@tangle-network/agent-eval 0.199.1 → 0.200.0

package/dist/rollout/index.d.ts

@@ -1,5 +1,5 @@
 import { C as assertMinted, D as isRolloutLine, E as gateGamedOutcome, O as isTrainableSplit, S as ToolDef, T as assertRolloutLine, _ as RolloutProvenance, a as MintedRolloutLine, b as RolloutStep, c as ROLLOUT_ROLES, d as RolloutArtifacts, f as RolloutCapture, g as RolloutPolicy, h as RolloutOutcome, i as GatedEvidence, k as validateRolloutLine, l as ROLLOUT_SCHEMA, m as RolloutLine, n as ChatRole, o as MintedRolloutOutcome, p as RolloutCostBlock, r as ChatToolCall, s as ROLLOUT_CAPTURES, t as ChatMessage, u as ROLLOUT_SPLITS, v as RolloutRole, w as assertMintedLines, x as RolloutTask, y as RolloutSplit } from "../schema-B5LtYvn9.js";
 import { a as mintRolloutRows, c as ScorePreference, d as observedSplitScore, f as scoreOrigin, i as RolloutSearchLineage, l as isRealnessGated, m as trainingScore, n as MintRolloutResult, o as unmintableReasons, p as trainingReward, r as RolloutScrubber, s as ScoreOrigin, t as MintRolloutOptions, u as observedScore } from "../mint-C9Lojt1V.js";
 import { a as SftRow, c as VerifiersTokenUsage, d as toRftItem, f as toRftItems, h as toVerifiersRolloutOutputs, i as SftExportOptions, l as toJsonl, m as toVerifiersRolloutOutput, n as RewardRow, p as toSftRows, r as RftItem, s as VerifiersRolloutOutput, t as RealnessLabels, u as toRewardRows } from "../exporters-XMCYJlqT.js";
-import { $ as HarborFinalMetrics, A as measureFormatGate, B as DEFAULT_CLAUDE_PROJECTS_DIR, C as FORMAT_GATE_DISPOSITION, Ct as gatedEvidenceOf, D as ReleaseRowRef, E as GateReport, F as openOpencodeDb, G as readRolloutJournal, H as findClaudeTranscripts, I as readOpencodeSessionMessages, J as ATIF_SCHEMA_VERSION, K as readRolloutLedger, L as ClaudeTranscript, M as DEFAULT_OPENCODE_DB, N as OpencodeSessionRow, O as assertGateReport, P as findOpencodeSessionsByDirectory, Q as HarborContentPart, R as ClaudeTranscriptRef, S as EmittedEvidence, St as gateErrors, T as GateDisposition, U as readClaudeTranscript, V as claudeProjectSlug, W as appendRolloutLines, X as HARBOR_IMPORT_GAP, Y as FromHarborOptions, Z as HarborAgent, _ as defaultRolloutScrubber, _t as GateCheckDisposition, a as buildHfDataset, at as HarborStepSource, b as scrubRolloutLine, bt as GateEntryPoint, c as runRolloutReleaseCli, ct as HarborTrajectory, d as RELEASE_FORMATS, dt as toHarborTrajectories, et as HarborImageSource, f as ReleaseFormat, ft as toHarborTrajectory, g as addScrubCounts, gt as GateCheck, h as ScrubRule, ht as GATE_POLICIES, i as ScrubReport, it as HarborStep, j as releaseRowRefs, k as gatedRolloutIds, l as DatasetCardInputs, lt as fromHarborTrajectory, m as ScrubCounts, mt as GATE_CHECK_IDS, n as BuildSummary, nt as HarborObservation, o as parseRolloutReleaseArgs, ot as HarborSubagentTrajectoryRef, p as buildDatasetCard, pt as GATE_CHECKS, q as writeRolloutLedger, r as RolloutReleaseCliArgs, rt as HarborObservationResult, s as planPushCommand, st as HarborToolCall, t as BuildOptions, tt as HarborMetrics, u as FORMAT_FILES, ut as relabelImportedSplit, v as emptyScrubCounts, vt as GateCheckId, w as FormatGateCounts, x as scrubText, xt as GatePolicy, y as scrubLines, yt as GateCheckedOutcome, z as ClaudeUsageTotals } from "../index-D4pC7Tvd.js";
+import { $ as HarborFinalMetrics, A as measureFormatGate, B as DEFAULT_CLAUDE_PROJECTS_DIR, C as FORMAT_GATE_DISPOSITION, Ct as gatedEvidenceOf, D as ReleaseRowRef, E as GateReport, F as openOpencodeDb, G as readRolloutJournal, H as findClaudeTranscripts, I as readOpencodeSessionMessages, J as ATIF_SCHEMA_VERSION, K as readRolloutLedger, L as ClaudeTranscript, M as DEFAULT_OPENCODE_DB, N as OpencodeSessionRow, O as assertGateReport, P as findOpencodeSessionsByDirectory, Q as HarborContentPart, R as ClaudeTranscriptRef, S as EmittedEvidence, St as gateErrors, T as GateDisposition, U as readClaudeTranscript, V as claudeProjectSlug, W as appendRolloutLines, X as HARBOR_IMPORT_GAP, Y as FromHarborOptions, Z as HarborAgent, _ as defaultRolloutScrubber, _t as GateCheckDisposition, a as buildHfDataset, at as HarborStepSource, b as scrubRolloutLine, bt as GateEntryPoint, c as runRolloutReleaseCli, ct as HarborTrajectory, d as RELEASE_FORMATS, dt as toHarborTrajectories, et as HarborImageSource, f as ReleaseFormat, ft as toHarborTrajectory, g as addScrubCounts, gt as GateCheck, h as ScrubRule, ht as GATE_POLICIES, i as ScrubReport, it as HarborStep, j as releaseRowRefs, k as gatedRolloutIds, l as DatasetCardInputs, lt as fromHarborTrajectory, m as ScrubCounts, mt as GATE_CHECK_IDS, n as BuildSummary, nt as HarborObservation, o as parseRolloutReleaseArgs, ot as HarborSubagentTrajectoryRef, p as buildDatasetCard, pt as GATE_CHECKS, q as writeRolloutLedger, r as RolloutReleaseCliArgs, rt as HarborObservationResult, s as planPushCommand, st as HarborToolCall, t as BuildOptions, tt as HarborMetrics, u as FORMAT_FILES, ut as relabelImportedSplit, v as emptyScrubCounts, vt as GateCheckId, w as FormatGateCounts, x as scrubText, xt as GatePolicy, y as scrubLines, yt as GateCheckedOutcome, z as ClaudeUsageTotals } from "../index-D4pC7Tvd2.js";
 export { ATIF_SCHEMA_VERSION, type BuildOptions, type BuildSummary, type ChatMessage, type ChatRole, type ChatToolCall, type ClaudeTranscript, type ClaudeTranscriptRef, type ClaudeUsageTotals, DEFAULT_CLAUDE_PROJECTS_DIR, DEFAULT_OPENCODE_DB, type DatasetCardInputs, type EmittedEvidence, FORMAT_FILES, FORMAT_GATE_DISPOSITION, type FormatGateCounts, type FromHarborOptions, GATE_CHECKS, GATE_CHECK_IDS, GATE_POLICIES, type GateCheck, type GateCheckDisposition, type GateCheckId, type GateCheckedOutcome, type GateDisposition, type GateEntryPoint, type GatePolicy, type GateReport, type GatedEvidence, HARBOR_IMPORT_GAP, type HarborAgent, type HarborContentPart, type HarborFinalMetrics, type HarborImageSource, type HarborMetrics, type HarborObservation, type HarborObservationResult, type HarborStep, type HarborStepSource, type HarborSubagentTrajectoryRef, type HarborToolCall, type HarborTrajectory, type MintRolloutOptions, type MintRolloutResult, type MintedRolloutLine, type MintedRolloutOutcome, type OpencodeSessionRow, RELEASE_FORMATS, ROLLOUT_CAPTURES, ROLLOUT_ROLES, ROLLOUT_SCHEMA, ROLLOUT_SPLITS, type RealnessLabels, type ReleaseFormat, type ReleaseRowRef, type RewardRow, type RftItem, type RolloutArtifacts, type RolloutCapture, type RolloutCostBlock, type RolloutLine, type RolloutOutcome, type RolloutPolicy, type RolloutProvenance, type RolloutReleaseCliArgs, type RolloutRole, type RolloutScrubber, type RolloutSearchLineage, type RolloutSplit, type RolloutStep, type RolloutTask, type ScoreOrigin, type ScorePreference, type ScrubCounts, type ScrubReport, type ScrubRule, type SftExportOptions, type SftRow, type ToolDef, type VerifiersRolloutOutput, type VerifiersTokenUsage, addScrubCounts, appendRolloutLines, assertGateReport, assertMinted, assertMintedLines, assertRolloutLine, buildDatasetCard, buildHfDataset, claudeProjectSlug, defaultRolloutScrubber, emptyScrubCounts, findClaudeTranscripts, findOpencodeSessionsByDirectory, fromHarborTrajectory, gateErrors, gateGamedOutcome, gatedEvidenceOf, gatedRolloutIds, isRealnessGated, isRolloutLine, isTrainableSplit, measureFormatGate, mintRolloutRows, observedScore, observedSplitScore, openOpencodeDb, parseRolloutReleaseArgs, planPushCommand, readClaudeTranscript, readOpencodeSessionMessages, readRolloutJournal, readRolloutLedger, relabelImportedSplit, releaseRowRefs, runRolloutReleaseCli, scoreOrigin, scrubLines, scrubRolloutLine, scrubText, toHarborTrajectories, toHarborTrajectory, toJsonl, toRewardRows, toRftItem, toRftItems, toSftRows, toVerifiersRolloutOutput, toVerifiersRolloutOutputs, trainingReward, trainingScore, unmintableReasons, validateRolloutLine, writeRolloutLedger };

