@tangle-network/agent-eval 0.221.8 → 0.222.0

package/dist/{llm-judge-Cybq29bL.js → llm-judge-cgbhTKQp.js}

@@ -1,25 +1,29 @@
-import { i as JudgeError, s as ValidationError } from "./errors-BcOs5MKW.js";
-import { a as hashCanonical, i as compareCodeUnits, r as canonicalString, t as LEDGER_HASH_PATTERN } from "./canonical-Dq2hOQ8L.js";
-import { i as modelHasSnapshot, s as searchCellRunId } from "./run-record-DhEJgoPz.js";
-import { t as mulberry32 } from "./random-DDDcC7YU.js";
-import { o as summarizeNumberSeries, s as weightedComposite, t as confidenceInterval } from "./descriptive-CVNy48rR.js";
-import { r as minimumPairsForPairedDeltaTest } from "./paired-promotion-decision-D0SHb-ZL.js";
-import { a as heldoutSignificance, i as dimensionRegressions, o as pairHoldout } from "./statistical-heldout-Bksv92Th.js";
-import { n as contentHash } from "./verdict-cache-CL_7tePD.js";
-import { r as projectCampaignCellQuality, t as campaignCellCostProvenance } from "./run-record-CysKLB93.js";
-import { n as CostAccountingIncompleteError, t as CostLedger } from "./cost-ledger-BGZB6yY2.js";
-import { A as surfaceHashMatches, D as surfaceContentHash, E as renderSurfaceDiff, F as assertCampaignSplitIdentity, I as assertCompleteCampaign, L as campaignCoverage, O as surfaceDispatchRef, P as assertCampaignDesign, R as campaignScenarioIdentity, V as formatCoverageFailures, b as FinalEvidenceConflictError, f as campaignBreakdown, j as BackendIntegrityError, k as surfaceHash, m as campaignMeanCompositeOrNull, p as campaignMeanComposite, x as FinalEvidenceError, z as campaignSplitDigest } from "./campaign-evidence-DunzcJGv.js";
-import { t as deepFreezeCanonicalJson } from "./deep-freeze-CurcOQg3.js";
-import { M as fsCampaignStorage, h as isExternalTextCandidate, j as createRunCostLedger } from "./external-optimizer-subprocess-B_ukPtRD.js";
-import { r as summarizeEvaluationUnits, t as defineEvaluationClaim } from "./claim-BPVuyHuy.js";
-import { d as detectPersonalData, s as redactText, t as REDACTION_VERSION, u as detectCredential } from "./redact-mtkTiJVk.js";
-import { t as detectRewardHacking } from "./reward-hacking-ByuETaIb.js";
-import { t as assertProposalFindings } from "./proposal-findings-DqLUWsAG.js";
-import { i as runSearch, m as searchExpansionIndex, n as paretoFrontier, o as searchPolicyView, p as SEARCH_KERNEL_SOURCE, r as paretoFrontierWithCrowding, s as EXPANSION_OPERATORS, w as estimateNode } from "./pareto-BiBnQXAE.js";
-import { a as validateSearchLedgerEvent, d as searchNodeId, f as searchTaskSetDigest, i as replaySearchLedgerText, n as openSearchLedger, u as searchCellId } from "./search-ledger-CZRqtFbW.js";
-import { n as searchProposerView, t as renderSearchSummary } from "./search-summary-D4bwYnAE.js";
-import { i as stripFencedJson, n as isRecord } from "./json-BhdsJaOQ.js";
-import { a as maximumChargeForLlmRequest, i as costReceiptFromLlmError, r as costReceiptFromLlm } from "./llm-client-b1ysZVIU.js";
+import { s as ValidationError } from "./errors-DeEEV9Vx.js";
+import { a as hashCanonical, i as compareCodeUnits, r as canonicalString, t as LEDGER_HASH_PATTERN } from "./canonical-CicJlINA.js";
+import { i as modelHasSnapshot, s as searchCellRunId } from "./run-record-DEjx43cI.js";
+import { t as mulberry32 } from "./random-DrzcgC7F.js";
