lowIdx)]!,\n    confidence,\n    resamples,\n  }\n}\n"],"mappings":";;;;;;;;;;;;;AAuCA,SAAgB,qBACd,cACA,UAAmC,CAAC,GACZ;CACxB,MAAM,aAAa,QAAQ,cAAc;CACzC,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,EAAE,aAAa,KAAK,aAAa,IACnC,MAAM,IAAI,gBACR,0DAA0D,YAC5D;CAEF,IAAI,CAAC,OAAO,UAAU,SAAS,KAAK,YAAY,GAC9C,MAAM,IAAI,gBACR,mEAAmE,WACrE;CAEF,IAAI,CAAC,aAAa,QAChB,MAAM,IAAI,gBAAgB,yDAAyD;CAGrF,MAAM,uBAAO,IAAI,IAA4C;CAC7D,MAAM,SAAmB,CAAC;CAC1B,KAAK,MAAM,CAAC,OAAO,gBAAgB,aAAa,QAAQ,GAAG;EACzD,IAAI,OAAO,aAAa,YAAY,YAAY,YAAY,YAAY,IACtE,MAAM,IAAI,gBAAgB,qCAAqC,MAAM,oBAAoB;EAE3F,IAAI,CAAC,OAAO,SAAS,YAAY,KAAK,GACpC,MAAM,IAAI,gBAAgB,qCAAqC,MAAM,sBAAsB;EAE7F,MAAM,OAAO,KAAK,IAAI,YAAY,OAAO,KAAK;GAAE,KAAK;GAAG,OAAO;EAAE;EACjE,KAAK,OAAO,YAAY;EACxB,KAAK,SAAS;EACd,KAAK,IAAI,YAAY,SAAS,IAAI;EAClC,OAAO,KAAK,YAAY,KAAK;CAC/B;CAEA,MAAM,QAAQ,CAAC,GAAG,KAAK,OAAO,CAAC;CAE/B,MAAM,OADQ,MAAM,QAAQ,KAAK,SAAS,MAAM,KAAK,KAAK,CACzC,IAAI,aAAa;CAClC,MAAM,IAAI,MAAM;CAChB,MAAM,MAAM,QAAQ,QAAQ,MAAM,MAAM;CACxC,MAAM,UAAU,IAAI,MAAc,SAAS;CAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;EAClC,IAAI,MAAM;EACV,IAAI,QAAQ;EACZ,KAAK,IAAI,OAAO,GAAG,OAAO,GAAG,QAAQ;GACnC,MAAM,OAAO,MAAM,KAAK,MAAM,IAAI,IAAI,CAAC;GACvC,OAAO,KAAK;GACZ,SAAS,KAAK;EAChB;EACA,QAAQ,KAAK,MAAM;CACrB;CACA,QAAQ,MAAM,GAAG,MAAM,IAAI,CAAC;CAC5B,MAAM,QAAQ,IAAI;CAClB,MAAM,SAAS,KAAK,MAAO,QAAQ,IAAK,SAAS;CACjD,MAAM,UAAU,KAAK,IAAI,YAAY,GAAG,KAAK,MAAM,IAAI,QAAQ,KAAK,SAAS,IAAI,CAAC;CAClF,OAAO;EACL,GAAG,aAAa;EAChB,UAAU;EACV;EACA,KAAK,QAAQ;EACb,MAAM,QAAQ,KAAK,IAAI,SAAS,MAAM;EACtC;EACA;CACF;AACF"}

package/dist/completion-verifier-7KEfroqx.d.ts

@@ -0,0 +1,168 @@
+import { o as CostLedgerHandle } from "./cost-ledger-Czro9wmo.js";
+import { Z as RawProviderSink, x as ChatClient } from "./types-0KglnQ9o.js";
+import { a as CheckerIdentity, n as VerdictCertification, t as DefaultVerdict, x as VerificationStrategySource } from "./verdict-D_qiBQ6_.js";
+/** An artifact an agent produced, as completion and produced-state checks read it. */
+interface ValidatedArtifact {
+  /** Logical kind — validators type-guard on this */
+  kind: 'file' | 'json' | 'text' | 'binary' | string;
+  /** Filesystem-style path, optional */
+  path?: string;
+  /** String content for text/json/file kinds */
+  content?: string;
+  /** Binary content (if kind === 'binary') */
+  bytes?: Uint8Array;
+  /** Caller-supplied metadata (mimeType, sha256, size, etc.) */
+  metadata?: Record<string, unknown>;
+}
+/** What kind of produced state can satisfy a requirement structurally. */
+type SatisfiedBy = 'artifact' | 'proposal' | 'tool-call' | 'any';
+interface CompletionRequirement {
+  /** Stable id from the task gold (e.g. a persona's `expected_requirements[].req_id`). */
+  reqId: string;
+  /** Human-readable description of the required deliverable. */
+  title: string;
+  /** Optional kind/category hint, matched against a produced item's kind. */
+  category?: string;
+  /** What produced state satisfies this requirement. Defaults to 'any'. */
+  satisfiedBy?: SatisfiedBy;
+}
+interface TaskGold {
+  taskId: string;
+  requirements: CompletionRequirement[];
+}
+interface ProducedProposal {
+  id: string;
+  title: string;
+  status: 'pending' | 'approved' | 'rejected';
+  /** Optional persisted body — when present, enables a correctness check. */
+  content?: string;
+}
+/** Everything observable about what a run actually produced. */
+interface ProducedState {
+  /** Persisted vault artifacts. Reuses the shared `ValidatedArtifact` shape. */
+  artifacts: ValidatedArtifact[];
+  /** Proposals / filings the agent created. */
+  proposals: ProducedProposal[];
+  /** Names of tools the agent invoked. */
+  toolCalls: string[];
+}
+interface RequirementCheck {
+  reqId: string;
+  title: string;
+  /** A produced item of the right kind matched the requirement, non-empty. */
+  structurallyPresent: boolean;
+  /**
+   * Whether the matched item actually fulfils the requirement. `null` when
+   * not structurally present, when the matched item carries no content
+   * to assess, or when the correctness check itself failed (`unmeasured`).
+   */
+  correct: boolean | null;
+  /** structurallyPresent && !unmeasured && correct !== false. */
+  satisfied: boolean;
+  /**
+   * Set when the correctness check itself errored (LLM call failure or an
+   * unparseable response after retry). The requirement's fulfilment is
+   * UNKNOWN — `correct` stays null, `satisfied` is false, and
+   * `completionVerdict` excludes the row from `completionRate`'s
+   * denominator. Never folded into a zero: a synthetic zero is
+   * indistinguishable from a real failure (see `JudgeParseError`).
+   */
+  unmeasured?: true;
+  /** Why the correctness check could not be measured (present iff `unmeasured`). */
+  unmeasuredReason?: string;
+  /** Human-readable evidence for the verdict. */
+  evidence: string[];
+}
+/** Extends the substrate verdict spine: `valid` = `fullyComplete` and
+ *  `score` = `completionRate` — derived in `completionVerdict()`, the one
+ *  place those equalities hold by construction. */
+interface CompletionVerdict extends DefaultVerdict {
+  taskId: string;
+  requirements: RequirementCheck[];
+  /** satisfied / MEASURABLE requirements (unmeasured rows leave the denominator). */
+  completionRate: number;
+  /** Every measurable requirement satisfied (false when anything is unmeasured). */
+  fullyComplete: boolean;
+  /** Requirements whose correctness check errored — reported, never scored as zero. */
+  unmeasuredCount: number;
+}
+/**
+ * Construct a `CompletionVerdict` from the per-requirement checks, deriving
+ * `completionRate` / `fullyComplete` and the spine fields (`valid` =
+ * `fullyComplete`, `score` = `completionRate`) in one place. Throws on zero
+ * requirements — a verdict over nothing is a misconfiguration, mirroring
+ * `verifyCompletion`'s gold-spec guard.
+ */
+declare function completionVerdict(input: {
+  taskId: string;
+  requirements: RequirementCheck[];
+  /** What certified the correctness stage, when anything did. Omitted =
+   *  an uncertified verdict — the honest default for a bare checker. */
+  certification?: VerdictCertification;
+}): CompletionVerdict;
+/**
+ * What a correctness checker declares about itself so the completion
+ * verdict can carry a certification: which strategy member it discharges,
+ * its exact identity, and the steps its answers rest on unverified.
+ */
+interface CorrectnessCheckerAttestation {
+  strategy: VerificationStrategySource;
+  checker: CheckerIdentity;
+  assumptions: string[];
+}
+/**
+ * Decides whether a produced item's content actually fulfils a requirement.
+ * Injected so the structural verifier stays pure and unit-testable; the
+ * production implementation is `createLlmCorrectnessChecker`.
+ *
+ * `attestation` is optional metadata on the function value: a checker that
+ * carries one yields certified completion verdicts; a bare function yields
+ * the same verdict uncertified. A plain arrow function remains a valid
+ * checker.
+ */
+interface CorrectnessChecker {
+  (requirement: CompletionRequirement, content: string): Promise<{
+    correct: boolean;
+    reason: string;
+  }>;
+  attestation?: CorrectnessCheckerAttestation;
+}
+/**
+ * Verify whether a run completed the task. `checkCorrectness` is injected —
+ * `createLlmCorrectnessChecker` for production, a deterministic stub in tests.
+ *
+ * Throws on a gold spec with no requirements: an eval task that requires
+ * nothing is a misconfiguration, not a vacuously-complete task.
+ */
+declare function verifyCompletion(gold: TaskGold, state: ProducedState, checkCorrectness: CorrectnessChecker): Promise<CompletionVerdict>;
+interface LlmCorrectnessCheckerOpts {
+  model?: string;
+  /** Optional ledger for direct use. */
+  costLedger?: CostLedgerHandle;
+  costPhase?: string;
+  costTags?: Record<string, string>;
+  signal?: AbortSignal;
+  /** Max chars of artifact content sent to the checker. */
+  maxContentChars?: number;
+  /**
+   * Checker LLM calls per requirement before giving up (parse failures and
+   * call errors both consume attempts). The failure then surfaces as an
+   * `unmeasured` requirement, never a zero.
+   */
+  maxAttempts?: number;
+  /**
+   * Forensic capture of every checker request/response/error — without it a
+   * checker failure is unauditable (the agent-turn raws never contain the
+   * checker's own calls). Same sink contract as `LlmClient`.
+   */
+  rawSink?: RawProviderSink;
+}
+/**
+ * Production `CorrectnessChecker` — one LLM call per matched artifact,
+ * deterministic (temperature 0), structured JSON out. Judges fulfilment
+ * only: a plan, a gesture, or a description of what should be done does not
+ * fulfil a requirement — the artifact must BE the deliverable.