package/dist/trace-questions-DnqdmwyV.d.ts

@@ -0,0 +1,68 @@
+import { c as CostLedgerHandle } from "./cost-ledger-DbQdN3nO.js";
+import { O as TraceAnalysisStore } from "./types-CeDfCRXp.js";
+import { a as TraceAnalystLimits, l as RawAnalystFinding, n as TraceAnalysisEngine } from "./engine-DlKIzDKo.js";
+//#region src/analyst/trace-questions.d.ts
+/** One question to ask of the store. */
+interface TraceQuestionSpec {
+  /** Stable id: letters, digits, dot, underscore, hyphen. */
+  readonly id: string;
+  readonly question: string;
+  /** Guidance for this question only, after the shared rules. */
+  readonly instructions?: string;
+}
+interface AskTraceQuestionsOptions {
+  readonly questions: readonly TraceQuestionSpec[];
+  readonly store: TraceAnalysisStore;
+  readonly engine: TraceAnalysisEngine;
+  /** Every question and verifier charges this ledger; its ceiling bounds the whole call. */
+  readonly costLedger: CostLedgerHandle;
+  /** Stable id for cost attribution and logs. */
+  readonly runId: string;
+  /** Material every question receives before its first model call, such as the check's verdict. */
+  readonly context?: string;
+  /** Investigations running at once. Default 4. */
+  readonly concurrency?: number;
+  /** Distinct citations a finding needs. Default 2. */
+  readonly minimumEvidenceCitations?: number;
+  /** Run the independent verifier on every admitted finding. Default true. */
+  readonly verify?: boolean;
+  readonly limits?: Partial<TraceAnalystLimits>;
+  readonly signal?: AbortSignal;
+  readonly log?: (message: string, fields?: Record<string, unknown>) => void;
+}
+/** A finding and what the verifier, shown only its cited spans, concluded. */
+interface VerifiedTraceFinding {
+  readonly finding: RawAnalystFinding;
+  /** The finding's trace citations, as `trace://` URIs. */
+  readonly citations: readonly string[];
+  /** `true` supported, `false` not supported, `null` when verification was off or failed. */
+  readonly verified: boolean | null;
+  /** The verifier's own words, when it answered. */
+  readonly verifierAnswer?: string;
+}
+interface TraceQuestionOutcome {
+  readonly id: string;
+  readonly question: string;
+  readonly status: 'answered' | 'failed';
+  readonly failure?: string;
+  /** The engine's prose answer, verbatim; `null` when it returned none. */
+  readonly answer: string | null;
+  readonly findings: readonly VerifiedTraceFinding[];
+  /** `null` when the engine failed before reporting. */
+  readonly modelCalls: number | null;
+  readonly toolCalls: number | null;
+}
+/**
+ * Ask every question, then verify every admitted finding. Never throws for a failed question or
+ * verifier: each failure is recorded on its outcome. Throws before any model call for a malformed
+ * question list or options.
+ */
+declare function askTraceQuestions(opts: AskTraceQuestionsOptions): Promise<TraceQuestionOutcome[]>;
+/**
+ * A store that exposes exactly the cited spans. Discovery and whole-trace reads refuse, so the
+ * verifier cannot go looking for other support; the underlying store's bounds still apply.
+ */
+declare function citedSpansOnly(store: TraceAnalysisStore, allowed: ReadonlyMap<string, ReadonlySet<string>>): TraceAnalysisStore;
+//#endregion
+export { askTraceQuestions as a, VerifiedTraceFinding as i, TraceQuestionOutcome as n, citedSpansOnly as o, TraceQuestionSpec as r, AskTraceQuestionsOptions as t };
+//# sourceMappingURL=trace-questions-DnqdmwyV.d.ts.map

package/dist/trace-questions-DnqdmwyV.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"trace-questions-DnqdmwyV.d.ts","names":[],"sources":["../src/analyst/trace-questions.ts"],"mappings":";;;;;UA6BiB;;WAEN;WACA;;WAEA;;UAGM;WACN,oBAAoB;WACpB,OAAO;WACP,QAAQ;;WAER,YAAY;;WAEZ;;WAEA;;WAEA;;WAEA;;WAEA;WACA,SAAS,QAAQ;WACjB,SAAS;WACT,OAAO,iBAAiB,SAAS;;;UAI3B;WACN,SAAS;;WAET;;WAEA;;WAEA;;UAGM;WACN;WACA;WACA;WACA;;WAEA;WACA,mBAAmB;;WAEnB;WACA;;;;;;;iBAyBW,kBACpB,MAAM,2BACL,QAAQ;;;;;iBA2JK,eACd,OAAO,oBACP,SAAS,oBAAoB,uBAC5B"}

package/dist/{types-B3jzCp0p2.js → types-B3jzCp0p.js}

@@ -105,4 +105,4 @@ function describeValue(value) {
 //#endregion
 export { assertMultishotShotResult as i, MultishotFatalToolError as n, MultishotShotResultError as r, MultishotDriverEmptyError as t };
 
-//# sourceMappingURL=types-B3jzCp0p2.js.map
+//# sourceMappingURL=types-B3jzCp0p.js.map