+import { o as summarizeNumberSeries, s as weightedComposite, t as confidenceInterval } from "./descriptive-cPwITVfy.js";
+import { t as mean$1 } from "./mean-CdDYG-us.js";
+import { r as minimumPairsForPairedDeltaTest } from "./paired-promotion-decision-FGvHA4p_.js";
+import { a as heldoutSignificance, i as dimensionRegressions, o as pairHoldout } from "./statistical-heldout-TP6ndL0B.js";
+import { d as mapConcurrent } from "./journal-C1uLJi0w.js";
+import { n as contentHash } from "./verdict-cache-BVH-fl18.js";
+import { r as projectCampaignCellQuality, t as campaignCellCostProvenance } from "./run-record-CZy4-TnZ.js";
+import { n as CostAccountingIncompleteError, t as CostLedger } from "./cost-ledger-CS8R3IZj.js";
+import { A as campaignSplitDigest, D as assertCompleteCampaign, E as assertCampaignSplitIdentity, M as formatCoverageFailures, O as campaignCoverage, S as BackendIntegrityError, T as assertCampaignDesign, b as FinalEvidenceError, f as campaignBreakdown, k as campaignScenarioIdentity, m as campaignMeanCompositeOrNull, p as campaignMeanComposite, y as FinalEvidenceConflictError } from "./campaign-evidence-DN8rWuBj.js";
+import { t as deepFreezeCanonicalJson } from "./deep-freeze-CK8-2W7p.js";
+import { M as fsCampaignStorage, h as isExternalTextCandidate, j as createRunCostLedger } from "./external-optimizer-subprocess-U616hV_l.js";
+import { c as surfaceDispatchRef, l as surfaceHash, o as renderSurfaceDiff, s as surfaceContentHash, u as surfaceHashMatches } from "./json-recovery-DK4MDm5r.js";
+import { r as summarizeEvaluationUnits, t as defineEvaluationClaim } from "./claim-Dod_mjk3.js";
+import { d as detectPersonalData, s as redactText, t as REDACTION_VERSION, u as detectCredential } from "./redact-D6APJS1d.js";
+import { t as detectRewardHacking } from "./reward-hacking-DZzHZlS1.js";
+import { t as assertProposalFindings } from "./proposal-findings-WFQrxnfe.js";
+import { i as runSearch, m as searchExpansionIndex, n as paretoFrontier, o as searchPolicyView, p as SEARCH_KERNEL_SOURCE, r as paretoFrontierWithCrowding, s as EXPANSION_OPERATORS, w as estimateNode } from "./pareto-B9udAt2i.js";
+import { a as validateSearchLedgerEvent, d as searchNodeId, f as searchTaskSetDigest, i as replaySearchLedgerText, n as openSearchLedger, u as searchCellId } from "./search-ledger-BsVi-Ngv.js";
+import { n as searchProposerView, t as renderSearchSummary } from "./search-summary-Ct7SrmVc.js";
+import { i as stripFencedJson, n as isRecord } from "./json-DtvZNAU9.js";
+import { a as maximumChargeForLlmRequest, i as costReceiptFromLlmError, r as costReceiptFromLlm } from "./llm-client-kJ9EOqjt.js";
+import { t as JudgeParseError } from "./judges-BLxg5p8E.js";
 import { createHash } from "node:crypto";
 import { z } from "zod";
 import { writeFileSync } from "node:fs";
@@ -27,7 +31,7 @@ import { basename, dirname, isAbsolute, join, relative, resolve, sep } from "nod
 import { homedir, tmpdir } from "node:os";
 import { execSync } from "node:child_process";
 import { fileURLToPath, pathToFileURL } from "node:url";
-//#region src/campaign/campaign-manifest.ts
+// src/campaign/campaign-manifest.ts
 /**
 * Campaign identity: the manifest hash over (scenarios, judges, dispatch,
 * seed, reps) that keys resumability, and the dispatch-ref rules feeding it.