+ */
+declare function createLlmCorrectnessChecker(chat: ChatClient, opts?: LlmCorrectnessCheckerOpts): CorrectnessChecker;
+export { ProducedProposal as a, SatisfiedBy as c, createLlmCorrectnessChecker as d, verifyCompletion as f, LlmCorrectnessCheckerOpts as i, TaskGold as l, CompletionVerdict as n, ProducedState as o, CorrectnessChecker as r, RequirementCheck as s, CompletionRequirement as t, completionVerdict as u };
+//# sourceMappingURL=completion-verifier-7KEfroqx.d.ts.map

package/dist/completion-verifier-7KEfroqx.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"completion-verifier-7KEfroqx.d.ts","names":[],"sources":["../src/artifact-validator.ts","../src/completion-verifier.ts"],"mappings":";;;;UACiB;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER,WAAW;;;KCiCD;UAEK;;EAEf;;EAEA;;EAEA;;EAEA,cAAc;;UAGC;EACf;EACA,cAAc;;UAGC;EACf;EACA;EACA;;EAEA;;;UAIe;;EAEf,WAAW;;EAEX,WAAW;;EAEX;;UAGe;EACf;EACA;;EAEA;;;;;;EAMA;;EAEA;;;;;;;;;EASA;;EAEA;;EAEA;;;;;UAMe,0BAA0B;EACzC;EACA,cAAc;;EAEd;;EAEA;;EAEA;;;;;;;;;iBAUc,kBAAkB;EAChC;EACA,cAAc;;;EAGd,gBAAgB;IACd;;;;;;UAqCa;EACf,UAAU;EACV,SAAS;EACT;;;;;;;;;;;;UAae;GAEb,aAAa,uBACb,kBACC;IAAU;IAAkB;;EAC/B,cAAc;;;;;;;;;iBAsLM,iBACpB,MAAM,UACN,OAAO,eACP,kBAAkB,qBACjB,QAAQ;UAyGM;EACf;;EAEA,aAAa;EACb;EACA,WAAW;EACX,SAAS;;EAET;;;;;;EAMA;;;;;;EAMA,UAAU;;;;;;;;iBA2CI,4BACd,MAAM,YACN,OAAM,4BACL"}

package/dist/{completion-verifier-asaV1nSj.js → completion-verifier-rqbMh5-4.js}

@@ -1,160 +1,10 @@
-import { s as ValidationError } from "./errors-BcOs5MKW.js";
-import { r as canonicalString } from "./canonical-Dq2hOQ8L.js";
-import { n as JudgeParseError } from "./llm-judge-Cybq29bL.js";
-import { t as CostLedger } from "./cost-ledger-BGZB6yY2.js";
-import { v as recoverTruncatedJson } from "./campaign-evidence-DunzcJGv.js";
-import { t as certificationEvidenceDigest } from "./verdict-BGQmxQ8A.js";
-import { a as maximumChargeForLlmRequest, c as ModelSubstitutionError, i as costReceiptFromLlmError, r as costReceiptFromLlm, u as assertServedModel } from "./llm-client-b1ysZVIU.js";
-import { createHash, randomUUID } from "node:crypto";
-import { harnessSupportsModel } from "@tangle-network/agent-interface";
-//#region src/agent-profile.ts
-/**
-* The agentic coding harnesses an eval sweeps by default — the ones we care about
-* ranking. This is the SINGLE source of that list; consumers import it instead of
-* re-declaring their own (a re-declared list is how the fleet drifts). Pass an
-* explicit `harnesses` (e.g. `harnessTypeSchema.options` for literally every known
-* harness) to widen beyond these.
-*/
-const CODING_HARNESSES = [
-	"opencode",
-	"claude-code",
-	"codex",
-	"kimi-code"
-];
-/** Model sentinel for a vendor-locked harness that supports none of the swept models:
-*  it carries no provider prefix, so `harnessSupportsModel` accepts it and the harness
-*  resolves it to its own native default model at runtime (e.g. kimi-code → its Kimi
-*  model). Lets `expandProfileAxes` snap-instead-of-drop without a per-harness flagship
-*  table that would rot as router catalogs change. */
-const HARNESS_NATIVE_MODEL = "default";
-/**
-* Expand a base profile across the harness × model matrix into the `AgentProfile[]`
-* that `runProfileMatrix` / `selfImprove` score — the ONE place "which harnesses ×
-* which models do we evaluate" lives, so no product hand-rolls its own harness list
-* or column→profile mapping (the pattern that let those copies drift and silently
-* break the harness pivot).
-*
-* Each cell clones `base`, sets the canonical top-level `harness` and `model.default`,
-* and stamps `metadata.harness` + `metadata.harnessModel` for matrix grouping. Both
-* metadata fields are hash-bearing, so every cell gets a distinct `agentProfileId` row
-* and results join back by harness/model via {@link harnessAxisOf} with no
-* hand-recomputed key. A vendor-locked harness snaps to its family's swept models — or
-* its native default ({@link HARNESS_NATIVE_MODEL}) when it supports none — so every
-* requested harness runs; `keepIncompatible` forces every pair verbatim.
-*
-* Omit `harnesses`/`models` to sweep the full default set — the "turn it on for
-* everything we care about" switch, identical in shape whether one harness or all.
-*/
-function expandProfileAxes(spec) {
-	const harnesses = spec.harnesses ?? CODING_HARNESSES;
-	if (harnesses.length === 0) throw new ValidationError("expandProfileAxes: no harnesses to sweep");
-	const baseModel = spec.base.model?.default;
-	const models = spec.models ?? (baseModel ? [baseModel] : []);
-	if (models.length === 0) throw new ValidationError("expandProfileAxes: no models to sweep — base profile has no model.default and none were supplied");
-	const out = [];
-	const seen = /* @__PURE__ */ new Set();
-	for (const harness of harnesses) {
-		const supported = spec.keepIncompatible ? models : models.filter((model) => harnessSupportsModel(harness, model));
-		const effective = supported.length > 0 ? supported : [HARNESS_NATIVE_MODEL];
-		for (const model of effective) {
-			const profile = {
-				...spec.base,
-				harness,
-				name: `${spec.base.name ?? "agent"}/${harness}/${model}`,
-				model: {
-					...spec.base.model,
-					default: model
-				},
-				metadata: {
-					...spec.base.metadata ?? {},
-					harness,
-					harnessModel: model
-				}
-			};
-			const id = agentProfileId(profile);
-			if (seen.has(id)) continue;
-			seen.add(id);
-			out.push(profile);
-		}
-	}
-	if (out.length === 0) throw new ValidationError(`expandProfileAxes: produced no profiles (harnesses=[${harnesses.join(", ")}], models=[${models.join(", ")}]).`);
-	return out;
-}
-/**
-* Read the (harness, model) a matrix cell ran under, off a profile or a result row's
-* profile — the join-back for a `byHarness` pivot. Returns undefined when the profile
-* wasn't produced by {@link expandProfileAxes}. Callers group `result.byProfile` by
-* this instead of recomputing an id (recomputing the wrong key is what broke the pivot
-* in the hand-rolled copies).
-*/
-function harnessAxisOf(profile) {
-	const m = profile.metadata;
-	const harness = m?.harness;
-	const model = m?.harnessModel;
-	if (typeof harness === "string" && typeof model === "string") return {
-		harness,
-		model
-	};
-}
-/**
-* Collision-resistant, path-safe, human-readable profile id for eval artifacts.
-* Scorecard joins still use `agentProfileHash`; this id is for run ids, matrix
-* keys, and directory names where two profiles must not collapse onto one row.
-* The suffix is the first 64 bits of the behaviour hash, enough for ordinary
-* eval matrices while keeping filenames readable.
-*/
-function agentProfileId(profile) {
-	return `${pathSafeProfileLabel(agentProfileDisplayLabel(profile)) ?? "profile"}-${agentProfileHash(profile).slice(0, 16)}`;
-}
-/**
-* Model snapshot used for `RunRecord.model`. Eval surfaces require a concrete
-* model id because run records reject bare/missing model aliases.
-*/
-function agentProfileModelId(profile) {
-	const model = profile.model?.default?.trim();
-	if (!model) {
-		const label = agentProfileDisplayLabel(profile) ?? "unnamed profile";
-		throw new ValidationError(`AgentProfile "${label}" has no model.default — cannot record eval run`);
-	}
-	return model;
-}
-function agentProfileDisplayLabel(profile) {
-	return profile.name?.trim() || profile.version?.trim() || void 0;
-}
-function pathSafeProfileLabel(label) {
-	return label?.trim().replace(/[^A-Za-z0-9._-]+/g, "-").replace(/-+/g, "-").replace(/^-|-$/g, "") || void 0;
-}
-function compact(input) {
-	const out = {};
-	for (const [key, value] of Object.entries(input)) if (value !== void 0) out[key] = value;
-	return out;
-}
-/**
-* Deterministic behaviour identity for the canonical
-* `@tangle-network/agent-interface` AgentProfile.
-*
-* `name` and `description` are labels and do not affect the hash. Profile
-* `version`, prompt, model hints, tools, resources, hooks, modes, permissions,
-* and extensions do affect the hash. Resource array order is hash-bearing
-* because mount order can change agent behaviour. Undefined fields are treated
-* as absent; explicit `null` fields remain hash-bearing.
-*/
-function agentProfileHash(profile) {
-	const model = agentProfileModelId(profile);
-	const behaviour = compact({
-		...profile,
-		name: void 0,
-		description: void 0,
-		tags: profile.tags ? [...profile.tags].sort() : void 0,
-		model: compact({
-			...profile.model,
-			default: model
-		})
-	});
-	return createHash("sha256").update(canonicalString(behaviour)).digest("hex");
-}
-//#endregion
-//#region src/completion-verifier.ts
+import { t as CostLedger } from "./cost-ledger-CS8R3IZj.js";
+import { n as recoverTruncatedJson } from "./json-recovery-DK4MDm5r.js";
+import { t as certificationEvidenceDigest } from "./verdict-C8TcdwfD.js";
+import { a as maximumChargeForLlmRequest, d as assertServedModel, i as costReceiptFromLlmError, l as ModelSubstitutionError, r as costReceiptFromLlm } from "./llm-client-kJ9EOqjt.js";
+import { t as JudgeParseError } from "./judges-BLxg5p8E.js";
+import { randomUUID } from "node:crypto";
+// src/completion-verifier.ts
 /**
 * Completion verifier — the task-completion oracle.