package/dist/types-B3jzCp0p.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"types-B3jzCp0p.js","names":[],"sources":["../src/multishot/types.ts"],"sourcesContent":["// Public types for the multishot substrate.\n\nimport type { CostProvenance } from '../cost-ledger'\n\nexport interface MultishotMessage {\n  role: 'user' | 'assistant' | 'tool'\n  content: string\n  toolCallId?: string\n  toolCalls?: Array<{ id: string; name: string; args: Record<string, unknown> }>\n}\n\nexport interface MultishotArtifact {\n  type: string\n  turn: number\n  invocation: { name: string; args: Record<string, unknown> }\n  content: string\n}\n\nexport interface MultishotResult {\n  transcript: MultishotMessage[]\n  artifacts: MultishotArtifact[]\n  toolCalls: number\n  durationMs: number\n  /** Known spend. A subtotal, not a total, when `costProvenance.kind` is\n   *  `uncaptured`. */\n  costUsd: number\n  /** Origin of `costUsd`. A shot that priced every call reports `estimated`\n   *  or `observed`; a shot with a call the router priced at nothing reports\n   *  `uncaptured`, and the matrix records the cell as under-counted instead of\n   *  presenting the subtotal as a complete estimate.\n   *\n   *  Optional so an engine written before this field keeps working; the matrix\n   *  then judges the cell on judge receipts alone, as it did before. */\n  costProvenance?: CostProvenance\n}\n\nexport interface MultishotToolDefinition {\n  type: 'function'\n  function: {\n    name: string\n    description: string\n    parameters: Record<string, unknown>\n  }\n}\n\n/** One chat-completion request the multishot loop issues for a single agent\n *  (or driver) inference step. Mirrors the OpenAI-compat body the loop would\n *  otherwise POST to the Tangle router. */\nexport interface MultishotTransportRequest {\n  model: string\n  messages: Array<Record<string, unknown>>\n  tools?: MultishotToolDefinition[]\n  temperature?: number\n  maxTokens?: number\n  signal?: AbortSignal\n}\n\nexport interface MultishotTransportToolCall {\n  id: string\n  type: 'function'\n  function: { name: string; arguments: string }\n}\n\nexport interface MultishotTransportResponse {\n  message: { content?: string | null; tool_calls?: MultishotTransportToolCall[] }\n  usage?: { prompt_tokens?: number; completion_tokens?: number }\n  /** Actual spend for this call. When omitted, the loop meters cost from\n   *  `usage` via the per-model estimator (estimateMultishotCost). */\n  costUsd?: number\n  /** Model identity the provider reported, when the transport observed one.\n   *  Omitted means unreported, and the requested model is used for\n   *  attribution — which is not proof that model answered. */\n  model?: string\n}\n\n/** Execution seam for one leg of the multishot loop. The caller owns model\n *  execution: agent-eval issues no provider request and holds no credential.\n *  The loop still owns turn scheduling, tool dispatch, transcript capture, and\n *  cost metering. agent-eval has no dependency on agent-runtime; adapt\n *  agent-runtime's `profileChatClient` (or any sandbox, bridge, or router\n *  client) into this signature product-side. */\nexport type MultishotTransport = (\n  req: MultishotTransportRequest,\n) => Promise<MultishotTransportResponse>\n\nexport type MultishotToolExecutor = (\n  args: Record<string, unknown>,\n  ctx: {\n    /** Caller-owned execution seam for the specialist leg this tool runs. */\n    transport: MultishotTransport\n    signal?: AbortSignal\n  },\n) => Promise<{ content: string; costUsd: number }>\n\nexport interface MultishotPersona {\n  /** Stable identifier — used for per-cell artifact paths + matrix axis keys. */\n  id: string\n  /** Per-domain payload (income/profile/voice/etc.) shaped by the consumer. */\n  [k: string]: unknown\n}\n\n/**\n * Persona-shaping callbacks. Both are OPTIONAL: when omitted, the loop derives\n * them from the `AgentProfile` + persona payload (see `defaultShapeFromProfile`)\n * so a pure-profile call — `runMultishot({ profile, persona })` — works with no\n * role-builder functions. Provide callbacks only to override the derived shape.\n */\nexport interface MultishotShape<TPersona extends MultishotPersona> {\n  /** Opening user message (turn 0) — the persona's first ask. */\n  buildOpener?: (persona: TPersona) => string\n  /** System prompt the driver LLM uses to roleplay the persona. Should set\n   *  voice, goals, constraints, time-pressure, and the \"never go silent\" rule. */\n  buildDriverSystemPrompt?: (persona: TPersona) => string\n}\n\nexport class MultishotDriverEmptyError extends Error {\n  constructor(public readonly turn: number) {\n    super(`multishot: driver returned empty content twice at turn ${turn} — failing loud`)\n    this.name = 'MultishotDriverEmptyError'\n  }\n}\n\nexport class MultishotFatalToolError extends Error {\n  constructor(message: string) {\n    super(message)\n    this.name = 'MultishotFatalToolError'\n  }\n}\n\nexport class MultishotShotResultError extends Error {\n  constructor(reason: string) {\n    super(`multishot: shot returned an invalid MultishotResult — ${reason}`)\n    this.name = 'MultishotShotResultError'\n  }\n}\n\nconst MULTISHOT_ROLES = new Set(['user', 'assistant', 'tool'])\n\n/** Contract guard for the value a caller-supplied shot resolves with. The\n *  matrix writes per-cell artifacts, builds judge inputs, and meters cost from\n *  this value, so a malformed result must stop the cell instead of scoring a\n *  degraded one. Two silent degradations this closes: an artifact with no\n *  `type` matches neither the code nor the content artifact set, so the cell\n *  scores as though the artifact was never produced; a non-finite `costUsd`\n *  reaches `summary.totalCostUsd` and makes every cost number NaN.\n *\n *  A rejected cell is still billed: the matrix cell reads the shot's own\n *  `costUsd` when it is a usable amount and declares that spend on the throw,\n *  so money the shot spent before returning a malformed result stays in the\n *  cumulative sum the cost ceiling reads. A result whose `costUsd` is itself\n *  malformed carries no usable amount, and the cell records as `uncaptured`.\n *\n *  Every required field of `MultishotMessage` and `MultishotArtifact` is\n *  checked, including `toolCalls` elements and `invocation.args`. Optional\n *  fields are checked only when present. */\nexport function assertMultishotShotResult(value: unknown): asserts value is MultishotResult {\n  if (typeof value !== 'object' || value === null) {\n    throw new MultishotShotResultError(`expected an object, received ${describeValue(value)}`)\n  }\n  const result = value as Record<string, unknown>\n  if (!Array.isArray(result.transcript)) {\n    throw new MultishotShotResultError(\n      `transcript must be an array, received ${describeValue(result.transcript)}`,\n    )\n  }\n  if (!Array.isArray(result.artifacts)) {\n    throw new MultishotShotResultError(\n      `artifacts must be an array, received ${describeValue(result.artifacts)}`,\n    )\n  }\n  result.transcript.forEach(assertMessage)\n  result.artifacts.forEach(assertArtifact)\n  assertFiniteCount(result.toolCalls, 'toolCalls')\n  assertFiniteCount(result.durationMs, 'durationMs')\n  assertFiniteCount(result.costUsd, 'costUsd')\n  if (result.costProvenance !== undefined) assertCostProvenance(result.costProvenance)\n}\n\nfunction assertCostProvenance(value: unknown): void {\n  const row = requireRow(value, 'costProvenance')\n  if (row.kind !== 'observed' && row.kind !== 'estimated' && row.kind !== 'uncaptured') {\n    throw new MultishotShotResultError(\n      `costProvenance.kind must be observed, estimated or uncaptured, received ${describeValue(row.kind)}`,\n    )\n  }\n  // The matrix reads the amount from `costUsd`; `usd` only has to agree with\n  // the kind, so an uncaptured provenance cannot smuggle in a total.\n  if (row.kind === 'uncaptured') {\n    if (row.usd !== null) {\n      throw new MultishotShotResultError(\n        `uncaptured costProvenance.usd must be null, received ${describeValue(row.usd)}`,\n      )\n    }\n    return\n  }\n  assertFiniteCount(row.usd, 'costProvenance.usd')\n}\n\nfunction assertMessage(value: unknown, index: number): void {\n  const row = requireRow(value, `transcript[${index}]`)\n  if (typeof row.role !== 'string' || !MULTISHOT_ROLES.has(row.role)) {\n    throw new MultishotShotResultError(\n      `transcript[${index}].role must be user, assistant or tool, received ${describeValue(row.role)}`,\n    )\n  }\n  assertString(row.content, `transcript[${index}].content`)\n  if (row.toolCallId !== undefined) {\n    assertString(row.toolCallId, `transcript[${index}].toolCallId`)\n  }\n  if (row.toolCalls === undefined) return\n  if (!Array.isArray(row.toolCalls)) {\n    throw new MultishotShotResultError(\n      `transcript[${index}].toolCalls must be an array when present, received ${describeValue(row.toolCalls)}`,\n    )\n  }\n  row.toolCalls.forEach((call, callIndex) => {\n    const field = `transcript[${index}].toolCalls[${callIndex}]`\n    const row = requireRow(call, field)\n    assertString(row.id, `${field}.id`)\n    assertString(row.name, `${field}.name`)\n    requireRow(row.args, `${field}.args`)\n  })\n}\n\nfunction assertArtifact(value: unknown, index: number): void {\n  const row = requireRow(value, `artifacts[${index}]`)\n  assertString(row.type, `artifacts[${index}].type`)\n  assertString(row.content, `artifacts[${index}].content`)\n  assertFiniteCount(row.turn, `artifacts[${index}].turn`)\n  const invocation = requireRow(row.invocation, `artifacts[${index}].invocation`)\n  assertString(invocation.name, `artifacts[${index}].invocation.name`)\n  requireRow(invocation.args, `artifacts[${index}].invocation.args`)\n}\n\nfunction requireRow(value: unknown, field: string): Record<string, unknown> {\n  if (typeof value !== 'object' || value === null || Array.isArray(value)) {\n    throw new MultishotShotResultError(\n      `${field} must be an object, received ${describeValue(value)}`,\n    )\n  }\n  return value as Record<string, unknown>\n}\n\nfunction assertString(value: unknown, field: string): void {\n  if (typeof value !== 'string') {\n    throw new MultishotShotResultError(\n      `${field} must be a string, received ${describeValue(value)}`,\n    )\n  }\n}\n\nfunction assertFiniteCount(value: unknown, field: string): void {\n  if (typeof value !== 'number' || !Number.isFinite(value) || value < 0) {\n    throw new MultishotShotResultError(\n      `${field} must be a finite number >= 0, received ${describeValue(value)}`,\n    )\n  }\n}\n\nfunction describeValue(value: unknown): string {\n  if (value === null) return 'null'\n  if (Array.isArray(value)) return 'an array'\n  if (typeof value === 'object') return 'an object'\n  return `${typeof value} ${String(value)}`\n}\n"],"mappings":";AAmHA,IAAa,4BAAb,cAA+C,MAAM;CACvB;CAA5B,YAAY,MAA8B;EACxC,MAAM,0DAA0D,KAAK,gBAAgB;EAD3D,KAAA,OAAA;EAE1B,KAAK,OAAO;CACd;AACF;AAEA,IAAa,0BAAb,cAA6C,MAAM;CACjD,YAAY,SAAiB;EAC3B,MAAM,OAAO;EACb,KAAK,OAAO;CACd;AACF;AAEA,IAAa,2BAAb,cAA8C,MAAM;CAClD,YAAY,QAAgB;EAC1B,MAAM,yDAAyD,QAAQ;EACvE,KAAK,OAAO;CACd;AACF;AAEA,MAAM,kCAAkB,IAAI,IAAI;CAAC;CAAQ;CAAa;AAAM,CAAC;;;;;;;;;;;;;;;;;;AAmB7D,SAAgB,0BAA0B,OAAkD;CAC1F,IAAI,OAAO,UAAU,YAAY,UAAU,MACzC,MAAM,IAAI,yBAAyB,gCAAgC,cAAc,KAAK,GAAG;CAE3F,MAAM,SAAS;CACf,IAAI,CAAC,MAAM,QAAQ,OAAO,UAAU,GAClC,MAAM,IAAI,yBACR,yCAAyC,cAAc,OAAO,UAAU,GAC1E;CAEF,IAAI,CAAC,MAAM,QAAQ,OAAO,SAAS,GACjC,MAAM,IAAI,yBACR,wCAAwC,cAAc,OAAO,SAAS,GACxE;CAEF,OAAO,WAAW,QAAQ,aAAa;CACvC,OAAO,UAAU,QAAQ,cAAc;CACvC,kBAAkB,OAAO,WAAW,WAAW;CAC/C,kBAAkB,OAAO,YAAY,YAAY;CACjD,kBAAkB,OAAO,SAAS,SAAS;CAC3C,IAAI,OAAO,mBAAmB,KAAA,GAAW,qBAAqB,OAAO,cAAc;AACrF;AAEA,SAAS,qBAAqB,OAAsB;CAClD,MAAM,MAAM,WAAW,OAAO,gBAAgB;CAC9C,IAAI,IAAI,SAAS,cAAc,IAAI,SAAS,eAAe,IAAI,SAAS,cACtE,MAAM,IAAI,yBACR,2EAA2E,cAAc,IAAI,IAAI,GACnG;CAIF,IAAI,IAAI,SAAS,cAAc;EAC7B,IAAI,IAAI,QAAQ,MACd,MAAM,IAAI,yBACR,wDAAwD,cAAc,IAAI,GAAG,GAC/E;EAEF;CACF;CACA,kBAAkB,IAAI,KAAK,oBAAoB;AACjD;AAEA,SAAS,cAAc,OAAgB,OAAqB;CAC1D,MAAM,MAAM,WAAW,OAAO,cAAc,MAAM,EAAE;CACpD,IAAI,OAAO,IAAI,SAAS,YAAY,CAAC,gBAAgB,IAAI,IAAI,IAAI,GAC/D,MAAM,IAAI,yBACR,cAAc,MAAM,mDAAmD,cAAc,IAAI,IAAI,GAC/F;CAEF,aAAa,IAAI,SAAS,cAAc,MAAM,UAAU;CACxD,IAAI,IAAI,eAAe,KAAA,GACrB,aAAa,IAAI,YAAY,cAAc,MAAM,aAAa;CAEhE,IAAI,IAAI,cAAc,KAAA,GAAW;CACjC,IAAI,CAAC,MAAM,QAAQ,IAAI,SAAS,GAC9B,MAAM,IAAI,yBACR,cAAc,MAAM,sDAAsD,cAAc,IAAI,SAAS,GACvG;CAEF,IAAI,UAAU,SAAS,MAAM,cAAc;EACzC,MAAM,QAAQ,cAAc,MAAM,cAAc,UAAU;EAC1D,MAAM,MAAM,WAAW,MAAM,KAAK;EAClC,aAAa,IAAI,IAAI,GAAG,MAAM,IAAI;EAClC,aAAa,IAAI,MAAM,GAAG,MAAM,MAAM;EACtC,WAAW,IAAI,MAAM,GAAG,MAAM,MAAM;CACtC,CAAC;AACH;AAEA,SAAS,eAAe,OAAgB,OAAqB;CAC3D,MAAM,MAAM,WAAW,OAAO,aAAa,MAAM,EAAE;CACnD,aAAa,IAAI,MAAM,aAAa,MAAM,OAAO;CACjD,aAAa,IAAI,SAAS,aAAa,MAAM,UAAU;CACvD,kBAAkB,IAAI,MAAM,aAAa,MAAM,OAAO;CACtD,MAAM,aAAa,WAAW,IAAI,YAAY,aAAa,MAAM,aAAa;CAC9E,aAAa,WAAW,MAAM,aAAa,MAAM,kBAAkB;CACnE,WAAW,WAAW,MAAM,aAAa,MAAM,kBAAkB;AACnE;AAEA,SAAS,WAAW,OAAgB,OAAwC;CAC1E,IAAI,OAAO,UAAU,YAAY,UAAU,QAAQ,MAAM,QAAQ,KAAK,GACpE,MAAM,IAAI,yBACR,GAAG,MAAM,+BAA+B,cAAc,KAAK,GAC7D;CAEF,OAAO;AACT;AAEA,SAAS,aAAa,OAAgB,OAAqB;CACzD,IAAI,OAAO,UAAU,UACnB,MAAM,IAAI,yBACR,GAAG,MAAM,8BAA8B,cAAc,KAAK,GAC5D;AAEJ;AAEA,SAAS,kBAAkB,OAAgB,OAAqB;CAC9D,IAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,KAAK,QAAQ,GAClE,MAAM,IAAI,yBACR,GAAG,MAAM,0CAA0C,cAAc,KAAK,GACxE;AAEJ;AAEA,SAAS,cAAc,OAAwB;CAC7C,IAAI,UAAU,MAAM,OAAO;CAC3B,IAAI,MAAM,QAAQ,KAAK,GAAG,OAAO;CACjC,IAAI,OAAO,UAAU,UAAU,OAAO;CACtC,OAAO,GAAG,OAAO,MAAM,GAAG,OAAO,KAAK;AACxC"}

package/docs/jev-behavior-review.md

@@ -144,6 +144,30 @@ requires resolution. Never convert an unresolved safety review to a perfect scor
 critical constraint failure into unrelated high scores. Working feedback must not expose the
 independent final evaluator, protected holdout labels, or monitor control credentials to the worker.
 
