@tangle-network/agent-eval 0.91.0 → 0.92.0

package/dist/campaign/index.d.ts

@@ -7,7 +7,7 @@ export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, k
 import { L as LlmClientOptions } from '../llm-client-BeEcAokY.js';
 import { T as TraceAnalystKindSpec } from '../kind-factory-5b7xXXOr.js';
 import { c as AnalystFinding } from '../types-2VVIL04s.js';
-import { S as SignedManifest, A as AgentProfile, B as BackendIntegrityReport, C as CompletionRequirement, R as RuntimeEventLike, a as CompletionVerdict, P as ProducedState, b as CorrectnessChecker } from '../pre-registration-DELOEJ8v.js';
+import { S as SignedManifest, A as AgentProfile, B as BackendIntegrityReport, C as CompletionRequirement, R as RuntimeEventLike, a as CompletionVerdict, P as ProducedState, b as CorrectnessChecker } from '../pre-registration-mAnCugl9.js';
 import { E as EProcessState, a as PairedBootstrapResult } from '../statistics-C7PozGrZ.js';
 import { A as AgentEvalError } from '../errors-CzMUYo7b.js';
 import { a as RunSplitTag, R as RunRecord } from '../run-record-e7vj1uZQ.js';

package/dist/campaign/index.js

@@ -15,7 +15,7 @@ import {
   agentProfileHash,
   extractProducedState,
   verifyCompletion
-} from "../chunk-FIUKOSWI.js";
+} from "../chunk-S6OZEZQK.js";
 import {
   buildLoopProvenanceRecord,
   campaignBreakdown,

package/dist/{chunk-FIUKOSWI.js → chunk-S6OZEZQK.js}

@@ -38,15 +38,36 @@ var STOPWORDS = /* @__PURE__ */ new Set([
   "with",
   "by"
 ]);
+var REQUIREMENT_FORM_STOPWORDS = /* @__PURE__ */ new Set([
+  "generated",
+  "generate",
+  "view",
+  "render",
+  "rendered",
+  "persisted",
+  "persist",
+  "artifact",
+  "file",
+  "document",
+  "note",
+  "proposal",
+  "deliverable",
+  "output",
+  "created",
+  "create",
+  "produce",
+  "produced",
+  "flag"
+]);
 var MATCH_THRESHOLD = 0.5;
 var MIN_CONTENT_CHARS = 50;
-function tokens(s) {
+function tokens(s, extraStop) {
   return new Set(
-    s.toLowerCase().split(/[^a-z0-9]+/).filter((t) => t.length > 1 && !STOPWORDS.has(t))
+    s.toLowerCase().split(/[^a-z0-9]+/).filter((t) => t.length > 1 && !STOPWORDS.has(t) && !extraStop?.has(t))
   );
 }
 function tokenRecall(requirementText, candidateText) {
-  const req = tokens(requirementText);
+  const req = tokens(requirementText, REQUIREMENT_FORM_STOPWORDS);
   if (req.size === 0) return 0;
   const cand = tokens(candidateText);
   let hit = 0;
@@ -58,7 +79,10 @@ function artifactCandidates(req, reqIndex, artifacts) {
   const out = [];
   artifacts.forEach((a, i) => {
     if ((a.content ?? "").trim().length < MIN_CONTENT_CHARS) return;
-    let score = tokenRecall(reqText, `${a.path ?? ""} ${a.kind}`);
+    let score = tokenRecall(
+      reqText,
+      `${a.path ?? ""} ${a.kind} ${(a.content ?? "").slice(0, 4e3)}`
+    );
     if (req.category && a.kind && req.category.toLowerCase() === a.kind.toLowerCase()) {
       score = Math.max(score, 1);
     }
@@ -78,15 +102,16 @@ function proposalCandidates(req, reqIndex, proposals) {
   const out = [];
   for (const p of proposals) {
     if (p.status !== "approved") continue;
-    const score = tokenRecall(reqText, p.title);
+    const body = (p.content ?? "").trim();
+    if (body.length < MIN_CONTENT_CHARS) continue;
+    const score = tokenRecall(reqText, `${p.title} ${body}`);
     if (score < MATCH_THRESHOLD) continue;
-    const body = p.content ?? "";
     out.push({
       reqIndex,
       itemKey: `proposal:${p.id}`,
       score,
       evidence: `approved proposal '${p.title}' matched (token recall ${score.toFixed(2)})`,
-      content: body.trim().length >= MIN_CONTENT_CHARS ? body : null
+      content: body
     });
   }
   return out;
@@ -278,7 +303,12 @@ function extractProducedState(events) {
       });
     } else if (ev.type === "proposal_created") {
       const p = ev;
-      proposals.push({ id: p.proposalId, title: p.title, status: p.status ?? "pending" });
+      proposals.push({
+        id: p.proposalId,
+        title: p.title,
+        status: p.status ?? "pending",
+        ...p.content !== void 0 ? { content: p.content } : {}
+      });
     }
   }
   return { artifacts, proposals, toolCalls };
@@ -309,4 +339,4 @@ export {
   extractProducedState,
   agentProfileHash
 };
-//# sourceMappingURL=chunk-FIUKOSWI.js.map
+//# sourceMappingURL=chunk-S6OZEZQK.js.map