 *
@@ -533,7 +383,6 @@ function createLlmCorrectnessChecker(chat, opts = {}) {
 	};
 	return checker;
 }
-//#endregion
-export { HARNESS_NATIVE_MODEL as a, agentProfileModelId as c, CODING_HARNESSES as i, expandProfileAxes as l, createLlmCorrectnessChecker as n, agentProfileHash as o, verifyCompletion as r, agentProfileId as s, completionVerdict as t, harnessAxisOf as u };
+export { createLlmCorrectnessChecker as n, verifyCompletion as r, completionVerdict as t };
 
-//# sourceMappingURL=completion-verifier-asaV1nSj.js.map
+//# sourceMappingURL=completion-verifier-rqbMh5-4.js.map

package/dist/completion-verifier-rqbMh5-4.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"completion-verifier-rqbMh5-4.js","names":[],"sources":["../src/completion-verifier.ts"],"sourcesContent":["/**\n * Completion verifier — the task-completion oracle.\n *\n * Answers the only eval question that is not a proxy: did the agent actually\n * COMPLETE the task — produce every required deliverable, persisted and\n * correct — rather than describe what should be done. A fluent transcript\n * that never produces the artifact scores zero here.\n *\n * Per requirement, a two-stage check:\n *   1. Structural — a produced item (vault artifact / approved proposal /\n *      tool call) of the right kind is matched against the requirement and\n *      carries non-empty content. Deterministic; no LLM.\n *   2. Correctness — only if structurally present AND the matched item\n *      carries content, one targeted check decides whether that item\n *      actually fulfils the requirement. A hallucinated artifact fails here;\n *      an absent one already failed stage 1.\n *\n * `completionRate` is satisfied / MEASURABLE requirements (unmeasured rows —\n * checker failures — are excluded from the denominator, never scored as\n * zeros). Quality dimensions are meaningless on an incomplete task — callers\n * gate on `fullyComplete` / `completionRate` before scoring quality.\n */\n\nimport { randomUUID } from 'node:crypto'\nimport type { ChatClient, ChatRequest } from './analyst/chat-client'\nimport type { ValidatedArtifact } from './artifact-validator'\nimport { CostLedger, type CostLedgerHandle } from './cost-ledger'\nimport { assertServedModel, ModelSubstitutionError } from './integrity/served-model'\nimport { recoverTruncatedJson } from './json-recovery'\nimport { JudgeParseError } from './judges'\nimport {\n  costReceiptFromLlm,\n  costReceiptFromLlmError,\n  maximumChargeForLlmRequest,\n} from './llm-client'\nimport type { RawProviderEvent, RawProviderSink } from './trace/raw-provider-sink'\nimport {\n  certificationEvidenceDigest,\n  type DefaultVerdict,\n  type VerdictCertification,\n} from './verdict'\nimport type { CheckerIdentity, VerificationStrategySource } from './verification-strategy'\n\n/** What kind of produced state can satisfy a requirement structurally. */\nexport type SatisfiedBy = 'artifact' | 'proposal' | 'tool-call' | 'any'\n\nexport interface CompletionRequirement {\n  /** Stable id from the task gold (e.g. a persona's `expected_requirements[].req_id`). */\n  reqId: string\n  /** Human-readable description of the required deliverable. */\n  title: string\n  /** Optional kind/category hint, matched against a produced item's kind. */\n  category?: string\n  /** What produced state satisfies this requirement. Defaults to 'any'. */\n  satisfiedBy?: SatisfiedBy\n}\n\nexport interface TaskGold {\n  taskId: string\n  requirements: CompletionRequirement[]\n}\n\nexport interface ProducedProposal {\n  id: string\n  title: string\n  status: 'pending' | 'approved' | 'rejected'\n  /** Optional persisted body — when present, enables a correctness check. */\n  content?: string\n}\n\n/** Everything observable about what a run actually produced. */\nexport interface ProducedState {\n  /** Persisted vault artifacts. Reuses the shared `ValidatedArtifact` shape. */\n  artifacts: ValidatedArtifact[]\n  /** Proposals / filings the agent created. */\n  proposals: ProducedProposal[]\n  /** Names of tools the agent invoked. */\n  toolCalls: string[]\n}\n\nexport interface RequirementCheck {\n  reqId: string\n  title: string\n  /** A produced item of the right kind matched the requirement, non-empty. */\n  structurallyPresent: boolean\n  /**\n   * Whether the matched item actually fulfils the requirement. `null` when\n   * not structurally present, when the matched item carries no content\n   * to assess, or when the correctness check itself failed (`unmeasured`).\n   */\n  correct: boolean | null\n  /** structurallyPresent && !unmeasured && correct !== false. */\n  satisfied: boolean\n  /**\n   * Set when the correctness check itself errored (LLM call failure or an\n   * unparseable response after retry). The requirement's fulfilment is\n   * UNKNOWN — `correct` stays null, `satisfied` is false, and\n   * `completionVerdict` excludes the row from `completionRate`'s\n   * denominator. Never folded into a zero: a synthetic zero is\n   * indistinguishable from a real failure (see `JudgeParseError`).\n   */\n  unmeasured?: true\n  /** Why the correctness check could not be measured (present iff `unmeasured`). */\n  unmeasuredReason?: string\n  /** Human-readable evidence for the verdict. */\n  evidence: string[]\n}\n\n/** Extends the substrate verdict spine: `valid` = `fullyComplete` and\n *  `score` = `completionRate` — derived in `completionVerdict()`, the one\n *  place those equalities hold by construction. */\nexport interface CompletionVerdict extends DefaultVerdict {\n  taskId: string\n  requirements: RequirementCheck[]\n  /** satisfied / MEASURABLE requirements (unmeasured rows leave the denominator). */\n  completionRate: number\n  /** Every measurable requirement satisfied (false when anything is unmeasured). */\n  fullyComplete: boolean\n  /** Requirements whose correctness check errored — reported, never scored as zero. */\n  unmeasuredCount: number\n}\n\n/**\n * Construct a `CompletionVerdict` from the per-requirement checks, deriving\n * `completionRate` / `fullyComplete` and the spine fields (`valid` =\n * `fullyComplete`, `score` = `completionRate`) in one place. Throws on zero\n * requirements — a verdict over nothing is a misconfiguration, mirroring\n * `verifyCompletion`'s gold-spec guard.\n */\nexport function completionVerdict(input: {\n  taskId: string\n  requirements: RequirementCheck[]\n  /** What certified the correctness stage, when anything did. Omitted =\n   *  an uncertified verdict — the honest default for a bare checker. */\n  certification?: VerdictCertification\n}): CompletionVerdict {\n  if (input.requirements.length === 0) {\n    throw new Error(\n      `completionVerdict: task '${input.taskId}' has no requirement checks — nothing to derive a verdict from`,\n    )\n  }\n  const measurable = input.requirements.filter((r) => !r.unmeasured)\n  const unmeasuredCount = input.requirements.length - measurable.length\n  if (measurable.length === 0) {\n    // Every check errored: this is an infrastructure failure, not a scored\n    // run. A 0-rate verdict here would be a fabricated measurement.\n    throw new Error(\n      `completionVerdict: task '${input.taskId}' has no measurable requirements — all ${input.requirements.length} correctness checks failed (${input.requirements[0]?.unmeasuredReason ?? 'unknown reason'})`,\n    )\n  }\n  const satisfiedCount = measurable.filter((r) => r.satisfied).length\n  const completionRate = satisfiedCount / measurable.length\n  // A run with unmeasured rows can still report a rate over what WAS\n  // measured, but must not claim full completion.\n  const fullyComplete = unmeasuredCount === 0 && satisfiedCount === measurable.length\n  return {\n    taskId: input.taskId,\n    requirements: input.requirements,\n    completionRate,\n    fullyComplete,\n    unmeasuredCount,\n    valid: fullyComplete,\n    score: completionRate,\n    ...(input.certification === undefined ? {} : { certification: input.certification }),\n  }\n}\n\n/**\n * What a correctness checker declares about itself so the completion\n * verdict can carry a certification: which strategy member it discharges,\n * its exact identity, and the steps its answers rest on unverified.\n */\nexport interface CorrectnessCheckerAttestation {\n  strategy: VerificationStrategySource\n  checker: CheckerIdentity\n  assumptions: string[]\n}\n\n/**\n * Decides whether a produced item's content actually fulfils a requirement.\n * Injected so the structural verifier stays pure and unit-testable; the\n * production implementation is `createLlmCorrectnessChecker`.\n *\n * `attestation` is optional metadata on the function value: a checker that\n * carries one yields certified completion verdicts; a bare function yields\n * the same verdict uncertified. A plain arrow function remains a valid\n * checker.