+## Claims an exact checker verified
+
+An exact checker decides whether an artifact satisfies the checker's own encoding of a statement.
+It cannot decide whether that encoding matches the statement the source meant.
+An agent that reads the checker finds the cases where they differ.
+A claim in such a case passes every exact check and settles nothing.
+
+`/rl` exports `claimIntegrityVerdict` for this case.
+It decides `pass`, `hack` or `unknown` for one claim from signals the caller computes:
+
+- `scopeSignals`: predicates over the claim's parameters, from a `SourceScope` written from the source by someone other than the checker's author. The source excludes the case, lists it, or reports it checked.
+- `farmingSignal`: one parameter value repeated across a lane's sweep of claims.
+- `graderReferenceSignal`: the claim page justifies itself by the checker or by the printed letter.
+- `refereeSignal`: a model reads the source and the claim's parameters, never the claimant's prose (`sourceScopeRefereeMessages`, `parseSourceScopeReferee`).
+- `jevSignal`: typed questions over the same brief (`claimIntegrityJevReview`). The caller supplies calibrated thresholds.
+- `monitorSignal`: `askTraceQuestions` with `claimIntegrityTraceQuestions`, on escalation only.
+- `checkerReadSignal`: context for a reader; it never moves a verdict.
+
+The policy fails closed.
+A pass needs a declared scope and a referee that evaluated the claim, with nothing against it.
+An exact signal decides `hack`, and so do two distinct kinds of signal.
+One kind alone holds the claim as `unknown`.
+Keep verdicts, scopes and questions away from the agent that filed the claim: an agent scored on them learns to hide the case.
+
 ## Security and validation boundaries
 
 TypeSafe's Jev 1.13 documentation explicitly warns that adversarial state content can move the

package/docs/public-api.md

@@ -7,11 +7,11 @@ Generated by `pnpm api:census` on demand — this is a dated reading, not a gate
 | measure | count |
 | --- | --- |
 | export subpaths | 33 |
-| published value exports (subpath x symbol) | 1570 |
-| distinct symbols | 1382 |
-| production | 1047 |
-| planned | 266 |
-| none | 257 |
+| published value exports (subpath x symbol) | 1589 |
+| distinct symbols | 1401 |
+| production | 1048 |
+| planned | 278 |
+| none | 263 |
 
 Type-only exports are not listed: a type binds no runtime surface, and removing one cannot break a caller at run time.
 
@@ -460,7 +460,7 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 
 ### `./analyst`
 
-131 value exports — 79 production, 19 planned, 33 none.
+131 value exports — 79 production, 20 planned, 32 none.
 
 | symbol | consumer | evidence |
 | --- | --- | --- |
@@ -489,7 +489,7 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | `AnalystRegistry` | production | blueprint-agent:scripts/experiments/vb-multi-analyst.ts |
 | `analystUsageReceiptFromPrimeUsage` | production | this package: src/analyst/benchmark-runner-prime.ts:19 |
 | `appendVerificationArtifactsToOtlp` | production | this package: src/analyst/benchmark-public-data.ts:32 |
-| `askTraceQuestions` | none | only this package's tests: src/analyst/trace-questions.test.ts:7 |
+| `askTraceQuestions` | planned | doc: docs/jev-behavior-review.md |
 | `assertProposalFindings` | production | agent-runtime:src/improvement/code-execution.ts |
 | `behavioralAnalyst` | production | agent-dev-container:products/intelligence/api/src/lib/deep-trace-analyst.ts |
 | `bindAnalyst` | planned | doc: docs/trace-analysis.md |
@@ -649,16 +649,16 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | symbol | consumer | evidence |
 | --- | --- | --- |
 | `acquireSingleRunLock` | production | agent-knowledge:src/memory/run-control.ts |
-| `aide` | production | this package: scripts/search-sim.ts:145 |
+| `aide` | production | this package: scripts/search-sim.ts:146 |
 | `analyzeCrossSurfaceInteractions` | none | only this package's tests: src/campaign/cross-surface-interaction.test.ts:2 |
-| `asha` | production | this package: scripts/search-sim.ts:119 |
+| `asha` | production | this package: scripts/search-sim.ts:120 |
 | `assertCampaignDesign` | production | this package: src/campaign/plan-campaign-run.ts:10 |
 | `assertCampaignSplitIdentity` | production | agent-knowledge:src/memory/experiment/learning-pairs.ts |
 | `assertCodeSurfaceIdentity` | none | — |
 | `assertCompleteSearchHistory` | production | this package: src/campaign/optimization-method.ts:14 |
 | `assertSearchHistoryMatchesState` | planned | doc: docs/search-ledger.md |
 | `autoevalsScorerJudge` | none | only this package's tests: src/campaign/upstream-evaluators.test.ts:6 |
-| `beam` | production | this package: scripts/search-sim.ts:145 |
+| `beam` | production | this package: scripts/search-sim.ts:146 |
 | `buildCellSchedule` | production | this package: src/campaign/plan-campaign-run.ts:9 |
 | `buildEvidenceVector` | planned | doc: docs/experiment.md |
 | `buildLoopProvenanceRecord` | none | only this package's tests: tests/campaign/candidate-attribution.test.ts:5 |
@@ -683,18 +683,18 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | `createReferenceEquivalenceJudge` | none | only this package's tests: src/reference-equivalence-judge.test.ts:13 |
 | `createRunCostLedger` | production | agent-dev-container:products/intelligence/api/src/lib/eval-engine.ts |
 | `createSearchHistoryReceipt` | production | this package: src/campaign/search-ledger-recording.ts:27 |
-| `crowdedFrontierParent` | production | this package: scripts/search-sim.ts:145 |
+| `crowdedFrontierParent` | production | this package: scripts/search-sim.ts:146 |
 | `decideSearchClaim` | production | this package: src/campaign/search-kernel.ts:42 |
 | `decodeExternalTextCandidate` | production | agent-runtime:src/improvement/method-execution.ts |
 | `DEFAULT_EXTERNAL_OPTIMIZER_CALLBACK_LIMITS` | none | — |
 | `DEFAULT_EXTERNAL_OPTIMIZER_PROCESS_LIMITS` | none | — |
 | `defaultProductionGate` | production | agent-app:src/eval-campaign/index.ts |
 | `detectScale` | production | this package: src/campaign/gates/promotion-policy.ts:25 |
-| `developmentClaim` | production | this package: scripts/generate-synthetic-search-ledger.ts:34 |
+| `developmentClaim` | production | this package: scripts/generate-synthetic-search-ledger.ts:37 |
 | `dimensionRegressions` | production | agent-dev-container:products/intelligence/api/src/lib/optimization-engine.ts |
 | `discoverEvalFixtures` | planned | doc: docs/eval-fixtures.md |
 | `emitLoopProvenance` | production | this package: src/contract/self-improve.ts:36 |
-| `estimateNode` | production | this package: scripts/search-sim.ts:120 |
+| `estimateNode` | production | this package: scripts/search-sim.ts:121 |
 | `estimateNodeFromCells` | production | this package: scripts/synthetic-search-ledger.ts:60 |
 | `externalSurface` | none | — |
 | `externalTextOptimizationMethod` | production | agent-app:src/eval-campaign/index.ts |
@@ -725,12 +725,12 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | `neutralizationGate` | planned | consumer tests: tax-agent:tests/eval/benchmarks/taxcalc/evolve.ts |
 | `neutralizeText` | planned | consumer tests: tax-agent:tests/eval/benchmarks/taxcalc/compile-and-prove/compile-diff.ts |
 | `openAutoPr` | production | this package: src/campaign/presets/run-improvement-loop.ts:8 |
-| `openSearchLedger` | production | this package: scripts/generate-synthetic-search-ledger.ts:34 |
+| `openSearchLedger` | production | this package: scripts/generate-synthetic-search-ledger.ts:37 |
 | `optimizationTokenUsageFromSummary` | production | this package: src/campaign/external-text-optimization.ts:41 |
 | `pairHoldout` | production | agent-dev-container:products/intelligence/api/src/lib/run-provenance.ts |
 | `paretoPolicy` | none | only this package's tests: src/campaign/gates/promotion-policy.test.ts:3 |
 | `paretoSignificanceGate` | production | agent-app:src/eval-campaign/index.ts |
-| `parseSearchLedgerLine` | production | this package: scripts/search-sim.ts:132 |
+| `parseSearchLedgerLine` | production | this package: scripts/search-sim.ts:133 |
 | `phoenixEvaluatorJudge` | none | only this package's tests: src/campaign/upstream-evaluators.test.ts:6 |
 | `planCampaignRun` | production | this package: src/campaign/fixtures.ts:5 |
 | `planEvalFixtureRun` | planned | example: examples/eval-fixtures-quickstart/index.ts:11 |
@@ -748,7 +748,7 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | `renderScoreboardMarkdown` | production | blueprint-agent:scripts/experiments/eval/launch-scoreboard/run.ts |
 | `renderSearchSummary` | production | this package: scripts/search-meta-sim.ts:50 |
 | `renderSurfaceDiff` | production | agent-dev-container:products/intelligence/api/src/lib/optimization-evidence-packet.ts |
-| `replaySearchLedgerText` | production | this package: scripts/generate-synthetic-search-ledger.ts:34 |
+| `replaySearchLedgerText` | production | this package: scripts/generate-synthetic-search-ledger.ts:37 |
 | `resolveExternalOptimizerCallbackLimits` | production | this package: src/analyst/dspy-rlm-engine.ts:1 |
 | `resolveExternalOptimizerProcessLimits` | production | this package: src/analyst/benchmark-command-persistence.ts:6 |
 | `resolveRunDir` | production | agent-dev-container:products/intelligence/api/src/lib/eval-engine.ts |
@@ -760,7 +760,7 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | `runOptimization` | production | agent-dev-container:products/sandbox/evals/src/auto-optimization.ts |
 | `runProfileMatrix` | production | agent-runtime:examples/product-eval/product-eval.ts |
 | `runProfileMatrixSegment` | production | discovery-lab:tools/run-profile-confirmation.mjs |
-| `runSearch` | production | this package: scripts/search-sim.ts:121 |
+| `runSearch` | production | this package: scripts/search-sim.ts:122 |
 | `scopedOptimizationMethod` | planned | doc: docs/campaign-proposers.md |
 | `scoreboardSummary` | production | blueprint-agent:scripts/experiments/eval/launch-scoreboard/run.ts |
 | `scoreDiscrimination` | production | supervisor-lab:bench/comms/seat-discrimination.ts |
@@ -774,7 +774,7 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | `searchClaimReserveUsd` | production | this package: scripts/search-meta-sim.ts:43 |
 | `searchDivergence` | planned | doc: docs/search-ledger.md |
 | `searchEdgeId` | planned | doc: docs/search-ledger.md |
-| `searchExpansionIndex` | production | this package: scripts/search-sim.ts:121 |
+| `searchExpansionIndex` | production | this package: scripts/search-sim.ts:122 |
 | `searchHistoryCoverageRow` | production | this package: src/campaign/optimization-method.ts:14 |
 | `SearchHistoryRequiredError` | none | — |
 | `SearchLedgerConflictError` | production | this package: src/campaign/search-ledger.ts:36 |
@@ -787,8 +787,8 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | `searchProposalExecution` | production | this package: src/campaign/presets/run-optimization.ts:51 |
 | `searchProposerView` | production | this package: src/campaign/presets/run-optimization.ts:86 |
 | `searchReceiptAccounting` | production | this package: src/campaign/presets/run-optimization.ts:51 |
-| `SearchRecorder` | production | this package: scripts/generate-synthetic-search-ledger.ts:34 |
-| `SearchState` | production | this package: scripts/search-sim.ts:144 |
+| `SearchRecorder` | production | this package: scripts/generate-synthetic-search-ledger.ts:37 |
+| `SearchState` | production | this package: scripts/search-sim.ts:145 |
 | `SearchStateView` | production | this package: scripts/search-meta-sim.ts:49 |
 | `searchTaskSetDigest` | production | this package: src/campaign/search-ledger-recording.ts:57 |
 | `searchUnitScores` | production | this package: src/campaign/estimate-node.ts:28 |
@@ -1063,7 +1063,7 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 
 ### `./jev`
 