package/dist/chunk-S6OZEZQK.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/completion-verifier.ts","../src/produced-state.ts","../src/agent-profile.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 / total. Quality dimensions are meaningless\n * on an incomplete task — callers gate on `fullyComplete` / `completionRate`\n * before scoring quality.\n */\n\nimport type { TCloud } from '@tangle-network/tcloud'\nimport type { Artifact } from './artifact-validator'\nimport type { DefaultVerdict } from './verdict'\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 `Artifact` shape. */\n  artifacts: Artifact[]\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, or when the matched item carries no content\n   * to assess.\n   */\n  correct: boolean | null\n  /** structurallyPresent && correct !== false. */\n  satisfied: boolean\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 / total requirements. */\n  completionRate: number\n  /** Every requirement satisfied. */\n  fullyComplete: boolean\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}): 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 satisfiedCount = input.requirements.filter((r) => r.satisfied).length\n  const completionRate = satisfiedCount / input.requirements.length\n  const fullyComplete = satisfiedCount === input.requirements.length\n  return {\n    taskId: input.taskId,\n    requirements: input.requirements,\n    completionRate,\n    fullyComplete,\n    valid: fullyComplete,\n    score: completionRate,\n  }\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 */\nexport type CorrectnessChecker = (\n  requirement: CompletionRequirement,\n  content: string,\n) => Promise<{ correct: boolean; reason: string }>\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: Artifact[],\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\n    if (match) {\n      evidence.push(match.evidence)\n      if (match.content !== null) {\n        const r = await checkCorrectness(req, match.content)\n        correct = r.correct\n        evidence.push(`correctness: ${r.correct ? 'pass' : 'fail'} — ${r.reason}`)\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 satisfied = structurallyPresent && correct !== false\n    requirements.push({\n      reqId: req.reqId,\n      title: req.title,\n      structurallyPresent,\n      correct,\n      satisfied,\n      evidence,\n    })\n  }\n\n  return completionVerdict({ taskId: gold.taskId, requirements })\n}\n\nexport interface LlmCorrectnessCheckerOpts {\n  model?: string\n  /** Max chars of artifact content sent to the checker. */\n  maxContentChars?: number\n}\n\n/** Parse the correctness checker's model response. Fails loud on a bad shape. */\nexport function parseCorrectnessResponse(raw: string): { correct: boolean; reason: string } {\n  const match = raw.match(/\\{[\\s\\S]*\\}/)\n  if (!match) {\n    throw new Error(`correctness checker: no JSON object in model response: ${raw.slice(0, 200)}`)\n  }\n  const parsed = JSON.parse(match[0]) as { correct?: unknown; reason?: unknown }\n  if (typeof parsed.correct !== 'boolean') {\n    throw new Error(`correctness checker: 'correct' is not a boolean in: ${match[0].slice(0, 200)}`)\n  }\n  return { correct: parsed.correct, reason: typeof parsed.reason === 'string' ? parsed.reason : '' }\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  tc: TCloud,\n  opts: LlmCorrectnessCheckerOpts = {},\n): CorrectnessChecker {\n  const model = opts.model ?? 'claude-sonnet-4-6'\n  const maxContentChars = opts.maxContentChars ?? 8000\n  return async (requirement, content) => {\n    const resp = await tc.chat({\n      model,\n      messages: [\n        {\n          role: 'system',\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',\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    })\n    const raw =\n      (resp as { choices?: { message?: { content?: string } }[] }).choices?.[0]?.message?.content ??\n      ''\n    return parseCorrectnessResponse(raw)\n  }\n}\n\n/** Stopwords for requirement-title tokenization — drops the imperative verbs\n *  ('review', 'update', …) common to deliverable titles so recall keys on the\n *  substantive nouns, not the boilerplate ask. */\nconst TITLE_STOPWORDS = new Set([\n  'the',\n  'a',\n  'an',\n  'and',\n  'or',\n  'for',\n  'to',\n  'of',\n  'in',\n  'on',\n  'with',\n  'review',\n  'update',\n  'new',\n  'proposed',\n])\n\n/**\n * Deterministic `CorrectnessChecker` — the no-LLM counterpart to\n * `createLlmCorrectnessChecker`. A produced item fulfils a requirement when its\n * content is substantive (≥ `minContentLength` chars) AND recalls ≥ `minRecall`\n * of the requirement title's significant tokens. No network.\n *\n * Polarity-blind: token recall credits a negation that contains the\n * requirement's tokens (\"I will NOT produce the comparison\" recalls every token\n * of \"produce the comparison\"). The structural match stage is ALSO lexical, so\n * pairing the two collapses to a single gameable gate. Use this only as an\n * opt-in structural pre-filter or for tasks whose requirements have no polarity\n * to invert; for produced-state grading the correctness checker MUST be semantic\n * (`createLlmCorrectnessChecker`). See the anti-game fixtures in the test suite.\n */\nexport function createTokenRecallChecker(\n  opts: { minRecall?: number; minContentLength?: number } = {},\n): CorrectnessChecker {\n  const minRecall = opts.minRecall ?? 0.5\n  const minLen = opts.minContentLength ?? 120\n  return async (requirement, content) => {\n    const body = content.trim()\n    if (body.length < minLen)\n      return {\n        correct: false,\n        reason: `content too thin (${body.length} chars) to be the deliverable`,\n      }\n    const titleTokens = requirement.title\n      .toLowerCase()\n      .split(/[^a-z0-9]+/)\n      .filter((t) => t.length > 2 && !TITLE_STOPWORDS.has(t))\n    if (titleTokens.length === 0)\n      return {\n        correct: true,\n        reason: 'requirement title has no significant tokens — structural match accepted',\n      }\n    const lower = body.toLowerCase()\n    const hits = titleTokens.filter((t) => lower.includes(t)).length\n    const recall = hits / titleTokens.length\n    return recall >= minRecall\n      ? {\n          correct: true,\n          reason: `content recalls ${hits}/${titleTokens.length} requirement tokens`,\n        }\n      : {\n          correct: false,\n          reason: `content recalls only ${hits}/${titleTokens.length} requirement tokens`,\n        }\n  }\n}\n","/**\n * Produced-state extraction — normalize a run's runtime event stream into the\n * typed `ProducedState` the completion oracle consumes.\n *\n * `ProducedState` answers \"what did the agent actually produce\" — vault\n * artifacts, proposals, tool calls. The runtime emits these as a stream of\n * events; this module is the single normalization point from that stream to\n * the shape `verifyCompletion` expects.\n *\n * Input is structurally typed (`RuntimeEventLike`) so this module does not\n * depend on agent-runtime — agent-runtime's `RuntimeStreamEvent` satisfies it\n * structurally. The `content` on `ArtifactEventLike` and the whole\n * `proposal_created` variant are the runtime-side enrichments this contract\n * requires; the runtime emits them, this module consumes them.\n */\n\nimport type { Artifact } from './artifact-validator'\nimport type { ProducedProposal, ProducedState } from './completion-verifier'\n\n/** A tool the agent invoked. */\nexport interface ToolCallEventLike {\n  type: 'tool_call'\n  toolName: string\n}\n\n/**\n * An artifact the agent produced. `content` is the enriched field — the\n * runtime's base `artifact` event carries only metadata; the completion\n * oracle needs the body to verify the deliverable, so the runtime emits it.\n */\nexport interface ArtifactEventLike {\n  type: 'artifact'\n  artifactId: string\n  name?: string\n  mimeType?: string\n  uri?: string\n  content?: string\n}\n\n/** A proposal / filing the agent created. */\nexport interface ProposalEventLike {\n  type: 'proposal_created'\n  proposalId: string\n  title: string\n  status?: 'pending' | 'approved' | 'rejected'\n  // body of the proposal (e.g. a submit_proposal `description`). When present,\n  // the completion oracle correctness-checks it like artifact content; absent,\n  // the proposal is graded presence-only.\n  content?: string\n}\n\n/**\n * The subset of runtime stream events `extractProducedState` consumes.\n * agent-runtime's full `RuntimeStreamEvent` union satisfies this structurally;\n * the `{ type: string }` catch-all keeps the input permissive so callers can\n * pass the whole unfiltered telemetry stream — unrecognized events are skipped.\n */\nexport type RuntimeEventLike =\n  | ToolCallEventLike\n  | ArtifactEventLike\n  | ProposalEventLike\n  | { type: string }\n\nfunction artifactKind(mimeType: string | undefined): string {\n  if (!mimeType) return 'file'\n  if (mimeType.includes('json')) return 'json'\n  if (mimeType.startsWith('text/')) return 'text'\n  return 'file'\n}\n\n/**\n * Normalize a run's runtime event stream into `ProducedState`.\n *\n * Pure and total — unrecognized event types are skipped. `toolCalls` is\n * deduplicated by name in first-seen order (completion cares about a tool's\n * presence, not its call count). An artifact with neither a name nor a uri\n * still yields an entry keyed by its `artifactId` so it is never silently\n * dropped; an artifact with no `content` yields empty content, which the\n * completion oracle's structural check then rejects on its own.\n */\nexport function extractProducedState(events: readonly RuntimeEventLike[]): ProducedState {\n  const artifacts: Artifact[] = []\n  const proposals: ProducedProposal[] = []\n  const toolCalls: string[] = []\n  const seenTools = new Set<string>()\n\n  for (const ev of events) {\n    if (ev.type === 'tool_call') {\n      const name = (ev as ToolCallEventLike).toolName\n      if (name && !seenTools.has(name)) {\n        seenTools.add(name)\n        toolCalls.push(name)\n      }\n    } else if (ev.type === 'artifact') {\n      const a = ev as ArtifactEventLike\n      artifacts.push({\n        kind: artifactKind(a.mimeType),\n        path: a.name ?? a.uri ?? a.artifactId,\n        content: a.content ?? '',\n      })\n    } else if (ev.type === 'proposal_created') {\n      const p = ev as ProposalEventLike\n      proposals.push({\n        id: p.proposalId,\n        title: p.title,\n        status: p.status ?? 