@tangle-network/agent-eval 0.200.0 → 0.201.0

package/dist/matrix/index.d.ts

@@ -1,4 +1,4 @@
 import { p as CostProvenance } from "../cost-ledger-DbQdN3nO.js";
 import { t as DefaultVerdict } from "../verdict-E4eRNf7-.js";
-import { a as MatrixCell, c as CellSpend, i as MatrixAxis, l as readCellSpend, n as AxisSummary, o as MatrixResult, r as CellResult, s as RunAgentMatrixOptions, t as runAgentMatrix, u as withCellSpend } from "../index-DNgf5gyG2.js";
+import { a as MatrixCell, c as CellSpend, i as MatrixAxis, l as readCellSpend, n as AxisSummary, o as MatrixResult, r as CellResult, s as RunAgentMatrixOptions, t as runAgentMatrix, u as withCellSpend } from "../index-DNgf5gyG.js";
 export { type AxisSummary, type CellResult, type CellSpend, type CostProvenance, type DefaultVerdict, type MatrixAxis, type MatrixCell, type MatrixResult, type RunAgentMatrixOptions, readCellSpend, runAgentMatrix, withCellSpend };

package/dist/matrix/index.js

@@ -1,2 +1,2 @@
-import { n as readCellSpend, r as withCellSpend, t as runAgentMatrix } from "../matrix-pWinRMPl.js";
+import { n as readCellSpend, r as withCellSpend, t as runAgentMatrix } from "../matrix-DnlHn5x1.js";
 export { readCellSpend, runAgentMatrix, withCellSpend };

package/dist/{matrix-_NLSWsKM.d.ts → matrix--IZKROxD.d.ts}

@@ -1,6 +1,6 @@
 import { p as CostProvenance } from "./cost-ledger-DbQdN3nO.js";
 import { w as JudgeScore } from "./types-DfvQOe91.js";
-import { o as MatrixResult } from "./index-DNgf5gyG2.js";
+import { o as MatrixResult } from "./index-DNgf5gyG.js";
 import { AgentProfile } from "@tangle-network/agent-interface";
 //#region src/multishot/types.d.ts
 interface MultishotMessage {
@@ -398,4 +398,4 @@ interface RunMultishotMatrixResult {
 declare function runMultishotMatrix<TPersona extends MultishotPersona>(opts: RunMultishotMatrixOptions<TPersona>): Promise<RunMultishotMatrixResult>;
 //#endregion
 export { MultishotToolExecutor as A, MultishotFatalToolError as C, MultishotShape as D, MultishotResult as E, assertMultishotShotResult as F, MultishotTransportRequest as M, MultishotTransportResponse as N, MultishotShotResultError as O, MultishotTransportToolCall as P, MultishotDriverEmptyError as S, MultishotPersona as T, JudgeRunResult as _, MultishotCellOutput as a, runJudge as b, RunMultishotMatrixResult as c, MultishotShot as d, RunMultishotOptions as f, JudgeDimension as g, JudgeConfig as h, ConversationJudgeInput as i, MultishotTransport as j, MultishotToolDefinition as k, computeCellComposite as l, DEFAULT_JUDGE_MODEL as m, CellCompositeInput as n, MultishotJudges as o, runMultishot as p, CellCompositeScore as r, RunMultishotMatrixOptions as s, ArtifactJudgeInput as t, runMultishotMatrix as u, renderDimensions as v, MultishotMessage as w, MultishotArtifact as x, renderJsonFooter as y };
-//# sourceMappingURL=matrix-_NLSWsKM.d.ts.map
+//# sourceMappingURL=matrix--IZKROxD.d.ts.map

package/dist/{matrix-_NLSWsKM.d.ts.map → matrix--IZKROxD.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"matrix-_NLSWsKM.d.ts","names":[],"sources":["../src/multishot/types.ts","../src/multishot/judges.ts","../src/multishot/multishot.ts","../src/multishot/matrix.ts"],"mappings":";;;;;UAIiB;EACf;EACA;EACA;EACA,YAAY;IAAQ;IAAY;IAAc,MAAM;;;UAGrC;EACf;EACA;EACA;IAAc;IAAc,MAAM;;EAClC;;UAGe;EACf,YAAY;EACZ,WAAW;EACX;EACA;;;EAGA;;;;;;;;EAQA,iBAAiB;;UAGF;EACf;EACA;IACE;IACA;IACA,YAAY;;;;;;UAOC;EACf;EACA,UAAU,MAAM;EAChB,QAAQ;EACR;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA;IAAY;IAAc;;;UAGX;EACf;IAAW;IAAyB,aAAa;;EACjD;IAAU;IAAwB;;;;EAGlC;;;;EAIA;;;;;;;;KASU,sBACV,KAAK,8BACF,QAAQ;KAED,yBACV,MAAM,yBACN;;EAEE,WAAW;EACX,SAAS;MAER;EAAU;EAAiB;;UAEf;;EAEf;;GAEC;;;;;;;;UASc,eAAe,iBAAiB;;EAE/C,eAAe,SAAS;;;EAGxB,2BAA2B,SAAS;;cAGzB,kCAAkC;WACjB;EAA5B,YAA4B;;cAMjB,gCAAgC;EAC3C,YAAY;;cAMD,iCAAiC;EAC5C,YAAY;;;;;;;;;;;;;;;;;;;iBAyBE,0BAA0B,yBAAyB,SAAS;;;cCvI/D;UAEI;;EAEf;;EAEA;;UAGe,YAAY;;EAE3B;;;EAGA,WAAW;;EAEX;;EAEA,YAAY;;EAEZ;;;EAGA,cAAc,OAAO;;EAErB;;UAGe;;EAEf,OAAO;;EAEP,MAAM;;iBAGc,SAAS,QAC7B,OAAO,YAAY,SACnB,OAAO,SACN,QAAQ;;;iBAsIK,iBAAiB,eAAe;;iBAKhC,iBAAiB,eAAe;;;UCxK/B,oBAAoB,iBAAiB;EACpD,SAAS;EACT,SAAS;;;EAGT,QAAQ,eAAe;;EAEvB,QAAQ;;EAER,gBAAgB,eAAe;;;;EAI/B,mBAAmB;EACnB;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,gBAAgB;;;;EAIhB,iBAAiB;;;EAGjB,gBAAgB;EAChB,SAAS;;;;;;;;;;;KAYC,cAAc,iBAAiB,qBACzC,MAAM,oBAAoB,cACvB,QAAQ;;;;;;;;;;;;;;;;;iBA2BS,aAAa,iBAAiB,kBAClD,MAAM,oBAAoB,YACzB,QAAQ;;;UClFM,uBAAuB,iBAAiB;EACvD,YAAY;EACZ,SAAS;;UAGM,mBAAmB,iBAAiB;EACnD,UAAU;EACV,SAAS;;UAGM,gBAAgB,iBAAiB;;EAEhD,cAAc,YAAY,uBAAuB;;EAEjD,aAAa,YAAY,mBAAmB;;EAE5C,iBAAiB,YAAY,mBAAmB;;EAEhD;;EAEA;;UAGe;EACf;EACA,cAAc;EACd;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;EAEF;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;;UAIa,0BAA0B,iBAAiB;;EAE1D,UAAU;IAAQ;IAAY,OAAO;;;EAErC,UAAU;;;EAGV,QAAQ,eAAe;;EAEvB,QAAQ,gBAAgB;;EAExB,QAAQ;;EAER,gBAAgB,eAAe;;EAE/B,mBAAmB;;EAEnB;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,gBAAgB;;EAEhB,iBAAiB;;;EAGjB,gBAAgB;;;;;;;;;;;;;;;EAehB,UAAU,cAAc;;;;;UAMT;EACf;EACA;EACA;;UAoBe;EACf,cAAc;;EAEd,cAAc,cAAc;;EAE5B,iBAAiB,cAAc;;;;;;;iBAQjB,qBAAqB,OAAO;EAC1C;EACA;EACA;EACA;;UAuBe;EACf,QAAQ,aAAa;;iBAGD,mBAAmB,iBAAiB,kBACxD,MAAM,0BAA0B,YAC/B,QAAQ"}
+{"version":3,"file":"matrix--IZKROxD.d.ts","names":[],"sources":["../src/multishot/types.ts","../src/multishot/judges.ts","../src/multishot/multishot.ts","../src/multishot/matrix.ts"],"mappings":";;;;;UAIiB;EACf;EACA;EACA;EACA,YAAY;IAAQ;IAAY;IAAc,MAAM;;;UAGrC;EACf;EACA;EACA;IAAc;IAAc,MAAM;;EAClC;;UAGe;EACf,YAAY;EACZ,WAAW;EACX;EACA;;;EAGA;;;;;;;;EAQA,iBAAiB;;UAGF;EACf;EACA;IACE;IACA;IACA,YAAY;;;;;;UAOC;EACf;EACA,UAAU,MAAM;EAChB,QAAQ;EACR;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA;IAAY;IAAc;;;UAGX;EACf;IAAW;IAAyB,aAAa;;EACjD;IAAU;IAAwB;;;;EAGlC;;;;EAIA;;;;;;;;KASU,sBACV,KAAK,8BACF,QAAQ;KAED,yBACV,MAAM,yBACN;;EAEE,WAAW;EACX,SAAS;MAER;EAAU;EAAiB;;UAEf;;EAEf;;GAEC;;;;;;;;UASc,eAAe,iBAAiB;;EAE/C,eAAe,SAAS;;;EAGxB,2BAA2B,SAAS;;cAGzB,kCAAkC;WACjB;EAA5B,YAA4B;;cAMjB,gCAAgC;EAC3C,YAAY;;cAMD,iCAAiC;EAC5C,YAAY;;;;;;;;;;;;;;;;;;;iBAyBE,0BAA0B,yBAAyB,SAAS;;;cCvI/D;UAEI;;EAEf;;EAEA;;UAGe,YAAY;;EAE3B;;;EAGA,WAAW;;EAEX;;EAEA,YAAY;;EAEZ;;;EAGA,cAAc,OAAO;;EAErB;;UAGe;;EAEf,OAAO;;EAEP,MAAM;;iBAGc,SAAS,QAC7B,OAAO,YAAY,SACnB,OAAO,SACN,QAAQ;;;iBAsIK,iBAAiB,eAAe;;iBAKhC,iBAAiB,eAAe;;;UCxK/B,oBAAoB,iBAAiB;EACpD,SAAS;EACT,SAAS;;;EAGT,QAAQ,eAAe;;EAEvB,QAAQ;;EAER,gBAAgB,eAAe;;;;EAI/B,mBAAmB;EACnB;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,gBAAgB;;;;EAIhB,iBAAiB;;;EAGjB,gBAAgB;EAChB,SAAS;;;;;;;;;;;KAYC,cAAc,iBAAiB,qBACzC,MAAM,oBAAoB,cACvB,QAAQ;;;;;;;;;;;;;;;;;iBA2BS,aAAa,iBAAiB,kBAClD,MAAM,oBAAoB,YACzB,QAAQ;;;UClFM,uBAAuB,iBAAiB;EACvD,YAAY;EACZ,SAAS;;UAGM,mBAAmB,iBAAiB;EACnD,UAAU;EACV,SAAS;;UAGM,gBAAgB,iBAAiB;;EAEhD,cAAc,YAAY,uBAAuB;;EAEjD,aAAa,YAAY,mBAAmB;;EAE5C,iBAAiB,YAAY,mBAAmB;;EAEhD;;EAEA;;UAGe;EACf;EACA,cAAc;EACd;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;EAEF;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;;UAIa,0BAA0B,iBAAiB;;EAE1D,UAAU;IAAQ;IAAY,OAAO;;;EAErC,UAAU;;;EAGV,QAAQ,eAAe;;EAEvB,QAAQ,gBAAgB;;EAExB,QAAQ;;EAER,gBAAgB,eAAe;;EAE/B,mBAAmB;;EAEnB;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,gBAAgB;;EAEhB,iBAAiB;;;EAGjB,gBAAgB;;;;;;;;;;;;;;;EAehB,UAAU,cAAc;;;;;UAMT;EACf;EACA;EACA;;UAoBe;EACf,cAAc;;EAEd,cAAc,cAAc;;EAE5B,iBAAiB,cAAc;;;;;;;iBAQjB,qBAAqB,OAAO;EAC1C;EACA;EACA;EACA;;UAuBe;EACf,QAAQ,aAAa;;iBAGD,mBAAmB,iBAAiB,kBACxD,MAAM,0BAA0B,YAC/B,QAAQ"}

package/dist/{matrix-pWinRMPl.js → matrix-DnlHn5x1.js}

@@ -368,11 +368,13 @@ async function runAgentMatrix(opts) {
 		}
 		if (inFlight === 0) resolveAll?.();
 	};
-	if (opts.signal) if (opts.signal.aborted) {
-		cellsUnscheduled = planned.length;
-		cursor = planned.length;
-		resolveAll?.();
-	} else opts.signal.addEventListener("abort", onAbort, { once: true });
+	if (opts.signal) {
+		if (opts.signal.aborted) {
+			cellsUnscheduled = planned.length;
+			cursor = planned.length;
+			resolveAll?.();
+		} else opts.signal.addEventListener("abort", onAbort, { once: true });
+	}
 	if (planned.length === 0) resolveAll?.();
 	else pump();
 	await done;
@@ -409,4 +411,4 @@ async function runAgentMatrix(opts) {
 //#endregion
 export { readCellSpend as n, withCellSpend as r, runAgentMatrix as t };
 
-//# sourceMappingURL=matrix-pWinRMPl.js.map
+//# sourceMappingURL=matrix-DnlHn5x1.js.map