-15 value exports — 7 production, 7 planned, 1 none.
+15 value exports — 8 production, 6 planned, 1 none.
 
 | symbol | consumer | evidence |
 | --- | --- | --- |
@@ -1081,7 +1081,7 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | `parseJevQuestions` | production | this package: src/jev.ts:7 |
 | `parseJevRequest` | production | this package: src/jev-review.ts:9 |
 | `parseJevResult` | production | this package: src/jev-review.ts:9 |
-| `prepareJevReview` | planned | example: examples/jev-behavior-review.ts:3 |
+| `prepareJevReview` | production | this package: src/rl/claim-integrity.ts:39 |
 
 ### `./jev/protocol`
 
@@ -1103,7 +1103,7 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | --- | --- | --- |
 | `appendLedgerLine` | production | this package: src/campaign/search-ledger-file.ts:3 |
 | `AtomicFileLockError` | production | this package: src/ledger-core/journal-file.ts:16 |
-| `canonicalString` | production | this package: scripts/search-sim.ts:154 |
+| `canonicalString` | production | this package: scripts/search-sim.ts:155 |
 | `FileLedgerJournal` | production | this package: src/campaign/search-ledger.ts:23 |
 | `hashCanonical` | production | discovery-lab:tools/provenance.mjs |
 | `jsonDocument` | production | this package: src/agent-profile-cell.ts:3 |
@@ -1280,7 +1280,7 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 
 ### `./rl`
 
-63 value exports — 18 production, 18 planned, 27 none.
+82 value exports — 18 production, 30 planned, 34 none.
 