@@ -61,8 +65,7 @@ function dispatchRefFor(dispatch, override) {
 	if (typeof ref !== "string" || ref.trim().length === 0) throw new Error("runCampaign: dispatchRef must be a non-empty string when provided");
 	return ref;
 }
-//#endregion
-//#region src/campaign/cell-aggregates.ts
+// src/campaign/cell-aggregates.ts
 function computeAggregates(cells, judges, seed, cost) {
 	const cellQuality = cells.map((cell) => ({
 		cell,
@@ -134,8 +137,7 @@ function aggregate(samples, seed) {
 		distribution
 	};
 }
-//#endregion
-//#region src/campaign/cell-schedule.ts
+// src/campaign/cell-schedule.ts
 /**
 * The campaign cell schedule: the (scenario × rep) fan-out with seed-group
 * aware per-cell seeds, plus the cell directory / cache-path / cost-tag
@@ -191,8 +193,7 @@ function stableCostTagsFor(opts, slot) {
 		rep: String(slot.rep)
 	};
 }
-//#endregion
-//#region src/campaign/cell-cache.ts
+// src/campaign/cell-cache.ts
 function assertScheduleCachesReusable(args) {
 	const blocked = [];
 	for (const slot of args.schedule) {
@@ -298,8 +299,7 @@ function invalidCachedCellsError(cells) {
 	const details = cells.map((cell) => `${cell.cellId} (${cell.reason}${cell.detail ? `: ${cell.detail}` : ""})`).join(", ");
 	return new CostAccountingIncompleteError(`runCampaign: cached cell(s) require explicit paid re-dispatch: ${details}; refusing to begin campaign. Inspect planCampaignRun, then set rerunInvalidCachedCells: true to rerun only these cells while retaining valid caches, or resumable: false when a full rerun is intended.`);
 }
-//#endregion
-//#region src/campaign/judge-cell.ts
+// src/campaign/judge-cell.ts
 /**
 * Judge scoring for one cell, with the paid-call accounting guard: a judge
 * that reports an LLM call without a CostLedger receipt fails the cell.
@@ -329,8 +329,7 @@ function assertReportedJudgeCallRecorded(judgeName, value, input, previousCallId
 function hasLlmCall(value) {
 	return typeof value === "object" && value !== null && "llmCall" in value && value.llmCall !== void 0;
 }
-//#endregion
-//#region src/campaign/execute-cell.ts
+// src/campaign/execute-cell.ts
 /**
 * One campaign cell, end to end: cache reuse, dispatch with per-cell abort +
 * timeout, judge scoring, cost settlement proof, failure receipts, and the
@@ -639,7 +638,7 @@ async function executeCell(args) {
 	const cachePath = join(cellDir, "cached-result.json");
 	if (args.resumable) {
 		const reused = reuseCachedCell(args, cachePath, stableCostTags);
-		if (reused) return reused;
+		if (reused && (reused.cell.error || !args.cellRetry?.retryResult?.(reused.cell.artifact))) return reused;
 	}
 	const startMs = Date.now();
 	const run = openAttempt(args, cellDir, attemptCostTags);
@@ -694,12 +693,13 @@ async function executeCell(args) {
 		...failure?.judge ? { errorJudge: failure.judge } : {},
 		error: errorMessage
 	};
-	const retry = failure ? recordCellFailure(args, run, failure, cell, costTags, fatalCellError) : false;
+	let retry = failure ? recordCellFailure(args, run, failure, cell, costTags, fatalCellError) : false;
 	await trace.flush();
 	const resultPath = join(attemptDir, "result.json");
 	storage.write(resultPath, JSON.stringify(cell));
 	artifactsByPath[`${args.slot.cellId}/${attemptPrefix}/result.json`] = resultPath;
-	if (!errorMessage && args.resumable) storage.write(cachePath, JSON.stringify(cell));
+	if (!failure) retry = args.cellRetry !== void 0 && args.attempt < args.cellRetry.attempts && !args.signal.aborted && args.cellRetry.retryResult?.(cell.artifact) === true;
+	if (!errorMessage && args.resumable && !retry) storage.write(cachePath, JSON.stringify(cell));