\n */\nexport interface CorrectnessChecker {\n  (\n    requirement: CompletionRequirement,\n    content: string,\n  ): Promise<{ correct: boolean; reason: string }>\n  attestation?: CorrectnessCheckerAttestation\n}\n\nconst STOPWORDS = new Set([\n  'the',\n  'a',\n  'an',\n  'of',\n  'for',\n  'and',\n  'or',\n  'to',\n  'in',\n  'on',\n  'with',\n  'by',\n])\n\n// Deliverable-FORM vocabulary — words that name the SHAPE of an output, not its\n// domain content. A correct \"swap-comparison view persisted as a ui/** artifact\"\n// is an OpenUI JSON whose body says nothing about \"artifact\" / \"persisted\" /\n// \"view\"; the discriminative tokens are the domain nouns (swap, comparison).\n// Stripped from the REQUIREMENT side of structural recall so a deliverable is\n// matched on what it IS about, not on the boilerplate describing its form. The\n// correctness checker strips the same class via TITLE_STOPWORDS. Anti-game holds:\n// the distinctive domain tokens remain, so an off-topic item still fails.\nconst REQUIREMENT_FORM_STOPWORDS = new Set([\n  'generated',\n  'generate',\n  'view',\n  'render',\n  'rendered',\n  'persisted',\n  'persist',\n  'artifact',\n  'file',\n  'document',\n  'note',\n  'proposal',\n  'deliverable',\n  'output',\n  'created',\n  'create',\n  'produce',\n  'produced',\n  'flag',\n])\n\nconst MATCH_THRESHOLD = 0.5\nconst MIN_CONTENT_CHARS = 50\n\nfunction tokens(s: string, extraStop?: Set<string>): Set<string> {\n  return new Set(\n    s\n      .toLowerCase()\n      .split(/[^a-z0-9]+/)\n      .filter((t) => t.length > 1 && !STOPWORDS.has(t) && !extraStop?.has(t)),\n  )\n}\n\n/**\n * Recall of the requirement's tokens within a candidate's identifying text.\n * Recall, not Jaccard — a candidate's path/id legitimately carries extra\n * tokens the requirement does not name. The requirement side drops\n * deliverable-FORM vocabulary so recall keys on the distinctive domain tokens.\n */\nfunction tokenRecall(requirementText: string, candidateText: string): number {\n  const req = tokens(requirementText, REQUIREMENT_FORM_STOPWORDS)\n  if (req.size === 0) return 0\n  const cand = tokens(candidateText)\n  let hit = 0\n  for (const t of req) if (cand.has(t)) hit++\n  return hit / req.size\n}\n\ninterface Candidate {\n  reqIndex: number\n  /** Unique key for a produced item — each item satisfies at most one requirement. */\n  itemKey: string\n  score: number\n  evidence: string\n  /** Content to correctness-check, or null when the matched item has none. */\n  content: string | null\n}\n\nfunction artifactCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  artifacts: ValidatedArtifact[],\n): Candidate[] {\n  const reqText = `${req.title} ${req.category ?? ''}`\n  const out: Candidate[] = []\n  artifacts.forEach((a, i) => {\n    if ((a.content ?? '').trim().length < MIN_CONTENT_CHARS) return\n    // Match against the artifact CONTENT too, not just its path + kind — a\n    // generated view / note whose path is generic still satisfies a requirement\n    // when its body covers it (e.g. an OpenUI comparison grounded in the on-file\n    // figures). Bounded slice keeps the recall text cheap; MATCH_THRESHOLD holds.\n    let score = tokenRecall(\n      reqText,\n      `${a.path ?? ''} ${a.kind} ${(a.content ?? '').slice(0, 4000)}`,\n    )\n    if (req.category && a.kind && req.category.toLowerCase() === a.kind.toLowerCase()) {\n      score = Math.max(score, 1)\n    }\n    if (score < MATCH_THRESHOLD) return\n    out.push({\n      reqIndex,\n      itemKey: `artifact:${i}`,\n      score,\n      evidence: `artifact '${a.path ?? a.kind}' matched (token recall ${score.toFixed(2)})`,\n      content: a.content ?? null,\n    })\n  })\n  return out\n}\n\nfunction proposalCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  proposals: ProducedProposal[],\n): Candidate[] {\n  const reqText = `${req.title} ${req.category ?? ''}`\n  const out: Candidate[] = []\n  for (const p of proposals) {\n    // Pending or rejected work is not a completed deliverable.\n    if (p.status !== 'approved') continue\n    // A proposal needs an assessable BODY to be a deliverable. A bare title is\n    // not completion: correctness cannot be judged on it, so a title-only match\n    // would auto-pass the oracle (structurallyPresent && correct===null →\n    // satisfied) with no verifiable content. Tool calls are the only\n    // legitimately content-less deliverable (`toolCallCandidates`).\n    const body = (p.content ?? '').trim()\n    if (body.length < MIN_CONTENT_CHARS) continue\n    // Match against the body as well as the (often short) title — a refusal /\n    // flag / analysis proposal whose title is a label still satisfies a\n    // descriptively-worded requirement when its content covers it. MATCH_THRESHOLD\n    // + the requirement's distinctive tokens keep an off-topic proposal out;\n    // correctness (a SEMANTIC checker, NOT this lexical pass) then judges\n    // polarity/fulfilment, so a negation that merely contains the tokens fails.\n    // Structural and correctness must use different evidence or the two-stage\n    // check collapses to one lexical gate.\n    const score = tokenRecall(reqText, `${p.title} ${body}`)\n    if (score < MATCH_THRESHOLD) continue\n    out.push({\n      reqIndex,\n      itemKey: `proposal:${p.id}`,\n      score,\n      evidence: `approved proposal '${p.title}' matched (token recall ${score.toFixed(2)})`,\n      content: body,\n    })\n  }\n  return out\n}\n\nfunction toolCallCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  toolCalls: string[],\n): Candidate[] {\n  const out: Candidate[] = []\n  toolCalls.forEach((name, i) => {\n    const score = tokenRecall(req.title, name)\n    if (score < MATCH_THRESHOLD) return\n    out.push({\n      reqIndex,\n      itemKey: `tool:${i}`,\n      score,\n      evidence: `tool call '${name}' matched (token recall ${score.toFixed(2)})`,\n      content: null,\n    })\n  })\n  return out\n}\n\n/**\n * Verify whether a run completed the task. `checkCorrectness` is injected —\n * `createLlmCorrectnessChecker` for production, a deterministic stub in tests.\n *\n * Throws on a gold spec with no requirements: an eval task that requires\n * nothing is a misconfiguration, not a vacuously-complete task.\n */\nexport async function verifyCompletion(\n  gold: TaskGold,\n  state: ProducedState,\n  checkCorrectness: CorrectnessChecker,\n): Promise<CompletionVerdict> {\n  if (gold.requirements.length === 0) {\n    throw new Error(\n      `verifyCompletion: task '${gold.taskId}' has no requirements — malformed gold spec`,\n    )\n  }\n\n  // Collect every above-threshold (requirement, produced-item) candidate, then\n  // assign greedily by descending score: each requirement and each produced\n  // item is used at most once. One deliverable fulfils one requirement.\n  const candidates: Candidate[] = []\n  gold.requirements.forEach((req, i) => {\n    const by = req.satisfiedBy ?? 'any'\n    if (by === 'artifact' || by === 'any') {\n      candidates.push(...artifactCandidates(req, i, state.artifacts))\n    }\n    if (by === 'proposal' || by === 'any') {\n      candidates.push(...proposalCandidates(req, i, state.proposals))\n    }\n    if (by === 'tool-call' || by === 'any') {\n      candidates.push(...toolCallCandidates(req, i, state.toolCalls))\n    }\n  })\n  candidates.sort((a, b) => b.score - a.score)\n\n  const assigned = new Map<number, Candidate>()\n  const itemTaken = new Set<string>()\n  for (const c of candidates) {\n    if (assigned.has(c.reqIndex) || itemTaken.has(c.itemKey)) continue\n    assigned.set(c.reqIndex, c)\n    itemTaken.add(c.itemKey)\n  }\n\n  const requirements: RequirementCheck[] = []\n  for (let i = 0; i < gold.requirements.length; i++) {\n    const req = gold.requirements[i]!\n    const match = assigned.get(i)\n    const evidence: string[] = []\n    let correct: boolean | null = null\n    let unmeasuredReason: string | undefined\n\n    if (match) {\n      evidence.push(match.evidence)\n      if (match.content !== null) {\n        try {\n          const r = await checkCorrectness(req, match.content)\n          correct = r.correct\n          evidence.push(`correctness: ${r.correct ? 'pass' : 'fail'} — ${r.reason}`)\n        } catch (err) {\n          // The CHECKER failed, not the requirement. Recording this as a\n          // zero would fabricate a model failure out of an infrastructure\n          // one; the requirement is unmeasured and leaves the denominator.\n          unmeasuredReason =\n            err instanceof JudgeParseError\n              ? `checker response unparseable after retry: ${err.raw.slice(0, 200)}`\n              : `checker call failed: ${err instanceof Error ? err.message : String(err)}`\n          evidence.push(`correctness: UNMEASURED — ${unmeasuredReason}`)\n        }\n      } else {\n        evidence.push('correctness: not assessed — matched item carries no content')\n      }\n    } else {\n      const by = req.satisfiedBy ?? 'any'\n      const kind = by === 'any' ? 'artifact/proposal/tool-call' : by\n      evidence.push(`no produced ${kind} matched this requirement`)\n    }\n\n    const structurallyPresent = match !