'pending',\n        ...(p.content !== undefined ? { content: p.content } : {}),\n      })\n    }\n  }\n\n  return { artifacts, proposals, toolCalls }\n}\n","/**\n * @stable\n *\n * AgentProfile — the eval harness's unit of variation.\n *\n * A profile pins everything that changes agent behaviour for a benchmark\n * cell: the model, the active skills, the prompt version, the available\n * tools. Vary the profile — swap a model, add a skill — and re-run the suite\n * to benchmark the change. The scorecard keys a cell on\n * `(scenarioId, profileHash)`, so the model is not a separate axis: it lives\n * inside the profile, and two profiles with the same model but different\n * skills are different cells.\n *\n * `agentProfileHash` is the profile's behaviour identity. Two profiles that\n * produce the same agent behaviour share a hash (and a scorecard cell);\n * reordering `skills` or `tools` does not change it; the human-facing `id`\n * label does not affect it.\n */\n\nimport { createHash } from 'node:crypto'\nimport { ValidationError } from './errors'\nimport { canonicalize } from './pre-registration'\n\nexport interface AgentProfile {\n  /** Human-facing label, e.g. `sonnet-legal-skills-v3`. Not part of the hash. */\n  id: string\n  /** Model snapshot id this profile pins, e.g. `claude-sonnet-4-6@2025-04-15`. */\n  model: string\n  /** Skill ids/versions active in this profile — the primary behaviour lever. */\n  skills?: string[]\n  /** Prompt version identifier. */\n  promptVersion?: string\n  /** Tool ids available to the agent. */\n  tools?: string[]\n  /** Any other behaviour-bearing knobs that should fingerprint into the hash. */\n  metadata?: Record<string, string | number | boolean>\n}\n\n/**\n * Deterministic behaviour identity of a profile — a sha256 over the\n * behaviour-bearing fields. `skills` and `tools` are order-insensitive; the\n * `id` label is excluded. Throws on a profile with no `model` — an unkeyable\n * profile must fail loud rather than collapse into a blank-model cell.\n */\nexport function agentProfileHash(profile: AgentProfile): string {\n  if (typeof profile.model !== 'string' || profile.model.trim().length === 0) {\n    throw new ValidationError(`AgentProfile \"${profile.id}\" has no model — cannot hash`)\n  }\n  const behaviour = {\n    model: profile.model.trim(),\n    skills: [...(profile.skills ?? [])].sort(),\n    promptVersion: profile.promptVersion ?? null,\n    tools: [...(profile.tools ?? [])].sort(),\n    metadata: profile.metadata ?? {},\n  }\n  return createHash('sha256')\n    .update(JSON.stringify(canonicalize(behaviour)))\n    .digest('hex')\n}\n"],"mappings":";;;;;;;;AAmGO,SAAS,kBAAkB,OAGZ;AACpB,MAAI,MAAM,aAAa,WAAW,GAAG;AACnC,UAAM,IAAI;AAAA,MACR,4BAA4B,MAAM,MAAM;AAAA,IAC1C;AAAA,EACF;AACA,QAAM,iBAAiB,MAAM,aAAa,OAAO,CAAC,MAAM,EAAE,SAAS,EAAE;AACrE,QAAM,iBAAiB,iBAAiB,MAAM,aAAa;AAC3D,QAAM,gBAAgB,mBAAmB,MAAM,aAAa;AAC5D,SAAO;AAAA,IACL,QAAQ,MAAM;AAAA,IACd,cAAc,MAAM;AAAA,IACpB;AAAA,IACA;AAAA,IACA,OAAO;AAAA,IACP,OAAO;AAAA,EACT;AACF;AAYA,IAAM,YAAY,oBAAI,IAAI;AAAA,EACxB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAUD,IAAM,6BAA6B,oBAAI,IAAI;AAAA,EACzC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAED,IAAM,kBAAkB;AACxB,IAAM,oBAAoB;AAE1B,SAAS,OAAO,GAAW,WAAsC;AAC/D,SAAO,IAAI;AAAA,IACT,EACG,YAAY,EACZ,MAAM,YAAY,EAClB,OAAO,CAAC,MAAM,EAAE,SAAS,KAAK,CAAC,UAAU,IAAI,CAAC,KAAK,CAAC,WAAW,IAAI,CAAC,CAAC;AAAA,EAC1E;AACF;AAQA,SAAS,YAAY,iBAAyB,eAA+B;AAC3E,QAAM,MAAM,OAAO,iBAAiB,0BAA0B;AAC9D,MAAI,IAAI,SAAS,EAAG,QAAO;AAC3B,QAAM,OAAO,OAAO,aAAa;AACjC,MAAI,MAAM;AACV,aAAW,KAAK,IAAK,KAAI,KAAK,IAAI,CAAC,EAAG;AACtC,SAAO,MAAM,IAAI;AACnB;AAYA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,UAAU,GAAG,IAAI,KAAK,IAAI,IAAI,YAAY,EAAE;AAClD,QAAM,MAAmB,CAAC;AAC1B,YAAU,QAAQ,CAAC,GAAG,MAAM;AAC1B,SAAK,EAAE,WAAW,IAAI,KAAK,EAAE,SAAS,kBAAmB;AAKzD,QAAI,QAAQ;AAAA,MACV;AAAA,MACA,GAAG,EAAE,QAAQ,EAAE,IAAI,EAAE,IAAI,KAAK,EAAE,WAAW,IAAI,MAAM,GAAG,GAAI,CAAC;AAAA,IAC/D;AACA,QAAI,IAAI,YAAY,EAAE,QAAQ,IAAI,SAAS,YAAY,MAAM,EAAE,KAAK,YAAY,GAAG;AACjF,cAAQ,KAAK,IAAI,OAAO,CAAC;AAAA,IAC3B;AACA,QAAI,QAAQ,gBAAiB;AAC7B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,YAAY,CAAC;AAAA,MACtB;AAAA,MACA,UAAU,aAAa,EAAE,QAAQ,EAAE,IAAI,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MAClF,SAAS,EAAE,WAAW;AAAA,IACxB,CAAC;AAAA,EACH,CAAC;AACD,SAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,UAAU,GAAG,IAAI,KAAK,IAAI,IAAI,YAAY,EAAE;AAClD,QAAM,MAAmB,CAAC;AAC1B,aAAW,KAAK,WAAW;AAEzB,QAAI,EAAE,WAAW,WAAY;AAM7B,UAAM,QAAQ,EAAE,WAAW,IAAI,KAAK;AACpC,QAAI,KAAK,SAAS,kBAAmB;AASrC,UAAM,QAAQ,YAAY,SAAS,GAAG,EAAE,KAAK,IAAI,IAAI,EAAE;AACvD,QAAI,QAAQ,gBAAiB;AAC7B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,YAAY,EAAE,EAAE;AAAA,MACzB;AAAA,MACA,UAAU,sBAAsB,EAAE,KAAK,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MAClF,SAAS;AAAA,IACX,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,MAAmB,CAAC;AAC1B,YAAU,QAAQ,CAAC,MAAM,MAAM;AAC7B,UAAM,QAAQ,YAAY,IAAI,OAAO,IAAI;AACzC,QAAI,QAAQ,gBAAiB;AAC7B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,QAAQ,CAAC;AAAA,MAClB;AAAA,MACA,UAAU,cAAc,IAAI,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MACvE,SAAS;AAAA,IACX,CAAC;AAAA,EACH,CAAC;AACD,SAAO;AACT;AASA,eAAsB,iBACpB,MACA,OACA,kBAC4B;AAC5B,MAAI,KAAK,aAAa,WAAW,GAAG;AAClC,UAAM,IAAI;AAAA,MACR,2BAA2B,KAAK,MAAM;AAAA,IACxC;AAAA,EACF;AAKA,QAAM,aAA0B,CAAC;AACjC,OAAK,aAAa,QAAQ,CAAC,KAAK,MAAM;AACpC,UAAM,KAAK,IAAI,eAAe;AAC9B,QAAI,OAAO,cAAc,OAAO,OAAO;AACrC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AACA,QAAI,OAAO,cAAc,OAAO,OAAO;AACrC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AACA,QAAI,OAAO,eAAe,OAAO,OAAO;AACtC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AAAA,EACF,CAAC;AACD,aAAW,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AAE3C,QAAM,WAAW,oBAAI,IAAuB;AAC5C,QAAM,YAAY,oBAAI,IAAY;AAClC,aAAW,KAAK,YAAY;AAC1B,QAAI,SAAS,IAAI,EAAE,QAAQ,KAAK,UAAU,IAAI,EAAE,OAAO,EAAG;AAC1D,aAAS,IAAI,EAAE,UAAU,CAAC;AAC1B,cAAU,IAAI,EAAE,OAAO;AAAA,EACzB;AAEA,QAAM,eAAmC,CAAC;AAC1C,WAAS,IAAI,GAAG,IAAI,KAAK,aAAa,QAAQ,KAAK;AACjD,UAAM,MAAM,KAAK,aAAa,CAAC;AAC/B,UAAM,QAAQ,SAAS,IAAI,CAAC;AAC5B,UAAM,WAAqB,CAAC;AAC5B,QAAI,UAA0B;AAE9B,QAAI,OAAO;AACT,eAAS,KAAK,MAAM,QAAQ;AAC5B,UAAI,MAAM,YAAY,MAAM;AAC1B,cAAM,IAAI,MAAM,iBAAiB,KAAK,MAAM,OAAO;AACnD,kBAAU,EAAE;AACZ,iBAAS,KAAK,gBAAgB,EAAE,UAAU,SAAS,MAAM,WAAM,EAAE,MAAM,EAAE;AAAA,MAC3E,OAAO;AACL,iBAAS,KAAK,kEAA6D;AAAA,MAC7E;AAAA,IACF,OAAO;AACL,YAAM,KAAK,IAAI,eAAe;AAC9B,YAAM,OAAO,OAAO,QAAQ,gCAAgC;AAC5D,eAAS,KAAK,eAAe,IAAI,2BAA2B;AAAA,IAC9D;AAEA,UAAM,sBAAsB,UAAU;AACtC,UAAM,YAAY,uBAAuB,YAAY;AACrD,iBAAa,KAAK;AAAA,MAChB,OAAO,IAAI;AAAA,MACX,OAAO,IAAI;AAAA,MACX;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAAA,EACH;AAEA,SAAO,kBAAkB,EAAE,QAAQ,KAAK,QAAQ,aAAa,CAAC;AAChE;AASO,SAAS,yBAAyB,KAAmD;AAC1F,QAAM,QAAQ,IAAI,MAAM,aAAa;AACrC,MAAI,CAAC,OAAO;AACV,UAAM,IAAI,MAAM,0DAA0D,IAAI,MAAM,GAAG,GAAG,CAAC,EAAE;AAAA,EAC/F;AACA,QAAM,SAAS,KAAK,MAAM,MAAM,CAAC,CAAC;AAClC,MAAI,OAAO,OAAO,YAAY,WAAW;AACvC,UAAM,IAAI,MAAM,uDAAuD,MAAM,CAAC,EAAE,MAAM,GAAG,GAAG,CAAC,EAAE;AAAA,EACjG;AACA,SAAO,EAAE,SAAS,OAAO,SAAS,QAAQ,OAAO,OAAO,WAAW,WAAW,OAAO,SAAS,GAAG;AACnG;AAQO,SAAS,4BACd,IACA,OAAkC,CAAC,GACf;AACpB,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,kBAAkB,KAAK,mBAAmB;AAChD,SAAO,OAAO,aAAa,YAAY;AACrC,UAAM,OAAO,MAAM,GAAG,KAAK;AAAA,MACzB;AAAA,MACA,UAAU;AAAA,QACR;AAAA,UACE,MAAM;AAAA,UACN,SACE;AAAA,QACJ;AAAA,QACA;AAAA,UACE,MAAM;AAAA,UACN,SAAS,gBAAgB,YAAY,KAAK;AAAA,EACxC,YAAY,WAAW,aAAa,YAAY,QAAQ;AAAA,IAAO,EACjE;AAAA;AAAA,EAAyB,QAAQ,MAAM,GAAG,eAAe,CAAC;AAAA,QAC5D;AAAA,MACF;AAAA,MACA,aAAa;AAAA,MACb,WAAW;AAAA,IACb,CAAC;AACD,UAAM,MACH,KAA4D,UAAU,CAAC,GAAG,SAAS,WACpF;AACF,WAAO,yBAAyB,GAAG;AAAA,EACrC;AACF;AAKA,IAAM,kBAAkB,oBAAI,IAAI;AAAA,EAC9B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAgBM,SAAS,yBACd,OAA0D,CAAC,GACvC;AACpB,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,SAAS,KAAK,oBAAoB;AACxC,SAAO,OAAO,aAAa,YAAY;AACrC,UAAM,OAAO,QAAQ,KAAK;AAC1B,QAAI,KAAK,SAAS;AAChB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,QAAQ,qBAAqB,KAAK,MAAM;AAAA,MAC1C;AACF,UAAM,cAAc,YAAY,MAC7B,YAAY,EACZ,MAAM,YAAY,EAClB,OAAO,CAAC,MAAM,EAAE,SAAS,KAAK,CAAC,gBAAgB,IAAI,CAAC,CAAC;AACxD,QAAI,YAAY,WAAW;AACzB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,QAAQ;AAAA,MACV;AACF,UAAM,QAAQ,KAAK,YAAY;AAC/B,UAAM,OAAO,YAAY,OAAO,CAAC,MAAM,MAAM,SAAS,CAAC,CAAC,EAAE;AAC1D,UAAM,SAAS,OAAO,YAAY;AAClC,WAAO,UAAU,YACb;AAAA,MACE,SAAS;AAAA,MACT,QAAQ,mBAAmB,IAAI,IAAI,YAAY,MAAM;AAAA,IACvD,IACA;AAAA,MACE,SAAS;AAAA,MACT,QAAQ,wBAAwB,IAAI,IAAI,YAAY,MAAM;AAAA,IAC5D;AAAA,EACN;AACF;;;AC/bA,SAAS,aAAa,UAAsC;AAC1D,MAAI,CAAC,SAAU,QAAO;AACtB,MAAI,SAAS,SAAS,MAAM,EAAG,QAAO;AACtC,MAAI,SAAS,WAAW,OAAO,EAAG,QAAO;AACzC,SAAO;AACT;AAYO,SAAS,qBAAqB,QAAoD;AACvF,QAAM,YAAwB,CAAC;AAC/B,QAAM,YAAgC,CAAC;AACvC,QAAM,YAAsB,CAAC;AAC7B,QAAM,YAAY,oBAAI,IAAY;AAElC,aAAW,MAAM,QAAQ;AACvB,QAAI,GAAG,SAAS,aAAa;AAC3B,YAAM,OAAQ,GAAyB;AACvC,UAAI,QAAQ,CAAC,UAAU,IAAI,IAAI,GAAG;AAChC,kBAAU,IAAI,IAAI;AAClB,kBAAU,KAAK,IAAI;AAAA,MACrB;AAAA,IACF,WAAW,GAAG,SAAS,YAAY;AACjC,YAAM,IAAI;AACV,gBAAU,KAAK;AAAA,QACb,MAAM,aAAa,EAAE,QAAQ;AAAA,QAC7B,MAAM,EAAE,QAAQ,EAAE,OAAO,EAAE;AAAA,QAC3B,SAAS,EAAE,WAAW;AAAA,MACxB,CAAC;AAAA,IACH,WAAW,GAAG,SAAS,oBAAoB;AACzC,YAAM,IAAI;AACV,gBAAU,KAAK;AAAA,QACb,IAAI,EAAE;AAAA,QACN,OAAO,EAAE;AAAA,QACT,QAAQ,EAAE,UAAU;AAAA,QACpB,GAAI,EAAE,YAAY,SAAY,EAAE,SAAS,EAAE,QAAQ,IAAI,CAAC;AAAA,MAC1D,CAAC;AAAA,IACH;AAAA,EACF;AAEA,SAAO,EAAE,WAAW,WAAW,UAAU;AAC3C;;;AC7FA,SAAS,kBAAkB;AAyBpB,SAAS,iBAAiB,SAA+B;AAC9D,MAAI,OAAO,QAAQ,UAAU,YAAY,QAAQ,MAAM,KAAK,EAAE,WAAW,GAAG;AAC1E,UAAM,IAAI,gBAAgB,iBAAiB,QAAQ,EAAE,mCAA8B;AAAA,EACrF;AACA,QAAM,YAAY;AAAA,IAChB,OAAO,QAAQ,MAAM,KAAK;AAAA,IAC1B,QAAQ,CAAC,GAAI,QAAQ,UAAU,CAAC,CAAE,EAAE,KAAK;AAAA,IACzC,eAAe,QAAQ,iBAAiB;AAAA,IACxC,OAAO,CAAC,GAAI,QAAQ,SAAS,CAAC,CAAE,EAAE,KAAK;AAAA,IACvC,UAAU,QAAQ,YAAY,CAAC;AAAA,EACjC;AACA,SAAO,WAAW,QAAQ,EACvB,OAAO,KAAK,UAAU,aAAa,SAAS,CAAC,CAAC,EAC9C,OAAO,KAAK;AACjB;","names":[]}