 	if (fatalCellError !== void 0) throw fatalCellError;
 	return {
 		cell,
@@ -802,8 +802,7 @@ function defaultBuildTraceWriter(storage) {
 		};
 	};
 }
-//#endregion
-//#region src/campaign/run-dir.ts
+// src/campaign/run-dir.ts
 /** The shared, out-of-repo root for campaign/benchmark run bundles. Keeping run
 *  outputs here means they never land in a repo working tree (no per-repo
 *  gitignore, no clutter, no accidental commits). Layout:
@@ -821,8 +820,7 @@ function resolveRunDir(runDir, repo) {
 	const r = repo && repo.trim().length > 0 ? repo : basename(process.cwd());
 	return join(tangleTracesRoot(), r, "runs", runDir);
 }
-//#endregion
-//#region src/campaign/plan-campaign-run.ts
+// src/campaign/plan-campaign-run.ts
 /**
 * Plan a campaign WITHOUT dispatching: computes the manifest hash and the per-cell
 * run-vs-cached schedule so callers can preview cost and resumability before spending.
@@ -910,8 +908,17 @@ function planCampaignRun(opts) {
 		cells
 	};
 }
-//#endregion
-//#region src/campaign/run-campaign.ts
+// src/campaign/run-campaign.ts
+/**
+* `runCampaign` — Pass A substrate primitive. ONE function that orchestrates
+* scenarios → dispatch → artifacts → judges → aggregates, with full
+* reproducibility (seed + manifest hash), cell-level resumability, bootstrap
+* CIs, and the `LabeledScenarioStore` capture flywheel.
+*
+* Improvement loops (optimizer / gate / autoOnPromote) ride on top of this
+* primitive but live in `presets/run-improvement-loop.ts`. This file keeps
+* the core orchestrator minimal — Phase 1 of the Pass A track.
+*/
 /**
 * Core campaign orchestrator: fan scenarios through dispatch, score with judges, aggregate bootstrap CIs, and persist reproducible `CampaignResult` records.
 */
@@ -930,6 +937,7 @@ async function runCampaign(opts) {
 	if (cellRetry !== void 0) {
 		if (!Number.isSafeInteger(cellRetry.attempts) || cellRetry.attempts < 1) throw new Error("runCampaign: cellRetry.attempts must be a positive safe integer (total attempts, including the first)");
 		if (typeof cellRetry.retryable !== "function") throw new Error("runCampaign: cellRetry.retryable must be a function");
+		if (cellRetry.retryResult !== void 0 && typeof cellRetry.retryResult !== "function") throw new Error("runCampaign: cellRetry.retryResult must be a function");
 	}
 	if (typeof opts.runDir !== "string" || opts.runDir.trim().length === 0) throw new Error("runCampaign: runDir is required and must be a non-empty string");
 	opts.runDir = resolveRunDir(opts.runDir, opts.repo);
@@ -971,18 +979,12 @@ async function runCampaign(opts) {
 	if (opts.signal?.aborted) campaignAbort.abort(opts.signal.reason);
 	else opts.signal?.addEventListener("abort", onOwnerAbort, { once: true });
 	const campaignSignal = campaignAbort.signal;
-	const lanes = [];
-	let nextIdx = 0;
 	const cellsRef = cells;
-	let firstLaneError;
+	let firstLaneFailure;
 	let firstCellFailure;
-	for (let i = 0; i < maxConcurrency; i++) lanes.push((async () => {
-		try {
-			while (true) {
-				if (campaignSignal.aborted) return;
-				const myIdx = nextIdx++;
-				if (myIdx >= schedule.length) return;
-				const slot = schedule[myIdx];
+	try {
+		await mapConcurrent(schedule, maxConcurrency, async (slot, _index, signal) => {
+			try {
 				let attempt = 1;
 				let result;
 				while (true) {
@@ -994,7 +996,7 @@ async function runCampaign(opts) {
 						now,
 						storage,
 						buildTraceWriter: opts.buildTraceWriter ?? defaultBuildTraceWriter(storage),
-						signal: campaignSignal,
+						signal,
 						dispatchTimeoutMs: opts.dispatchTimeoutMs,