== undefined\n    const unmeasured = unmeasuredReason !== undefined\n    const satisfied = structurallyPresent && !unmeasured && correct !== false\n    requirements.push({\n      reqId: req.reqId,\n      title: req.title,\n      structurallyPresent,\n      correct,\n      satisfied,\n      ...(unmeasured ? { unmeasured: true as const, unmeasuredReason } : {}),\n      evidence,\n    })\n  }\n\n  // The verdict blends the deterministic structural stage with the injected\n  // correctness stage, so the certification names the checker's own member\n  // and carries the structural stage's lexical nature as an assumption. A\n  // checker without an attestation yields an uncertified verdict.\n  const attestation = checkCorrectness.attestation\n  const certification: VerdictCertification | undefined = attestation\n    ? {\n        strategy: attestation.strategy,\n        checker: attestation.checker,\n        assumptions: [\n          'structural matching is lexical token recall over produced items — the correctness stage only sees items it matched',\n          ...attestation.assumptions,\n        ],\n        evidenceDigest: certificationEvidenceDigest({ taskId: gold.taskId, requirements }),\n      }\n    : undefined\n\n  return completionVerdict({\n    taskId: gold.taskId,\n    requirements,\n    ...(certification === undefined ? {} : { certification }),\n  })\n}\n\nexport interface LlmCorrectnessCheckerOpts {\n  model?: string\n  /** Optional ledger for direct use. */\n  costLedger?: CostLedgerHandle\n  costPhase?: string\n  costTags?: Record<string, string>\n  signal?: AbortSignal\n  /** Max chars of artifact content sent to the checker. */\n  maxContentChars?: number\n  /**\n   * Checker LLM calls per requirement before giving up (parse failures and\n   * call errors both consume attempts). The failure then surfaces as an\n   * `unmeasured` requirement, never a zero.\n   */\n  maxAttempts?: number\n  /**\n   * Forensic capture of every checker request/response/error — without it a\n   * checker failure is unauditable (the agent-turn raws never contain the\n   * checker's own calls). Same sink contract as `LlmClient`.\n   */\n  rawSink?: RawProviderSink\n}\n\n/**\n * Parse the correctness checker's model response. Tolerates a response\n * truncated mid-JSON (max_tokens cap) by auto-closing the prefix — the\n * verdict boolean usually lands in the first few tokens, so a recovered\n * prefix with a boolean `correct` is a real measurement, not a guess.\n * Fails loud (JudgeParseError) when no boolean verdict is recoverable.\n */\nfunction parseCorrectnessResponse(raw: string): { correct: boolean; reason: string } {\n  const readVerdict = (candidate: unknown): { correct: boolean; reason: string } | null => {\n    if (candidate === null || typeof candidate !== 'object') return null\n    const { correct, reason } = candidate as { correct?: unknown; reason?: unknown }\n    if (typeof correct !== 'boolean') return null\n    return { correct, reason: typeof reason === 'string' ? reason : '' }\n  }\n\n  const match = raw.match(/\\{[\\s\\S]*\\}/)\n  if (match) {\n    try {\n      const strict = readVerdict(JSON.parse(match[0]))\n      if (strict) return strict\n    } catch {\n      // fall through to truncation recovery\n    }\n  }\n  // The strict path needs a closing `}`; a cap-hit response has none. Take\n  // everything from the first `{` and auto-close it.\n  const start = raw.indexOf('{')\n  if (start !== -1) {\n    const recovered = readVerdict(recoverTruncatedJson(raw.slice(start)))\n    if (recovered) return recovered\n  }\n  throw new JudgeParseError('correctness-checker', raw)\n}\n\n/**\n * Production `CorrectnessChecker` — one LLM call per matched artifact,\n * deterministic (temperature 0), structured JSON out. Judges fulfilment\n * only: a plan, a gesture, or a description of what should be done does not\n * fulfil a requirement — the artifact must BE the deliverable.\n */\nexport function createLlmCorrectnessChecker(\n  chat: ChatClient,\n  opts: LlmCorrectnessCheckerOpts = {},\n): CorrectnessChecker {\n  const model = opts.model ?? 'claude-sonnet-4-6'\n  const maxContentChars = opts.maxContentChars ?? 8000\n  const maxAttempts = opts.maxAttempts ?? 2\n  const costLedger = opts.costLedger ?? new CostLedger()\n  const sink = opts.rawSink\n  const record = async (event: RawProviderEvent): Promise<void> => {\n    // Forensic capture is best-effort; the verdict is the system of record.\n    try {\n      await sink?.record(event)\n    } catch {\n      // Intentionally swallowed.\n    }\n  }\n  const checker: CorrectnessChecker = async (requirement, content) => {\n    const request = {\n      model,\n      messages: [\n        {\n          role: 'system' as const,\n          content:\n            'You verify whether a produced work artifact actually fulfils a stated requirement. Judge fulfilment only — is the deliverable substantively present and on-point — not polish. A plan to do it later, a vague gesture, or a description of what should be done does NOT fulfil a requirement; the artifact must BE the deliverable. Respond with a single JSON object: {\"correct\": boolean, \"reason\": string (<= 30 words)}.',\n        },\n        {\n          role: 'user' as const,\n          content: `Requirement: ${requirement.title}\\n${\n            requirement.category ? `Category: ${requirement.category}\\n` : ''\n          }\\nProduced artifact:\\n${content.slice(0, maxContentChars)}`,\n        },\n      ],\n      temperature: 0,\n      maxTokens: 200,\n    } satisfies ChatRequest\n    let lastErr: unknown\n    for (let attempt = 0; attempt < maxAttempts; attempt++) {\n      const started = Date.now()\n      await record({\n        eventId: randomUUID(),\n        provider: chat.transport,\n        model,\n        endpoint: '/chat',\n        baseUrl: '',\n        attemptIndex: attempt,\n        direction: 'request',\n        timestamp: started,\n        requestBody: request,\n        redactedFields: [],\n      })\n      try {\n        const paid = await costLedger.runPaidCall({\n          channel: 'verifier',\n          phase: opts.costPhase ?? 'completion.correctness',\n          actor: 'correctness-checker',\n          model,\n          maximumCharge:\n            chat.maximumAttempts === undefined\n              ? undefined\n              : maximumChargeForLlmRequest(request, { maximumAttempts: chat.maximumAttempts }),\n          tags: {\n            ...opts.costTags,\n            requirementId: requirement.reqId,\n            attempt: String(attempt),\n          },\n          signal: opts.signal,\n          execute: (signal, callId) => chat.chat(request, { signal, idempotencyKey: callId }),\n          receipt: costReceiptFromLlm,\n          receiptFromError: costReceiptFromLlmError,\n        })\n        if (!paid.succeeded) throw paid.error\n        const resp = paid.value\n        // Hold the transport to its own word: a verdict produced by a model\n        // other than the one requested is not this model's verdict. A\n        // transport that echoes no id cannot be made to prove identity here —\n        // callers needing that proof enable it at the client\n        // (`LlmClientOptions.assertServedModel`) or gate on `assertModelsServed`.\n        assertServedModel(model, resp.servedModel, {\n          allowUnreported: true,\n          context: `correctness checker for requirement ${requirement.reqId}`,\n        })\n        const raw = resp.content\n        await record({\n          eventId: randomUUID(),\n          provider: chat.transport,\n          model,\n          endpoint: '/chat',\n          baseUrl: '',\n          attemptIndex: attempt,\n          direction: 'response',\n          timestamp: Date.now(),\n          durationMs: Date.now() - started,\n          responseBody: resp,\n          redactedFields: [],\n        })\n        return parseCorrectnessResponse(raw)\n      } catch (err) {\n        lastErr = err\n        await record({\n          eventId: randomUUID(),\n          provider: chat.transport,\n          model,\n          endpoint: '/chat',\n          baseUrl: '',\n          attemptIndex: attempt,\n          direction: 'error',\n          timestamp: Date.now(),\n          durationMs: Date.now() - started,\n          errorMessage: err instanceof Error ? err.message : String(err),\n          redactedFields: [],\n        })\n        // Substitution is a verdict about identity, not a transient fault:\n        // another attempt reaches the same wrong model and spends more tokens.