package/dist/index.d.ts

@@ -14,8 +14,8 @@ import { A as AgentEvalError, J as JudgeError, a as ConfigError } from './errors
 export { b as AgentEvalErrorCode, C as CaptureIntegrityError, N as NotFoundError, R as ReplayError, V as ValidationError, c as VerificationError } from './errors-CzMUYo7b.js';
 import { b as FeedbackLabel, F as FeedbackTrajectoryStore, a as FeedbackTrajectory } from './feedback-trajectory-D9OVLrg9.js';
 export { c as FeedbackArtifactType, d as FeedbackAttempt, e as FeedbackLabelKind, f as FeedbackLabelSource, g as FeedbackOptimizerRow, h as FeedbackOutcome, i as FeedbackReplayAdapter, j as FeedbackReplayResult, k as FeedbackSeverity, l as FeedbackSplitPolicy, m as FeedbackTask, n as FeedbackTrajectoryFilter, o as FileSystemFeedbackTrajectoryStore, I as InMemoryFeedbackTrajectoryStore, P as PreferenceMemoryEntry, p as ProposedSideEffect, q as assignFeedbackSplit, r as controlRunToFeedbackTrajectory, s as createFeedbackTrajectory, t as feedbackTrajectoriesToDatasetScenarios, u as feedbackTrajectoriesToOptimizerRows, v as feedbackTrajectoryToDatasetScenario, w as feedbackTrajectoryToOptimizerRow, x as parseFeedbackTrajectoriesJsonl, y as renderPreferenceMemoryMarkdown, z as replayFeedbackTrajectories, A as replayFeedbackTrajectory, B as serializeFeedbackTrajectoriesJsonl, C as summarizePreferenceMemory, D as withAssignedFeedbackSplit } from './feedback-trajectory-D9OVLrg9.js';
-import { b as CorrectnessChecker, A as AgentProfile$1 } from './pre-registration-DELOEJ8v.js';
-export { c as ArtifactCheckArtifact, d as ArtifactEventLike, e as ArtifactValidator, f as BackendIntegrityError, B as BackendIntegrityReport, C as CompletionRequirement, a as CompletionVerdict, H as HypothesisManifest, g as HypothesisResult, L as LlmCorrectnessCheckerOpts, h as ProducedProposal, P as ProducedState, i as ProposalEventLike, j as RequirementCheck, R as RuntimeEventLike, k as SatisfiedBy, S as SignedManifest, l as SignedManifestAlgo, T as TaskGold, m as ToolCallEventLike, V as ValidationContext, n as ValidationIssue, o as ValidationResult, p as agentProfileHash, q as assertRealBackend, r as byteLengthRange, s as canonicalize, t as completionVerdict, u as composeValidators, v as containsAll, w as createLlmCorrectnessChecker, x as createTokenRecallChecker, y as evaluateHypothesis, z as extractProducedState, D as hashJson, E as jsonHasKeys, F as parseCorrectnessResponse, G as regexMatch, I as signManifest, J as summarizeBackendIntegrity, K as verifyCompletion, M as verifyManifest } from './pre-registration-DELOEJ8v.js';
+import { b as CorrectnessChecker, A as AgentProfile$1 } from './pre-registration-mAnCugl9.js';
+export { c as ArtifactCheckArtifact, d as ArtifactEventLike, e as ArtifactValidator, f as BackendIntegrityError, B as BackendIntegrityReport, C as CompletionRequirement, a as CompletionVerdict, H as HypothesisManifest, g as HypothesisResult, L as LlmCorrectnessCheckerOpts, h as ProducedProposal, P as ProducedState, i as ProposalEventLike, j as RequirementCheck, R as RuntimeEventLike, k as SatisfiedBy, S as SignedManifest, l as SignedManifestAlgo, T as TaskGold, m as ToolCallEventLike, V as ValidationContext, n as ValidationIssue, o as ValidationResult, p as agentProfileHash, q as assertRealBackend, r as byteLengthRange, s as canonicalize, t as completionVerdict, u as composeValidators, v as containsAll, w as createLlmCorrectnessChecker, x as createTokenRecallChecker, y as evaluateHypothesis, z as extractProducedState, D as hashJson, E as jsonHasKeys, F as parseCorrectnessResponse, G as regexMatch, I as signManifest, J as summarizeBackendIntegrity, K as verifyCompletion, M as verifyManifest } from './pre-registration-mAnCugl9.js';
 export { DataAcquisitionPlan, KnowledgeAcquisitionMode, KnowledgeBundle, KnowledgeFallbackPolicy, KnowledgeFreshness, KnowledgeImportance, KnowledgeReadinessReport, KnowledgeRecommendedAction, KnowledgeRequirement, KnowledgeRequirementCategory, KnowledgeResponsibleSurface, KnowledgeSensitivity, ScoreKnowledgeReadinessOptions, UserQuestion, acquisitionPlansForKnowledgeGaps, blockingKnowledgeEval, knowledgeReadinessTracePayload, scoreKnowledgeReadiness, userQuestionsForKnowledgeGaps } from './knowledge/index.js';
 import { h as ReleaseConfidenceThresholds, f as ReleaseConfidenceScorecard } from './release-report-euXIV_Sk.js';
 export { A as ActionableSideInfo, o as AsiSeverity, B as BootstrapOptions, a as BootstrapResult, J as JudgeReplayGateArgs, R as ReleaseConfidenceAxis, b as ReleaseConfidenceAxisName, c as ReleaseConfidenceInput, d as ReleaseConfidenceIssue, e as ReleaseConfidenceMetrics, g as ReleaseConfidenceStatus, i as ReleaseTraceEvidence, j as RenderReleaseReportOptions, V as Verdict, k as assertReleaseConfidence, l as bootstrapCi, m as evaluateReleaseConfidence, n as judgeReplayGate, r as renderReleaseReport } from './release-report-euXIV_Sk.js';