 						dispatchShutdownTimeoutMs,
 						costLedger,
@@ -1014,30 +1016,31 @@ async function runCampaign(opts) {
 					attempt += 1;
 				}
 				cellsRef.push(result.cell);
-				if (opts.labeledStore && opts.labeledStore !== "off" && !result.cell.error) await captureToStore({
-					store: opts.labeledStore,
-					cell: result.cell,
+				await opts.onCellSettled?.(result.cell, slot.scenario);
+				if (opts.labeledStore && opts.labeledStore !== "off" && !result.cell.error) await opts.labeledStore.observe({
 					scenario: slot.scenario,
-					opts,
-					now
-				}).catch((err) => {
-					console.warn(`[runCampaign] capture failed for ${result.cell.cellId}: ${err instanceof Error ? err.message : String(err)}`);
+					artifact: result.cell.artifact,
+					judgeScores: result.cell.judgeScores,
+					source: opts.captureSource ?? "eval-run",
+					sourceVersionHash: opts.captureSourceVersionHash ?? "unknown",
+					capturedAt: now().toISOString(),
+					redactionStatus: "raw"
+				}).catch((error) => {
+					console.warn(`[runCampaign] capture failed for ${result.cell.cellId}: ${error instanceof Error ? error.message : String(error)}`);
 				});
-				if (opts.abortOnCellError && result.failure) return;
-			}
-		} catch (error) {
-			if (firstLaneError === void 0) {
-				firstLaneError = error;
+			} catch (error) {
+				firstLaneFailure ??= { cause: error };
 				campaignAbort.abort(error);
+				throw error;
 			}
-			throw error;
-		}
-	})());
-	const laneResults = await Promise.allSettled(lanes);
-	opts.signal?.removeEventListener("abort", onOwnerAbort);
-	const failedLane = laneResults.find((result) => result.status === "rejected");
-	if (firstCellFailure) throw firstCellFailure.cause;
-	if (failedLane) throw firstLaneError ?? failedLane.reason;
+		}, campaignSignal);
+	} catch (error) {
+		if (firstCellFailure) throw firstCellFailure.cause;
+		if (firstLaneFailure) throw firstLaneFailure.cause;
+		if (!opts.signal?.aborted) throw error;
+	} finally {
+		opts.signal?.removeEventListener("abort", onOwnerAbort);
+	}
 	const endedAt = now();
 	cellsRef.sort((a, b) => compareCodeUnits(a.cellId, b.cellId));
 	const aggregates = computeAggregates(cellsRef, judges, seed, costLedger.summary({ tags: { runDir: opts.runDir } }));
@@ -1056,19 +1059,7 @@ async function runCampaign(opts) {
 		scenarios: opts.scenarios.map(campaignScenarioIdentity)
 	};
 }
-async function captureToStore(args) {
-	await args.store.observe({
-		scenario: args.scenario,
-		artifact: args.cell.artifact,
-		judgeScores: args.cell.judgeScores,
-		source: args.opts.captureSource ?? "eval-run",
-		sourceVersionHash: args.opts.captureSourceVersionHash ?? "unknown",
-		capturedAt: args.now().toISOString(),
-		redactionStatus: "raw"
-	});
-}
-//#endregion
-//#region src/campaign/presets/run-eval.ts
+// src/campaign/presets/run-eval.ts
 /**
 * `runEval` — the simplest preset over `runCampaign`. No optimizer, no
 * gate, no auto-PR. Just: run scenarios through dispatch, score with
@@ -1082,8 +1073,7 @@ async function captureToStore(args) {
 async function runEval(opts) {
 	return runCampaign(opts);
 }
-//#endregion
-//#region src/campaign/final-evidence.ts
+// src/campaign/final-evidence.ts
 /** Capture identities before an asynchronous search can mutate its caller's options. */
 function captureFinalEvidencePolicy(policy) {
 	return Object.freeze({
@@ -1151,8 +1141,7 @@ async function exposeFinalEvidence(policy, claim, scenarios, surfaces) {
 		record: result.value.record
 	};
 }
-//#endregion
-//#region src/canary.ts
+// src/canary.ts
 /**
 * Run all configured canaries against a chronological run list.