\n        if (err instanceof ModelSubstitutionError) throw err\n      }\n    }\n    throw lastErr instanceof Error ? lastErr : new Error(String(lastErr))\n  }\n  checker.attestation = {\n    strategy: 'judge',\n    checker: { name: 'llm-correctness-checker', version: model },\n    assumptions: [\n      'served-model identity is accepted when the transport reports none (assertServedModel allowUnreported)',\n    ],\n  }\n  return checker\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiIA,SAAgB,kBAAkB,OAMZ;CACpB,IAAI,MAAM,aAAa,WAAW,GAChC,MAAM,IAAI,MACR,4BAA4B,MAAM,OAAO,+DAC3C;CAEF,MAAM,aAAa,MAAM,aAAa,QAAQ,MAAM,CAAC,EAAE,UAAU;CACjE,MAAM,kBAAkB,MAAM,aAAa,SAAS,WAAW;CAC/D,IAAI,WAAW,WAAW,GAGxB,MAAM,IAAI,MACR,4BAA4B,MAAM,OAAO,yCAAyC,MAAM,aAAa,OAAO,8BAA8B,MAAM,aAAa,EAAE,EAAE,oBAAoB,iBAAiB,EACxM;CAEF,MAAM,iBAAiB,WAAW,QAAQ,MAAM,EAAE,SAAS,CAAC,CAAC;CAC7D,MAAM,iBAAiB,iBAAiB,WAAW;CAGnD,MAAM,gBAAgB,oBAAoB,KAAK,mBAAmB,WAAW;CAC7E,OAAO;EACL,QAAQ,MAAM;EACd,cAAc,MAAM;EACpB;EACA;EACA;EACA,OAAO;EACP,OAAO;EACP,GAAI,MAAM,kBAAkB,KAAA,IAAY,CAAC,IAAI,EAAE,eAAe,MAAM,cAAc;CACpF;AACF;AA+BA,MAAM,4BAAY,IAAI,IAAI;CACxB;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAUD,MAAM,6CAA6B,IAAI,IAAI;CACzC;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAED,MAAM,kBAAkB;AACxB,MAAM,oBAAoB;AAE1B,SAAS,OAAO,GAAW,WAAsC;CAC/D,OAAO,IAAI,IACT,EACG,YAAY,CAAC,CACb,MAAM,YAAY,CAAC,CACnB,QAAQ,MAAM,EAAE,SAAS,KAAK,CAAC,UAAU,IAAI,CAAC,KAAK,CAAC,WAAW,IAAI,CAAC,CAAC,CAC1E;AACF;;;;;;;AAQA,SAAS,YAAY,iBAAyB,eAA+B;CAC3E,MAAM,MAAM,OAAO,iBAAiB,0BAA0B;CAC9D,IAAI,IAAI,SAAS,GAAG,OAAO;CAC3B,MAAM,OAAO,OAAO,aAAa;CACjC,IAAI,MAAM;CACV,KAAK,MAAM,KAAK,KAAK,IAAI,KAAK,IAAI,CAAC,GAAG;CACtC,OAAO,MAAM,IAAI;AACnB;AAYA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,UAAU,GAAG,IAAI,MAAM,GAAG,IAAI,YAAY;CAChD,MAAM,MAAmB,CAAC;CAC1B,UAAU,SAAS,GAAG,MAAM;EAC1B,KAAK,EAAE,WAAW,GAAA,CAAI,KAAK,CAAC,CAAC,SAAS,mBAAmB;EAKzD,IAAI,QAAQ,YACV,SACA,GAAG,EAAE,QAAQ,GAAG,GAAG,EAAE,KAAK,IAAI,EAAE,WAAW,GAAA,CAAI,MAAM,GAAG,GAAI,GAC9D;EACA,IAAI,IAAI,YAAY,EAAE,QAAQ,IAAI,SAAS,YAAY,MAAM,EAAE,KAAK,YAAY,GAC9E,QAAQ,KAAK,IAAI,OAAO,CAAC;EAE3B,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,YAAY;GACrB;GACA,UAAU,aAAa,EAAE,QAAQ,EAAE,KAAK,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACnF,SAAS,EAAE,WAAW;EACxB,CAAC;CACH,CAAC;CACD,OAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,UAAU,GAAG,IAAI,MAAM,GAAG,IAAI,YAAY;CAChD,MAAM,MAAmB,CAAC;CAC1B,KAAK,MAAM,KAAK,WAAW;EAEzB,IAAI,EAAE,WAAW,YAAY;EAM7B,MAAM,QAAQ,EAAE,WAAW,GAAA,CAAI,KAAK;EACpC,IAAI,KAAK,SAAS,mBAAmB;EASrC,MAAM,QAAQ,YAAY,SAAS,GAAG,EAAE,MAAM,GAAG,MAAM;EACvD,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,YAAY,EAAE;GACvB;GACA,UAAU,sBAAsB,EAAE,MAAM,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACnF,SAAS;EACX,CAAC;CACH;CACA,OAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,MAAmB,CAAC;CAC1B,UAAU,SAAS,MAAM,MAAM;EAC7B,MAAM,QAAQ,YAAY,IAAI,OAAO,IAAI;EACzC,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,QAAQ;GACjB;GACA,UAAU,cAAc,KAAK,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACxE,SAAS;EACX,CAAC;CACH,CAAC;CACD,OAAO;AACT;;;;;;;;AASA,eAAsB,iBACpB,MACA,OACA,kBAC4B;CAC5B,IAAI,KAAK,aAAa,WAAW,GAC/B,MAAM,IAAI,MACR,2BAA2B,KAAK,OAAO,4CACzC;CAMF,MAAM,aAA0B,CAAC;CACjC,KAAK,aAAa,SAAS,KAAK,MAAM;EACpC,MAAM,KAAK,IAAI,eAAe;EAC9B,IAAI,OAAO,cAAc,OAAO,OAC9B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;EAEhE,IAAI,OAAO,cAAc,OAAO,OAC9B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;EAEhE,IAAI,OAAO,eAAe,OAAO,OAC/B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;CAElE,CAAC;CACD,WAAW,MAAM,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;CAE3C,MAAM,2BAAW,IAAI,IAAuB;CAC5C,MAAM,4BAAY,IAAI,IAAY;CAClC,KAAK,MAAM,KAAK,YAAY;EAC1B,IAAI,SAAS,IAAI,EAAE,QAAQ,KAAK,UAAU,IAAI,EAAE,OAAO,GAAG;EAC1D,SAAS,IAAI,EAAE,UAAU,CAAC;EAC1B,UAAU,IAAI,EAAE,OAAO;CACzB;CAEA,MAAM,eAAmC,CAAC;CAC1C,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,aAAa,QAAQ,KAAK;EACjD,MAAM,MAAM,KAAK,aAAa;EAC9B,MAAM,QAAQ,SAAS,IAAI,CAAC;EAC5B,MAAM,WAAqB,CAAC;EAC5B,IAAI,UAA0B;EAC9B,IAAI;EAEJ,IAAI,OAAO;GACT,SAAS,KAAK,MAAM,QAAQ;GAC5B,IAAI,MAAM,YAAY,MACpB,IAAI;IACF,MAAM,IAAI,MAAM,iBAAiB,KAAK,MAAM,OAAO;IACnD,UAAU,EAAE;IACZ,SAAS,KAAK,gBAAgB,EAAE,UAAU,SAAS,OAAO,KAAK,EAAE,QAAQ;GAC3E,SAAS,KAAK;IAIZ,mBACE,eAAe,kBACX,6CAA6C,IAAI,IAAI,MAAM,GAAG,GAAG,MACjE,wBAAwB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;IAC7E,SAAS,KAAK,6BAA6B,kBAAkB;GAC/D;QAEA,SAAS,KAAK,6DAA6D;EAE/E,OAAO;GACL,MAAM,KAAK,IAAI,eAAe;GAC9B,MAAM,OAAO,OAAO,QAAQ,gCAAgC;GAC5D,SAAS,KAAK,eAAe,KAAK,0BAA0B;EAC9D;EAEA,MAAM,sBAAsB,UAAU,KAAA;EACtC,MAAM,aAAa,qBAAqB,KAAA;EACxC,MAAM,YAAY,uBAAuB,CAAC,cAAc,YAAY;EACpE,aAAa,KAAK;GAChB,OAAO,IAAI;GACX,OAAO,IAAI;GACX;GACA;GACA;GACA,GAAI,aAAa;IAAE,YAAY;IAAe;GAAiB,IAAI,CAAC;GACpE;EACF,CAAC;CACH;CAMA,MAAM,cAAc,iBAAiB;CACrC,MAAM,gBAAkD,cACpD;EACE,UAAU,YAAY;EACtB,SAAS,YAAY;EACrB,aAAa,CACX,sHACA,GAAG,YAAY,WACjB;EACA,gBAAgB,4BAA4B;GAAE,QAAQ,KAAK;GAAQ;EAAa,CAAC;CACnF,IACA,KAAA;CAEJ,OAAO,kBAAkB;EACvB,QAAQ,KAAK;EACb;EACA,GAAI,kBAAkB,KAAA,IAAY,CAAC,IAAI,EAAE,cAAc;CACzD,CAAC;AACH;;;;;;;;AAgCA,SAAS,yBAAyB,KAAmD;CACnF,MAAM,eAAe,cAAoE;EACvF,IAAI,cAAc,QAAQ,OAAO,cAAc,UAAU,OAAO;EAChE,MAAM,EAAE,SAAS,WAAW;EAC5B,IAAI,OAAO,YAAY,WAAW,OAAO;EACzC,OAAO;GAAE;GAAS,QAAQ,OAAO,WAAW,WAAW,SAAS;EAAG;CACrE;CAEA,MAAM,QAAQ,IAAI,MAAM,aAAa;CACrC,IAAI,OACF,IAAI;EACF,MAAM,SAAS,YAAY,KAAK,MAAM,MAAM,EAAE,CAAC;EAC/C,IAAI,QAAQ,OAAO;CACrB,QAAQ,CAER;CAIF,MAAM,QAAQ,IAAI,QAAQ,GAAG;CAC7B,IAAI,UAAU,IAAI;EAChB,MAAM,YAAY,YAAY,qBAAqB,IAAI,MAAM,KAAK,CAAC,CAAC;EACpE,IAAI,WAAW,OAAO;CACxB;CACA,MAAM,IAAI,gBAAgB,uBAAuB,GAAG;AACtD;;;;;;;AAQA,SAAgB,4BACd,MACA,OAAkC,CAAC,GACf;CACpB,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,kBAAkB,KAAK,mBAAmB;CAChD,MAAM,cAAc,KAAK,eAAe;CACxC,MAAM,aAAa,KAAK,cAAc,IAAI,WAAW;CACrD,MAAM,OAAO,KAAK;CAClB,MAAM,SAAS,OAAO,UAA2C;EAE/D,IAAI;GACF,MAAM,MAAM,OAAO,KAAK;EAC1B,QAAQ,CAER;CACF;CACA,MAAM,UAA8B,OAAO,aAAa,YAAY;EAClE,MAAM,UAAU;GACd;GACA,UAAU,CACR;IACE,MAAM;IACN,SACE;GACJ,GACA;IACE,MAAM;IACN,SAAS,gBAAgB,YAAY,MAAM,IACzC,YAAY,WAAW,aAAa,YAAY,SAAS,MAAM,GAChE,wBAAwB,QAAQ,MAAM,GAAG,eAAe;GAC3D,CACF;GACA,aAAa;GACb,WAAW;EACb;EACA,IAAI;EACJ,KAAK,IAAI,UAAU,GAAG,UAAU,aAAa,WAAW;GACtD,MAAM,UAAU,KAAK,IAAI;GACzB,MAAM,OAAO;IACX,SAAS,WAAW;IACpB,UAAU,KAAK;IACf;IACA,UAAU;IACV,SAAS;IACT,cAAc;IACd,WAAW;IACX,WAAW;IACX,aAAa;IACb,gBAAgB,CAAC;GACnB,CAAC;GACD,IAAI;IACF,MAAM,OAAO,MAAM,WAAW,YAAY;KACxC,SAAS;KACT,OAAO,KAAK,aAAa;KACzB,OAAO;KACP;KACA,eACE,KAAK,oBAAoB,KAAA,IACrB,KAAA,IACA,2BAA2B,SAAS,EAAE,iBAAiB,KAAK,gBAAgB,CAAC;KACnF,MAAM;MACJ,GAAG,KAAK;MACR,eAAe,YAAY;MAC3B,SAAS,OAAO,OAAO;KACzB;KACA,QAAQ,KAAK;KACb,UAAU,QAAQ,WAAW,KAAK,KAAK,SAAS;MAAE;MAAQ,gBAAgB;KAAO,CAAC;KAClF,SAAS;KACT,kBAAkB;IACpB,CAAC;IACD,IAAI,CAAC,KAAK,WAAW,MAAM,KAAK;IAChC,MAAM,OAAO,KAAK;IAMlB,kBAAkB,OAAO,KAAK,aAAa;KACzC,iBAAiB;KACjB,SAAS,uCAAuC,YAAY;IAC9D,CAAC;IACD,MAAM,MAAM,KAAK;IACjB,MAAM,OAAO;KACX,SAAS,WAAW;KACpB,UAAU,KAAK;KACf;KACA,UAAU;KACV,SAAS;KACT,cAAc;KACd,WAAW;KACX,WAAW,KAAK,IAAI;KACpB,YAAY,KAAK,IAAI,IAAI;KACzB,cAAc;KACd,gBAAgB,CAAC;IACnB,CAAC;IACD,OAAO,yBAAyB,GAAG;GACrC,SAAS,KAAK;IACZ,UAAU;IACV,MAAM,OAAO;KACX,SAAS,WAAW;KACpB,UAAU,KAAK;KACf;KACA,UAAU;KACV,SAAS;KACT,cAAc;KACd,WAAW;KACX,WAAW,KAAK,IAAI;KACpB,YAAY,KAAK,IAAI,IAAI;KACzB,cAAc,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;KAC7D,gBAAgB,CAAC;IACnB,CAAC;IAGD,IAAI,eAAe,wBAAwB,MAAM;GACnD;EACF;EACA,MAAM,mBAAmB,QAAQ,UAAU,IAAI,MAAM,OAAO,OAAO,CAAC;CACtE;CACA,QAAQ,cAAc;EACpB,UAAU;EACV,SAAS;GAAE,MAAM;GAA2B,SAAS;EAAM;EAC3D,aAAa,CACX,uGACF;CACF;CACA,OAAO;AACT"}

package/dist/contract/index.d.ts

@@ -1,14 +1,14 @@
-import { o as CostLedgerHandle } from "../cost-ledger-C1W3ICmP.js";
-import { S as makeProposalFinding, i as AnalystFinding, v as ProposalFinding, y as ProposalFindingOrigin } from "../types-6Lfdlshp.js";
-import { A as createChatClient, T as CreateChatClientOpts, x as ChatClient } from "../types-B_IFcKp5.js";
-import { $ as paretoPolicy, G as AxisVerdict, J as ObjectiveSource, K as BuildEvidenceVectorOptions, Q as buildEvidenceVector, W as AxisEvidence, X as PromotionObjective, Y as ParetoSignificanceGateOptions, Z as PromotionPolicy, et as paretoSignificanceGate, q as EvidenceVector } from "../statistical-heldout-BjzSOmLH.js";
-import { $r as OptimizationMethodResult, Bi as runCampaign, Br as OpenAICompatibleOptimizerModel, Cr as composeGate, Di as runImprovementLoop, Dr as skillOptOptimizationMethod, Ei as RunImprovementLoopResult, Er as SkillOptTrainerConfig, Fi as RunEvalOptions, Fr as GepaEngineRun, Gr as ExternalTextOptimizerResult, Hr as externalTextOptimizationMethod, Ii as runEval, Ir as GepaOptimizationMethodConfig, Kr as ExternalOptimizationExample, Li as CampaignCellFailureReceipt, Lr as GepaOptimizationRecipe, Mr as compareOptimizationMethods, Nr as GepaAdaptiveEngineRun, Or as CompareOptimizationMethodsOptions, Pr as GepaEngineOptions, Qr as OptimizationMethodProvenance, Ri as CampaignCellRetryPolicy, Rr as GepaRunnerCommand, Ti as RunImprovementLoopOptions, Tr as SkillOptRunnerCommand, Ur as ExternalTextOptimizationMethodConfig, Vr as OptimizerModelBudget, Wr as ExternalTextOptimizerContext, Xi as FinalEvidencePolicy, Yr as OptimizationMethod, Zi as FinalEvidenceUse, Zr as OptimizationMethodInput, _i as ReferenceEquivalenceJudgeResult, ar as DefaultProductionGateCheck, bi as LlmJudgeOptions, cr as defaultProductionGate, gi as ReferenceEquivalenceJudgeOptions, hi as ReferenceEquivalenceJudgeInput, ir as heldOutGate, kr as OptimizationMethodComparison, na as campaignSplitDigest, ni as ComparisonCost, or as DefaultProductionGateOptions, qr as ExternalTextEvaluationResponse, ri as OptimizationTokenUsage, rr as HeldOutGateOptions, sr as DefaultProductionRewardHackingOptions, ti as OptimizationPackageSource, vi as ReferenceEquivalenceScenario, wi as transientDispatchFailure, wr as SkillOptOptimizationMethodConfig, xi as llmJudge, yi as LlmJudgeDimension, zi as RunCampaignOptions, zr as gepaOptimizationMethod } from "../index-CdEVr8-O.js";