package/dist/{matrix-pWinRMPl.js.map → matrix-DnlHn5x1.js.map}

@@ -1 +1 @@
-{"version":3,"file":"matrix-pWinRMPl.js","names":[],"sources":["../src/matrix/cell-spend.ts","../src/matrix/aggregation.ts","../src/matrix/runner.ts"],"sourcesContent":["/**\n * Spend carrier for a cell that fails after it has already spent money.\n *\n * `runCell` reports what a cell cost by returning a `CellResult`. A throw\n * carries no result, so the runner cannot know what the cell spent before it\n * failed: the cumulative sum the cost ceiling reads, and `totalCostUsd`, both\n * miss that money. A throw that carries a `CellSpend` closes the gap — the\n * runner bills the failed cell for the spend the carrier declares, and records\n * a cell that declared nothing as `uncaptured` instead of a zero that reads\n * like \"this cell spent nothing\".\n *\n * The carrier is read structurally, never by `instanceof`, so a throw that\n * crosses a package or realm boundary still bills.\n */\n\n/** Property key the carrier occupies on a thrown object. Stable across\n *  versions — two copies of this package must agree on it. */\nconst CELL_SPEND_KEY = '__agentEvalCellSpend'\n\n/** Money and wall time a cell consumed before it threw. */\nexport interface CellSpend {\n  /** KNOWN subtotal in USD. When `kind` is `uncaptured` the cell spent this\n   *  much AND an unknown amount more, so the value must not be read as the\n   *  cell's total. */\n  costUsd: number\n  /** Wall time the cell consumed before the throw, in milliseconds. */\n  durationMs: number\n  /** `observed` = provider-reported amounts, `estimated` = computed from token\n   *  prices, `uncaptured` = `costUsd` is a subtotal and part of the spend\n   *  could not be measured. */\n  kind: 'observed' | 'estimated' | 'uncaptured'\n}\n\n/**\n * Return `error` carrying `spend`, for a throw site inside `runCell`:\n * `throw withCellSpend(err, { costUsd, durationMs, kind })`.\n *\n * A value that cannot hold the carrier — a thrown string or number, or a\n * frozen, sealed, or otherwise non-extensible object — is wrapped in an\n * `Error` that keeps the original as `cause`, so the spend is never dropped.\n * A later call overwrites an earlier carrier — the outer frame knows more\n * than the inner one.\n *\n * Throws `TypeError` on a non-finite or negative amount. A poisoned amount\n * would reach `cumulativeCost` and disable the cost ceiling for the whole run,\n * because `NaN >= ceiling` is false.\n */\nexport function withCellSpend(error: unknown, spend: CellSpend): unknown {\n  assertAmount(spend.costUsd, 'costUsd')\n  assertAmount(spend.durationMs, 'durationMs')\n  if (spend.kind !== 'observed' && spend.kind !== 'estimated' && spend.kind !== 'uncaptured') {\n    throw new TypeError(\n      `withCellSpend: kind must be observed, estimated or uncaptured, received ${String(spend.kind)}`,\n    )\n  }\n  const carrier: CellSpend = {\n    costUsd: spend.costUsd,\n    durationMs: spend.durationMs,\n    kind: spend.kind,\n  }\n  if (typeof error === 'object' && error !== null && attach(error, carrier)) return error\n  // A primitive cannot hold a property, and a frozen or sealed object refuses\n  // one. Carry the spend on a fresh Error that keeps the original as `cause`\n  // rather than lose it — a dropped carrier bills the cell as uncaptured,\n  // which is the under-billing this module exists to prevent.\n  const wrapper = new Error(`cell failed: ${messageOf(error)}`, { cause: error })\n  attach(wrapper, carrier)\n  return wrapper\n}\n\n/** True when the carrier landed. `defineProperty` throws on a non-extensible\n *  target, so the caller must be able to fall back. */\nfunction attach(target: object, carrier: CellSpend): boolean {\n  try {\n    Object.defineProperty(target, CELL_SPEND_KEY, {\n      value: carrier,\n      enumerable: false,\n      configurable: true,\n      writable: true,\n    })\n    return true\n  } catch {\n    return false\n  }\n}\n\nfunction messageOf(error: unknown): string {\n  if (error instanceof Error) return error.message\n  return String(error)\n}\n\n/** Read the spend a thrown value carries. `undefined` when the value carries\n *  none, or carries one that does not satisfy `CellSpend` — the runner records\n *  both as `uncaptured`. */\nexport function readCellSpend(error: unknown): CellSpend | undefined {\n  if (typeof error !== 'object' || error === null) return undefined\n  const raw = (error as Record<string, unknown>)[CELL_SPEND_KEY]\n  if (typeof raw !== 'object' || raw === null) return undefined\n  const candidate = raw as Record<string, unknown>\n  if (!isAmount(candidate.costUsd) || !isAmount(candidate.durationMs)) return undefined\n  const kind = candidate.kind\n  if (kind !== 'observed' && kind !== 'estimated' && kind !== 'uncaptured') return undefined\n  return { costUsd: candidate.costUsd, durationMs: candidate.durationMs, kind }\n}\n\n/** A number the matrix can add to a running cost or duration: finite and not\n *  negative. Anything else would poison a sum the cost ceiling reads. */\nexport function isAmount(value: unknown): value is number {\n  return typeof value === 'number' && Number.isFinite(value) && value >= 0\n}\n\nfunction assertAmount(value: number, field: string): void {\n  if (!isAmount(value)) {\n    throw new TypeError(\n      `withCellSpend: ${field} must be a finite number >= 0, received ${String(value)}`,\n    )\n  }\n}\n","/**\n * Per-axis aggregation of cell runs into `AxisSummary` rows.\n *\n * Pure: consumes the final `cells: [{cell, runs}]` array and returns the\n * `byAxis` table. Error runs contribute 0 to passRate and meanScore. Cost and\n * duration always count — the budget was spent regardless. A run whose cost is\n * a subtotal rather than a total raises `costUncapturedCells` for its bucket,\n * so a per-axis cost row says when it under-counts.\n */\n\nimport type { AxisSummary, CellResult, MatrixAxis, MatrixCell, MatrixResult } from './types'\n\ninterface Row<Output> {\n  cell: MatrixCell\n  result: CellResult<Output>\n}\n\nfunction flattenRuns<Output>(cells: MatrixResult<Output>['cells']): Row<Output>[] {\n  const rows: Row<Output>[] = []\n  for (const { cell, runs } of cells) {\n    for (const result of runs) rows.push({ cell, result })\n  }\n  return rows\n}\n\nfunction quantile(sorted: number[], q: number): number {\n  if (sorted.length === 0) return 0\n  if (sorted.length === 1) return sorted[0] as number\n  const pos = (sorted.length - 1) * q\n  const lo = Math.floor(pos)\n  const hi = Math.ceil(pos)\n  if (lo === hi) return sorted[lo] as number\n  const frac = pos - lo\n  return (sorted[lo] as number) * (1 - frac) + (sorted[hi] as number) * frac\n}\n\nexport function summariseRows<Output>(\n  rows: Row<Output>[],\n  axisName: string,\n  axisValue: string,\n): AxisSummary {\n  if (rows.length === 0) {\n    return {\n      axisName,\n      axisValue,\n      cells: 0,\n      passRate: 0,\n      meanScore: 0,\n      p50Score: 0,\n      p90Score: 0,\n      totalCostUsd: 0,\n      costUncapturedCells: 0,\n      meanDurationMs: 0,\n    }\n  }\n  let pass = 0\n  let scoreSum = 0\n  let costSum = 0\n  let durSum = 0\n  let costUncapturedCells = 0\n  const scores: number[] = []\n  for (const { result } of rows) {\n    const errored = result.error !== undefined\n    const score = errored ? 0 : result.verdict.score\n    const valid = !errored && result.verdict.valid\n    if (valid) pass++\n    scoreSum += score\n    scores.push(score)\n    costSum += result.costUsd\n    durSum += result.durationMs\n    if (result.costProvenance?.kind === 'uncaptured') costUncapturedCells++\n  }\n  scores.sort((a, b) => a - b)\n  return {\n    axisName,\n    axisValue,\n    cells: rows.length,\n    passRate: pass / rows.length,\n    meanScore: scoreSum / rows.length,\n    p50Score: quantile(scores, 0.5),\n    p90Score: quantile(scores, 0.9),\n    totalCostUsd: costSum,\n    costUncapturedCells,\n    meanDurationMs: durSum / rows.length,\n  }\n}\n\nfunction bucketBy<Output>(\n  rows: Row<Output>[],\n  axisName: string,\n  labelFor: (id: string) => string,\n): Record<string, AxisSummary> {\n  const buckets = new Map<string, Row<Output>[]>()\n  for (const row of rows) {\n    const slot = row.cell.axes[axisName]\n    if (!slot) continue\n    const id = slot.id\n    let arr = buckets.get(id)\n    if (!arr) {\n      arr = []\n      buckets.set(id, arr)\n    }\n    arr.push(row)\n  }\n  const out: Record<string, AxisSummary> = {}\n  // Sorted keys for deterministic JSON serialisation.\n  for (const id of [...buckets.keys()].sort()) {\n    out[id] = summariseRows(buckets.get(id) as Row<Output>[], axisName, labelFor(id))\n  }\n  return out\n}\n\nexport function buildByAxis<Output>(\n  cells: MatrixResult<Output>['cells'],\n  axes: MatrixAxis<unknown>[],\n  aggregateBy: string[],\n): Record<string, Record<string, AxisSummary>> {\n  const rows = flattenRuns(cells)\n  const byName = new Map(axes.map((a) => [a.name, a]))\n  const byAxis: Record<string, Record<string, AxisSummary>> = {}\n  for (const name of aggregateBy) {\n    const axis = byName.get(name)\n    const labelFor = (id: string): string => {\n      if (!axis?.label) return id\n      const found = axis.values.find((v) => v.id === id)\n      if (!found) return id\n      return axis.label(found.value, id)\n    }\n    byAxis[name] = bucketBy(rows, name, labelFor)\n  }\n  return byAxis\n}\n","/**\n * N-axis cartesian runner.\n *\n * Expansion order: cartesian over `axes` in declared order, then `reps` as the\n * inner-most dim → `ordinal = (cartIdx * reps) + rep`. The returned\n * `cells[]` is sorted by `ordinal` so concurrent execution does not reorder\n * the output.\n *\n * Scheduling is a sliding window of in-flight promises capped at\n * `maxConcurrency`. The window stops admitting new cells when the cost\n * ceiling trips or the abort signal fires; in-flight cells finish.\n */\n\nimport { buildByAxis } from './aggregation'\nimport { isAmount, readCellSpend } from './cell-spend'\nimport type {\n  CellResult,\n  CostProvenance,\n  MatrixAxis,\n  MatrixCell,\n  MatrixResult,\n  RunAgentMatrixOptions,\n} from './types'\n\ninterface BaseCell {\n  axes: Record<string, { id: string; value: unknown }>\n}\n\nfunction cartesian(axes: MatrixAxis<unknown>[]): BaseCell[] {\n  // Empty axes (`values=[]`) collapse the whole product to zero cells. An\n  // empty `axes` array yields a single empty-axes cell — degenerate but\n  // valid (caller is iterating only reps).\n  if (axes.length === 0) return [{ axes: {} }]\n  for (const a of axes) if (a.values.length === 0) return []\n  const out: BaseCell[] = []\n  const idx = new Array(axes.length).fill(0)\n  while (true) {\n    const slot: Record<string, { id: string; value: unknown }> = {}\n    for (let i = 0; i < axes.length; i++) {\n      const axis = axes[i] as MatrixAxis<unknown>\n      const v = axis.values[idx[i] as number] as { id: string; value: unknown }\n      slot[axis.name] = { id: v.id, value: v.value }\n    }\n    out.push({ axes: slot })\n    // Increment like an odometer, left-most axis is fastest.