package/dist/index.js

@@ -41,7 +41,7 @@ import {
   extractProducedState,
   parseCorrectnessResponse,
   verifyCompletion
-} from "./chunk-FIUKOSWI.js";
+} from "./chunk-S6OZEZQK.js";
 import {
   DEFAULT_MUTATION_PRIMITIVES,
   DEFAULT_RED_TEAM_CORPUS,

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.91.0",
+    "version": "0.92.0",
     "description": "HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: 1.0.0. Bumps on breaking changes to request/response schemas.",
     "contact": {
       "name": "Tangle Network",

package/dist/{pre-registration-DELOEJ8v.d.ts → pre-registration-mAnCugl9.d.ts}

@@ -277,9 +277,15 @@ declare function createLlmCorrectnessChecker(tc: TCloud, opts?: LlmCorrectnessCh
  * Deterministic `CorrectnessChecker` — the no-LLM counterpart to
  * `createLlmCorrectnessChecker`. A produced item fulfils a requirement when its
  * content is substantive (≥ `minContentLength` chars) AND recalls ≥ `minRecall`
- * of the requirement title's significant tokens. No network — the default gate
- * for apps and tests without an LLM judge. Pass to `verifyCompletion` as the
- * checker.
+ * of the requirement title's significant tokens. No network.
+ *
+ * Polarity-blind: token recall credits a negation that contains the
+ * requirement's tokens ("I will NOT produce the comparison" recalls every token
+ * of "produce the comparison"). The structural match stage is ALSO lexical, so
+ * pairing the two collapses to a single gameable gate. Use this only as an
+ * opt-in structural pre-filter or for tasks whose requirements have no polarity
+ * to invert; for produced-state grading the correctness checker MUST be semantic
+ * (`createLlmCorrectnessChecker`). See the anti-game fixtures in the test suite.
  */
 declare function createTokenRecallChecker(opts?: {
     minRecall?: number;
@@ -326,6 +332,7 @@ interface ProposalEventLike {
     proposalId: string;
     title: string;
     status?: 'pending' | 'approved' | 'rejected';
+    content?: string;
 }
 /**
  * The subset of runtime stream events `extractProducedState` consumes.

package/docs/concepts.md

@@ -183,6 +183,7 @@ release decision.
 ## Where to go next
 
 - **Confused by "GEPA / HALO / trace analysis / drivers everywhere"?** → [self-improvement-map.md](./self-improvement-map.md) — one loop, four roles, the seven-driver catalog (production vs bench-only), and why `gepa-refine` is the same loop on a test bench.
+- **Which `run*` primitive do I use, and how do I grade produced state?** → [eval-surface-map.md](./eval-surface-map.md) — the campaign/matrix/optimization/gate primitives as a pick-by-"use-when" table, plus the produced-state grading composition (verifyCompletion-as-judge — there is no persona-dispatch wrapper) and the in-band body contract.
 - **Need the layman feature map?** → [feature-guide.md](./feature-guide.md) — what each primitive does, when to use it, integration patterns, and guardrails.
 - **Just want to score a string against a rubric?** → [wire-protocol.md](./wire-protocol.md) — HTTP/RPC interface, pluggable from any language.
 - **Need a reusable driver/worker/evaluator loop?** → [control-runtime.md](./control-runtime.md) — generic runtime plus coding, browser, computer-use, and research integration patterns.

package/docs/eval-surface-map.md

@@ -0,0 +1,59 @@
+# Eval surface map — which primitive, when
+
+The eval surface is a small set of orthogonal primitives. They compose; they do
+not overlap. If two seem interchangeable, read the "use when" — the distinction
+is real and load-bearing. **Do not add a new wrapper to bridge two of these; the
+composition point already exists (see Produced-state grading below).**
+
+## The run\* primitives
+
+| Primitive | Use when | Returns |
+|---|---|---|
+| `runCampaign` | The measurement primitive. Run a dispatch over scenarios × seeds × reps, score each with judges, aggregate. Caller owns the dispatch. | `CampaignResult` |
+| `runEval` | The simplest preset over `runCampaign` — just score, no loop, no gate. The 80% "I want a scorecard" case. | `CampaignResult` |
+| `runProfileMatrix` | Factor the SAME scenarios across N agent **profiles** (models / prompt variants), with RunRecord stamping + a real-backend integrity guard. | `RunRecord[]` |
+| `runOptimization` | GENERATE: baseline → N generations of propose → measure → rank → promote. No release gate. | generations + winner |
+| `runImprovementLoop` | The release-gate shell around `runOptimization`: adds a held-out re-score + a promotion gate (+ optional auto-PR). | gate decision + winner |
+| `runEvalCampaign` | Inversion-of-control variant of `runCampaign` — the runner is handed a pre-wired trace/sink/emitter and integrity gating as a precondition. Use when you need full capture by construction. | `CampaignResult` + records |
+| `runAgentMatrix` | The bare N-axis cartesian scheduler with concurrency control. The layer beneath the eval surface — reach for it only when you need raw scheduling, not eval semantics. | cell results |
+
+Mental model: **measure** (`runCampaign`/`runEval`) → **factor** (`runProfileMatrix`) →
+**generate** (`runOptimization`) → **gate** (`runImprovementLoop`). `runEvalCampaign`
+is `runCampaign` with capture inverted; `runAgentMatrix` is the scheduler underneath.
+
+Merging any two of these conflates distinct mental models (measure ≠ search ≠
+release-gate). Keep them separate; pick by the table.
+
+## Produced-state grading — there is NO persona-dispatch wrapper
+
+To grade what an agent actually **produced** (filed the proposal, wrote the
+artifact) rather than what it said, the composition point is a **judge that wraps
+`verifyCompletion`** — not a dedicated runner. The pipeline:
+
+```
+runtime/app-tool events ──► extractProducedState(events) ──► ProducedState
+                                                                  │
+                            verifyCompletion(taskGold, state, correctnessChecker)
+                                                                  │
+                            inject as a JudgeConfig into runProfileMatrix / runCampaign
+```
+
+`extractProducedState` is a pure function over the produced-event stream; the
+judge calls it inline. This is why **`runProducedStatePersonaDispatch` does not
+exist and should not be built** — it would be a fourth layer over a composition
+that is already one judge. (Archetype: `playback.ts` `scoreUserStory`.)
+
+### The in-band body contract
+
+Produced events carry their **body in-band** — the grader never reaches into a
+product database to recover it:
+
+- `artifact` events carry `content` (the persisted file body).
+- `proposal_created` events carry `content` (the `submit_proposal` description) —
+  same role, same field name. A title-only filing omits it; a content-less
+  proposal is graded presence-only (and, by the completion oracle's rule, does
+  not count as a completed deliverable).
+
+A consumer that finds itself re-fetching a deliverable's body from its own DB to
+grade it is working around a thin event — fix the event (carry `content`), don't
+add an enrichment band-aid.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-eval",
-  "version": "0.91.0",
+  "version": "0.92.0",
   "description": "Evaluate and improve AI agents from runs, traces, judges, and feedback. Compare candidates, cluster failures, measure lift, and gate releases.",
   "homepage": "https://github.com/tangle-network/agent-eval#readme",
   "repository": {