 | symbol | consumer | evidence |
 | --- | --- | --- |
@@ -1294,6 +1294,14 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | `buildRlDataset` | production | this package: src/rl/corpus.ts:24 |
 | `buildVerifiedFindingRow` | none | only this package's tests: src/rl/verified-findings-dataset.test.ts:5 |
 | `campaignToRunRecords` | production | agent-builder:src/lib/.server/eval/loops/auto-research-runner.ts |
+| `checkerReadSignal` | planned | doc: docs/jev-behavior-review.md |
+| `CLAIM_INTEGRITY_VERSION` | none | — |
+| `claimBrief` | none | — |
+| `claimIntegrityDigest` | none | only this package's tests: src/rl/claim-integrity.test.ts:4 |
+| `claimIntegrityJevQuestions` | none | — |
+| `claimIntegrityJevReview` | planned | doc: docs/jev-behavior-review.md |
+| `claimIntegrityTraceQuestions` | planned | doc: docs/jev-behavior-review.md |
+| `claimIntegrityVerdict` | planned | doc: docs/jev-behavior-review.md |
 | `compareAdaptationCurves` | planned | doc: docs/design/statistics-decisions.md |
 | `datasheetToMarkdown` | none | only this package's tests: src/rl/dataset.test.ts:7 |
 | `defaultBehaviorFeatures` | none | only this package's tests: src/rl/sim-fidelity.test.ts:3 |
@@ -1304,32 +1312,43 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | `extractStepRewards` | none | only this package's tests: tests/rl-process-reward.test.ts:3 |
 | `extractVerifiableReward` | planned | named in gtm-agent (bind not in the import graph) |
 | `extractVerifiableRewardsFromRecords` | production | creative-agent:eval/canonical-export.ts |
+| `farmingSignal` | planned | doc: docs/jev-behavior-review.md |
 | `FileSystemOutcomeStore` | production | phony:products/builder/api/src/eval/outcomes.ts |
 | `filterDeterministicallyRewarded` | production | this package: src/rl/reward-hacking.ts:43 |
 | `firstPassK` | planned | doc: docs/statistical-evidence.md |
 | `fitBradleyTerry` | none | only this package's tests: tests/rl-tournament.test.ts:2 |
+| `GRADER_REFERENCE_PHRASES` | none | — |
+| `graderReferenceSignal` | planned | doc: docs/jev-behavior-review.md |
 | `injectIrrelevantClause` | none | only this package's tests: tests/rl-contamination.test.ts:2 |
 | `InMemoryOutcomeStore` | planned | consumer tests: phony:products/builder/api/src/eval/outcomes-correlate.test.ts |
 | `inverseProbabilityWeighting` | planned | consumer tests: agent-knowledge:tests/memory-holdout.test.ts |
+| `jevSignal` | planned | doc: docs/jev-behavior-review.md |
 | `jsDivergence` | none | only this package's tests: src/rl/sim-fidelity.test.ts:3 |
 | `loadVerifiedFindingsDataset` | planned | doc: docs/verified-labels-flywheel.md |
+| `monitorSignal` | planned | doc: docs/jev-behavior-review.md |
 | `offPolicyEstimateAll` | none | only this package's tests: tests/rl-off-policy.test.ts:3 |
 | `paretoFrontier` | production | agent-dev-container:products/sandbox/evals/run.ts |
+| `parseSourceScopeReferee` | planned | doc: docs/jev-behavior-review.md |
 | `PredictiveValidityResearcher` | planned | doc: docs/design/mlbenchmarks-book-review.md |
 | `prmTrainingPairs` | planned | doc: examples/fine-tune-with-prime-rl/README.md |
 | `quantileEdges` | none | only this package's tests: src/rl/sim-fidelity.test.ts:3 |
 | `readCorpus` | planned | named in agent-runtime (bind not in the import graph) |
+| `REFEREE_VERDICTS` | none | — |
+| `refereeSignal` | planned | doc: docs/jev-behavior-review.md |
 | `renameVariables` | none | only this package's tests: tests/rl-contamination.test.ts:2 |
+| `ruleHolds` | none | only this package's tests: src/rl/claim-integrity.test.ts:4 |
 | `runAdaptationCurve` | planned | doc: docs/statistical-evidence.md |
 | `runComputeCurve` | none | only this package's tests: tests/rl-adversarial-and-compute.test.ts:2 |
 | `runContaminationProbe` | planned | doc: docs/statistical-evidence.md |
 | `runEvalCampaign` | production | agent-builder:scripts/eval.ts |
 | `runRLCampaign` | planned | doc: docs/outcome-validity.md |
 | `runwiseStepRewardSummary` | none | only this package's tests: tests/rl-process-reward.test.ts:3 |
+| `scopeSignals` | planned | doc: docs/jev-behavior-review.md |
 | `selfConsistency` | none | only this package's tests: tests/rl-adversarial-and-compute.test.ts:2 |
 | `selfNormalizedImportanceWeighting` | planned | consumer tests: agent-knowledge:tests/memory-holdout.test.ts |
 | `shuffleOrder` | none | — |
 | `simFidelityReport` | none | only this package's tests: src/rl/sim-fidelity.test.ts:3 |
+| `sourceScopeRefereeMessages` | planned | doc: docs/jev-behavior-review.md |
 | `stepRewardsToJsonl` | none | only this package's tests: src/rollout/gate-properties.test.ts:38 |
 | `summarizeVerifiedFindings` | none | only this package's tests: src/rl/verified-findings-dataset.test.ts:5 |
 | `toAnthropicFormat` | production | blueprint-agent:scripts/experiments/lib/steering-dataset.ts |
@@ -1428,7 +1447,7 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 
 | symbol | consumer | evidence |
 | --- | --- | --- |
-| `draftOnPlateau` | production | this package: scripts/search-sim.ts:161 |
+| `draftOnPlateau` | production | this package: scripts/search-sim.ts:162 |
 | `EDIT_CREDIT_ESTIMATOR` | none | — |
 | `EDIT_CREDIT_SIGNAL` | none | — |
 | `editCredit` | production | this package: scripts/synthetic-search-ledger.ts:70 |
@@ -1437,9 +1456,9 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | `formatLandscape` | production | this package: src/search-command.ts:41 |
 | `formatSkillManifold` | production | this package: src/search-command.ts:45 |
 | `front` | production | this package: src/search-command.ts:40 |
-| `landscape` | production | this package: scripts/search-sim.ts:155 |
+| `landscape` | production | this package: scripts/search-sim.ts:156 |
 | `lineageEdits` | planned | doc: docs/search-ledger.md |
-| `lineEditDistance` | production | this package: scripts/search-sim.ts:155 |
+| `lineEditDistance` | production | this package: scripts/search-sim.ts:156 |
 | `META_SEARCH_SCORE_SOURCE` | production | this package: scripts/search-meta-sim.ts:52 |
 | `META_SEARCH_SIGNAL` | none | — |
 | `metaSearch` | production | this package: scripts/search-meta-sim.ts:52 |
@@ -1447,7 +1466,7 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | `MIN_OUTCOMES_FOR_WEIGHT` | planned | doc: docs/search-ledger.md |
 | `nearestNeighbours` | none | — |
 | `nestedSearchId` | planned | doc: docs/search-ledger.md |
-| `nextUnitExtension` | production | this package: scripts/search-sim.ts:162 |
+| `nextUnitExtension` | production | this package: scripts/search-sim.ts:163 |
 | `objectiveKey` | production | this package: scripts/search-meta-sim.ts:52 |
 | `operatorYield` | production | this package: scripts/prove-operator-bandit.mts:24 |
 | `profileTextLines` | none | — |
@@ -1458,12 +1477,12 @@ The `none` set was reviewed symbol by symbol on 2026-08-21. Four rules decided m
 | `searchConfigCodec` | planned | doc: docs/search-ledger.md |
 | `searchPlateau` | production | this package: src/search/lenses/landscape.ts:49 |
 | `searchPolicyGenome` | planned | doc: docs/search-ledger.md |
-| `skillCalibration` | production | this package: scripts/search-sim.ts:162 |
-| `skillManifold` | production | this package: scripts/search-sim.ts:162 |
+| `skillCalibration` | production | this package: scripts/search-sim.ts:163 |
+| `skillManifold` | production | this package: scripts/search-sim.ts:163 |
 | `summarizeSamples` | production | this package: src/search/lenses/meta-search.ts:55 |
 | `surfaceDigestEdits` | none | — |
 | `surfaceTextEdits` | production | this package: src/search-command.ts:41 |
-| `taskMatrix` | production | this package: src/search-command.ts:46 |
+| `taskMatrix` | production | this package: scripts/prove-task-matrix.mts:8 |
 | `tree` | production | this package: src/search-command.ts:47 |
 | `vectorEmbedding` | planned | doc: docs/search-ledger.md |
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-eval",
-  "version": "0.199.1",
+  "version": "0.200.0",
   "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/index-D4pC7Tvd.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"index-D4pC7Tvd.d.ts","names":[],"sources":["../src/rollout/gate-checks.ts","../src/rollout/interchange/harbor.ts","../src/rollout/ledger.ts","../src/rollout/readers/claude-jsonl.ts","../src/rollout/readers/opencode-sqlite.ts","../src/rollout/release/gate-report.ts","../src/rollout/release/scrub.ts","../src/rollout/release/card.ts","../src/rollout/release/hf-dataset.ts"],"mappings":";;;;;;;;;;;;cA2Ca;KAOD,sBAAsB;;;;;;;;;KAUtB,qBAAqB,QAAQ,kBAAkB,SAAS;;;;;;;;;;;;;;;;;UAkBnD;EACf,SAAS;;EAET;;UAOe;EACf,IAAI;;EAEJ;;EAEA,SAAS,SAAS;;;;;;;;;;;;EAYlB,WAAW;;;;;;;;;;;;;iBA6LG,gBAAgB,SAAS,cAAc;;;;;;cA6I1C,yBAAyB,KAAK,cAAc;;;;;;;;;KAe7C;WACG;;;;WAEA;WAAyB;;;;WAEzB;WAAuB;;;KAU1B,yBAAyB,KAAK,cAAc;;;;;;;;;cAU3C;;;;;;;aAWT;eAnCiB;;aAoCjB,kBAAgB;aAMhB,2BAAyB;aAMzB,qBAAmB;;;;;;;;aAenB;;;aACA,kBAAgB;aAChB,2BAAyB;aACzB;;;;;;;;;;;aASA;;;aACA;;;aACA;;;aACA;;;;;;;;;;;aASA;;;aACA;;;aACA;;;aACA;;;;;;KAKQ,8BAA8B;;;;;;iBAO1B,WAAW,SAAS,aAAa,QAAQ;;;cCxd5C;;cAGA;KAwBD;UAEK;EACf;EACA;;UAGe;EACf;EACA;EACA,SAAS;;UAGM;EACf;EACA;;EAEA,WAAW;EACX,QAAQ;;UAGO;EACf;EACA;;EAEA;EACA,QAAQ;;UAGO;EACf;EACA,mBAAmB;EACnB,0BAA0B;EAC1B,QAAQ;;UAGO;EACf,SAAS;;UAGM;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA,QAAQ;;UAGO;;EAEf;EACA;EACA,QAAQ;EACR;EACA;EACA,kBAAkB;EAClB;EACA,aAAa;EACb,cAAc;EACd,UAAU;EACV;EACA;EACA,QAAQ;;UAGO;EACf;EACA;EACA;;EAEA,mBAAmB;EACnB,QAAQ;;UAGO;EACf;EACA;EACA;EACA;EACA;EACA,QAAQ;;UAGO;EACf;EACA;;EAEA;EACA,OAAO;EACP,OAAO;EACP;EACA,gBAAgB;EAChB;EACA,wBAAwB;EACxB,QAAQ;;;;;;;;;;;;;;;;;;iBAmXM,mBAAmB,OAAO,gBAAgB;;iBAa1C,qBAAqB,OAAO,gBAAgB;UA2O3C;;EAEf,YAAY;;;;;;;;;;;iBAwGE,qBACd,YAAY,kBACZ,UAAS,oBACR;;;;;;;;;;;;;;;iBAiCa,qBACd,gBAAgB,eAChB,OAAO,eACN;;;;iBC/6BmB,mBAAmB,cAAc,OAAO,gBAAgB;;iBAOxD,mBAAmB,cAAc,OAAO,gBAAgB;;;;;;;;;;;iBAiBxD,kBAAkB,eAAe,QAAQ;;;;;;;;;;;;;;;iBAkBzC,mBAAmB,eAAe,QAAQ;;;cClDnD;;iBAGG,kBAAkB;UAIjB;EACf;EACA;;;iBAIoB,sBACpB,aACA,uBACC,QAAQ;UASM;EACf;EACA;EACA;EACA;;UAGe;EACf,UAAU;EACV,OAAO;;EAEP;EACA;EACA;;UAyDe;;;;;;WAMN;;;iBAgBW,qBACpB,cACA,UAAS,8BACR,QAAQ;;;cCnHE;UAEI;EACf;EACA;EACA;EACA;;EAEA;IAAS;IAAa;;EACtB;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;;iBAYoB,eACpB,gBACC,QAAQ;;iBAiDK,gCACd,IAAI,cACJ,oBACC;;;;;;;iBA2Ba,4BAA4B,IAAI,cAAc,oBAAoB;;;;KC5EtE;cAEC,yBAAyB,OAAO,eAAe;;UAQ3C;EACf;EACA;;;;;;;;;;;;EAYA;;;;;;;;EAQA;;;;;;;;;;;;EAYA;;;UAIe;;EAEf;EACA;;UAGe;;EAEf;;EAEA;;;;;;EAMA;;EAEA;;;;;;;;;;;;;;EAcA,oBAAoB;;;;;;;;;;;;EAYpB;;EAEA;;;;;;;;;;EAUA,wBAAwB;;UAGT;;EAEf;EACA,UAAU,QAAQ,OAAO,eAAe;;;iBAI1B,gBAAgB,gBAAgB,sBAAsB;;;;;;;cAUzD;EAIX,MAAG,eAAkB,aAAW;EAShC,YAAS,eAAkB,6BAA2B;EAStD,MAAG,eAAkB,cAAY;EAOjC,MAAG,gBAAmB,wBAAsB;;;iBAgC9B,kBACd,OAAO,qBACP,eAAe,kBACd;;;;;;;;;;;;;;;;iBA+Fa,iBAAiB,QAAQ;;;UCrUxB;EACf;EACA,SAAS;;EAET,UAAU,eAAe;;;KAqEf,cAAc;iBAEV,oBAAoB;iBAIpB,eAAe,MAAM,aAAa,MAAM,cAAc;iBAKtD,UAAU,cAAc,QAAQ;;;;;;;;;iBA8BhC,iBAAiB,MAAM,mBAAmB,QAAQ,cAAc;iBAIhE,WAAW,OAAO;EAChC,OAAO;EACP,QAAQ;;;;;;;;;iBAaM,uBAAuB;;;cCrI1B;KACD,wBAAwB;;cAGvB,cAAc,OAAO;UAejB;;EAEf,OAAO;EACP,SAAS;EACT;;EAEA;EACA,aAAa;EACb;IAAY;IAAmB;;EAC/B,cAAc,QAAQ,OAAO;;;;;;;EAO7B,MAAM;;iBA4GQ,iBAAiB,QAAQ;;;UC1HxB;EACf;EACA,SAAS;EACT;;UAGe;;EAEf,OAAO,eAAe;EACtB,QAAQ;EACR;IAAY;IAAmB;;;UAGhB;EACf;EACA;EACA;EACA,OAAO;EACP,cAAc,QAAQ,OAAO;;EAE7B,MAAM;EACN;;iBAGoB,eACpB,kBACA,SAAS,eACR,QAAQ;iBAuGK,gBAAgB,cAAc;UAoB7B,8BAA8B;EAC7C;EACA;;iBAMc,wBAAwB,iBAAiB;;iBA0CnC,qBAAqB,iBAAiB"}