\n    let i = 0\n    while (i < axes.length) {\n      const next = (idx[i] as number) + 1\n      const axis = axes[i] as MatrixAxis<unknown>\n      if (next < axis.values.length) {\n        idx[i] = next\n        break\n      }\n      idx[i] = 0\n      i++\n    }\n    if (i === axes.length) break\n  }\n  return out\n}\n\nfunction makeMatrixId(): string {\n  // Stable id-like string: time + 8 random hex chars. Avoids node:crypto\n  // import to keep the matrix dep-free.\n  const t = Date.now().toString(36)\n  let r = ''\n  for (let i = 0; i < 8; i++) r += Math.floor(Math.random() * 16).toString(16)\n  return `mtx_${t}_${r}`\n}\n\n/** Fresh per result — a shared object would let one consumer's mutation reach\n *  every other cell in the run. */\nfunction uncaptured(): CostProvenance {\n  return { kind: 'uncaptured', usd: null }\n}\n\nfunction makeErrorResult<Output>(err: unknown): CellResult<Output> {\n  const e = err as { message?: string; name?: string }\n  const spend = readCellSpend(err)\n  return {\n    output: undefined as unknown as Output,\n    verdict: { valid: false, score: 0 },\n    costUsd: spend?.costUsd ?? 0,\n    durationMs: spend?.durationMs ?? 0,\n    costProvenance:\n      spend === undefined || spend.kind === 'uncaptured'\n        ? uncaptured()\n        : { kind: spend.kind, usd: spend.costUsd },\n    error: {\n      message: typeof e?.message === 'string' ? e.message : String(err),\n      kind: typeof e?.name === 'string' ? e.name : 'Error',\n    },\n  }\n}\n\n/** Make a result safe to sum.\n *\n *  A `costUsd` that cannot be summed fails the cell: `NaN` in `cumulativeCost`\n *  makes `>= costCeiling` false forever, so one bad cell would disable the\n *  ceiling for every cell after it.\n *\n *  A bad `durationMs` reaches only `meanDurationMs`, never the ceiling, so it\n *  cannot justify throwing away a real verdict and a real cost. It is reported\n *  as 0 and warned about, and the rest of the result stands. A wall clock that\n *  steps back mid-cell makes `Date.now() - startedAt` negative, so this is a\n *  reachable state, not only a caller bug. */\nfunction billableResult<Output>(\n  result: CellResult<Output>,\n  onBadDuration: (reported: unknown) => void,\n): CellResult<Output> {\n  if (!isAmount(result.costUsd)) {\n    return makeErrorResult<Output>(\n      new RangeError(\n        `runCell reported costUsd=${String(result.costUsd)}; a cell cost must be a finite number >= 0 so the cost ceiling stays enforceable`,\n      ),\n    )\n  }\n  if (isAmount(result.durationMs)) return result\n  onBadDuration(result.durationMs)\n  return { ...result, durationMs: 0 }\n}\n\nexport async function runAgentMatrix<Output>(\n  opts: RunAgentMatrixOptions<Output>,\n): Promise<MatrixResult<Output>> {\n  const startedAt = Date.now()\n  const reps = Math.max(1, opts.reps ?? 1)\n  const maxConcurrency = Math.max(1, opts.maxConcurrency ?? 4)\n  const costCeiling = opts.costCeiling ?? Number.POSITIVE_INFINITY\n  const aggregateBy = opts.aggregateBy ?? opts.axes.map((a) => a.name)\n  if (opts.maxCellCostUsd !== undefined && !isAmount(opts.maxCellCostUsd)) {\n    throw new RangeError(\n      `runAgentMatrix: maxCellCostUsd must be a finite number >= 0, received ${String(opts.maxCellCostUsd)}`,\n    )\n  }\n\n  const base = cartesian(opts.axes)\n  const filtered = opts.filter\n    ? base.filter((c) => (opts.filter as (b: BaseCell) => boolean)(c))\n    : base\n  const filteredOut = base.length - filtered.length\n\n  const planned: MatrixCell[] = []\n  for (let i = 0; i < filtered.length; i++) {\n    for (let r = 0; r < reps; r++) {\n      planned.push({\n        axes: (filtered[i] as BaseCell).axes,\n        rep: r,\n        ordinal: i * reps + r,\n      })\n    }\n  }\n\n  const cellRecords: Array<{ cell: MatrixCell; runs: CellResult<Output>[] }> = []\n  // What the ceiling reads. It runs ahead of the reported total whenever a\n  // cell's cost is a subtotal and `maxCellCostUsd` bounds what it could have\n  // been — a budget must assume the worst about spend it cannot see.\n  let cumulativeCost = 0\n  let costCeilingReached = false\n  let runsExecuted = 0\n  let cellsUnscheduled = 0\n  let costUncapturedCells = 0\n  let badDurationCells = 0\n\n  const aborted = (): boolean => opts.signal?.aborted === true\n\n  // Per-run abort controller forwards the external signal so cell executors\n  // see cancellation. We don't expose it on `MatrixCell` — the signature on\n  // `runCell` per the public API is `(cell) => Promise<...>`. Executors that\n  // need cancellation use the external signal directly via closure.\n\n  let inFlight = 0\n  let cursor = 0\n  let resolveAll: (() => void) | undefined\n  const done = new Promise<void>((res) => {\n    resolveAll = res\n  })\n\n  const pump = (): void => {\n    while (inFlight < maxConcurrency && cursor < planned.length) {\n      if (aborted() || costCeilingReached) {\n        // Drain remaining as unscheduled.\n        const left = planned.length - cursor\n        cellsUnscheduled += left\n        cursor = planned.length\n        break\n      }\n      const cell = planned[cursor++] as MatrixCell\n      inFlight++\n      // Lazily allocate the record so cells appear in `cells[]` in any\n      // arrival order; we sort by ordinal at the end.\n      const record = { cell, runs: [] as CellResult<Output>[] }\n      cellRecords.push(record)\n      const promise: Promise<CellResult<Output>> = (async () => {\n        try {\n          return await opts.runCell(cell)\n        } catch (err) {\n          return makeErrorResult<Output>(err)\n        }\n      })()\n      promise.then((settled) => {\n        const result = billableResult(settled, (reported) => {\n          badDurationCells++\n          if (badDurationCells === 1) {\n            // eslint-disable-next-line no-console\n            console.warn(\n              `[matrix] a cell reported durationMs=${String(reported)}; recorded as 0 — meanDurationMs under-reports this run`,\n            )\n          }\n        })\n        record.runs.push(result)\n        runsExecuted++\n        const uncapturedCost = result.costProvenance?.kind === 'uncaptured'\n        cumulativeCost +=\n          uncapturedCost && opts.maxCellCostUsd !== undefined\n            ? Math.max(result.costUsd, opts.maxCellCostUsd)\n            : result.costUsd\n        if (uncapturedCost) {\n          costUncapturedCells++\n          if (costUncapturedCells === 1) {\n            // eslint-disable-next-line no-console\n            console.warn(\n              \"[matrix] a cell's cost is a subtotal, not a total — totalCostUsd and the cost ceiling under-count this run\",\n            )\n          }\n        }\n        if (cumulativeCost >= costCeiling && !costCeilingReached) {\n          costCeilingReached = true\n          // eslint-disable-next-line no-console\n          console.warn('[matrix] cost ceiling reached')\n        }\n        try {\n          opts.onCellComplete?.(cell, result)\n        } catch {\n          // onCellComplete is observational — swallow throws so a noisy\n          // callback can't tank the run.\n        }\n        inFlight--\n        if (cursor < planned.length) {\n          pump()\n        } else if (inFlight === 0) {\n          resolveAll?.()\n        }\n      })\n    }\n    if (cursor >= planned.length && inFlight === 0) resolveAll?.()\n  }\n\n  const onAbort = (): void => {\n    // External abort: stop scheduling. In-flight cells finish; their\n    // executors observe `opts.signal.aborted` directly via closure.\n    if (cursor < planned.length) {\n      cellsUnscheduled += planned.length - cursor\n      cursor = planned.length\n    }\n    if (inFlight === 0) resolveAll?.()\n  }\n  if (opts.signal) {\n    if (opts.signal.aborted) {\n      cellsUnscheduled = planned.length\n      cursor = planned.length\n      resolveAll?.()\n    } else {\n      opts.signal.addEventListener('abort', onAbort, { once: true })\n    }\n  }\n\n  if (planned.length === 0) {\n    resolveAll?.()\n  } else {\n    pump()\n  }\n\n  await done\n  if (opts.signal) opts.signal.removeEventListener('abort', onAbort)\n\n  cellRecords.sort((a, b) => a.cell.ordinal - b.cell.ordinal)\n\n  let pass = 0\n  let scoreSum = 0\n  let totalCost = 0\n  let runCount = 0\n  for (const { runs } of cellRecords) {\n    for (const r of runs) {\n      runCount++\n      const errored = r.error !== undefined\n      if (!errored && r.verdict.valid) pass++\n      scoreSum += errored ? 0 : r.verdict.score\n      totalCost += r.costUsd\n    }\n  }\n\n  const byAxis = buildByAxis(cellRecords, opts.axes, aggregateBy)\n\n  return {\n    cells: cellRecords,\n    byAxis,\n    summary: {\n      totalCells: planned.length,\n      runsExecuted,\n      cellsSkipped: cellsUnscheduled + filteredOut * reps,\n      overallPassRate: runCount === 0 ? 0 : pass / runCount,\n      overallMeanScore: runCount === 0 ? 0 : scoreSum / runCount,\n      totalCostUsd: totalCost,\n      costUncapturedCells,\n      ceilingChargedUsd: cumulativeCost,\n      durationMs: Date.now() - startedAt,\n    },\n    matrixId: makeMatrixId(),\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAiBA,MAAM,iBAAiB;;;;;;;;;;;;;;;AA8BvB,SAAgB,cAAc,OAAgB,OAA2B;CACvE,aAAa,MAAM,SAAS,SAAS;CACrC,aAAa,MAAM,YAAY,YAAY;CAC3C,IAAI,MAAM,SAAS,cAAc,MAAM,SAAS,eAAe,MAAM,SAAS,cAC5E,MAAM,IAAI,UACR,2EAA2E,OAAO,MAAM,IAAI,GAC9F;CAEF,MAAM,UAAqB;EACzB,SAAS,MAAM;EACf,YAAY,MAAM;EAClB,MAAM,MAAM;CACd;CACA,IAAI,OAAO,UAAU,YAAY,UAAU,QAAQ,OAAO,OAAO,OAAO,GAAG,OAAO;CAKlF,MAAM,UAAU,IAAI,MAAM,gBAAgB,UAAU,KAAK,KAAK,EAAE,OAAO,MAAM,CAAC;CAC9E,OAAO,SAAS,OAAO;CACvB,OAAO;AACT;;;AAIA,SAAS,OAAO,QAAgB,SAA6B;CAC3D,IAAI;EACF,OAAO,eAAe,QAAQ,gBAAgB;GAC5C,OAAO;GACP,YAAY;GACZ,cAAc;GACd,UAAU;EACZ,CAAC;EACD,OAAO;CACT,QAAQ;EACN,OAAO;CACT;AACF;AAEA,SAAS,UAAU,OAAwB;CACzC,IAAI,iBAAiB,OAAO,OAAO,MAAM;CACzC,OAAO,OAAO,KAAK;AACrB;;;;AAKA,SAAgB,cAAc,OAAuC;CACnE,IAAI,OAAO,UAAU,YAAY,UAAU,MAAM,OAAO,KAAA;CACxD,MAAM,MAAO,MAAkC;CAC/C,IAAI,OAAO,QAAQ,YAAY,QAAQ,MAAM,OAAO,KAAA;CACpD,MAAM,YAAY;CAClB,IAAI,CAAC,SAAS,UAAU,OAAO,KAAK,CAAC,SAAS,UAAU,UAAU,GAAG,OAAO,KAAA;CAC5E,MAAM,OAAO,UAAU;CACvB,IAAI,SAAS,cAAc,SAAS,eAAe,SAAS,cAAc,OAAO,KAAA;CACjF,OAAO;EAAE,SAAS,UAAU;EAAS,YAAY,UAAU;EAAY;CAAK;AAC9E;;;AAIA,SAAgB,SAAS,OAAiC;CACxD,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,KAAK,SAAS;AACzE;AAEA,SAAS,aAAa,OAAe,OAAqB;CACxD,IAAI,CAAC,SAAS,KAAK,GACjB,MAAM,IAAI,UACR,kBAAkB,MAAM,0CAA0C,OAAO,KAAK,GAChF;AAEJ;;;ACpGA,SAAS,YAAoB,OAAqD;CAChF,MAAM,OAAsB,CAAC;CAC7B,KAAK,MAAM,EAAE,MAAM,UAAU,OAC3B,KAAK,MAAM,UAAU,MAAM,KAAK,KAAK;EAAE;EAAM;CAAO,CAAC;CAEvD,OAAO;AACT;AAEA,SAAS,SAAS,QAAkB,GAAmB;CACrD,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,IAAI,OAAO,WAAW,GAAG,OAAO,OAAO;CACvC,MAAM,OAAO,OAAO,SAAS,KAAK;CAClC,MAAM,KAAK,KAAK,MAAM,GAAG;CACzB,MAAM,KAAK,KAAK,KAAK,GAAG;CACxB,IAAI,OAAO,IAAI,OAAO,OAAO;CAC7B,MAAM,OAAO,MAAM;CACnB,OAAQ,OAAO,OAAkB,IAAI,QAAS,OAAO,MAAiB;AACxE;AAEA,SAAgB,cACd,MACA,UACA,WACa;CACb,IAAI,KAAK,WAAW,GAClB,OAAO;EACL;EACA;EACA,OAAO;EACP,UAAU;EACV,WAAW;EACX,UAAU;EACV,UAAU;EACV,cAAc;EACd,qBAAqB;EACrB,gBAAgB;CAClB;CAEF,IAAI,OAAO;CACX,IAAI,WAAW;CACf,IAAI,UAAU;CACd,IAAI,SAAS;CACb,IAAI,sBAAsB;CAC1B,MAAM,SAAmB,CAAC;CAC1B,KAAK,MAAM,EAAE,YAAY,MAAM;EAC7B,MAAM,UAAU,OAAO,UAAU,KAAA;EACjC,MAAM,QAAQ,UAAU,IAAI,OAAO,QAAQ;EAE3C,IADc,CAAC,WAAW,OAAO