package/dist/chunk-FIUKOSWI.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/completion-verifier.ts","../src/produced-state.ts","../src/agent-profile.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 / total. Quality dimensions are meaningless\n * on an incomplete task — callers gate on `fullyComplete` / `completionRate`\n * before scoring quality.\n */\n\nimport type { TCloud } from '@tangle-network/tcloud'\nimport type { Artifact } from './artifact-validator'\nimport type { DefaultVerdict } from './verdict'\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 `Artifact` shape. */\n  artifacts: Artifact[]\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, or when the matched item carries no content\n   * to assess.\n   */\n  correct: boolean | null\n  /** structurallyPresent && correct !== false. */\n  satisfied: boolean\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 / total requirements. */\n  completionRate: number\n  /** Every requirement satisfied. */\n  fullyComplete: boolean\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}): 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 satisfiedCount = input.requirements.filter((r) => r.satisfied).length\n  const completionRate = satisfiedCount / input.requirements.length\n  const fullyComplete = satisfiedCount === input.requirements.length\n  return {\n    taskId: input.taskId,\n    requirements: input.requirements,\n    completionRate,\n    fullyComplete,\n    valid: fullyComplete,\n    score: completionRate,\n  }\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 */\nexport type CorrectnessChecker = (\n  requirement: CompletionRequirement,\n  content: string,\n) => Promise<{ correct: boolean; reason: string }>\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\nconst MATCH_THRESHOLD = 0.5\nconst MIN_CONTENT_CHARS = 50\n\nfunction tokens(s: 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)),\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.\n */\nfunction tokenRecall(requirementText: string, candidateText: string): number {\n  const req = tokens(requirementText)\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: Artifact[],\n): Candidate[] {\n  const reqText = `${req.title} ${req.category ?? ''}`\n  const out: Candidate[] = []\n  artifacts.forEach((a, i) => {\n    if ((a.content ?? '').trim().length < MIN_CONTENT_CHARS) return\n    let score = tokenRecall(reqText, `${a.path ?? ''} ${a.kind}`)\n    if (req.category && a.kind && req.category.toLowerCase() === a.kind.toLowerCase()) {\n      score = Math.max(score, 1)\n    }\n    if (score < MATCH_THRESHOLD) return\n    out.push({\n      reqIndex,\n      itemKey: `artifact:${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} ${req.category ?? ''}`\n  const out: Candidate[] = []\n  for (const p of proposals) {\n    // Pending or rejected work is not a completed deliverable.\n    if (p.status !== 'approved') continue\n    const score = tokenRecall(reqText, p.title)\n    if (score < MATCH_THRESHOLD) continue\n    const body = p.content ?? ''\n    out.push({\n      reqIndex,\n      itemKey: `proposal:${p.id}`,\n      score,\n      evidence: `approved proposal '${p.title}' matched (token recall ${score.toFixed(2)})`,\n      content: body.trim().length >= MIN_CONTENT_CHARS ? body : null,\n    })\n  }\n  return out\n}\n\nfunction toolCallCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  toolCalls: string[],\n): Candidate[] {\n  const out: Candidate[] = []\n  toolCalls.forEach((name, i) => {\n    const score = tokenRecall(req.title, name)\n    if (score < MATCH_THRESHOLD) return\n    out.push({\n      reqIndex,\n      itemKey: `tool:${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\n    if (match) {\n      evidence.push(match.evidence)\n      if (match.content !== null) {\n        const r = await checkCorrectness(req, match.content)\n        correct = r.correct\n        evidence.push(`correctness: ${r.correct ? 'pass' : 'fail'} — ${r.reason}`)\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 satisfied = structurallyPresent && correct !== false\n    requirements.push({\n      reqId: req.reqId,\n      title: req.title,\n      structurallyPresent,\n      correct,\n      satisfied,\n      evidence,\n    })\n  }\n\n  return completionVerdict({ taskId: gold.taskId, requirements })\n}\n\nexport interface LlmCorrectnessCheckerOpts {\n  model?: string\n  /** Max chars of artifact content sent to the checker. */\n  maxContentChars?: number\n}\n\n/** Parse the correctness checker's model response. Fails loud on a bad shape. */\nexport function parseCorrectnessResponse(raw: string): { correct: boolean; reason: string } {\n  const match = raw.match(/\\{[\\s\\S]*\\}/)\n  if (!match) {\n    throw new Error(`correctness checker: no JSON object in model response: ${raw.slice(0, 200)}`)\n  }\n  const parsed = JSON.parse(match[0]) as { correct?: unknown; reason?: unknown }\n  if (typeof parsed.correct !== 'boolean') {\n    throw new Error(`correctness checker: 'correct' is not a boolean in: ${match[0].slice(0, 200)}`)\n  }\n  return { correct: parsed.correct, reason: typeof parsed.reason === 'string' ? parsed.reason : '' }\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  tc: TCloud,\n  opts: LlmCorrectnessCheckerOpts = {},\n): CorrectnessChecker {\n  const model = opts.model ?? 'claude-sonnet-4-6'\n  const maxContentChars = opts.maxContentChars ?? 8000\n  return async (requirement, content) => {\n    const resp = await tc.chat({\n      model,\n      messages: [\n        {\n          role: 'system',\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',\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    })\n    const raw =\n      (resp as { choices?: { message?: { content?: string } }[] }).choices?.[0]?.message?.content ??\n      ''\n    return parseCorrectnessResponse(raw)\n  }\n}\n\n/** Stopwords for requirement-title tokenization — drops the imperative verbs\n *  ('review', 'update', …) common to deliverable titles so recall keys on the\n *  substantive nouns, not the boilerplate ask. */\nconst TITLE_STOPWORDS = new Set([\n  'the',\n  'a',\n  'an',\n  'and',\n  'or',\n  'for',\n  'to',\n  'of',\n  'in',\n  'on',\n  'with',\n  'review',\n  'update',\n  'new',\n  'proposed',\n])\n\n/**\n * Deterministic `CorrectnessChecker` — the no-LLM counterpart to\n * `createLlmCorrectnessChecker`. A produced item fulfils a requirement when its\n * content is substantive (≥ `minContentLength` chars) AND recalls ≥ `minRecall`\n * of the requirement title's significant tokens. No network — the default gate\n * for apps and tests without an LLM judge. Pass to `verifyCompletion` as the\n * checker.\n */\nexport function createTokenRecallChecker(\n  opts: { minRecall?: number; minContentLength?: number } = {},\n): CorrectnessChecker {\n  const minRecall = opts.minRecall ?? 0.5\n  const minLen = opts.minContentLength ?? 