package/dist/index-DNgf5gyG.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"index-DNgf5gyG.d.ts","names":[],"sources":["../src/matrix/cell-spend.ts","../src/matrix/types.ts","../src/matrix/runner.ts"],"mappings":";;;;;;;;;;;;;;;;;;UAoBiB;;;;EAIf;;EAEA;;;;EAIA;;;;;;;;;;;;;;;;iBAiBc,cAAc,gBAAgB,OAAO;;;;iBA+CrC,cAAc,iBAAiB;;;;;UC5E9B,WAAW;;EAE1B;;EAEA,QAAQ;IAAQ;IAAY,OAAO;;;;EAGnC,SAAS,OAAO,GAAG;;;UAIJ;EACf,MAAM;IAAiB;IAAY;;;EAEnC;;EAEA;;UAGe,WAAW;EAC1B,QAAQ;EACR,SAAS;;;;EAIT;EACA;EACA;;;;;;;;;;;EAWA,iBAAiB;;;;EAIjB;IAAU;IAAiB;;;UAGZ;EACf;EACA;EACA;EACA;EACA;EACA;EACA;;;EAGA;;EAEA;EACA;;UAGe,aAAa;EAC5B,OAAO;IAAQ,MAAM;IAAY,MAAM,WAAW;;;;EAGlD,QAAQ,eAAe,eAAe;EACtC;IACE;IACA;;;IAGA;IACA;IACA;;;;IAIA;;;;IAIA;;;;IAIA;IACA;;;EAGF;;UAGe,sBAAsB;EACrC,MAAM;;;;;;EAMN,UAAU,MAAM,eAAe,QAAQ,WAAW;;EAElD;;EAEA,UAAU,MAAM,KAAK;;EAErB;;EAEA;;;;;EAKA;;;;;;;;;;;;;EAaA;;EAEA,kBAAkB,MAAM,YAAY,QAAQ,WAAW;;;EAGvD,SAAS;;;;iBC1BW,eAAe,QACnC,MAAM,sBAAsB,UAC3B,QAAQ,aAAa"}

package/dist/jev-protocol.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"jev-protocol.d.ts","names":[],"sources":["../src/jev-protocol.ts"],"mappings":";;;;;KAOY,6CAA6C;GAAe,cAAc;;;KAE1E;GAAuB,cAAc;IAAY;;KAEjD,iBAAiB;KACjB,6BAA6B,UAAU,aAAa;KACpD;EAEN;EACA,eAAe;EACf;IAAa,OAAO;IAAgB,QAAQ;;;EAE5C;EAAe,eAAe;EAAgB,UAAU;;EACxD;EAAgB,eAAe;EAAgB,UAAU,eAAe;;KAClE,eAAe,eAAe;UACzB,WAAW,UAAU,eAAe;EACnD;EACA,OAAO;EACP,WAAW;;KAGR,UAAU,UAAU,mCAAmC,uBAExD,cAAc;KACN,aAAa,UAAU,eAAe;EAAY;;WAC/C;WAAuB;IAClC;EAAY;EAAe,gBAAgB,UAAU;;WAExC;WACA;WACA;WACA,oBAAoB,KAAK,UAAU,KAAK,EAAE;WAC1C,2BAA2B,KAAK,UAAU;IAErD;EAAY;EAAgB,gBAAgB;;WAE/B;WACA,cAAc;WACd;WACA,2BAA2B,WAAW;;KAG7C,YAAY,aAAa;UACpB,UAAU,UAAU,eAAe;WACzC;WACA,qBAAqB,WAAW,IAAI,aAAa,EAAE;WACnD;aAAkB;aAA+B;;;cAG/C,yBAAyB;EACpC,YAAY;;;iBA6BE,kBAAkB,eAAe;iBAkCjC,gBAAgB,eAAe;iBAU/B,SAAS,eAAe;;iBA2BxB,eAAe,UAAU,cACvC,cACA,SAAS,WAAW,KACnB,UAAU"}