,QAAQ,OAC9B;EACX,YAAY;EACZ,OAAO,KAAK,KAAK;EACjB,WAAW,OAAO;EAClB,UAAU,OAAO;EACjB,IAAI,OAAO,gBAAgB,SAAS,cAAc;CACpD;CACA,OAAO,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3B,OAAO;EACL;EACA;EACA,OAAO,KAAK;EACZ,UAAU,OAAO,KAAK;EACtB,WAAW,WAAW,KAAK;EAC3B,UAAU,SAAS,QAAQ,EAAG;EAC9B,UAAU,SAAS,QAAQ,EAAG;EAC9B,cAAc;EACd;EACA,gBAAgB,SAAS,KAAK;CAChC;AACF;AAEA,SAAS,SACP,MACA,UACA,UAC6B;CAC7B,MAAM,0BAAU,IAAI,IAA2B;CAC/C,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,OAAO,IAAI,KAAK,KAAK;EAC3B,IAAI,CAAC,MAAM;EACX,MAAM,KAAK,KAAK;EAChB,IAAI,MAAM,QAAQ,IAAI,EAAE;EACxB,IAAI,CAAC,KAAK;GACR,MAAM,CAAC;GACP,QAAQ,IAAI,IAAI,GAAG;EACrB;EACA,IAAI,KAAK,GAAG;CACd;CACA,MAAM,MAAmC,CAAC;CAE1C,KAAK,MAAM,MAAM,CAAC,GAAG,QAAQ,KAAK,CAAC,CAAC,CAAC,KAAK,GACxC,IAAI,MAAM,cAAc,QAAQ,IAAI,EAAE,GAAoB,UAAU,SAAS,EAAE,CAAC;CAElF,OAAO;AACT;AAEA,SAAgB,YACd,OACA,MACA,aAC6C;CAC7C,MAAM,OAAO,YAAY,KAAK;CAC9B,MAAM,SAAS,IAAI,IAAI,KAAK,KAAK,MAAM,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC;CACnD,MAAM,SAAsD,CAAC;CAC7D,KAAK,MAAM,QAAQ,aAAa;EAC9B,MAAM,OAAO,OAAO,IAAI,IAAI;EAC5B,MAAM,YAAY,OAAuB;GACvC,IAAI,CAAC,MAAM,OAAO,OAAO;GACzB,MAAM,QAAQ,KAAK,OAAO,MAAM,MAAM,EAAE,OAAO,EAAE;GACjD,IAAI,CAAC,OAAO,OAAO;GACnB,OAAO,KAAK,MAAM,MAAM,OAAO,EAAE;EACnC;EACA,OAAO,QAAQ,SAAS,MAAM,MAAM,QAAQ;CAC9C;CACA,OAAO;AACT;;;;;;;;;;;;;;;ACvGA,SAAS,UAAU,MAAyC;CAI1D,IAAI,KAAK,WAAW,GAAG,OAAO,CAAC,EAAE,MAAM,CAAC,EAAE,CAAC;CAC3C,KAAK,MAAM,KAAK,MAAM,IAAI,EAAE,OAAO,WAAW,GAAG,OAAO,CAAC;CACzD,MAAM,MAAkB,CAAC;CACzB,MAAM,MAAM,IAAI,MAAM,KAAK,MAAM,CAAC,CAAC,KAAK,CAAC;CACzC,OAAO,MAAM;EACX,MAAM,OAAuD,CAAC;EAC9D,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;GACpC,MAAM,OAAO,KAAK;GAClB,MAAM,IAAI,KAAK,OAAO,IAAI;GAC1B,KAAK,KAAK,QAAQ;IAAE,IAAI,EAAE;IAAI,OAAO,EAAE;GAAM;EAC/C;EACA,IAAI,KAAK,EAAE,MAAM,KAAK,CAAC;EAEvB,IAAI,IAAI;EACR,OAAO,IAAI,KAAK,QAAQ;GACtB,MAAM,OAAQ,IAAI,KAAgB;GAElC,IAAI,OADS,KAAK,EACH,CAAC,OAAO,QAAQ;IAC7B,IAAI,KAAK;IACT;GACF;GACA,IAAI,KAAK;GACT;EACF;EACA,IAAI,MAAM,KAAK,QAAQ;CACzB;CACA,OAAO;AACT;AAEA,SAAS,eAAuB;CAG9B,MAAM,IAAI,KAAK,IAAI,CAAC,CAAC,SAAS,EAAE;CAChC,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,KAAK,KAAK,MAAM,KAAK,OAAO,IAAI,EAAE,CAAC,CAAC,SAAS,EAAE;CAC3E,OAAO,OAAO,EAAE,GAAG;AACrB;;;AAIA,SAAS,aAA6B;CACpC,OAAO;EAAE,MAAM;EAAc,KAAK;CAAK;AACzC;AAEA,SAAS,gBAAwB,KAAkC;CACjE,MAAM,IAAI;CACV,MAAM,QAAQ,cAAc,GAAG;CAC/B,OAAO;EACL,QAAQ,KAAA;EACR,SAAS;GAAE,OAAO;GAAO,OAAO;EAAE;EAClC,SAAS,OAAO,WAAW;EAC3B,YAAY,OAAO,cAAc;EACjC,gBACE,UAAU,KAAA,KAAa,MAAM,SAAS,eAClC,WAAW,IACX;GAAE,MAAM,MAAM;GAAM,KAAK,MAAM;EAAQ;EAC7C,OAAO;GACL,SAAS,OAAO,GAAG,YAAY,WAAW,EAAE,UAAU,OAAO,GAAG;GAChE,MAAM,OAAO,GAAG,SAAS,WAAW,EAAE,OAAO;EAC/C;CACF;AACF;;;;;;;;;;;;AAaA,SAAS,eACP,QACA,eACoB;CACpB,IAAI,CAAC,SAAS,OAAO,OAAO,GAC1B,OAAO,gCACL,IAAI,WACF,4BAA4B,OAAO,OAAO,OAAO,EAAE,iFACrD,CACF;CAEF,IAAI,SAAS,OAAO,UAAU,GAAG,OAAO;CACxC,cAAc,OAAO,UAAU;CAC/B,OAAO;EAAE,GAAG;EAAQ,YAAY;CAAE;AACpC;AAEA,eAAsB,eACpB,MAC+B;CAC/B,MAAM,YAAY,KAAK,IAAI;CAC3B,MAAM,OAAO,KAAK,IAAI,GAAG,KAAK,QAAQ,CAAC;CACvC,MAAM,iBAAiB,KAAK,IAAI,GAAG,KAAK,kBAAkB,CAAC;CAC3D,MAAM,cAAc,KAAK,eAAe,OAAO;CAC/C,MAAM,cAAc,KAAK,eAAe,KAAK,KAAK,KAAK,MAAM,EAAE,IAAI;CACnE,IAAI,KAAK,mBAAmB,KAAA,KAAa,CAAC,SAAS,KAAK,cAAc,GACpE,MAAM,IAAI,WACR,yEAAyE,OAAO,KAAK,cAAc,GACrG;CAGF,MAAM,OAAO,UAAU,KAAK,IAAI;CAChC,MAAM,WAAW,KAAK,SAClB,KAAK,QAAQ,MAAO,KAAK,OAAoC,CAAC,CAAC,IAC/D;CACJ,MAAM,cAAc,KAAK,SAAS,SAAS;CAE3C,MAAM,UAAwB,CAAC;CAC/B,KAAK,IAAI,IAAI,GAAG,IAAI,SAAS,QAAQ,KACnC,KAAK,IAAI,IAAI,GAAG,IAAI,MAAM,KACxB,QAAQ,KAAK;EACX,MAAO,SAAS,EAAE,CAAc;EAChC,KAAK;EACL,SAAS,IAAI,OAAO;CACtB,CAAC;CAIL,MAAM,cAAuE,CAAC;CAI9E,IAAI,iBAAiB;CACrB,IAAI,qBAAqB;CACzB,IAAI,eAAe;CACnB,IAAI,mBAAmB;CACvB,IAAI,sBAAsB;CAC1B,IAAI,mBAAmB;CAEvB,MAAM,gBAAyB,KAAK,QAAQ,YAAY;CAOxD,IAAI,WAAW;CACf,IAAI,SAAS;CACb,IAAI;CACJ,MAAM,OAAO,IAAI,SAAe,QAAQ;EACtC,aAAa;CACf,CAAC;CAED,MAAM,aAAmB;EACvB,OAAO,WAAW,kBAAkB,SAAS,QAAQ,QAAQ;GAC3D,IAAI,QAAQ,KAAK,oBAAoB;IAEnC,MAAM,OAAO,QAAQ,SAAS;IAC9B,oBAAoB;IACpB,SAAS,QAAQ;IACjB;GACF;GACA,MAAM,OAAO,QAAQ;GACrB;GAGA,MAAM,SAAS;IAAE;IAAM,MAAM,CAAC;GAA0B;GACxD,YAAY,KAAK,MAAM;GAQvB,CAP8C,YAAY;IACxD,IAAI;KACF,OAAO,MAAM,KAAK,QAAQ,IAAI;IAChC,SAAS,KAAK;KACZ,OAAO,gBAAwB,GAAG;IACpC;GACF,EAAA,CACM,CAAC,CAAC,MAAM,YAAY;IACxB,MAAM,SAAS,eAAe,UAAU,aAAa;KACnD;KACA,IAAI,qBAAqB,GAEvB,QAAQ,KACN,uCAAuC,OAAO,QAAQ,EAAE,wDAC1D;IAEJ,CAAC;IACD,OAAO,KAAK,KAAK,MAAM;IACvB;IACA,MAAM,iBAAiB,OAAO,gBAAgB,SAAS;IACvD,kBACE,kBAAkB,KAAK,mBAAmB,KAAA,IACtC,KAAK,IAAI,OAAO,SAAS,KAAK,cAAc,IAC5C,OAAO;IACb,IAAI,gBAAgB;KAClB;KACA,IAAI,wBAAwB,GAE1B,QAAQ,KACN,4GACF;IAEJ;IACA,IAAI,kBAAkB,eAAe,CAAC,oBAAoB;KACxD,qBAAqB;KAErB,QAAQ,KAAK,+BAA+B;IAC9C;IACA,IAAI;KACF,KAAK,iBAAiB,MAAM,MAAM;IACpC,QAAQ,CAGR;IACA;IACA,IAAI,SAAS,QAAQ,QACnB,KAAK;SACA,IAAI,aAAa,GACtB,aAAa;GAEjB,CAAC;EACH;EACA,IAAI,UAAU,QAAQ,UAAU,aAAa,GAAG,aAAa;CAC/D;CAEA,MAAM,gBAAsB;EAG1B,IAAI,SAAS,QAAQ,QAAQ;GAC3B,oBAAoB,QAAQ,SAAS;GACrC,SAAS,QAAQ;EACnB;EACA,IAAI,aAAa,GAAG,aAAa;CACnC;CACA,IAAI,KAAK,QACP,IAAI,KAAK,OAAO,SAAS;EACvB,mBAAmB,QAAQ;EAC3B,SAAS,QAAQ;EACjB,aAAa;CACf,OACE,KAAK,OAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;CAIjE,IAAI,QAAQ,WAAW,GACrB,aAAa;MAEb,KAAK;CAGP,MAAM;CACN,IAAI,KAAK,QAAQ,KAAK,OAAO,oBAAoB,SAAS,OAAO;CAEjE,YAAY,MAAM,GAAG,MAAM,EAAE,KAAK,UAAU,EAAE,KAAK,OAAO;CAE1D,IAAI,OAAO;CACX,IAAI,WAAW;CACf,IAAI,YAAY;CAChB,IAAI,WAAW;CACf,KAAK,MAAM,EAAE,UAAU,aACrB,KAAK,MAAM,KAAK,MAAM;EACpB;EACA,MAAM,UAAU,EAAE,UAAU,KAAA;EAC5B,IAAI,CAAC,WAAW,EAAE,QAAQ,OAAO;EACjC,YAAY,UAAU,IAAI,EAAE,QAAQ;EACpC,aAAa,EAAE;CACjB;CAKF,OAAO;EACL,OAAO;EACP,QAJa,YAAY,aAAa,KAAK,MAAM,WAI5C;EACL,SAAS;GACP,YAAY,QAAQ;GACpB;GACA,cAAc,mBAAmB,cAAc;GAC/C,iBAAiB,aAAa,IAAI,IAAI,OAAO;GAC7C,kBAAkB,aAAa,IAAI,IAAI,WAAW;GAClD,cAAc;GACd;GACA,mBAAmB;GACnB,YAAY,KAAK,IAAI,IAAI;EAC3B;EACA,UAAU,aAAa;CACzB;AACF"}
+{"version":3,"file":"matrix-DnlHn5x1.js","names":[],"sources":["../src/matrix/cell-spend.ts","../src/matrix/aggregation.ts","../src/matrix/runner.ts"],"sourcesContent":["/**\n * Spend carrier for a cell that fails after it has already spent money.\n *\n * `runCell` reports what a cell cost by returning a `CellResult`. A throw\n * carries no result, so the runner cannot know what the cell spent before it\n * failed: the cumulative sum the cost ceiling reads, and `totalCostUsd`, both\n * miss that money. A throw that carries a `CellSpend` closes the gap — the\n * runner bills the failed cell for the spend the carrier declares, and records\n * a cell that declared nothing as `uncaptured` instead of a zero that reads\n * like \"this cell spent nothing\".\n *\n * The carrier is read structurally, never by `instanceof`, so a throw that\n * crosses a package or realm boundary still bills.\n */\n\n/** Property key the carrier occupies on a thrown object. Stable across\n *  versions — two copies of this package must agree on it. */\nconst CELL_SPEND_KEY = '__agentEvalCellSpend'\n\n/** Money and wall time a cell consumed before it threw. */\nexport interface CellSpend {\n  /** KNOWN subtotal in USD. When `kind` is `uncaptured` the cell spent this\n   *  much AND an unknown amount more, so the value must not be read as the\n   *  cell's total. */\n  costUsd: number\n  /** Wall time the cell consumed before the throw, in milliseconds. */\n  durationMs: number\n  /** `observed` = provider-reported amounts, `estimated` = computed from token\n   *  prices, `uncaptured` = `costUsd` is a subtotal and part of the spend\n   *  could not be measured. */\n  kind: 'observed' | 'estimated' | 'uncaptured'\n}\n\n/**\n * Return `error` carrying `spend`, for a throw site inside `runCell`:\n * `throw withCellSpend(err, { costUsd, durationMs, kind })`.\n *\n * A value that cannot hold the carrier — a thrown string or number, or a\n * frozen, sealed, or otherwise non-extensible object — is wrapped in an\n * `Error` that keeps the original as `cause`, so the spend is never dropped.\n * A later call overwrites an earlier carrier — the outer frame knows more\n * than the inner one.\n *\n * Throws `TypeError` on a non-finite or negative amount. A poisoned amount\n * would reach `cumulativeCost` and disable the cost ceiling for the whole run,\n * because `NaN >= ceiling` is false.\n */\nexport function withCellSpend(error: unknown, spend: CellSpend): unknown {\n  assertAmount(spend.costUsd, 'costUsd')\n  assertAmount(spend.durationMs, 'durationMs')\n  if (spend.kind !== 'observed' && spend.kind !== 'estimated' && spend.kind !== 'uncaptured') {\n    throw new TypeError(\n      `withCellSpend: kind must be observed, estimated or uncaptured, received ${String(spend.kind)}`,\n    )\n  }\n  const carrier: CellSpend = {\n    costUsd: spend.costUsd,\n    durationMs: spend.durationMs,\n    kind: spend.kind,\n  }\n  if (typeof error === 'object' && error !== null && attach(error, carrier)) return error\n  // A primitive cannot hold a property, and a frozen or sealed object refuses\n  // one. Carry the spend on a fresh Error that keeps the original as `cause`\n  // rather than lose it — a dropped carrier bills the cell as uncaptured,\n  // which is the under-billing this module exists to prevent.\n  const wrapper = new Error(`cell failed: ${messageOf(error)}`, { cause: error })\n  attach(wrapper, carrier)\n  return wrapper\n}\n\n/** True when the carrier landed. `defineProperty` throws on a non-extensible\n *  target, so the caller must be able to fall back. */\nfunction attach(target: object, carrier: CellSpend): boolean {\n  try {\n    Object.defineProperty(target, CELL_SPEND_KEY, {\n      value: carrier,\n      enumerable: false,\n      configurable: true,\n      writable: true,\n    })\n    return true\n  } catch {\n    return false\n  }\n}\n\nfunction messageOf(error: unknown): string {\n  if (error instanceof Error) return error.message\n  return String(error)\n}\n\n/** Read the spend a thrown value carries. `undefined` when the value carries\n *  none, or carries one that does not satisfy `CellSpend` — the runner records\n *  both as `uncaptured`. */\nexport function readCellSpend(error: unknown): CellSpend | undefined {\n  if (typeof error !== 'object' || error === null) return undefined\n  const raw = (error as Record<string, unknown>)[CELL_SPEND_KEY]\n  if (typeof raw !== 'object' || raw === null) return undefined\n  const candidate = raw as Record<string, unknown>\n  if (!isAmount(candidate.costUsd) || !isAmount(candidate.durationMs)) return undefined\n  const kind = candidate.kind\n  if (kind !== 'observed' && kind !== 'estimated' && kind !== 'uncaptured') return undefined\n  return { costUsd: candidate.costUsd, durationMs: candidate.durationMs, kind }\n}\n\n/** A number the matrix can add to a running cost or duration: finite and not\n *  negative. Anything else would poison a sum the cost ceiling reads. */\nexport function isAmount(value: unknown): value is number {\n  return typeof value === 'number' && Number.isFinite(value) && value >= 0\n}\n\nfunction assertAmount(value: number, field: string): void {\n  if (!isAmount(value)) {\n    throw new TypeError(\n      `withCellSpend: ${field} must be a finite number >= 0, received ${String(value)}`,\n    )\n  }\n}\n","/**\n * Per-axis aggregation of cell runs into `AxisSummary` rows.