-import { C as JudgeDimension, H as SurfaceProposer, M as OptimizerConfig, R as Scenario, S as JudgeConfig, V as SessionScript, _ as GateDecision, a as CampaignResult, b as GenerationRecord, c as CampaignTraceWriter, d as DispatchContext, f as DispatchFn, g as GateContribution, h as GateContext, i as CampaignCostMeter, j as MutableSurface, l as CodeSurface, m as GateCheckStatus, n as CampaignArtifactWriter, p as Gate, r as CampaignCellResult, t as CampaignAggregates, v as GateResult, w as JudgeScore, y as GenerationCandidate } from "../types-M6eF9m-r.js";
-import { Kt as fsCampaignStorage, Wt as CampaignStorage, qt as inMemoryCampaignStorage } from "../search-kernel-CctpiDd2.js";
-import { d as FileSystemOutcomeStore, f as FileSystemOutcomeStoreOptions, h as OutcomeStore, p as InMemoryOutcomeStore, u as DeploymentOutcome } from "../rubric-predictive-validity-Cd5SCtAf.js";
-import { n as HostedTenant } from "../client-hwo3cJjR.js";
-import { _ as ReleaseSummary, a as summarizeExecution, c as ExecutionInsight, d as InsightReport, f as InterRaterInsight, g as Recommendation, h as OutcomeCorrelationInsight, i as SummarizeExecutionOptions, l as FailureClassTally, m as LiftInsight, n as analyzeRuns, o as CostProvenanceSummary, p as JudgeInsight, r as ExecutionReport, s as ExecutionErrorOutcomeCell, t as AnalyzeRunsOptions, u as FailureClusterInsight, v as ScalarDistribution, y as TokenUsageInsight } from "../analyze-runs-DfX6dHP0.js";
-import { _ as SelfImproveMethodResult, a as DefinedAgentEval, c as SelfImproveMethodOptions, d as SelfImproveProposerOptions, f as SelfImproveProposerResult, g as SelfImproveMethodProvenance, h as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveOptions, m as SelfImproveRunError, n as AgentEvalEvaluateOptions, o as defineAgentEval, p as SelfImproveResult, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveProgressEvent } from "../define-agent-eval-JKt3gMQr.js";
-import { $ as AgentProfileImprovementExperimentExecutionInput, A as CodeAgentSessionActionSurface, B as AgentTraceIndex, C as CodeAgentSessionIntakeOptions, Ct as PairedMeasurementAdapter, D as CodeAgentSessionAction, E as ParsedCodeAgentJsonl, Et as SealCandidateBenchmarkSuiteOptions, F as observeCodeAgentSession, G as parseAgentTrace, H as AgentTraceRecord, I as AgentTraceContributor, J as EvalReportingSuiteOptions, K as partitionRunsByAuthoringModel, L as AgentTraceContributorType, M as CodeAgentSessionObservation, N as CodeAgentSessionSource, O as CodeAgentSessionActionKind, P as CodeAgentSessionTerminalStatus, Q as RunRecordRejection, R as AgentTraceConversation, S as CodeAgentSessionDiagnostic, St as PairedMeasurement, T as CodeAgentSessionMetrics, Tt as RunCandidateExperimentOptions, U as AuthoringProvenance, V as AgentTraceRange, W as PartitionByAuthoringModelResult, X as FromRunRecordDirOptions, Y as EvalReportingSuiteResult, Z as FromRunRecordDirResult, _ as fromPiSession, _t as verifyCandidateExperimentComparison, a as FromFeedbackTableOptions, at as runAgentProfileImprovementExperiment, b as streamCodeAgentJsonlFile, bt as CompareCandidateExperimentOptions, c as CodeAgentStoreObservation, ct as sealAgentProfileImprovementTask, d as CodeAgentStoreScope, dt as measuredComparisonFromCandidateExperiment, et as AgentProfileImprovementExperimentRun, f as CodeAgentStoreSession, ft as runCandidateExperiment, g as fromOpenCodeSession, gt as verifyCandidateExperiment, h as fromKimiCodeSession, ht as sealCandidateExperiment, i as FeedbackTableRow, it as measuredComparisonFromAgentProfileImprovementExperiment, j as CodeAgentSessionExecutionReceipt, k as CodeAgentSessionActionStatus, l as CodeAgentStoreRejection, lt as verifyAgentProfileImprovementExperimentComparison, m as fromCodexSession, mt as sealCandidateBenchmarkTask, n as fromOtelSpans, nt as RunAgentProfileImprovementExperimentOptions, o as FromFeedbackTableResult, ot as sealAgentProfileImprovementExperiment, p as fromClaudeCodeSession, pt as sealCandidateBenchmarkSuite, q as EvalReportingSuiteInput, r as FeedbackTableMeta, rt as SealAgentProfileImprovementSuiteOptions, s as fromFeedbackTable, st as sealAgentProfileImprovementSuite, t as FromOtelSpansOptions, tt as CompareAgentProfileImprovementExperimentOptions, u as CodeAgentStoreRejections, ut as evaluatePairedMeasurements, v as parseCodeAgentJsonl, vt as CandidateExperimentExecutionInput, w as CodeAgentSessionIntakeResult, wt as PairedMeasurementEvaluation, x as CodeAgentJsonlLine, xt as EvaluatePairedMeasurementsOptions, y as parseCodeAgentJsonlFile, yt as CandidateExperimentRun, z as AgentTraceFile } from "../index-jnuFLxOs.js";
-import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-CLG5YqGS.js";
+import { o as CostLedgerHandle } from "../cost-ledger-Czro9wmo.js";
+import { S as makeProposalFinding, i as AnalystFinding, v as ProposalFinding, y as ProposalFindingOrigin } from "../types-DzCO-S1m.js";
+import { A as createChatClient, T as CreateChatClientOpts, x as ChatClient } from "../types-0KglnQ9o.js";
+import { $ as paretoPolicy, G as AxisVerdict, J as ObjectiveSource, K as BuildEvidenceVectorOptions, Q as buildEvidenceVector, W as AxisEvidence, X as PromotionObjective, Y as ParetoSignificanceGateOptions, Z as PromotionPolicy, et as paretoSignificanceGate, q as EvidenceVector } from "../statistical-heldout-Cei036nE.js";
+import { Ar as externalTextOptimizationMethod, Br as OptimizationMethodProvenance, Ci as RunEvalOptions, Cr as GepaEngineRun, Di as RunCampaignOptions, Dr as gepaOptimizationMethod, Ei as CampaignCellRetryPolicy, Er as GepaRunnerCommand, Fr as ExternalTextEvaluationResponse, Gn as HeldOutGateOptions, Gr as OptimizationTokenUsage, Jn as DefaultProductionGateOptions, Kn as heldOutGate, Lr as OptimizationMethod, Mr as ExternalTextOptimizerContext, Nr as ExternalTextOptimizerResult, Oi as runCampaign, Or as OpenAICompatibleOptimizerModel, Pr as ExternalOptimizationExample, Ri as FinalEvidencePolicy, Sr as GepaEngineOptions, Ti as CampaignCellFailureReceipt, Tr as GepaOptimizationRecipe, Ur as OptimizationPackageSource, Vr as OptimizationMethodResult, Wi as campaignSplitDigest, Wr as ComparisonCost, Xn as defaultProductionGate, Yn as DefaultProductionRewardHackingOptions, _r as OptimizationMethodComparison, ai as ReferenceEquivalenceJudgeResult, br as compareOptimizationMethods, ci as LlmJudgeOptions, dr as composeGate, fi as transientDispatchFailure, fr as SkillOptOptimizationMethodConfig, gr as CompareOptimizationMethodsOptions, hi as runImprovementLoop, hr as skillOptOptimizationMethod, ii as ReferenceEquivalenceJudgeOptions, jr as ExternalTextOptimizationMethodConfig, kr as OptimizerModelBudget, li as llmJudge, mi as RunImprovementLoopResult, mr as SkillOptTrainerConfig, oi as ReferenceEquivalenceScenario, pi as RunImprovementLoopOptions, pr as SkillOptRunnerCommand, qn as DefaultProductionGateCheck, ri as ReferenceEquivalenceJudgeInput, si as LlmJudgeDimension, wi as runEval, wr as GepaOptimizationMethodConfig, xr as GepaAdaptiveEngineRun, zi as FinalEvidenceUse, zr as OptimizationMethodInput } from "../index-B7v2QHlD.js";
+import { C as JudgeDimension, H as SurfaceProposer, M as OptimizerConfig, R as Scenario, S as JudgeConfig, V as SessionScript, _ as GateDecision, a as CampaignResult, b as GenerationRecord, c as CampaignTraceWriter, d as DispatchContext, f as DispatchFn, g as GateContribution, h as GateContext, i as CampaignCostMeter, j as MutableSurface, l as CodeSurface, m as GateCheckStatus, n as CampaignArtifactWriter, p as Gate, r as CampaignCellResult, t as CampaignAggregates, v as GateResult, w as JudgeScore, y as GenerationCandidate } from "../types-C54Uk_a1.js";
+import { Kt as fsCampaignStorage, Wt as CampaignStorage, qt as inMemoryCampaignStorage } from "../search-kernel-CBdwFZ05.js";
+import { d as FileSystemOutcomeStore, f as FileSystemOutcomeStoreOptions, h as OutcomeStore, p as InMemoryOutcomeStore, u as DeploymentOutcome } from "../rubric-predictive-validity-B3RzbwOx.js";
+import { n as HostedTenant } from "../client-BIFwUg7I.js";
+import { _ as ReleaseSummary, a as summarizeExecution, c as ExecutionInsight, d as InsightReport, f as InterRaterInsight, g as Recommendation, h as OutcomeCorrelationInsight, i as SummarizeExecutionOptions, l as FailureClassTally, m as LiftInsight, n as analyzeRuns, o as CostProvenanceSummary, p as JudgeInsight, r as ExecutionReport, s as ExecutionErrorOutcomeCell, t as AnalyzeRunsOptions, u as FailureClusterInsight, v as ScalarDistribution, y as TokenUsageInsight } from "../analyze-runs-C-pP1sK9.js";