 * Runs MUST be sorted oldest-to-newest by the caller — the order of
@@ -1266,8 +1255,8 @@ function detectCalibrationDrift(runs, opts) {
 				alpha,
 				recentN: recent.length,
 				historyN: historical.length,
-				recentMean: mean$1(recent),
-				historyMean: mean$1(historical)
+				recentMean: mean(recent),
+				historyMean: mean(historical)
 			}
 		}],
 		evaluation: {
@@ -1504,12 +1493,11 @@ function chiSquareCritical(df, alpha) {
 	}
 	return TABLE[10][idx];
 }
-function mean$1(xs) {
+function mean(xs) {
 	if (xs.length === 0) return 0;
 	return xs.reduce((s, x) => s + x, 0) / xs.length;
 }
-//#endregion
-//#region src/dataset.ts
+// src/dataset.ts
 /** Locked holdouts — throws on mutate. Callers that need a mutable dataset fork it. */
 var HoldoutLockedError = class extends ValidationError {
 	constructor(datasetName) {
@@ -1645,8 +1633,7 @@ function seededShuffle(items, seed) {
 	}
 	return out;
 }
-//#endregion
-//#region src/red-team.ts
+// src/red-team.ts
 const REFUSAL_MARKERS = [
 	/\b(?:i\s+(?:can(?:not|'t)|won't|will\s+(?:not|never)|am\s+unable|refuse))\b/i,
 	/\b(?:sorry,?\s+(?:but\s+)?i\s+(?:can't|cannot))/i,
@@ -1955,8 +1942,7 @@ function excerptAt(source, at, needleLength) {
 	const end = Math.min(source.length, at + needleLength + 30);
 	return (start > 0 ? "…" : "") + source.slice(start, end) + (end < source.length ? "…" : "");
 }
-//#endregion
-//#region src/campaign/gates/default-production-checks.ts
+// src/campaign/gates/default-production-checks.ts
 function newFindings() {
 	const reasons = [];
 	const contributing = [];
@@ -2258,8 +2244,7 @@ function extractText(artifact) {
 		if (typeof rec.content === "string") return rec.content;
 	}
 }
-//#endregion
-//#region src/campaign/gates/default-production-gate.ts
+// src/campaign/gates/default-production-gate.ts
 /**
 * Opinionated production gate composing held-out significance, red-team, reward-hacking, and canary checks into a single `Gate.decide` decision.
 */
@@ -2314,8 +2299,7 @@ function verdictOf({ reasons, contributing, requiredUnavailable }, delta) {
 		delta
 	};
 }
-//#endregion
-//#region src/campaign/judge-snapshot.ts
+// src/campaign/judge-snapshot.ts
 /** Capture configuration and callbacks; receiver and closed-over state remain caller-owned. */
 function captureJudge(judge) {
 	const dimensions = judge.dimensions.map((dimension) => Object.freeze({
@@ -2337,8 +2321,7 @@ function captureJudge(judge) {
 		appliesTo: captured.appliesTo?.bind(judge)
 	});
 }
-//#endregion
-//#region src/campaign/auto-pr.ts
+// src/campaign/auto-pr.ts
 /**
 * `openAutoPr` — thin shell-out helper for the `runImprovementLoop` preset's
 * `autoOnPromote: 'pr'` mode. Substitutes for the per-product PR-opening
@@ -2464,8 +2447,7 @@ function quoteArg(arg) {
 	if (/^[a-zA-Z0-9_/\-:.@]+$/.test(arg)) return arg;
 	return `"${arg.replace(/"/g, "\\\"")}"`;
 }
-//#endregion
-//#region src/campaign/presets/run-final-comparison.ts
+// src/campaign/presets/run-final-comparison.ts
 /** Compare a search-selected surface without changing the search's selection. */
 async function runFinalComparison(opts) {
 	opts = {
@@ -2622,8 +2604,7 @@ async function runFinalComparison(opts) {
 		promotedDiff
 	};
 }
-//#endregion
-//#region src/campaign/allocation.ts
+// src/campaign/allocation.ts
 /**
 * Allocation: where a search spends rollouts.