\n *\n * Pure: consumes the final `cells: [{cell, runs}]` array and returns the\n * `byAxis` table. Error runs contribute 0 to passRate and meanScore. Cost and\n * duration always count — the budget was spent regardless. A run whose cost is\n * a subtotal rather than a total raises `costUncapturedCells` for its bucket,\n * so a per-axis cost row says when it under-counts.\n */\n\nimport type { AxisSummary, CellResult, MatrixAxis, MatrixCell, MatrixResult } from './types'\n\ninterface Row<Output> {\n  cell: MatrixCell\n  result: CellResult<Output>\n}\n\nfunction flattenRuns<Output>(cells: MatrixResult<Output>['cells']): Row<Output>[] {\n  const rows: Row<Output>[] = []\n  for (const { cell, runs } of cells) {\n    for (const result of runs) rows.push({ cell, result })\n  }\n  return rows\n}\n\nfunction quantile(sorted: number[], q: number): number {\n  if (sorted.length === 0) return 0\n  if (sorted.length === 1) return sorted[0] as number\n  const pos = (sorted.length - 1) * q\n  const lo = Math.floor(pos)\n  const hi = Math.ceil(pos)\n  if (lo === hi) return sorted[lo] as number\n  const frac = pos - lo\n  return (sorted[lo] as number) * (1 - frac) + (sorted[hi] as number) * frac\n}\n\nexport function summariseRows<Output>(\n  rows: Row<Output>[],\n  axisName: string,\n  axisValue: string,\n): AxisSummary {\n  if (rows.length === 0) {\n    return {\n      axisName,\n      axisValue,\n      cells: 0,\n      passRate: 0,\n      meanScore: 0,\n      p50Score: 0,\n      p90Score: 0,\n      totalCostUsd: 0,\n      costUncapturedCells: 0,\n      meanDurationMs: 0,\n    }\n  }\n  let pass = 0\n  let scoreSum = 0\n  let costSum = 0\n  let durSum = 0\n  let costUncapturedCells = 0\n  const scores: number[] = []\n  for (const { result } of rows) {\n    const errored = result.error !== undefined\n    const score = errored ? 0 : result.verdict.score\n    const valid = !errored && result.verdict.valid\n    if (valid) pass++\n    scoreSum += score\n    scores.push(score)\n    costSum += result.costUsd\n    durSum += result.durationMs\n    if (result.costProvenance?.kind === 'uncaptured') costUncapturedCells++\n  }\n  scores.sort((a, b) => a - b)\n  return {\n    axisName,\n    axisValue,\n    cells: rows.length,\n    passRate: pass / rows.length,\n    meanScore: scoreSum / rows.length,\n    p50Score: quantile(scores, 0.5),\n    p90Score: quantile(scores, 0.9),\n    totalCostUsd: costSum,\n    costUncapturedCells,\n    meanDurationMs: durSum / rows.length,\n  }\n}\n\nfunction bucketBy<Output>(\n  rows: Row<Output>[],\n  axisName: string,\n  labelFor: (id: string) => string,\n): Record<string, AxisSummary> {\n  const buckets = new Map<string, Row<Output>[]>()\n  for (const row of rows) {\n    const slot = row.cell.axes[axisName]\n    if (!slot) continue\n    const id = slot.id\n    let arr = buckets.get(id)\n    if (!arr) {\n      arr = []\n      buckets.set(id, arr)\n    }\n    arr.push(row)\n  }\n  const out: Record<string, AxisSummary> = {}\n  // Sorted keys for deterministic JSON serialisation.\n  for (const id of [...buckets.keys()].sort()) {\n    out[id] = summariseRows(buckets.get(id) as Row<Output>[], axisName, labelFor(id))\n  }\n  return out\n}\n\nexport function buildByAxis<Output>(\n  cells: MatrixResult<Output>['cells'],\n  axes: MatrixAxis<unknown>[],\n  aggregateBy: string[],\n): Record<string, Record<string, AxisSummary>> {\n  const rows = flattenRuns(cells)\n  const byName = new Map(axes.map((a) => [a.name, a]))\n  const byAxis: Record<string, Record<string, AxisSummary>> = {}\n  for (const name of aggregateBy) {\n    const axis = byName.get(name)\n    const labelFor = (id: string): string => {\n      if (!axis?.label) return id\n      const found = axis.values.find((v) => v.id === id)\n      if (!found) return id\n      return axis.label(found.value, id)\n    }\n    byAxis[name] = bucketBy(rows, name, labelFor)\n  }\n  return byAxis\n}\n","/**\n * N-axis cartesian runner.\n *\n * Expansion order: cartesian over `axes` in declared order, then `reps` as the\n * inner-most dim → `ordinal = (cartIdx * reps) + rep`. The returned\n * `cells[]` is sorted by `ordinal` so concurrent execution does not reorder\n * the output.\n *\n * Scheduling is a sliding window of in-flight promises capped at\n * `maxConcurrency`. The window stops admitting new cells when the cost\n * ceiling trips or the abort signal fires; in-flight cells finish.\n */\n\nimport { buildByAxis } from './aggregation'\nimport { isAmount, readCellSpend } from './cell-spend'\nimport type {\n  CellResult,\n  CostProvenance,\n  MatrixAxis,\n  MatrixCell,\n  MatrixResult,\n  RunAgentMatrixOptions,\n} from './types'\n\ninterface BaseCell {\n  axes: Record<string, { id: string; value: unknown }>\n}\n\nfunction cartesian(axes: MatrixAxis<unknown>[]): BaseCell[] {\n  // Empty axes (`values=[]`) collapse the whole product to zero cells. An\n  // empty `axes` array yields a single empty-axes cell — degenerate but\n  // valid (caller is iterating only reps).\n  if (axes.length === 0) return [{ axes: {} }]\n  for (const a of axes) if (a.values.length === 0) return []\n  const out: BaseCell[] = []\n  const idx = new Array(axes.length).fill(0)\n  while (true) {\n    const slot: Record<string, { id: string; value: unknown }> = {}\n    for (let i = 0; i < axes.length; i++) {\n      const axis = axes[i] as MatrixAxis<unknown>\n      const v = axis.values[idx[i] as number] as { id: string; value: unknown }\n      slot[axis.name] = { id: v.id, value: v.value }\n    }\n    out.push({ axes: slot })\n    // Increment like an odometer, left-most axis is fastest.\n    let i = 0\n    while (i < axes.length) {\n      const next = (idx[i] as number) + 1\n      const axis = axes[i] as MatrixAxis<unknown>\n      if (next < axis.values.length) {\n        idx[i] = next\n        break\n      }\n      idx[i] = 0\n      i++\n    }\n    if (i === axes.length) break\n  }\n  return out\n}\n\nfunction makeMatrixId(): string {\n  // Stable id-like string: time + 8 random hex chars. Avoids node:crypto\n  // import to keep the matrix dep-free.\n  const t = Date.now().toString(36)\n  let r = ''\n  for (let i = 0; i < 8; i++) r += Math.floor(Math.random() * 16).toString(16)\n  return `mtx_${t}_${r}`\n}\n\n/** Fresh per result — a shared object would let one consumer's mutation reach\n *  every other cell in the run. */\nfunction uncaptured(): CostProvenance {\n  return { kind: 'uncaptured', usd: null }\n}\n\nfunction makeErrorResult<Output>(err: unknown): CellResult<Output> {\n  const e = err as { message?: string; name?: string }\n  const spend = readCellSpend(err)\n  return {\n    output: undefined as unknown as Output,\n    verdict: { valid: false, score: 0 },\n    costUsd: spend?.costUsd ?? 0,\n    durationMs: spend?.durationMs ?? 0,\n    costProvenance:\n      spend === undefined || spend.kind === 'uncaptured'\n        ? uncaptured()\n        : { kind: spend.kind, usd: spend.costUsd },\n    error: {\n      message: typeof e?.message === 'string' ? e.message : String(err),\n      kind: typeof e?.name === 'string' ? e.name : 'Error',\n    },\n  }\n}\n\n/** Make a result safe to sum.\n *\n *  A `costUsd` that cannot be summed fails the cell: `NaN` in `cumulativeCost`\n *  makes `>= costCeiling` false forever, so one bad cell would disable the\n *  ceiling for every cell after it.\n *\n *  A bad `durationMs` reaches only `meanDurationMs`, never the ceiling, so it\n *  cannot justify throwing away a real verdict and a real cost. It is reported\n *  as 0 and warned about, and the rest of the result stands. A wall clock that\n *  steps back mid-cell makes `Date.now() - startedAt` negative, so this is a\n *  reachable state, not only a caller bug. */\nfunction billableResult<Output>(\n  result: CellResult<Output>,\n  onBadDuration: (reported: unknown) => void,\n): CellResult<Output> {\n  if (!isAmount(result.costUsd)) {\n    return makeErrorResult<Output>(\n      new RangeError(\n        `runCell reported costUsd=${String(result.costUsd)}; a cell cost must be a finite number >= 0 so the cost ceiling stays enforceable`,\n      ),\n    )\n  }\n  if (isAmount(result.durationMs)) return result\n  onBadDuration(result.durationMs)\n  return { ...result, durationMs: 0 }\n}\n\nexport async function runAgentMatrix<Output>(\n  opts: RunAgentMatrixOptions<Output>,\n): Promise<MatrixResult<Output>> {\n  const startedAt = Date.now()\n  const reps = Math.max(1, opts.reps ?? 1)\n  const maxConcurrency = Math.max(1, opts.maxConcurrency ?? 4)\n  const costCeiling = opts.costCeiling ?? Number.POSITIVE_INFINITY\n  const aggregateBy = opts.aggregateBy ?? opts.axes.map((a) => a.name)\n  if (opts.maxCellCostUsd !== undefined && !isAmount(opts.maxCellCostUsd)) {\n    throw new RangeError(\n      `runAgentMatrix: maxCellCostUsd must be a finite number >= 0, received ${String(opts.maxCellCostUsd)}`,\n    )\n  }\n\n  const base = cartesian(opts.axes)\n  const filtered = opts.filter\n    ? base.filter((c) => (opts.filter as (b: BaseCell) => boolean)(c))\n    : base\n  const filteredOut = base.length - filtered.length\n\n  const planned: MatrixCell[] = []\n  for (let i = 0; i < filtered.length; i++) {\n    for (let r = 0; r < reps; r++) {\n      planned.push({\n        axes: (filtered[i] as BaseCell).axes,\n        rep: r,\n        ordinal: i * reps + r,\n      })\n    }\n  }\n\n  const cellRecords: Array<{ cell: MatrixCell; runs: CellResult<Output>[] }> = []\n  // What the ceiling reads. It runs ahead of the reported total whenever a\n  // cell's cost is a subtotal and `maxCellCostUsd` bounds what it could have\n  // been — a budget must assume the worst about spend it cannot see.\n  let cumulativeCost = 0\n  let costCeilingReached = false\n  let runsExecuted = 0\n  let cellsUnscheduled = 0\n  let costUncapturedCells = 0\n  let badDurationCells = 0\n\n  const aborted = (): boolean => opts.signal?.aborted === true\n\n  // Per-run abort controller forwards the external signal so cell executors\n  // see cancellation. We don't expose it on `MatrixCell` — the signature on\n  // `runCell` per the public API is `(cell) => Promise<...>`. Executors that\n  // need cancellation use the external signal directly via closure.\n\n  let inFlight = 0\n  let cursor = 0\n  let resolveAll: (() => void) | undefined\n  const done = new Promise<void>((res) => {\n    resolveAll = res\n  })\n\n  const pump = (): void => {\n    while (inFlight < maxConcurrency && cursor < planned.length) {\n      if (aborted() || costCeilingReached) {\n        // Drain remaining as unscheduled.\n        const left = planned.length - cursor\n        cellsUnscheduled += left\n        cursor = planned.length\n        break\n      }\n      const cell = planned[cursor++] as MatrixCell\n      inFlight++\n      // Lazily allocate the record so cells appear in `cells[]` in any\n      // arrival order; we sort by ordinal at the end.