+import { _ as SelfImproveMethodResult, a as DefinedAgentEval, c as SelfImproveMethodOptions, d as SelfImproveProposerOptions, f as SelfImproveProposerResult, g as SelfImproveMethodProvenance, h as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveOptions, m as SelfImproveRunError, n as AgentEvalEvaluateOptions, o as defineAgentEval, p as SelfImproveResult, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveProgressEvent } from "../define-agent-eval-CAffiFbN.js";
+import { $ as AgentProfileImprovementExperimentExecutionInput, A as CodeAgentSessionActionSurface, B as AgentTraceIndex, C as CodeAgentSessionIntakeOptions, Ct as PairedMeasurementAdapter, D as CodeAgentSessionAction, E as ParsedCodeAgentJsonl, Et as SealCandidateBenchmarkSuiteOptions, F as observeCodeAgentSession, G as parseAgentTrace, H as AgentTraceRecord, I as AgentTraceContributor, J as EvalReportingSuiteOptions, K as partitionRunsByAuthoringModel, L as AgentTraceContributorType, M as CodeAgentSessionObservation, N as CodeAgentSessionSource, O as CodeAgentSessionActionKind, P as CodeAgentSessionTerminalStatus, Q as RunRecordRejection, R as AgentTraceConversation, S as CodeAgentSessionDiagnostic, St as PairedMeasurement, T as CodeAgentSessionMetrics, Tt as RunCandidateExperimentOptions, U as AuthoringProvenance, V as AgentTraceRange, W as PartitionByAuthoringModelResult, X as FromRunRecordDirOptions, Y as EvalReportingSuiteResult, Z as FromRunRecordDirResult, _ as fromPiSession, _t as verifyCandidateExperimentComparison, a as FromFeedbackTableOptions, at as runAgentProfileImprovementExperiment, b as streamCodeAgentJsonlFile, bt as CompareCandidateExperimentOptions, c as CodeAgentStoreObservation, ct as sealAgentProfileImprovementTask, d as CodeAgentStoreScope, dt as measuredComparisonFromCandidateExperiment, et as AgentProfileImprovementExperimentRun, f as CodeAgentStoreSession, ft as runCandidateExperiment, g as fromOpenCodeSession, gt as verifyCandidateExperiment, h as fromKimiCodeSession, ht as sealCandidateExperiment, i as FeedbackTableRow, it as measuredComparisonFromAgentProfileImprovementExperiment, j as CodeAgentSessionExecutionReceipt, k as CodeAgentSessionActionStatus, l as CodeAgentStoreRejection, lt as verifyAgentProfileImprovementExperimentComparison, m as fromCodexSession, mt as sealCandidateBenchmarkTask, n as fromOtelSpans, nt as RunAgentProfileImprovementExperimentOptions, o as FromFeedbackTableResult, ot as sealAgentProfileImprovementExperiment, p as fromClaudeCodeSession, pt as sealCandidateBenchmarkSuite, q as EvalReportingSuiteInput, r as FeedbackTableMeta, rt as SealAgentProfileImprovementSuiteOptions, s as fromFeedbackTable, st as sealAgentProfileImprovementSuite, t as FromOtelSpansOptions, tt as CompareAgentProfileImprovementExperimentOptions, u as CodeAgentStoreRejections, ut as evaluatePairedMeasurements, v as parseCodeAgentJsonl, vt as CandidateExperimentExecutionInput, w as CodeAgentSessionIntakeResult, wt as PairedMeasurementEvaluation, x as CodeAgentJsonlLine, xt as EvaluatePairedMeasurementsOptions, y as parseCodeAgentJsonlFile, yt as CandidateExperimentRun, z as AgentTraceFile } from "../index-aDPVYBIM.js";
+import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-B1DstPYg.js";
 export { type AgentEvalAgent, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, type AgentProfileImprovementExperimentExecutionInput, type AgentProfileImprovementExperimentRun, type AgentTraceContributor, type AgentTraceContributorType, type AgentTraceConversation, type AgentTraceFile, type AgentTraceIndex, type AgentTraceRange, type AgentTraceRecord, type AnalystFinding, type AnalyzeRunsOptions, type AuthoringProvenance, type AxisEvidence, type AxisVerdict, type BuildEvidenceVectorOptions, type CampaignAggregates, type CampaignArtifactWriter, type CampaignCellFailureReceipt, type CampaignCellResult, type CampaignCellRetryPolicy, type CampaignCostMeter, type CampaignResult, type CampaignStorage, type CampaignTraceWriter, type CandidateExperimentExecutionInput, type CandidateExperimentRun, type ChatClient, type CodeAgentJsonlLine, type CodeAgentSessionAction, type CodeAgentSessionActionKind, type CodeAgentSessionActionStatus, type CodeAgentSessionActionSurface, type CodeAgentSessionDiagnostic, type CodeAgentSessionExecutionReceipt, type CodeAgentSessionIntakeOptions, type CodeAgentSessionIntakeResult, type CodeAgentSessionMetrics, type CodeAgentSessionObservation, type CodeAgentSessionSource, type CodeAgentSessionTerminalStatus, type CodeAgentStoreObservation, type CodeAgentStoreRejection, type CodeAgentStoreRejections, type CodeAgentStoreScope, type CodeAgentStoreSession, type CodeSurface, type CompareAgentProfileImprovementExperimentOptions, type CompareCandidateExperimentOptions, type CompareOptimizationMethodsOptions, type ComparisonCost, type CostLedgerHandle, type CostProvenanceSummary, type CreateChatClientOpts, type DefaultAnalystRegistryOptions, type DefaultProductionGateCheck, type DefaultProductionGateOptions, type DefaultProductionRewardHackingOptions, type DefineAgentEvalOptions, type DefinedAgentEval, type DeploymentOutcome, type DispatchFn as Dispatch, type DispatchContext, type EvalReportingSuiteInput, type EvalReportingSuiteOptions, type EvalReportingSuiteResult, type EvaluatePairedMeasurementsOptions, type EvidenceVector, type ExecutionErrorOutcomeCell, type ExecutionInsight, type ExecutionReport, type ExternalOptimizationExample, type ExternalTextEvaluationResponse, type ExternalTextOptimizationMethodConfig, type ExternalTextOptimizerContext, type ExternalTextOptimizerResult, type FailureClassTally, type FailureClusterInsight, type FeedbackTableMeta, type FeedbackTableRow, FileSystemOutcomeStore, type FileSystemOutcomeStoreOptions, type FinalEvidencePolicy, type FinalEvidenceUse, type FromFeedbackTableOptions, type FromFeedbackTableResult, type FromOtelSpansOptions, type FromRunRecordDirOptions, type FromRunRecordDirResult, type Gate, type GateCheckStatus, type GateContext, type GateContribution, type GateDecision, type GateResult, type GenerationCandidate, type GenerationRecord, type GepaAdaptiveEngineRun, type GepaEngineOptions, type GepaEngineRun, type GepaOptimizationMethodConfig, type GepaOptimizationRecipe, type GepaRunnerCommand, type HeldOutGateOptions, type HostedTenant, InMemoryOutcomeStore, type InsightReport, type InterRaterInsight, type JudgeConfig, type JudgeDimension, type JudgeInsight, type JudgeScore, type LiftInsight, type LlmJudgeDimension, type LlmJudgeOptions, type MutableSurface, type ObjectiveSource, type OpenAICompatibleOptimizerModel, type OptimizationMethod, type OptimizationMethodComparison, type OptimizationMethodInput, type OptimizationMethodProvenance, type OptimizationMethodResult, type OptimizationPackageSource, type OptimizationTokenUsage, type OptimizerConfig, type OptimizerModelBudget, type OutcomeCorrelationInsight, type OutcomeStore, type PairedMeasurement, type PairedMeasurementAdapter, type PairedMeasurementEvaluation, type ParetoSignificanceGateOptions, type ParsedCodeAgentJsonl, type PartitionByAuthoringModelResult, type PromotionObjective, type PromotionPolicy, type ProposalFinding, type ProposalFindingOrigin, type Recommendation, type ReferenceEquivalenceJudgeInput, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceJudgeResult, type ReferenceEquivalenceScenario, type ReleaseSummary, type RunAgentProfileImprovementExperimentOptions, type RunCampaignOptions, type RunCandidateExperimentOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunRecordRejection, type ScalarDistribution, type Scenario, type SealAgentProfileImprovementSuiteOptions, type SealCandidateBenchmarkSuiteOptions, type SelfImproveBudget, type SelfImproveMethodOptions, type SelfImproveMethodProvenance, type SelfImproveMethodResult, type SelfImproveOptions, type SelfImproveProgressEvent, type SelfImproveProposerOptions, type SelfImproveProposerResult, type SelfImproveResult, SelfImproveRunError, type SessionScript, type SkillOptOptimizationMethodConfig, type SkillOptRunnerCommand, type SkillOptTrainerConfig, type SummarizeExecutionOptions, type SurfaceProposer, type TokenUsageInsight, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, defaultProductionGate, defineAgentEval, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiSession, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, makeProposalFinding, measuredComparisonFromAgentProfileImprovementExperiment, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, parseCodeAgentJsonlFile, partitionRunsByAuthoringModel, runAgentProfileImprovementExperiment, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, sealAgentProfileImprovementExperiment, sealAgentProfileImprovementSuite, sealAgentProfileImprovementTask, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, streamCodeAgentJsonlFile, summarizeExecution, transientDispatchFailure, verifyAgentProfileImprovementExperimentComparison, verifyCandidateExperiment, verifyCandidateExperimentComparison };