120\n  return async (requirement, content) => {\n    const body = content.trim()\n    if (body.length < minLen)\n      return {\n        correct: false,\n        reason: `content too thin (${body.length} chars) to be the deliverable`,\n      }\n    const titleTokens = requirement.title\n      .toLowerCase()\n      .split(/[^a-z0-9]+/)\n      .filter((t) => t.length > 2 && !TITLE_STOPWORDS.has(t))\n    if (titleTokens.length === 0)\n      return {\n        correct: true,\n        reason: 'requirement title has no significant tokens — structural match accepted',\n      }\n    const lower = body.toLowerCase()\n    const hits = titleTokens.filter((t) => lower.includes(t)).length\n    const recall = hits / titleTokens.length\n    return recall >= minRecall\n      ? {\n          correct: true,\n          reason: `content recalls ${hits}/${titleTokens.length} requirement tokens`,\n        }\n      : {\n          correct: false,\n          reason: `content recalls only ${hits}/${titleTokens.length} requirement tokens`,\n        }\n  }\n}\n","/**\n * Produced-state extraction — normalize a run's runtime event stream into the\n * typed `ProducedState` the completion oracle consumes.\n *\n * `ProducedState` answers \"what did the agent actually produce\" — vault\n * artifacts, proposals, tool calls. The runtime emits these as a stream of\n * events; this module is the single normalization point from that stream to\n * the shape `verifyCompletion` expects.\n *\n * Input is structurally typed (`RuntimeEventLike`) so this module does not\n * depend on agent-runtime — agent-runtime's `RuntimeStreamEvent` satisfies it\n * structurally. The `content` on `ArtifactEventLike` and the whole\n * `proposal_created` variant are the runtime-side enrichments this contract\n * requires; the runtime emits them, this module consumes them.\n */\n\nimport type { Artifact } from './artifact-validator'\nimport type { ProducedProposal, ProducedState } from './completion-verifier'\n\n/** A tool the agent invoked. */\nexport interface ToolCallEventLike {\n  type: 'tool_call'\n  toolName: string\n}\n\n/**\n * An artifact the agent produced. `content` is the enriched field — the\n * runtime's base `artifact` event carries only metadata; the completion\n * oracle needs the body to verify the deliverable, so the runtime emits it.\n */\nexport interface ArtifactEventLike {\n  type: 'artifact'\n  artifactId: string\n  name?: string\n  mimeType?: string\n  uri?: string\n  content?: string\n}\n\n/** A proposal / filing the agent created. */\nexport interface ProposalEventLike {\n  type: 'proposal_created'\n  proposalId: string\n  title: string\n  status?: 'pending' | 'approved' | 'rejected'\n}\n\n/**\n * The subset of runtime stream events `extractProducedState` consumes.\n * agent-runtime's full `RuntimeStreamEvent` union satisfies this structurally;\n * the `{ type: string }` catch-all keeps the input permissive so callers can\n * pass the whole unfiltered telemetry stream — unrecognized events are skipped.\n */\nexport type RuntimeEventLike =\n  | ToolCallEventLike\n  | ArtifactEventLike\n  | ProposalEventLike\n  | { type: string }\n\nfunction artifactKind(mimeType: string | undefined): string {\n  if (!mimeType) return 'file'\n  if (mimeType.includes('json')) return 'json'\n  if (mimeType.startsWith('text/')) return 'text'\n  return 'file'\n}\n\n/**\n * Normalize a run's runtime event stream into `ProducedState`.\n *\n * Pure and total — unrecognized event types are skipped. `toolCalls` is\n * deduplicated by name in first-seen order (completion cares about a tool's\n * presence, not its call count). An artifact with neither a name nor a uri\n * still yields an entry keyed by its `artifactId` so it is never silently\n * dropped; an artifact with no `content` yields empty content, which the\n * completion oracle's structural check then rejects on its own.\n */\nexport function extractProducedState(events: readonly RuntimeEventLike[]): ProducedState {\n  const artifacts: Artifact[] = []\n  const proposals: ProducedProposal[] = []\n  const toolCalls: string[] = []\n  const seenTools = new Set<string>()\n\n  for (const ev of events) {\n    if (ev.type === 'tool_call') {\n      const name = (ev as ToolCallEventLike).toolName\n      if (name && !seenTools.has(name)) {\n        seenTools.add(name)\n        toolCalls.push(name)\n      }\n    } else if (ev.type === 'artifact') {\n      const a = ev as ArtifactEventLike\n      artifacts.push({\n        kind: artifactKind(a.mimeType),\n        path: a.name ?? a.uri ?? a.artifactId,\n        content: a.content ?? '',\n      })\n    } else if (ev.type === 'proposal_created') {\n      const p = ev as ProposalEventLike\n      proposals.push({ id: p.proposalId, title: p.title, status: p.status ?? 'pending' })\n    }\n  }\n\n  return { artifacts, proposals, toolCalls }\n}\n","/**\n * @stable\n *\n * AgentProfile — the eval harness's unit of variation.\n *\n * A profile pins everything that changes agent behaviour for a benchmark\n * cell: the model, the active skills, the prompt version, the available\n * tools. Vary the profile — swap a model, add a skill — and re-run the suite\n * to benchmark the change. The scorecard keys a cell on\n * `(scenarioId, profileHash)`, so the model is not a separate axis: it lives\n * inside the profile, and two profiles with the same model but different\n * skills are different cells.\n *\n * `agentProfileHash` is the profile's behaviour identity. Two profiles that\n * produce the same agent behaviour share a hash (and a scorecard cell);\n * reordering `skills` or `tools` does not change it; the human-facing `id`\n * label does not affect it.\n */\n\nimport { createHash } from 'node:crypto'\nimport { ValidationError } from './errors'\nimport { canonicalize } from './pre-registration'\n\nexport interface AgentProfile {\n  /** Human-facing label, e.g. `sonnet-legal-skills-v3`. Not part of the hash. */\n  id: string\n  /** Model snapshot id this profile pins, e.g. `claude-sonnet-4-6@2025-04-15`. */\n  model: string\n  /** Skill ids/versions active in this profile — the primary behaviour lever. */\n  skills?: string[]\n  /** Prompt version identifier. */\n  promptVersion?: string\n  /** Tool ids available to the agent. */\n  tools?: string[]\n  /** Any other behaviour-bearing knobs that should fingerprint into the hash. */\n  metadata?: Record<string, string | number | boolean>\n}\n\n/**\n * Deterministic behaviour identity of a profile — a sha256 over the\n * behaviour-bearing fields. `skills` and `tools` are order-insensitive; the\n * `id` label is excluded. Throws on a profile with no `model` — an unkeyable\n * profile must fail loud rather than collapse into a blank-model cell.\n */\nexport function agentProfileHash(profile: AgentProfile): string {\n  if (typeof profile.model !== 'string' || profile.model.trim().length === 0) {\n    throw new ValidationError(`AgentProfile \"${profile.id}\" has no model — cannot hash`)\n  }\n  const behaviour = {\n    model: profile.model.trim(),\n    skills: [...(profile.skills ?? [])].sort(),\n    promptVersion: profile.promptVersion ?? null,\n    tools: [...