\n      const record = { cell, runs: [] as CellResult<Output>[] }\n      cellRecords.push(record)\n      const promise: Promise<CellResult<Output>> = (async () => {\n        try {\n          return await opts.runCell(cell)\n        } catch (err) {\n          return makeErrorResult<Output>(err)\n        }\n      })()\n      promise.then((settled) => {\n        const result = billableResult(settled, (reported) => {\n          badDurationCells++\n          if (badDurationCells === 1) {\n            // eslint-disable-next-line no-console\n            console.warn(\n              `[matrix] a cell reported durationMs=${String(reported)}; recorded as 0 — meanDurationMs under-reports this run`,\n            )\n          }\n        })\n        record.runs.push(result)\n        runsExecuted++\n        const uncapturedCost = result.costProvenance?.kind === 'uncaptured'\n        cumulativeCost +=\n          uncapturedCost && opts.maxCellCostUsd !== undefined\n            ? Math.max(result.costUsd, opts.maxCellCostUsd)\n            : result.costUsd\n        if (uncapturedCost) {\n          costUncapturedCells++\n          if (costUncapturedCells === 1) {\n            // eslint-disable-next-line no-console\n            console.warn(\n              \"[matrix] a cell's cost is a subtotal, not a total — totalCostUsd and the cost ceiling under-count this run\",\n            )\n          }\n        }\n        if (cumulativeCost >= costCeiling && !costCeilingReached) {\n          costCeilingReached = true\n          // eslint-disable-next-line no-console\n          console.warn('[matrix] cost ceiling reached')\n        }\n        try {\n          opts.onCellComplete?.(cell, result)\n        } catch {\n          // onCellComplete is observational — swallow throws so a noisy\n          // callback can't tank the run.\n        }\n        inFlight--\n        if (cursor < planned.length) {\n          pump()\n        } else if (inFlight === 0) {\n          resolveAll?.()\n        }\n      })\n    }\n    if (cursor >= planned.length && inFlight === 0) resolveAll?.()\n  }\n\n  const onAbort = (): void => {\n    // External abort: stop scheduling. In-flight cells finish; their\n    // executors observe `opts.signal.aborted` directly via closure.\n    if (cursor < planned.length) {\n      cellsUnscheduled += planned.length - cursor\n      cursor = planned.length\n    }\n    if (inFlight === 0) resolveAll?.()\n  }\n  if (opts.signal) {\n    if (opts.signal.aborted) {\n      cellsUnscheduled = planned.length\n      cursor = planned.length\n      resolveAll?.()\n    } else {\n      opts.signal.addEventListener('abort', onAbort, { once: true })\n    }\n  }\n\n  if (planned.length === 0) {\n    resolveAll?.()\n  } else {\n    pump()\n  }\n\n  await done\n  if (opts.signal) opts.signal.removeEventListener('abort', onAbort)\n\n  cellRecords.sort((a, b) => a.cell.ordinal - b.cell.ordinal)\n\n  let pass = 0\n  let scoreSum = 0\n  let totalCost = 0\n  let runCount = 0\n  for (const { runs } of cellRecords) {\n    for (const r of runs) {\n      runCount++\n      const errored = r.error !== undefined\n      if (!errored && r.verdict.valid) pass++\n      scoreSum += errored ? 0 : r.verdict.score\n      totalCost += r.costUsd\n    }\n  }\n\n  const byAxis = buildByAxis(cellRecords, opts.axes, aggregateBy)\n\n  return {\n    cells: cellRecords,\n    byAxis,\n    summary: {\n      totalCells: planned.length,\n      runsExecuted,\n      cellsSkipped: cellsUnscheduled + filteredOut * reps,\n      overallPassRate: runCount === 0 ? 0 : pass / runCount,\n      overallMeanScore: runCount === 0 ? 0 : scoreSum / runCount,\n      totalCostUsd: totalCost,\n      costUncapturedCells,\n      ceilingChargedUsd: cumulativeCost,\n      durationMs: Date.now() - startedAt,\n    },\n    matrixId: makeMatrixId(),\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAiBA,MAAM,iBAAiB;;;;;;;;;;;;;;;AA8BvB,SAAgB,cAAc,OAAgB,OAA2B;CACvE,aAAa,MAAM,SAAS,SAAS;CACrC,aAAa,MAAM,YAAY,YAAY;CAC3C,IAAI,MAAM,SAAS,cAAc,MAAM,SAAS,eAAe,MAAM,SAAS,cAC5E,MAAM,IAAI,UACR,2EAA2E,OAAO,MAAM,IAAI,GAC9F;CAEF,MAAM,UAAqB;EACzB,SAAS,MAAM;EACf,YAAY,MAAM;EAClB,MAAM,MAAM;CACd;CACA,IAAI,OAAO,UAAU,YAAY,UAAU,QAAQ,OAAO,OAAO,OAAO,GAAG,OAAO;CAKlF,MAAM,UAAU,IAAI,MAAM,gBAAgB,UAAU,KAAK,KAAK,EAAE,OAAO,MAAM,CAAC;CAC9E,OAAO,SAAS,OAAO;CACvB,OAAO;AACT;;;AAIA,SAAS,OAAO,QAAgB,SAA6B;CAC3D,IAAI;EACF,OAAO,eAAe,QAAQ,gBAAgB;GAC5C,OAAO;GACP,YAAY;GACZ,cAAc;GACd,UAAU;EACZ,CAAC;EACD,OAAO;CACT,QAAQ;EACN,OAAO;CACT;AACF;AAEA,SAAS,UAAU,OAAwB;CACzC,IAAI,iBAAiB,OAAO,OAAO,MAAM;CACzC,OAAO,OAAO,KAAK;AACrB;;;;AAKA,SAAgB,cAAc,OAAuC;CACnE,IAAI,OAAO,UAAU,YAAY,UAAU,MAAM,OAAO,KAAA;CACxD,MAAM,MAAO,MAAkC;CAC/C,IAAI,OAAO,QAAQ,YAAY,QAAQ,MAAM,OAAO,KAAA;CACpD,MAAM,YAAY;CAClB,IAAI,CAAC,SAAS,UAAU,OAAO,KAAK,CAAC,SAAS,UAAU,UAAU,GAAG,OAAO,KAAA;CAC5E,MAAM,OAAO,UAAU;CACvB,IAAI,SAAS,cAAc,SAAS,eAAe,SAAS,cAAc,OAAO,KAAA;CACjF,OAAO;EAAE,SAAS,UAAU;EAAS,YAAY,UAAU;EAAY;CAAK;AAC9E;;;AAIA,SAAgB,SAAS,OAAiC;CACxD,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,KAAK,SAAS;AACzE;AAEA,SAAS,aAAa,OAAe,OAAqB;CACxD,IAAI,CAAC,SAAS,KAAK,GACjB,MAAM,IAAI,UACR,kBAAkB,MAAM,0CAA0C,OAAO,KAAK,GAChF;AAEJ;;;ACpGA,SAAS,YAAoB,OAAqD;CAChF,MAAM,OAAsB,CAAC;CAC7B,KAAK,MAAM,EAAE,MAAM,UAAU,OAC3B,KAAK,MAAM,UAAU,MAAM,KAAK,KAAK;EAAE;EAAM;CAAO,CAAC;CAEvD,OAAO;AACT;AAEA,SAAS,SAAS,QAAkB,GAAmB;CACrD,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,IAAI,OAAO,WAAW,GAAG,OAAO,OAAO;CACvC,MAAM,OAAO,OAAO,SAAS,KAAK;CAClC,MAAM,KAAK,KAAK,MAAM,GAAG;CACzB,MAAM,KAAK,KAAK,KAAK,GAAG;CACxB,IAAI,OAAO,IAAI,OAAO,OAAO;CAC7B,MAAM,OAAO,MAAM;CACnB,OAAQ,OAAO,OAAkB,IAAI,QAAS,OAAO,MAAiB;AACxE;AAEA,SAAgB,cACd,MACA,UACA,WACa;CACb,IAAI,KAAK,WAAW,GAClB,OAAO;EACL;EACA;EACA,OAAO;EACP,UAAU;EACV,WAAW;EACX,UAAU;EACV,UAAU;EACV,cAAc;EACd,qBAAqB;EACrB,gBAAgB;CAClB;CAEF,IAAI,OAAO;CACX,IAAI,WAAW;CACf,IAAI,UAAU;CACd,IAAI,SAAS;CACb,IAAI,sBAAsB;CAC1B,MAAM,SAAmB,CAAC;CAC1B,KAAK,MAAM,EAAE,YAAY,MAAM;EAC7B,MAAM,UAAU,OAAO,UAAU,KAAA;EACjC,MAAM,QAAQ,UAAU,IAAI,OAAO,QAAQ;EAE3C,IADc,CAAC,WAAW,OAAO,QAAQ,OAC9B;EACX,YAAY;EACZ,OAAO,KAAK,KAAK;EACjB,WAAW,OAAO;EAClB,UAAU,OAAO;EACjB,IAAI,OAAO,gBAAgB,SAAS,cAAc;CACpD;CACA,OAAO,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3B,OAAO;EACL;EACA;EACA,OAAO,KAAK;EACZ,UAAU,OAAO,KAAK;EACtB,WAAW,WAAW,KAAK;EAC3B,UAAU,SAAS,QAAQ,EAAG;EAC9B,UAAU,SAAS,QAAQ,EAAG;EAC9B,cAAc;EACd;EACA,gBAAgB,SAAS,KAAK;CAChC;AACF;AAEA,SAAS,SACP,MACA,UACA,UAC6B;CAC7B,MAAM,0BAAU,IAAI,IAA2B;CAC/C,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,OAAO,IAAI,KAAK,KAAK;EAC3B,IAAI,CAAC,MAAM;EACX,MAAM,KAAK,KAAK;EAChB,IAAI,MAAM,QAAQ,IAAI,EAAE;EACxB,IAAI,CAAC,KAAK;GACR,MAAM,CAAC;GACP,QAAQ,IAAI,IAAI,GAAG;EACrB;EACA,IAAI,KAAK,GAAG;CACd;CACA,MAAM,MAAmC,CAAC;CAE1C,KAAK,MAAM,MAAM,CAAC,GAAG,QAAQ,KAAK,CAAC,CAAC,CAAC,KAAK,GACxC,IAAI,MAAM,cAAc,QAAQ,IAAI,EAAE,GAAoB,UAAU,SAAS,EAAE,CAAC;CAElF,OAAO;AACT;AAEA,SAAgB,YACd,OACA,MACA,aAC6C;CAC7C,MAAM,OAAO,YAAY,KAAK;CAC9B,MAAM,SAAS,IAAI,IAAI,KAAK,KAAK,MAAM,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC;CACnD,MAAM,SAAsD,CAAC;CAC7D,KAAK,MAAM,QAAQ,aAAa;EAC9B,MAAM,OAAO,OAAO,IAAI,IAAI;EAC5B,MAAM,YAAY,OAAuB;GACvC,IAAI,CAAC,MAAM,OAAO,OAAO;GACzB,MAAM,QAAQ,KAAK,OAAO,MAAM,MAAM,EAAE,OAAO,EAAE;GACjD,IAAI,CAAC,OAAO,OAAO;GACnB,OAAO,KAAK,MAAM,MAAM,OAAO,EAAE;EACnC;EACA,OAAO,QAAQ,SAAS,MAAM,MAAM,QAAQ;CAC9C;CACA,OAAO;AACT;;;;;;;;;;;;;;;ACvGA,SAAS,UAAU,MAAyC;CAI1D,IAAI,KAAK,WAAW,GAAG,OAAO,CAAC,EAAE,MAAM,CAAC,EAAE,CAAC;CAC3C,KAAK,MAAM,KAAK,MAAM,IAAI,EAAE,OAAO,WAAW,GAAG,OAAO,CAAC;CACzD,MAAM,MAAkB,CAAC;CACzB,MAAM,MAAM,IAAI,MAAM,KAAK,MAAM,CAAC,CAAC,KAAK,CAAC;CACzC,OAAO,MAAM;EACX,MAAM,OAAuD,CAAC;EAC9D,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;GACpC,MAAM,OAAO,KAAK;GAClB,MAAM,IAAI,KAAK,OAAO,IAAI;GAC1B,KAAK,KAAK,QAAQ;IAAE,IAAI,EAAE;IAAI,OAAO,EAAE;GAAM;EAC/C;EACA,IAAI,KAAK,EAAE,MAAM,KAAK,CAAC;EAEvB,IAAI,IAAI;EACR,OAAO,IAAI,KAAK,QAAQ;GACtB,MAAM,OAAQ,IAAI,KAAgB;GAElC,IAAI,OADS,KAAK,EACH,CAAC,OAAO,QAAQ;IAC7B,IAAI,KAAK;IACT;GACF;GACA,IAAI,KAAK;GACT;EACF;EACA,IAAI,MAAM,KAAK,QAAQ;CACzB;CACA,OAAO;AACT;AAEA,SAAS,eAAuB;CAG9B,MAAM,IAAI,KAAK,IAAI,CAAC,CAAC,SAAS,EAAE;CAChC,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,KAAK,KAAK,MAAM,KAAK,OAAO,IAAI,EAAE,CAAC,CAAC,SAAS,EAAE;CAC3E,OAAO,OAAO,EAAE,GAAG;AACrB;;;AAIA,SAAS,aAA6B;CACpC,OAAO;EAAE,MAAM;EAAc,KAAK;CAAK;AACzC;AAEA,SAAS,gBAAwB,KAAkC;CACjE,MAAM,IAAI;CACV,MAAM,QAAQ,cAAc,GAAG;CAC/B,OAAO;EACL,QAAQ,KAAA;EACR,SAAS;GAAE,OAAO;GAAO,OAAO;EAAE;EAClC,SAAS,OAAO,WAAW;EAC3B,YAAY,OAAO,cAAc;EACjC,gBACE,UAAU,KAAA,KAAa,MAAM,SAAS,eAClC,WAAW,IACX;GAAE,MAAM,MAAM;GAAM,KAAK,MAAM;EAAQ;EAC7C,OAAO;GACL,SAAS,OAAO,GAAG,YAAY,WAAW,EAAE,UAAU,OAAO,GAAG;GAChE,MAAM,OAAO,GAAG,SAAS,WAAW,EAAE,OAAO;EAC/C;CACF;AACF;;;;;;;;;;;;AAaA,SAAS,eACP,QACA,eACoB;CACpB,IAAI,CAAC,SAAS,OAAO,OAAO,GAC1B,OAAO,gCACL,IAAI,WACF,4BAA4B,OAAO,OAAO,OAAO,EAAE,iFACrD,CACF;CAEF,IAAI,SAAS,OAAO,UAAU,GAAG,OAAO;CACxC,cAAc,OAAO,UAAU;CAC/B,OAAO;EAAE,GAAG;EAAQ,YAAY;CAAE;AACpC;AAEA,eAAsB,eACpB,MAC+B;CAC/B,MAAM,YAAY,KAAK,IAAI;CAC3B,MAAM,OAAO,KAAK,IAAI,GAAG,KAAK,QAAQ,CAAC;CACvC,MAAM,iBAAiB,KAAK,IAAI,GAAG,KAAK,kBAAkB,CAAC;CAC3D,MAAM,cAAc,KAAK,eAAe,OAAO;CAC/C,MAAM,cAAc,KAAK,eAAe,KAAK,KAAK,KAAK,MAAM,EAAE,IAAI;CACnE,IAAI,KAAK,mBAAmB,KAAA,KAAa,CAAC,SAAS,KAAK,cAAc,GACpE,MAAM,IAAI,WACR,yEAAyE,OAAO,KAAK,cAAc,GACrG;CAGF,MAAM,OAAO,UAAU,KAAK,IAAI;CAChC,MAAM,WAAW,KAAK,SAClB,KAAK,QAAQ,MAAO,KAAK,OAAoC,CAAC,CAAC,IAC/D;CACJ,MAAM,cAAc,KAAK,SAAS,SAAS;CAE3C,MAAM,UAAwB,CAAC;CAC/B,KAAK,IAAI,IAAI,GAAG,IAAI,SAAS,QAAQ,KACnC,KAAK,IAAI,IAAI,GAAG,IAAI,MAAM,KACxB,QAAQ,KAAK;EACX,MAAO,SAAS,EAAE,CAAc;EAChC,KAAK;EACL,SAAS,IAAI,OAAO;CACtB,CAAC;CAIL,MAAM,cAAuE,CAAC;CAI9E,IAAI,iBAAiB;CACrB,IAAI,qBAAqB;CACzB,IAAI,eAAe;CACnB,IAAI,mBAAmB;CACvB,IAAI,sBAAsB;CAC1B,IAAI,mBAAmB;CAEvB,MAAM,gBAAyB,KAAK,QAAQ,YAAY;CAOxD,IAAI,WAAW;CACf,IAAI,SAAS;CACb,IAAI;CACJ,MAAM,OAAO,IAAI,SAAe,QAAQ;EACtC,aAAa;CACf,CAAC;CAED,MAAM,aAAmB;EACvB,OAAO,WAAW,kBAAkB,SAAS,QAAQ,QAAQ;GAC3D,IAAI,QAAQ,KAAK,oBAAoB;IAEnC,MAAM,OAAO,QAAQ,SAAS;IAC9B,oBAAoB;IACpB,SAAS,QAAQ;IACjB;GACF;GACA,MAAM,OAAO,QAAQ;GACrB;GAGA,MAAM,SAAS;IAAE;IAAM,MAAM,CAAC;GAA0B;GACxD,YAAY,KAAK,MAAM;GAQvB,CAP8C,YAAY;IACxD,IAAI;KACF,OAAO,MAAM,KAAK,QAAQ,IAAI;IAChC,SAAS,KAAK;KACZ,OAAO,gBAAwB,GAAG;IACpC;GACF,EAAA,CACM,CAAC,CAAC,MAAM,YAAY;IACxB,MAAM,SAAS,eAAe,UAAU,aAAa;KACnD;KACA,IAAI,qBAAqB,GAEvB,QAAQ,KACN,uCAAuC,OAAO,QAAQ,EAAE,wDAC1D;IAEJ,CAAC;IACD,OAAO,KAAK,KAAK,MAAM;IACvB;IACA,MAAM,iBAAiB,OAAO,gBAAgB,SAAS;IACvD,kBACE,kBAAkB,KAAK,mBAAmB,KAAA,IACtC,KAAK,IAAI,OAAO,SAAS,KAAK,cAAc,IAC5C,OAAO;IACb,IAAI,gBAAgB;KAClB;KACA,IAAI,wBAAwB,GAE1B,QAAQ,KACN,4GACF;IAEJ;IACA,IAAI,kBAAkB,eAAe,CAAC,oBAAoB;KACxD,qBAAqB;KAErB,QAAQ,KAAK,+BAA+B;IAC9C;IACA,IAAI;KACF,KAAK,iBAAiB,MAAM,MAAM;IACpC,QAAQ,CAGR;IACA;IACA,IAAI,SAAS,QAAQ,QACnB,KAAK;SACA,IAAI,aAAa,GACtB,aAAa;GAEjB,CAAC;EACH;EACA,IAAI,UAAU,QAAQ,UAAU,aAAa,GAAG,aAAa;CAC/D;CAEA,MAAM,gBAAsB;EAG1B,IAAI,SAAS,QAAQ,QAAQ;GAC3B,oBAAoB,QAAQ,SAAS;GACrC,SAAS,QAAQ;EACnB;EACA,IAAI,aAAa,GAAG,aAAa;CACnC;CACA,IAAI,KAAK,QAAQ;EACf,IAAI,KAAK,OAAO,SAAS;GACvB,mBAAmB,QAAQ;GAC3B,SAAS,QAAQ;GACjB,aAAa;EACf,OACE,KAAK,OAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;CAEjE;CAEA,IAAI,QAAQ,WAAW,GACrB,aAAa;MAEb,KAAK;CAGP,MAAM;CACN,IAAI,KAAK,QAAQ,KAAK,OAAO,oBAAoB,SAAS,OAAO;CAEjE,YAAY,MAAM,GAAG,MAAM,EAAE,KAAK,UAAU,EAAE,KAAK,OAAO;CAE1D,IAAI,OAAO;CACX,IAAI,WAAW;CACf,IAAI,YAAY;CAChB,IAAI,WAAW;CACf,KAAK,MAAM,EAAE,UAAU,aACrB,KAAK,MAAM,KAAK,MAAM;EACpB;EACA,MAAM,UAAU,EAAE,UAAU,KAAA;EAC5B,IAAI,CAAC,WAAW,EAAE,QAAQ,OAAO;EACjC,YAAY,UAAU,IAAI,EAAE,QAAQ;EACpC,aAAa,EAAE;CACjB;CAKF,OAAO;EACL,OAAO;EACP,QAJa,YAAY,aAAa,KAAK,MAAM,WAI5C;EACL,SAAS;GACP,YAAY,QAAQ;GACpB;GACA,cAAc,mBAAmB,cAAc;GAC/C,iBAAiB,aAAa,IAAI,IAAI,OAAO;GAC7C,kBAAkB,aAAa,IAAI,IAAI,WAAW;GAClD,cAAc;GACd;GACA,mBAAmB;GACnB,YAAY,KAAK,IAAI,IAAI;EAC3B;EACA,UAAU,aAAa;CACzB;AACF"}