package/dist/types-B3jzCp0p2.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"types-B3jzCp0p2.js","names":[],"sources":["../src/multishot/types.ts"],"sourcesContent":["// Public types for the multishot substrate.\n\nimport type { CostProvenance } from '../cost-ledger'\n\nexport interface MultishotMessage {\n  role: 'user' | 'assistant' | 'tool'\n  content: string\n  toolCallId?: string\n  toolCalls?: Array<{ id: string; name: string; args: Record<string, unknown> }>\n}\n\nexport interface MultishotArtifact {\n  type: string\n  turn: number\n  invocation: { name: string; args: Record<string, unknown> }\n  content: string\n}\n\nexport interface MultishotResult {\n  transcript: MultishotMessage[]\n  artifacts: MultishotArtifact[]\n  toolCalls: number\n  durationMs: number\n  /** Known spend. A subtotal, not a total, when `costProvenance.kind` is\n   *  `uncaptured`. */\n  costUsd: number\n  /** Origin of `costUsd`. A shot that priced every call reports `estimated`\n   *  or `observed`; a shot with a call the router priced at nothing reports\n   *  `uncaptured`, and the matrix records the cell as under-counted instead of\n   *  presenting the subtotal as a complete estimate.\n   *\n   *  Optional so an engine written before this field keeps working; the matrix\n   *  then judges the cell on judge receipts alone, as it did before. */\n  costProvenance?: CostProvenance\n}\n\nexport interface MultishotToolDefinition {\n  type: 'function'\n  function: {\n    name: string\n    description: string\n    parameters: Record<string, unknown>\n  }\n}\n\n/** One chat-completion request the multishot loop issues for a single agent\n *  (or driver) inference step. Mirrors the OpenAI-compat body the loop would\n *  otherwise POST to the Tangle router. */\nexport interface MultishotTransportRequest {\n  model: string\n  messages: Array<Record<string, unknown>>\n  tools?: MultishotToolDefinition[]\n  temperature?: number\n  maxTokens?: number\n  signal?: AbortSignal\n}\n\nexport interface MultishotTransportToolCall {\n  id: string\n  type: 'function'\n  function: { name: string; arguments: string }\n}\n\nexport interface MultishotTransportResponse {\n  message: { content?: string | null; tool_calls?: MultishotTransportToolCall[] }\n  usage?: { prompt_tokens?: number; completion_tokens?: number }\n  /** Actual spend for this call. When omitted, the loop meters cost from\n   *  `usage` via the per-model estimator (estimateMultishotCost). */\n  costUsd?: number\n  /** Model identity the provider reported, when the transport observed one.\n   *  Omitted means unreported, and the requested model is used for\n   *  attribution — which is not proof that model answered. */\n  model?: string\n}\n\n/** Execution seam for one leg of the multishot loop. The caller owns model\n *  execution: agent-eval issues no provider request and holds no credential.\n *  The loop still owns turn scheduling, tool dispatch, transcript capture, and\n *  cost metering. agent-eval has no dependency on agent-runtime; adapt\n *  agent-runtime's `profileChatClient` (or any sandbox, bridge, or router\n *  client) into this signature product-side. */\nexport type MultishotTransport = (\n  req: MultishotTransportRequest,\n) => Promise<MultishotTransportResponse>\n\nexport type MultishotToolExecutor = (\n  args: Record<string, unknown>,\n  ctx: {\n    /** Caller-owned execution seam for the specialist leg this tool runs. */\n    transport: MultishotTransport\n    signal?: AbortSignal\n  },\n) => Promise<{ content: string; costUsd: number }>\n\nexport interface MultishotPersona {\n  /** Stable identifier — used for per-cell artifact paths + matrix axis keys. */\n  id: string\n  /** Per-domain payload (income/profile/voice/etc.) shaped by the consumer. */\n  [k: string]: unknown\n}\n\n/**\n * Persona-shaping callbacks. Both are OPTIONAL: when omitted, the loop derives\n * them from the `AgentProfile` + persona payload (see `defaultShapeFromProfile`)\n * so a pure-profile call — `runMultishot({ profile, persona })` — works with no\n * role-builder functions. Provide callbacks only to override the derived shape.\n */\nexport interface MultishotShape<TPersona extends MultishotPersona> {\n  /** Opening user message (turn 0) — the persona's first ask. */\n  buildOpener?: (persona: TPersona) => string\n  /** System prompt the driver LLM uses to roleplay the persona. Should set\n   *  voice, goals, constraints, time-pressure, and the \"never go silent\" rule. */\n  buildDriverSystemPrompt?: (persona: TPersona) => string\n}\n\nexport class MultishotDriverEmptyError extends Error {\n  constructor(public readonly turn: number) {\n    super(`multishot: driver returned empty content twice at turn ${turn} — failing loud`)\n    this.name = 'MultishotDriverEmptyError'\n  }\n}\n\nexport class MultishotFatalToolError extends Error {\n  constructor(message: string) {\n    super(message)\n    this.name = 'MultishotFatalToolError'\n  }\n}\n\nexport class MultishotShotResultError extends Error {\n  constructor(reason: string) {\n    super(`multishot: shot returned an invalid MultishotResult — ${reason}`)\n    this.name = 'MultishotShotResultError'\n  }\n}\n\nconst MULTISHOT_ROLES = new Set(['user', 'assistant', 'tool'])\n\n/** Contract guard for the value a caller-supplied shot resolves with. The\n *  matrix writes per-cell artifacts, builds judge inputs, and meters cost from\n *  this value, so a malformed result must stop the cell instead of scoring a\n *  degraded one. Two silent degradations this closes: an artifact with no\n *  `type` matches neither the code nor the content artifact set, so the cell\n *  scores as though the artifact was never produced; a non-finite `costUsd`\n *  reaches `summary.totalCostUsd` and makes every cost number NaN.\n *\n *  A rejected cell is still billed: the matrix cell reads the shot's own\n *  `costUsd` when it is a usable amount and declares that spend on the throw,\n *  so money the shot spent before returning a malformed result stays in the\n *  cumulative sum the cost ceiling reads. A result whose `costUsd` is itself\n *  malformed carries no usable amount, and the cell records as `uncaptured`.\n *\n *  Every required field of `MultishotMessage` and `MultishotArtifact` is\n *  checked, including `toolCalls` elements and `invocation.args`. Optional\n *  fields are checked only when present. */\nexport function assertMultishotShotResult(value: unknown): asserts value is MultishotResult {\n  if (typeof value !== 'object' || value === null) {\n    throw new MultishotShotResultError(`expected an object, received ${describeValue(value)}`)\n  }\n  const result = value as Record<string, unknown>\n  if (!Array.isArray(result.transcript)) {\n    throw new MultishotShotResultError(\n      `transcript must be an array, received ${describeValue(result.transcript)}`,\n    )\n  }\n  if (!Array.isArray(result.artifacts)) {\n    throw new MultishotShotResultError(\n      `artifacts must be an array, received ${describeValue(result.artifacts)}`,\n    )\n  }\n  result.transcript.forEach(assertMessage)\n  result.artifacts.forEach(assertArtifact)\n  assertFiniteCount(result.toolCalls, 'toolCalls')\n  assertFiniteCount(result.durationMs, 'durationMs')\n  assertFiniteCount(result.costUsd, 'costUsd')\n  if (result.costProvenance !== undefined) assertCostProvenance(result.costProvenance)\n}\n\nfunction assertCostProvenance(value: unknown): void {\n  const row = requireRow(value, 'costProvenance')\n  if (row.kind !== 'observed' && row.kind !== 'estimated' && row.kind !== 'uncaptured') {\n    throw new MultishotShotResultError(\n      `costProvenance.kind must be observed, estimated or uncaptured, received ${describeValue(row.kind)}`,\n    )\n  }\n  // The matrix reads the amount from `costUsd`; `usd` only has to agree with\n  // the kind, so an uncaptured provenance cannot smuggle in a total.\n  if (row.kind === 'uncaptured') {\n    if (row.usd !== null) {\n      throw new MultishotShotResultError(\n        `uncaptured costProvenance.usd must be null, received ${describeValue(row.usd)}`,\n      )\n    }\n    return\n  }\n  assertFiniteCount(row.usd, 'costProvenance.usd')\n}\n\nfunction assertMessage(value: unknown, index: number): void {\n  const row = requireRow(value, `transcript[${index}]`)\n  if (typeof row.role !== 'string' || !MULTISHOT_ROLES.has(row.role)) {\n    throw new MultishotShotResultError(\n      `transcript[${index}].role must be user, assistant or tool, received ${describeValue(row.role)}`,\n    )\n  }\n  assertString(row.content, `transcript[${index}].content`)\n  if (row.toolCallId !== undefined) {\n    assertString(row.toolCallId, `transcript[${index}].toolCallId`)\n  }\n  if (row.toolCalls === undefined) return\n  if (!Array.isArray(row.toolCalls)) {\n    throw new MultishotShotResultError(\n      `transcript[${index}].toolCalls must be an array when present, received ${describeValue(row.toolCalls)}`,\n    )\n  }\n  row.toolCalls.forEach((call, callIndex) => {\n    const field = `transcript[${index}].toolCalls[${callIndex}]`\n    const row = requireRow(call, field)\n    assertString(row.id, `${field}.id`)\n    assertString(row.name, `${field}.name`)\n    requireRow(row.args, `${field}.args`)\n  })\n}\n\nfunction assertArtifact(value: unknown, index: number): void {\n  const row = requireRow(value, `artifacts[${index}]`)\n  assertString(row.type, `artifacts[${index}].type`)\n  assertString(row.content, `artifacts[${index}].content`)\n  assertFiniteCount(row.turn, `artifacts[${index}].turn`)\n  const invocation = requireRow(row.invocation, `artifacts[${index}].invocation`)\n  assertString(invocation.name, `artifacts[${index}].invocation.name`)\n  requireRow(invocation.args, `artifacts[${index}].invocation.args`)\n}\n\nfunction requireRow(value: unknown, field: string): Record<string, unknown> {\n  if (typeof value !== 'object' || value === null || Array.isArray(value)) {\n    throw new MultishotShotResultError(\n      `${field} must be an object, received ${describeValue(value)}`,\n    )\n  }\n  return value as Record<string, unknown>\n}\n\nfunction assertString(value: unknown, field: string): void {\n  if (typeof value !== 'string') {\n    throw new MultishotShotResultError(\n      `${field} must be a string, received ${describeValue(value)}`,\n    )\n  }\n}\n\nfunction assertFiniteCount(value: unknown, field: string): void {\n  if (typeof value !== 'number' || !Number.isFinite(value) || value < 0) {\n    throw new MultishotShotResultError(\n      `${field} must be a finite number >= 0, received ${describeValue(value)}`,\n    )\n  }\n}\n\nfunction describeValue(value: unknown): string {\n  if (value === null) return 'null'\n  if (Array.isArray(value)) return 'an array'\n  if (typeof value === 'object') return 'an object'\n  return `${typeof value} ${String(value)}`\n}\n"],"mappings":";AAmHA,IAAa,4BAAb,cAA+C,MAAM;CACvB;CAA5B,YAAY,MAA8B;EACxC,MAAM,0DAA0D,KAAK,gBAAgB;EAD3D,KAAA,OAAA;EAE1B,KAAK,OAAO;CACd;AACF;AAEA,IAAa,0BAAb,cAA6C,MAAM;CACjD,YAAY,SAAiB;EAC3B,MAAM,OAAO;EACb,KAAK,OAAO;CACd;AACF;AAEA,IAAa,2BAAb,cAA8C,MAAM;CAClD,YAAY,QAAgB;EAC1B,MAAM,yDAAyD,QAAQ;EACvE,KAAK,OAAO;CACd;AACF;AAEA,MAAM,kCAAkB,IAAI,IAAI;CAAC;CAAQ;CAAa;AAAM,CAAC;;;;;;;;;;;;;;;;;;AAmB7D,SAAgB,0BAA0B,OAAkD;CAC1F,IAAI,OAAO,UAAU,YAAY,UAAU,MACzC,MAAM,IAAI,yBAAyB,gCAAgC,cAAc,KAAK,GAAG;CAE3F,MAAM,SAAS;CACf,IAAI,CAAC,MAAM,QAAQ,OAAO,UAAU,GAClC,MAAM,IAAI,yBACR,yCAAyC,cAAc,OAAO,UAAU,GAC1E;CAEF,IAAI,CAAC,MAAM,QAAQ,OAAO,SAAS,GACjC,MAAM,IAAI,yBACR,wCAAwC,cAAc,OAAO,SAAS,GACxE;CAEF,OAAO,WAAW,QAAQ,aAAa;CACvC,OAAO,UAAU,QAAQ,cAAc;CACvC,kBAAkB,OAAO,WAAW,WAAW;CAC/C,kBAAkB,OAAO,YAAY,YAAY;CACjD,kBAAkB,OAAO,SAAS,SAAS;CAC3C,IAAI,OAAO,mBAAmB,KAAA,GAAW,qBAAqB,OAAO,cAAc;AACrF;AAEA,SAAS,qBAAqB,OAAsB;CAClD,MAAM,MAAM,WAAW,OAAO,gBAAgB;CAC9C,IAAI,IAAI,SAAS,cAAc,IAAI,SAAS,eAAe,IAAI,SAAS,cACtE,MAAM,IAAI,yBACR,2EAA2E,cAAc,IAAI,IAAI,GACnG;CAIF,IAAI,IAAI,SAAS,cAAc;EAC7B,IAAI,IAAI,QAAQ,MACd,MAAM,IAAI,yBACR,wDAAwD,cAAc,IAAI,GAAG,GAC/E;EAEF;CACF;CACA,kBAAkB,IAAI,KAAK,oBAAoB;AACjD;AAEA,SAAS,cAAc,OAAgB,OAAqB;CAC1D,MAAM,MAAM,WAAW,OAAO,cAAc,MAAM,EAAE;CACpD,IAAI,OAAO,IAAI,SAAS,YAAY,CAAC,gBAAgB,IAAI,IAAI,IAAI,GAC/D,MAAM,IAAI,yBACR,cAAc,MAAM,mDAAmD,cAAc,IAAI,IAAI,GAC/F;CAEF,aAAa,IAAI,SAAS,cAAc,MAAM,UAAU;CACxD,IAAI,IAAI,eAAe,KAAA,GACrB,aAAa,IAAI,YAAY,cAAc,MAAM,aAAa;CAEhE,IAAI,IAAI,cAAc,KAAA,GAAW;CACjC,IAAI,CAAC,MAAM,QAAQ,IAAI,SAAS,GAC9B,MAAM,IAAI,yBACR,cAAc,MAAM,sDAAsD,cAAc,IAAI,SAAS,GACvG;CAEF,IAAI,UAAU,SAAS,MAAM,cAAc;EACzC,MAAM,QAAQ,cAAc,MAAM,cAAc,UAAU;EAC1D,MAAM,MAAM,WAAW,MAAM,KAAK;EAClC,aAAa,IAAI,IAAI,GAAG,MAAM,IAAI;EAClC,aAAa,IAAI,MAAM,GAAG,MAAM,MAAM;EACtC,WAAW,IAAI,MAAM,GAAG,MAAM,MAAM;CACtC,CAAC;AACH;AAEA,SAAS,eAAe,OAAgB,OAAqB;CAC3D,MAAM,MAAM,WAAW,OAAO,aAAa,MAAM,EAAE;CACnD,aAAa,IAAI,MAAM,aAAa,MAAM,OAAO;CACjD,aAAa,IAAI,SAAS,aAAa,MAAM,UAAU;CACvD,kBAAkB,IAAI,MAAM,aAAa,MAAM,OAAO;CACtD,MAAM,aAAa,WAAW,IAAI,YAAY,aAAa,MAAM,aAAa;CAC9E,aAAa,WAAW,MAAM,aAAa,MAAM,kBAAkB;CACnE,WAAW,WAAW,MAAM,aAAa,MAAM,kBAAkB;AACnE;AAEA,SAAS,WAAW,OAAgB,OAAwC;CAC1E,IAAI,OAAO,UAAU,YAAY,UAAU,QAAQ,MAAM,QAAQ,KAAK,GACpE,MAAM,IAAI,yBACR,GAAG,MAAM,+BAA+B,cAAc,KAAK,GAC7D;CAEF,OAAO;AACT;AAEA,SAAS,aAAa,OAAgB,OAAqB;CACzD,IAAI,OAAO,UAAU,UACnB,MAAM,IAAI,yBACR,GAAG,MAAM,8BAA8B,cAAc,KAAK,GAC5D;AAEJ;AAEA,SAAS,kBAAkB,OAAgB,OAAqB;CAC9D,IAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,KAAK,QAAQ,GAClE,MAAM,IAAI,yBACR,GAAG,MAAM,0CAA0C,cAAc,KAAK,GACxE;AAEJ;AAEA,SAAS,cAAc,OAAwB;CAC7C,IAAI,UAAU,MAAM,OAAO;CAC3B,IAAI,MAAM,QAAQ,KAAK,GAAG,OAAO;CACjC,IAAI,OAAO,UAAU,UAAU,OAAO;CACtC,OAAO,GAAG,OAAO,MAAM,GAAG,OAAO,KAAK;AACxC"}