(profile.tools ?? [])].sort(),\n    metadata: profile.metadata ?? {},\n  }\n  return createHash('sha256')\n    .update(JSON.stringify(canonicalize(behaviour)))\n    .digest('hex')\n}\n"],"mappings":";;;;;;;;AAmGO,SAAS,kBAAkB,OAGZ;AACpB,MAAI,MAAM,aAAa,WAAW,GAAG;AACnC,UAAM,IAAI;AAAA,MACR,4BAA4B,MAAM,MAAM;AAAA,IAC1C;AAAA,EACF;AACA,QAAM,iBAAiB,MAAM,aAAa,OAAO,CAAC,MAAM,EAAE,SAAS,EAAE;AACrE,QAAM,iBAAiB,iBAAiB,MAAM,aAAa;AAC3D,QAAM,gBAAgB,mBAAmB,MAAM,aAAa;AAC5D,SAAO;AAAA,IACL,QAAQ,MAAM;AAAA,IACd,cAAc,MAAM;AAAA,IACpB;AAAA,IACA;AAAA,IACA,OAAO;AAAA,IACP,OAAO;AAAA,EACT;AACF;AAYA,IAAM,YAAY,oBAAI,IAAI;AAAA,EACxB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAED,IAAM,kBAAkB;AACxB,IAAM,oBAAoB;AAE1B,SAAS,OAAO,GAAwB;AACtC,SAAO,IAAI;AAAA,IACT,EACG,YAAY,EACZ,MAAM,YAAY,EAClB,OAAO,CAAC,MAAM,EAAE,SAAS,KAAK,CAAC,UAAU,IAAI,CAAC,CAAC;AAAA,EACpD;AACF;AAOA,SAAS,YAAY,iBAAyB,eAA+B;AAC3E,QAAM,MAAM,OAAO,eAAe;AAClC,MAAI,IAAI,SAAS,EAAG,QAAO;AAC3B,QAAM,OAAO,OAAO,aAAa;AACjC,MAAI,MAAM;AACV,aAAW,KAAK,IAAK,KAAI,KAAK,IAAI,CAAC,EAAG;AACtC,SAAO,MAAM,IAAI;AACnB;AAYA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,UAAU,GAAG,IAAI,KAAK,IAAI,IAAI,YAAY,EAAE;AAClD,QAAM,MAAmB,CAAC;AAC1B,YAAU,QAAQ,CAAC,GAAG,MAAM;AAC1B,SAAK,EAAE,WAAW,IAAI,KAAK,EAAE,SAAS,kBAAmB;AACzD,QAAI,QAAQ,YAAY,SAAS,GAAG,EAAE,QAAQ,EAAE,IAAI,EAAE,IAAI,EAAE;AAC5D,QAAI,IAAI,YAAY,EAAE,QAAQ,IAAI,SAAS,YAAY,MAAM,EAAE,KAAK,YAAY,GAAG;AACjF,cAAQ,KAAK,IAAI,OAAO,CAAC;AAAA,IAC3B;AACA,QAAI,QAAQ,gBAAiB;AAC7B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,YAAY,CAAC;AAAA,MACtB;AAAA,MACA,UAAU,aAAa,EAAE,QAAQ,EAAE,IAAI,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MAClF,SAAS,EAAE,WAAW;AAAA,IACxB,CAAC;AAAA,EACH,CAAC;AACD,SAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,UAAU,GAAG,IAAI,KAAK,IAAI,IAAI,YAAY,EAAE;AAClD,QAAM,MAAmB,CAAC;AAC1B,aAAW,KAAK,WAAW;AAEzB,QAAI,EAAE,WAAW,WAAY;AAC7B,UAAM,QAAQ,YAAY,SAAS,EAAE,KAAK;AAC1C,QAAI,QAAQ,gBAAiB;AAC7B,UAAM,OAAO,EAAE,WAAW;AAC1B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,YAAY,EAAE,EAAE;AAAA,MACzB;AAAA,MACA,UAAU,sBAAsB,EAAE,KAAK,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MAClF,SAAS,KAAK,KAAK,EAAE,UAAU,oBAAoB,OAAO;AAAA,IAC5D,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,MAAmB,CAAC;AAC1B,YAAU,QAAQ,CAAC,MAAM,MAAM;AAC7B,UAAM,QAAQ,YAAY,IAAI,OAAO,IAAI;AACzC,QAAI,QAAQ,gBAAiB;AAC7B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,QAAQ,CAAC;AAAA,MAClB;AAAA,MACA,UAAU,cAAc,IAAI,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MACvE,SAAS;AAAA,IACX,CAAC;AAAA,EACH,CAAC;AACD,SAAO;AACT;AASA,eAAsB,iBACpB,MACA,OACA,kBAC4B;AAC5B,MAAI,KAAK,aAAa,WAAW,GAAG;AAClC,UAAM,IAAI;AAAA,MACR,2BAA2B,KAAK,MAAM;AAAA,IACxC;AAAA,EACF;AAKA,QAAM,aAA0B,CAAC;AACjC,OAAK,aAAa,QAAQ,CAAC,KAAK,MAAM;AACpC,UAAM,KAAK,IAAI,eAAe;AAC9B,QAAI,OAAO,cAAc,OAAO,OAAO;AACrC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AACA,QAAI,OAAO,cAAc,OAAO,OAAO;AACrC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AACA,QAAI,OAAO,eAAe,OAAO,OAAO;AACtC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AAAA,EACF,CAAC;AACD,aAAW,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AAE3C,QAAM,WAAW,oBAAI,IAAuB;AAC5C,QAAM,YAAY,oBAAI,IAAY;AAClC,aAAW,KAAK,YAAY;AAC1B,QAAI,SAAS,IAAI,EAAE,QAAQ,KAAK,UAAU,IAAI,EAAE,OAAO,EAAG;AAC1D,aAAS,IAAI,EAAE,UAAU,CAAC;AAC1B,cAAU,IAAI,EAAE,OAAO;AAAA,EACzB;AAEA,QAAM,eAAmC,CAAC;AAC1C,WAAS,IAAI,GAAG,IAAI,KAAK,aAAa,QAAQ,KAAK;AACjD,UAAM,MAAM,KAAK,aAAa,CAAC;AAC/B,UAAM,QAAQ,SAAS,IAAI,CAAC;AAC5B,UAAM,WAAqB,CAAC;AAC5B,QAAI,UAA0B;AAE9B,QAAI,OAAO;AACT,eAAS,KAAK,MAAM,QAAQ;AAC5B,UAAI,MAAM,YAAY,MAAM;AAC1B,cAAM,IAAI,MAAM,iBAAiB,KAAK,MAAM,OAAO;AACnD,kBAAU,EAAE;AACZ,iBAAS,KAAK,gBAAgB,EAAE,UAAU,SAAS,MAAM,WAAM,EAAE,MAAM,EAAE;AAAA,MAC3E,OAAO;AACL,iBAAS,KAAK,kEAA6D;AAAA,MAC7E;AAAA,IACF,OAAO;AACL,YAAM,KAAK,IAAI,eAAe;AAC9B,YAAM,OAAO,OAAO,QAAQ,gCAAgC;AAC5D,eAAS,KAAK,eAAe,IAAI,2BAA2B;AAAA,IAC9D;AAEA,UAAM,sBAAsB,UAAU;AACtC,UAAM,YAAY,uBAAuB,YAAY;AACrD,iBAAa,KAAK;AAAA,MAChB,OAAO,IAAI;AAAA,MACX,OAAO,IAAI;AAAA,MACX;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAAA,EACH;AAEA,SAAO,kBAAkB,EAAE,QAAQ,KAAK,QAAQ,aAAa,CAAC;AAChE;AASO,SAAS,yBAAyB,KAAmD;AAC1F,QAAM,QAAQ,IAAI,MAAM,aAAa;AACrC,MAAI,CAAC,OAAO;AACV,UAAM,IAAI,MAAM,0DAA0D,IAAI,MAAM,GAAG,GAAG,CAAC,EAAE;AAAA,EAC/F;AACA,QAAM,SAAS,KAAK,MAAM,MAAM,CAAC,CAAC;AAClC,MAAI,OAAO,OAAO,YAAY,WAAW;AACvC,UAAM,IAAI,MAAM,uDAAuD,MAAM,CAAC,EAAE,MAAM,GAAG,GAAG,CAAC,EAAE;AAAA,EACjG;AACA,SAAO,EAAE,SAAS,OAAO,SAAS,QAAQ,OAAO,OAAO,WAAW,WAAW,OAAO,SAAS,GAAG;AACnG;AAQO,SAAS,4BACd,IACA,OAAkC,CAAC,GACf;AACpB,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,kBAAkB,KAAK,mBAAmB;AAChD,SAAO,OAAO,aAAa,YAAY;AACrC,UAAM,OAAO,MAAM,GAAG,KAAK;AAAA,MACzB;AAAA,MACA,UAAU;AAAA,QACR;AAAA,UACE,MAAM;AAAA,UACN,SACE;AAAA,QACJ;AAAA,QACA;AAAA,UACE,MAAM;AAAA,UACN,SAAS,gBAAgB,YAAY,KAAK;AAAA,EACxC,YAAY,WAAW,aAAa,YAAY,QAAQ;AAAA,IAAO,EACjE;AAAA;AAAA,EAAyB,QAAQ,MAAM,GAAG,eAAe,CAAC;AAAA,QAC5D;AAAA,MACF;AAAA,MACA,aAAa;AAAA,MACb,WAAW;AAAA,IACb,CAAC;AACD,UAAM,MACH,KAA4D,UAAU,CAAC,GAAG,SAAS,WACpF;AACF,WAAO,yBAAyB,GAAG;AAAA,EACrC;AACF;AAKA,IAAM,kBAAkB,oBAAI,IAAI;AAAA,EAC9B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAUM,SAAS,yBACd,OAA0D,CAAC,GACvC;AACpB,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,SAAS,KAAK,oBAAoB;AACxC,SAAO,OAAO,aAAa,YAAY;AACrC,UAAM,OAAO,QAAQ,KAAK;AAC1B,QAAI,KAAK,SAAS;AAChB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,QAAQ,qBAAqB,KAAK,MAAM;AAAA,MAC1C;AACF,UAAM,cAAc,YAAY,MAC7B,YAAY,EACZ,MAAM,YAAY,EAClB,OAAO,CAAC,MAAM,EAAE,SAAS,KAAK,CAAC,gBAAgB,IAAI,CAAC,CAAC;AACxD,QAAI,YAAY,WAAW;AACzB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,QAAQ;AAAA,MACV;AACF,UAAM,QAAQ,KAAK,YAAY;AAC/B,UAAM,OAAO,YAAY,OAAO,CAAC,MAAM,MAAM,SAAS,CAAC,CAAC,EAAE;AAC1D,UAAM,SAAS,OAAO,YAAY;AAClC,WAAO,UAAU,YACb;AAAA,MACE,SAAS;AAAA,MACT,QAAQ,mBAAmB,IAAI,IAAI,YAAY,MAAM;AAAA,IACvD,IACA;AAAA,MACE,SAAS;AAAA,MACT,QAAQ,wBAAwB,IAAI,IAAI,YAAY,MAAM;AAAA,IAC5D;AAAA,EACN;AACF;;;ACzYA,SAAS,aAAa,UAAsC;AAC1D,MAAI,CAAC,SAAU,QAAO;AACtB,MAAI,SAAS,SAAS,MAAM,EAAG,QAAO;AACtC,MAAI,SAAS,WAAW,OAAO,EAAG,QAAO;AACzC,SAAO;AACT;AAYO,SAAS,qBAAqB,QAAoD;AACvF,QAAM,YAAwB,CAAC;AAC/B,QAAM,YAAgC,CAAC;AACvC,QAAM,YAAsB,CAAC;AAC7B,QAAM,YAAY,oBAAI,IAAY;AAElC,aAAW,MAAM,QAAQ;AACvB,QAAI,GAAG,SAAS,aAAa;AAC3B,YAAM,OAAQ,GAAyB;AACvC,UAAI,QAAQ,CAAC,UAAU,IAAI,IAAI,GAAG;AAChC,kBAAU,IAAI,IAAI;AAClB,kBAAU,KAAK,IAAI;AAAA,MACrB;AAAA,IACF,WAAW,GAAG,SAAS,YAAY;AACjC,YAAM,IAAI;AACV,gBAAU,KAAK;AAAA,QACb,MAAM,aAAa,EAAE,QAAQ;AAAA,QAC7B,MAAM,EAAE,QAAQ,EAAE,OAAO,EAAE;AAAA,QAC3B,SAAS,EAAE,WAAW;AAAA,MACxB,CAAC;AAAA,IACH,WAAW,GAAG,SAAS,oBAAoB;AACzC,YAAM,IAAI;AACV,gBAAU,KAAK,EAAE,IAAI,EAAE,YAAY,OAAO,EAAE,OAAO,QAAQ,EAAE,UAAU,UAAU,CAAC;AAAA,IACpF;AAAA,EACF;AAEA,SAAO,EAAE,WAAW,WAAW,UAAU;AAC3C;;;ACpFA,SAAS,kBAAkB;AAyBpB,SAAS,iBAAiB,SAA+B;AAC9D,MAAI,OAAO,QAAQ,UAAU,YAAY,QAAQ,MAAM,KAAK,EAAE,WAAW,GAAG;AAC1E,UAAM,IAAI,gBAAgB,iBAAiB,QAAQ,EAAE,mCAA8B;AAAA,EACrF;AACA,QAAM,YAAY;AAAA,IAChB,OAAO,QAAQ,MAAM,KAAK;AAAA,IAC1B,QAAQ,CAAC,GAAI,QAAQ,UAAU,CAAC,CAAE,EAAE,KAAK;AAAA,IACzC,eAAe,QAAQ,iBAAiB;AAAA,IACxC,OAAO,CAAC,GAAI,QAAQ,SAAS,CAAC,CAAE,EAAE,KAAK;AAAA,IACvC,UAAU,QAAQ,YAAY,CAAC;AAAA,EACjC;AACA,SAAO,WAAW,QAAQ,EACvB,OAAO,KAAK,UAAU,aAAa,SAAS,CAAC,CAAC,EAC9C,OAAO,KAAK;AACjB;","names":[]}