package/dist/meta-eval/index.d.ts

@@ -8,17 +8,17 @@ import { a as ComputedInterval } from "../ast-hI-vjW6J.js";
 import { a as RubricPredictiveValidityReport, c as CorrelationInterval, i as RubricPredictiveValidityInput, l as OutcomeReduction, n as RubricOutcomeExclusion, o as RubricRanking, r as RubricOutcomePair, s as rubricPredictiveValidity, t as OutcomeMetricSpec } from "../rubric-predictive-validity-DIUwJJ9Y.js";
 import { z } from "zod";
 //#region src/meta-eval/correlation-study.d.ts
-interface EvalMetricSpec {
+export interface EvalMetricSpec {
   id: string;
   /** Extract a scalar from a run. Omit it and `id` must name one of
    *  `RUN_METRICS`; any other `id` is refused. */
   extract?: (run: Run, store: TraceStore) => Promise<number | null>;
 }
-interface OutcomePair {
+export interface OutcomePair {
   evalMetric: string;
   outcomeMetric: string;
 }
-interface CorrelationResult {
+export interface CorrelationResult {
   evalMetric: string;
   outcomeMetric: string;
   n: number;
@@ -37,7 +37,7 @@ interface CorrelationResult {
   /** Rough verdict: 'strong' ≥ 0.7, 'moderate' ≥ 0.4, else 'weak'. */
   verdict: 'strong' | 'moderate' | 'weak';
 }
-interface CorrelationStudyResult {
+export interface CorrelationStudyResult {
   pairs: CorrelationResult[];
   /** Declared pairs without an estimable correlation, including their usable sample count. */
   excludedPairs: Array<OutcomePair & {
@@ -47,7 +47,7 @@ interface CorrelationStudyResult {
   joinedSamples: number;
   skippedRuns: number;
 }
-interface CorrelationStudyOptions {
+export interface CorrelationStudyOptions {
   /** Only join outcomes captured within this window after run.startedAt. */
   maxCaptureLagMs?: number;
   /** Restrict to a subset of outcomes (cohort, region, source). */
@@ -60,10 +60,10 @@ interface CorrelationStudyOptions {
    *  paired observations, so the same study reproduces the same interval. */
   seed?: number;
 }
-declare function correlationStudy(traceStore: TraceStore, outcomeStore: OutcomeStore, evalMetrics: EvalMetricSpec[], outcomeMetricNames: string[], options?: CorrelationStudyOptions): Promise<CorrelationStudyResult>;
+export declare function correlationStudy(traceStore: TraceStore, outcomeStore: OutcomeStore, evalMetrics: EvalMetricSpec[], outcomeMetricNames: string[], options?: CorrelationStudyOptions): Promise<CorrelationStudyResult>;
 //#endregion
 //#region src/meta-eval/calibration.d.ts
-interface CalibrationBin {
+export interface CalibrationBin {
   lower: number;
   upper: number;
   n: number;
@@ -72,7 +72,7 @@ interface CalibrationBin {
   /** |outcomeMean − evalMean|; contributes to ECE weighted by n/total. */
   gap: number;
 }
-interface CalibrationReport {
+export interface CalibrationReport {
   evalMetric: string;
   outcomeMetric: string;
   n: number;
@@ -82,7 +82,7 @@ interface CalibrationReport {
   /** Largest observed difference between a bin's mean score and mean outcome. */
   maxGap: number;
 }
-interface CalibrationOptions {
+export interface CalibrationOptions {
   /** Positive integer; empty bins are omitted. Default 10. */
   bins?: number;
   /** Equal-width (fixed bin edges) or equal-frequency (quantile bins). */
@@ -93,13 +93,13 @@ interface CalibrationOptions {
     hi: number;
   };
 }
-interface CalibrationPair {
+export interface CalibrationPair {
   evalScore: number;
   outcome: number;
 }
-declare function calibrationCurve(traceStore: TraceStore, outcomeStore: OutcomeStore, evalMetric: EvalMetricSpec, outcomeMetric: string, options?: CalibrationOptions): Promise<CalibrationReport | null>;
+export declare function calibrationCurve(traceStore: TraceStore, outcomeStore: OutcomeStore, evalMetric: EvalMetricSpec, outcomeMetric: string, options?: CalibrationOptions): Promise<CalibrationReport | null>;
 /** Measure already joined observations without constructing trace and outcome stores. */
-declare function calibrationFromPairs(inputPairs: readonly CalibrationPair[], evalMetric: string, outcomeMetric: string, options?: CalibrationOptions): CalibrationReport | null;
+export declare function calibrationFromPairs(inputPairs: readonly CalibrationPair[], evalMetric: string, outcomeMetric: string, options?: CalibrationOptions): CalibrationReport | null;
 //#endregion
 //#region src/meta-eval/evaluator-admission.d.ts
 declare const policySchema: z.ZodObject<{
@@ -215,7 +215,7 @@ interface EvaluatorAdmissionReport {
  * A unit fails a class when any control in that class is misjudged.
  * The execution owner enforces auditor separation, fresh sampling, and evidence authenticity.
  */
-declare function auditEvaluator(input: EvaluatorAuditInput): EvaluatorAdmissionReport;
+export declare function auditEvaluator(input: EvaluatorAuditInput): EvaluatorAdmissionReport;
 declare const probabilityAuditSchema: z.ZodObject<{
   evaluatorDigest: z.ZodString;
   population: z.ZodString;
@@ -267,7 +267,7 @@ interface ProbabilityPolicyAuditReport extends EvaluatorAdmissionReport {
  * No threshold search, model calls, promotion, or assumption of calibrated probabilities.
  * Null scores and the open interval between thresholds remain unknown, never correct by default.
  */
-declare function auditProbabilityPolicy(input: ProbabilityPolicyAuditInput): ProbabilityPolicyAuditReport;
+export declare function auditProbabilityPolicy(input: ProbabilityPolicyAuditInput): ProbabilityPolicyAuditReport;
 //#endregion
 //#region src/meta-eval/plants.d.ts
 /**
@@ -275,7 +275,7 @@ declare function auditProbabilityPolicy(input: ProbabilityPolicyAuditInput): Pro
  * {@link CatchRateReport.byKind} because a grader is routinely sharp on one
  * and blind to another.
  */
-type PlantKind =
+export type PlantKind =
 /** A load-bearing value is altered: a number off by one, a comparison flipped. */
 'wrong-value' |
 /** The item carries its own check, and that check passes without testing the claim. */
@@ -285,8 +285,8 @@ type PlantKind =
 /** A copy of an item already in the set, which is owed a duplicate flag rather than a second grade. */
 'duplicate';
 /** What a working grader owes a seeded item. */
-type PlantExpectation = 'reject' | 'accept';
-interface Plant {
+export type PlantExpectation = 'reject' | 'accept';
+export interface Plant {
   /** Name of the plant record. Reported in `missedIds` and `missingIds`. */
   id: string;
   kind: PlantKind;
@@ -307,7 +307,7 @@ interface Plant {
  * cannot change what the manifest sealed, and `group` is dropped when it is
  * absent so the record always has a canonical JSON form.
  */
-declare function definePlant(input: {
+export declare function definePlant(input: {
   id: string;
   kind: PlantKind;
   item: GoldenItem;
@@ -317,17 +317,17 @@ declare function definePlant(input: {
  * A claim's executable evidence: the command a grader runs and the output it
  * must produce.
  */
-interface PlantEvidence {
+export interface PlantEvidence {
   /** The command a grader runs to test the claim. */
   check?: string;
   /** The output that command must produce for the claim to stand. */
   expect?: string;
 }
 /** Which evidence field {@link perturbEvidence} altered. */
-type PerturbedField = 'check' | 'expect';
+export type PerturbedField = 'check' | 'expect';
 /** How {@link perturbEvidence} altered it. */
-type PerturbationKind = 'number-off-by-one' | 'comparison-flipped';
-interface EvidencePerturbation {
+export type PerturbationKind = 'number-off-by-one' | 'comparison-flipped';
+export interface EvidencePerturbation {
   /** The evidence with exactly one value altered. */
   evidence: PlantEvidence;
   /** Which field changed. */
@@ -370,8 +370,8 @@ interface EvidencePerturbation {
  * rounds back to itself, and the plant is then a correct claim the grader is
  * right to verify.
  */
-declare function perturbEvidence(evidence: PlantEvidence): EvidencePerturbation | null;
-interface PerturbedPlant {
+export declare function perturbEvidence(evidence: PlantEvidence): EvidencePerturbation | null;
+export interface PerturbedPlant {
   plant: Plant;
   /** The perturbed evidence to write into the item the grader is handed. */
   evidence: PlantEvidence;
@@ -400,7 +400,7 @@ interface PerturbedPlant {
  * and in `missedIds`, the other in the join with the results — and a miss
  * nobody can look up is not a measurement.
  */
-declare function plantByPerturbation(input: {
+export declare function plantByPerturbation(input: {
   id: string;
   itemId: string;
   evidence: PlantEvidence;
@@ -408,7 +408,7 @@ declare function plantByPerturbation(input: {
   humanScore?: number;
   group?: string;
 }): PerturbedPlant | null;
-interface PlantManifest {
+export interface PlantManifest {
   /**
    * Digest over the seeded order, the plants, and the threshold. Publish it
    * before the grading run: a manifest edited afterwards to match the results
@@ -424,7 +424,7 @@ interface PlantManifest {
   /** The seeded plants. This is the answer key; keep it out of the graded workspace. */
   plants: Plant[];
 }
-interface SeededGradingSet {
+export interface SeededGradingSet {
   /**
    * The mixed set in seeded order. Operator-side: it still carries every
    * item's `humanScore`, so hand the graded policy the payload each `itemId`
@@ -433,7 +433,7 @@ interface SeededGradingSet {
   items: GoldenItem[];
   manifest: PlantManifest;
 }
-interface SeedPlantsOptions {
+export interface SeedPlantsOptions {
   /**
    * Fixes the mix order. The same dataset, plants, threshold, and seed always
    * produce the same seeded order and the same seal.
@@ -455,13 +455,13 @@ interface SeedPlantsOptions {
  * dataset item with no usable label, and a plant whose expectation the run's
  * `acceptThreshold` contradicts.
  */
-declare function seedPlants(dataset: readonly GoldenItem[], plants: readonly Plant[], options?: SeedPlantsOptions): SeededGradingSet;
+export declare function seedPlants(dataset: readonly GoldenItem[], plants: readonly Plant[], options?: SeedPlantsOptions): SeededGradingSet;
 /**
  * One grader outcome for one item. `score: null` says the grader ran and
  * declined to decide — the check never tested the item's defect. Any
  * `CandidateScore` from a judge run is already a valid outcome.
  */
-interface PlantOutcome {
+export interface PlantOutcome {
   itemId: string;
   score: number | null;
 }
@@ -469,8 +469,8 @@ interface PlantOutcome {
  * `evaluated` — a rate stands. `incomplete` — a seeded id had no result.
  * `not_evaluated` — nothing was seeded, or nothing seeded was decided.
  */
-type CatchRateStatus = 'evaluated' | 'incomplete' | 'not_evaluated';
-interface PlantKindCounts {
+export type CatchRateStatus = 'evaluated' | 'incomplete' | 'not_evaluated';
+export interface PlantKindCounts {
   seeded: number;
   caught: number;
   missed: number;
@@ -484,14 +484,14 @@ interface PlantKindCounts {
  * on reject-plants, and a `rejectionRate` of 1.0 here is what separates that
  * reflex from discrimination.
  */
-interface UnseededRejection {
+export interface UnseededRejection {
   n: number;
   decided: number;
   rejected: number;
   /** rejected / decided, or null when nothing unseeded was decided. */
   rejectionRate: number | null;
 }
-interface CatchRateReport {
+export interface CatchRateReport {
   status: CatchRateStatus;
   /** Why the status is not `evaluated`. Absent when it is. */
   reason?: string;
@@ -521,12 +521,12 @@ interface CatchRateReport {
  * the results and the manifest describe different runs, and a rate computed
  * across two runs is a fabrication.
  */
-declare function catchRate(results: readonly PlantOutcome[], manifest: PlantManifest): CatchRateReport;
+export declare function catchRate(results: readonly PlantOutcome[], manifest: PlantManifest): CatchRateReport;
 //#endregion
 //#region src/meta-eval/sentinel.d.ts
 declare const SENTINEL_METRIC_NAMES: readonly ['irr', 'calibrationKappa', 'sentinelPassRate'];
-type SentinelMetricName = (typeof SENTINEL_METRIC_NAMES)[number];
-interface SentinelMetrics {
+export type SentinelMetricName = (typeof SENTINEL_METRIC_NAMES)[number];
+export interface SentinelMetrics {
   /** Inter-rater reliability — ICC(2,1) from agreement instruments. */
   irr?: number;
   /** Weighted κ vs the human golden set (`calibrateJudge*`). */
@@ -534,7 +534,7 @@ interface SentinelMetrics {
   /** Fraction of a fixed sentinel golden set the judge still scores within tolerance. */
   sentinelPassRate?: number;
 }
-interface SentinelSnapshot {
+export interface SentinelSnapshot {
   /** Caller-supplied ISO-8601 timestamp — the module never reads a clock. */
   at: string;
   judgeId: string;
@@ -549,19 +549,19 @@ interface SentinelSnapshot {
   metrics: SentinelMetrics;
 }
 /** Identity + timestamp the caller supplies alongside an instrument output. */
-interface SnapshotMeta {
+export interface SnapshotMeta {
   at: string;
   judgeId: string;
   judgeModel: string;
 }
 /** Shape guard for snapshots crossing a boundary (store append, JSONL load, report input). */
-declare function validateSentinelSnapshot(snapshot: SentinelSnapshot, source?: string): void;
+export declare function validateSentinelSnapshot(snapshot: SentinelSnapshot, source?: string): void;
 /**
  * From `calibrateJudge` / `calibrateJudgeContinuous` output. Prefers the
  * un-rounded continuous κ_w when the report carries one (integer κ
  * discards information for [0,1] judges).
  */
-declare function snapshotFromCalibration(report: CalibrationResult | ContinuousCalibrationResult, meta: SnapshotMeta): SentinelSnapshot;
+export declare function snapshotFromCalibration(report: CalibrationResult | ContinuousCalibrationResult, meta: SnapshotMeta): SentinelSnapshot;
 /**
  * From `corpusInterRaterAgreement` (statistics.ts) or `continuousAgreement`
  * (judge-calibration.ts) output. IRR = ICC(2,1) — the reliability
@@ -569,8 +569,8 @@ declare function snapshotFromCalibration(report: CalibrationResult | ContinuousC
  * `meta.judgeId` names the ensemble and `meta.judgeModel` pins its
  * composition (e.g. a joined list of member model strings).
  */
-declare function snapshotFromAgreement(report: CorpusAgreementReport | ContinuousAgreement, meta: SnapshotMeta): SentinelSnapshot;
-interface SentinelSetOptions {
+export declare function snapshotFromAgreement(report: CorpusAgreementReport | ContinuousAgreement, meta: SnapshotMeta): SentinelSnapshot;
+export interface SentinelSetOptions {
   /** Max |judge − human| for an item to count as passed. Default 0.1 (tuned for [0,1] scores). */
   tolerance?: number;
 }
@@ -581,21 +581,21 @@ interface SentinelSetOptions {
  * human score. Items the judge didn't score are excluded from the
  * denominator; zero joined items is an error, not a 0% pass rate.
  */
-declare function snapshotFromSentinelSet(scores: CandidateScore[], golden: GoldenItem[], meta: SnapshotMeta, options?: SentinelSetOptions): SentinelSnapshot;
-interface SentinelStore {
+export declare function snapshotFromSentinelSet(scores: CandidateScore[], golden: GoldenItem[], meta: SnapshotMeta, options?: SentinelSetOptions): SentinelSnapshot;
+export interface SentinelStore {
   append(snapshot: SentinelSnapshot): Promise<void>;
   /** All snapshots (optionally for one judge), in append order. */
   history(judgeId?: string): Promise<SentinelSnapshot[]>;
 }
-declare function inMemorySentinelStore(initial?: SentinelSnapshot[]): SentinelStore;
+export declare function inMemorySentinelStore(initial?: SentinelSnapshot[]): SentinelStore;
 /**
  * JSONL store — one snapshot per line, appended atomically per call.
  * A corrupt or shape-invalid line is a loud error naming the file and
  * line number: a sentinel history that silently drops records would
  * defeat the drift detection it exists to provide.
  */
-declare function fileSentinelStore(path: string): SentinelStore;
-interface SentinelThresholds {
+export declare function fileSentinelStore(path: string): SentinelStore;
+export interface SentinelThresholds {
   /** Floor for irr — below it the judge ensemble is unreliable regardless of trend. Default 0.6. */
   minIrr?: number;
   /** Max tolerated calibrationKappa drop vs the series baseline. Default 0.15. */
@@ -605,14 +605,14 @@ interface SentinelThresholds {
   /** Newest snapshot older than this (days, vs asOf) = the sentinel itself went blind. Default 30. */
   staleAfterDays?: number;
 }
-interface JudgeSentinelOptions {
+export interface JudgeSentinelOptions {
   /** Caller-supplied ISO timestamp staleness is measured against. */
   asOf: string;
   thresholds?: SentinelThresholds;
   /** Forwarded to `analyzeSeries` (window, stableCv, driftRun). */
   convergence?: SeriesConvergenceOptions;
 }
-interface SentinelTrend {
+export interface SentinelTrend {
   judgeId: string;
   metric: SentinelMetricName;
   /** Verbatim `analyzeSeries` state for this judge×metric history. */
@@ -626,7 +626,7 @@ interface SentinelTrend {
   alarmed: boolean;
   reason?: string;
 }
-interface SentinelReport {
+export interface SentinelReport {
   perJudge: SentinelTrend[];
   alarms: string[];
   healthy: boolean;
@@ -638,8 +638,8 @@ interface SentinelReport {
    */
   insufficientHistory: string[];
 }
-declare function judgeSentinelReport(history: SentinelSnapshot[], opts: JudgeSentinelOptions): SentinelReport;
-interface EvalHealthStamp {
+export declare function judgeSentinelReport(history: SentinelSnapshot[], opts: JudgeSentinelOptions): SentinelReport;
+export interface EvalHealthStamp {
   healthy: boolean;
   alarms: string[];
 }
@@ -651,7 +651,7 @@ interface EvalHealthStamp {
  * unresolved drift / decay / silent-upgrade / staleness alarm — treat the
  * verdicts as carrying an integrity warning until recalibration clears it.
  */
-declare function evalHealthStamp(report: SentinelReport): EvalHealthStamp;
+export declare function evalHealthStamp(report: SentinelReport): EvalHealthStamp;
 //#endregion
-export { CalibrationBin, CalibrationOptions, CalibrationPair, CalibrationReport, type CalibrationResult, type CandidateScore, CatchRateReport, CatchRateStatus, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type CorrelationInterval, CorrelationResult, CorrelationStudyOptions, CorrelationStudyResult, DeploymentOutcome, EvalHealthStamp, EvalMetricSpec, type EvaluatorAdmissionPolicy, type EvaluatorAdmissionReport, type EvaluatorAuditInput, type EvaluatorAuditObservation, type EvaluatorErrorRate, EvidencePerturbation, FileSystemOutcomeStore, FileSystemOutcomeStoreOptions, type GoldenItem, InMemoryOutcomeStore, JudgeSentinelOptions, OutcomeFilter, OutcomeMetricSpec, OutcomePair, type OutcomeReduction, OutcomeStore, OutcomeStoreError, PerturbationKind, PerturbedField, PerturbedPlant, Plant, PlantEvidence, PlantExpectation, PlantKind, PlantKindCounts, PlantManifest, PlantOutcome, type PositionalBiasResult, type ProbabilityPolicyAuditInput, type ProbabilityPolicyAuditReport, RubricOutcomeExclusion, RubricOutcomePair, RubricPredictiveValidityInput, RubricPredictiveValidityReport, RubricRanking, SeedPlantsOptions, SeededGradingSet, type SelfPreferenceResult, SentinelMetricName, SentinelMetrics, SentinelReport, SentinelSetOptions, SentinelSnapshot, SentinelStore, SentinelThresholds, SentinelTrend, SnapshotMeta, UnseededRejection, type VerbosityBiasResult, auditEvaluator, auditProbabilityPolicy, calibrateJudge, calibrateJudgeContinuous, calibrationCurve, calibrationFromPairs, catchRate, continuousAgreement, correlationStudy, definePlant, evalHealthStamp, fileSentinelStore, inMemorySentinelStore, judgeSentinelReport, perturbEvidence, plantByPerturbation, positionalBias, rubricPredictiveValidity, seedPlants, selfPreference, snapshotFromAgreement, snapshotFromCalibration, snapshotFromSentinelSet, validateSentinelSnapshot, verbosityBias };
+export { type CalibrationResult, type CandidateScore, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type CorrelationInterval, DeploymentOutcome, type EvaluatorAdmissionPolicy, type EvaluatorAdmissionReport, type EvaluatorAuditInput, type EvaluatorAuditObservation, type EvaluatorErrorRate, FileSystemOutcomeStore, FileSystemOutcomeStoreOptions, type GoldenItem, InMemoryOutcomeStore, OutcomeFilter, OutcomeMetricSpec, type OutcomeReduction, OutcomeStore, OutcomeStoreError, type PositionalBiasResult, type ProbabilityPolicyAuditInput, type ProbabilityPolicyAuditReport, RubricOutcomeExclusion, RubricOutcomePair, RubricPredictiveValidityInput, RubricPredictiveValidityReport, RubricRanking, type SelfPreferenceResult, type VerbosityBiasResult, calibrateJudge, calibrateJudgeContinuous, continuousAgreement, positionalBias, rubricPredictiveValidity, selfPreference, verbosityBias };
 //# sourceMappingURL=index.d.ts.map

package/dist/meta-eval/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/meta-eval/correlation-study.ts","../../src/meta-eval/calibration.ts","../../src/meta-eval/evaluator-admission.ts","../../src/meta-eval/plants.ts","../../src/meta-eval/sentinel.ts"],"mappings":";;;;;;;;;;UAoBiB;EACf;;;EAGA,WAAW,KAAK,KAAK,OAAO,eAAe;;UAG5B;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA;IAAe;IAAe;;;EAE9B;IAAgB;IAAe;;;EAE/B;;UAGe;EACf,OAAO;;EAEP,eAAe,MACb;IACE;IACA;;EAGJ;EACA;;UAGe;;EAEf;;EAEA,gBAAgB;;EAEhB;;EAEA;;;EAGA;;iBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,aAAa,kBACb,8BACA,UAAS,0BACR,QAAQ;;;UChEM;EACf;EACA;EACA;EACA;EACA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,MAAM;;EAEN;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;IAAU;IAAY;;;UAGP;EACf;EACA;;iBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,YAAY,gBACZ,uBACA,UAAS,qBACR,QAAQ;;iBA0CK,qBACd,qBAAqB,mBACrB,oBACA,uBACA,UAAS,qBACR;;;cC9FG,cAAY,EAAA;;;;GAMP,EAAA,KAAA;;KAGC,2BAA2B,EAAE,aAAa;cAEhD,mBAAiB,EAAA;;;;;;;;;;;;;;;;;GASZ,EAAA,KAAA;;KAGC,4BAA4B,EAAE,aAAa;cAEjD,aAAW,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAeN,EAAA,KAAA;KAEC,sBAAsB,EAAE,aAAa;UAEhC;;EAEf;EACA;;EAEA;;EAEA;EACA;;EAEA;;EAEA,UAAU;EACV;EACA;;UAGe;EACf,iBAAiB;EACjB,cAAc;EACd,aAAa;EACb,cAAc;EACd;EACA;EACA,WAAW;EACX,QAAQ;;EAER;EACA;EACA;EACA;EACA,cAAc;EACd;IACE;IACA;IACA;IACA;IACA;IACA;;EAEF,YAAY;IAAQ;IAAY;IAA2B;;EAC3D,iBAAiB;EACjB,gBAAgB;;;;;;;iBAoEF,eAAe,OAAO,sBAAsB;cA2GtD,wBAAsB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAG1B,EAAA,KAAA;;KAGU,8BAA8B,EAAE,aAAa;UAExC,qCAAqC;EACpD;IACE,uBAAuB;IACvB,YAAY;;IAEZ,cAAc;MAAQ;MAAY;;;;;;;;;iBAStB,uBACd,OAAO,8BACN;;;;;;;;KCxOS;;;;;;;;;;KAkBA;UAOK;;EAEf;EACA,MAAM;;EAEN,MAAM;;EAEN,iBAAiB;;;;;;;;;;;;;;iBAeH,YAAY;EAC1B;EACA,MAAM;EACN,MAAM;EACN,iBAAiB;IACf;;;;;UAqDa;;EAEf;;EAEA;;;KAIU;;KAGA;UAEK;;EAEf,UAAU;;EAEV,OAAO;;EAEP,KAAK;;;;;;EAML;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAsJc,gBAAgB,UAAU,gBAAgB;UA8BzC;EACf,OAAO;;EAEP,UAAU;EACV,OAAO;EACP,KAAK;EACL;EACA;;;;;;;;;;;;;;;;;;;;;;iBAuBc,oBAAoB;EAClC;EACA;EACA,UAAU;;EAEV;EACA;IACE;UA0Ba;;;;;;EAMf,MAAM;;EAEN;;EAEA;;EAEA;;EAEA,QAAQ;;UAGO;;;;;;EAMf,OAAO;EACP,UAAU;;UAGK;;;;;EAKf;;;;;;EAMA;;;;;;;;;;;iBAYc,WACd,kBAAkB,cAClB,iBAAiB,SACjB,UAAS,oBACR;;;;;;UAqGc;EACf;EACA;;;;;;KAOU;UAEK;EACf;EACA;EACA;EACA;;EAEA;;;;;;;;UASe;EACf;EACA;EACA;;EAEA;;UAGe;EACf,QAAQ;;EAER;EACA;EACA;EACA;;;;;EAKA;;EAEA;;EAEA,QAAQ,QAAQ,OAAO,WAAW;;EAElC;;EAEA;EACA,UAAU;;;;;;;;;;iBAWI,UACd,kBAAkB,gBAClB,UAAU,gBACT;;;cC/lBG;KAEM,6BAA6B;UAExB;;EAEf;;EAEA;;EAEA;;UAGe;;EAEf;EACA;;EAEA;;;;;;;EAOA,SAAS;;;UAIM;EACf;EACA;EACA;;;iBAcc,yBAAyB,UAAU,kBAAkB;;;;;;iBAiCrD,wBACd,QAAQ,oBAAoB,6BAC5B,MAAM,eACL;;;;;;;;iBAmBa,sBACd,QAAQ,wBAAwB,qBAChC,MAAM,eACL;UAYc;;EAEf;;;;;;;;;iBAUc,wBACd,QAAQ,kBACR,QAAQ,cACR,MAAM,cACN,UAAS,qBACR;UA4Cc;EACf,OAAO,UAAU,mBAAmB;;EAEpC,QAAQ,mBAAmB,QAAQ;;iBAGrB,sBAAsB,UAAS,qBAA0B;;;;;;;iBAqBzD,kBAAkB,eAAe;UA0ChC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;;EAEf;EACA,aAAa;;EAEb,cAAc;;UAGC;EACf;EACA,QAAQ;;EAER,OAAO;;EAEP;;EAEA;;EAEA;EACA;EACA;;UAGe;EACf,UAAU;EACV;EACA;;;;;;;EAOA;;iBAKc,oBACd,SAAS,oBACT,MAAM,uBACL;UA6Gc;EACf;EACA;;;;;;;;;;iBAWc,gBAAgB,QAAQ,iBAAiB"}
+{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/meta-eval/correlation-study.ts","../../src/meta-eval/calibration.ts","../../src/meta-eval/evaluator-admission.ts","../../src/meta-eval/plants.ts","../../src/meta-eval/sentinel.ts"],"mappings":";;;;;;;;;;iBAoBiB;EACf;;;EAGA,WAAW,KAAK,KAAK,OAAO,eAAe;;iBAG5B;EACf;EACA;;iBAGe;EACf;EACA;EACA;EACA;EACA;;EAEA;IAAe;IAAe;;;EAE9B;IAAgB;IAAe;;;EAE/B;;iBAGe;EACf,OAAO;;EAEP,eAAe,MACb;IACE;IACA;;EAGJ;EACA;;iBAGe;;EAEf;;EAEA,gBAAgB;;EAEhB;;EAEA;;;EAGA;;wBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,aAAa,kBACb,8BACA,UAAS,0BACR,QAAQ;;;iBChEM;EACf;EACA;EACA;EACA;EACA;;EAEA;;iBAGe;EACf;EACA;EACA;EACA,MAAM;;EAEN;;EAEA;;iBAGe;;EAEf;;EAEA;;EAEA;IAAU;IAAY;;;iBAGP;EACf;EACA;;wBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,YAAY,gBACZ,uBACA,UAAS,qBACR,QAAQ;;wBA0CK,qBACd,qBAAqB,mBACrB,oBACA,uBACA,UAAS,qBACR;;;cC9FG,cAAY,EAAA;;;;GAMP,EAAA,KAAA;;KAGC,2BAA2B,EAAE,aAAa;cAEhD,mBAAiB,EAAA;;;;;;;;;;;;;;;;;GASZ,EAAA,KAAA;;KAGC,4BAA4B,EAAE,aAAa;cAEjD,aAAW,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAeN,EAAA,KAAA;KAEC,sBAAsB,EAAE,aAAa;UAEhC;;EAEf;EACA;;EAEA;;EAEA;EACA;;EAEA;;EAEA,UAAU;EACV;EACA;;UAGe;EACf,iBAAiB;EACjB,cAAc;EACd,aAAa;EACb,cAAc;EACd;EACA;EACA,WAAW;EACX,QAAQ;;EAER;EACA;EACA;EACA;EACA,cAAc;EACd;IACE;IACA;IACA;IACA;IACA;IACA;;EAEF,YAAY;IAAQ;IAAY;IAA2B;;EAC3D,iBAAiB;EACjB,gBAAgB;;;;;;;wBAoEF,eAAe,OAAO,sBAAsB;cA2GtD,wBAAsB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAG1B,EAAA,KAAA;;KAGU,8BAA8B,EAAE,aAAa;UAExC,qCAAqC;EACpD;IACE,uBAAuB;IACvB,YAAY;;IAEZ,cAAc;MAAQ;MAAY;;;;;;;;;wBAStB,uBACd,OAAO,8BACN;;;;;;;;YCxOS;;;;;;;;;;YAkBA;iBAOK;;EAEf;EACA,MAAM;;EAEN,MAAM;;EAEN,iBAAiB;;;;;;;;;;;;;;wBAeH,YAAY;EAC1B;EACA,MAAM;EACN,MAAM;EACN,iBAAiB;IACf;;;;;iBAqDa;;EAEf;;EAEA;;;YAIU;;YAGA;iBAEK;;EAEf,UAAU;;EAEV,OAAO;;EAEP,KAAK;;;;;;EAML;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;wBAsJc,gBAAgB,UAAU,gBAAgB;iBA8BzC;EACf,OAAO;;EAEP,UAAU;EACV,OAAO;EACP,KAAK;EACL;EACA;;;;;;;;;;;;;;;;;;;;;;wBAuBc,oBAAoB;EAClC;EACA;EACA,UAAU;;EAEV;EACA;IACE;iBA0Ba;;;;;;EAMf,MAAM;;EAEN;;EAEA;;EAEA;;EAEA,QAAQ;;iBAGO;;;;;;EAMf,OAAO;EACP,UAAU;;iBAGK;;;;;EAKf;;;;;;EAMA;;;;;;;;;;;wBAYc,WACd,kBAAkB,cAClB,iBAAiB,SACjB,UAAS,oBACR;;;;;;iBAqGc;EACf;EACA;;;;;;YAOU;iBAEK;EACf;EACA;EACA;EACA;;EAEA;;;;;;;;iBASe;EACf;EACA;EACA;;EAEA;;iBAGe;EACf,QAAQ;;EAER;EACA;EACA;EACA;;;;;EAKA;;EAEA;;EAEA,QAAQ,QAAQ,OAAO,WAAW;;EAElC;;EAEA;EACA,UAAU;;;;;;;;;;wBAWI,UACd,kBAAkB,gBAClB,UAAU,gBACT;;;cC/lBG;YAEM,6BAA6B;iBAExB;;EAEf;;EAEA;;EAEA;;iBAGe;;EAEf;EACA;;EAEA;;;;;;;EAOA,SAAS;;;iBAIM;EACf;EACA;EACA;;;wBAcc,yBAAyB,UAAU,kBAAkB;;;;;;wBAiCrD,wBACd,QAAQ,oBAAoB,6BAC5B,MAAM,eACL;;;;;;;;wBAmBa,sBACd,QAAQ,wBAAwB,qBAChC,MAAM,eACL;iBAYc;;EAEf;;;;;;;;;wBAUc,wBACd,QAAQ,kBACR,QAAQ,cACR,MAAM,cACN,UAAS,qBACR;iBA4Cc;EACf,OAAO,UAAU,mBAAmB;;EAEpC,QAAQ,mBAAmB,QAAQ;;wBAGrB,sBAAsB,UAAS,qBAA0B;;;;;;;wBAqBzD,kBAAkB,eAAe;iBA0ChC;;EAEf;;EAEA;;EAEA;;EAEA;;iBAGe;;EAEf;EACA,aAAa;;EAEb,cAAc;;iBAGC;EACf;EACA,QAAQ;;EAER,OAAO;;EAEP;;EAEA;;EAEA;EACA;EACA;;iBAGe;EACf,UAAU;EACV;EACA;;;;;;;EAOA;;wBAKc,oBACd,SAAS,oBACT,MAAM,uBACL;iBA6Gc;EACf;EACA;;;;;;;;;;wBAWc,gBAAgB,QAAQ,iBAAiB"}

package/dist/meta-eval/index.js

@@ -1,12 +1,12 @@
 import { n as CaptureIntegrityError, s as ValidationError } from "../errors-Dngq5h35.js";
 import { a as hashCanonical, i as compareCodeUnits } from "../canonical-BsZVgBhW.js";
 import { t as mulberry32 } from "../random-Dn5fPWkt.js";
-import { a as selfPreference, i as positionalBias, n as calibrateJudgeContinuous, o as verbosityBias, r as continuousAgreement, t as calibrateJudge } from "../judge-calibration-DYmaBtJr.js";
+import { a as selfPreference, i as positionalBias, n as calibrateJudgeContinuous, o as verbosityBias, r as continuousAgreement, t as calibrateJudge } from "../judge-calibration-C5zgtiJT.js";
 import { u as runMetricExtractor } from "../query-DlV0pdx4.js";
-import { r as computeInterval } from "../ast-CP9ae9B0.js";
+import { r as computeInterval } from "../ast-DJ6Bc7ab.js";
 import { t as analyzeSeries } from "../series-convergence-CjO2QdRW.js";
-import { a as reduceOutcomeMetric, i as hasVariation, n as assertUniqueObservationIds, o as validateObservationOptions, r as correlationSummary, t as rubricPredictiveValidity } from "../rubric-predictive-validity-CCK-1B7w.js";
-import { n as InMemoryOutcomeStore, r as OutcomeStoreError, t as FileSystemOutcomeStore } from "../outcome-store-BXlkwMPR.js";
+import { a as reduceOutcomeMetric, i as hasVariation, n as assertUniqueObservationIds, o as validateObservationOptions, r as correlationSummary, t as rubricPredictiveValidity } from "../rubric-predictive-validity-B0Pg-3I5.js";
+import { n as InMemoryOutcomeStore, r as OutcomeStoreError, t as FileSystemOutcomeStore } from "../outcome-store-DUcMj-3D.js";
 import { z } from "zod";
 //#region src/meta-eval/calibration.ts
 /**