 *
@@ -3005,8 +2986,7 @@ function nonNegativeInteger(owner, name, value) {
 function round(value) {
 	return Math.round(value * 1e4) / 1e4;
 }
-//#endregion
-//#region src/campaign/search-campaign-cell.ts
+// src/campaign/search-campaign-cell.ts
 /**
 * What `cell-settled` records for one campaign cell: `passed` with the
 * composite (and each judge's) when every judge scored it, else `errored`,
@@ -3135,8 +3115,7 @@ function searchProposalExecution(receipts, provider, source) {
 		source
 	};
 }
-//#endregion
-//#region src/campaign/search-history-receipt.ts
+// src/campaign/search-history-receipt.ts
 const SEARCH_HISTORY_RECEIPT_SCHEMA_VERSION = "2026-09";
 const SEARCH_HISTORY_RECEIPT_DIGEST_ALGORITHM = "rfc8785-sha256";
 function assertSearchHistoryAdmissionOptions(options) {
@@ -3300,8 +3279,7 @@ function nonEmpty(value, label) {
 	if (typeof value !== "string" || value.length === 0 || value.trim() !== value) throw new TypeError(`${label} must be a trimmed non-empty string`);
 	return value;
 }
-//#endregion
-//#region src/campaign/search-ledger-recording.ts
+// src/campaign/search-ledger-recording.ts
 /**
 * `SearchRecorder`: writes a search into its ledger as it happens.
 *
@@ -3697,8 +3675,7 @@ function canonicalHeader(header) {
 		occurredAt: "2000-01-01T00:00:00.000Z"
 	});
 }
-//#endregion
-//#region src/campaign/search-policy.ts
+// src/campaign/search-policy.ts
 /**
 * Expansion policies: which node a search expands next, with which operator,
 * and which node it currently keeps.
@@ -4097,7 +4074,7 @@ function hillClimb(spec) {
 					evidence: {
 						...chosen.evidence,
 						units: units.length,
-						...units.length > 0 ? { mean: mean(units.map((unit) => unit.mean)) } : {},
+						...units.length > 0 ? { mean: mean$1(units.map((unit) => unit.mean)) } : {},
 						isLeader: chosen.nodeId === kept ? 1 : 0
 					}
 				}
@@ -4153,7 +4130,7 @@ function paretoMembers(view) {
 		members.push({
 			nodeId,
 			ordinal,
-			mean: sign * mean([...objectives.values()]),
+			mean: sign * mean$1([...objectives.values()]),
 			objectives,
 			distance: 0
 		});
@@ -4175,20 +4152,13 @@ function compareCrowded(a, b) {
 	if (a.mean !== b.mean) return a.mean > b.mean ? -1 : 1;
 	return a.ordinal - b.ordinal;
 }
-function mean(values) {
-	let total = 0;
-	for (const value of values) total += value;
-	return total / values.length;
-}
-//#endregion
-//#region src/campaign/types.ts
+// src/campaign/types.ts
 /** Type guard: a proposal carrying its rationale vs a bare
 *  surface. The loop branches on this to populate `GenerationCandidate`. */
 function isProposedCandidate(value) {
 	return typeof value === "object" && value !== null && "surface" in value && "label" in value && "rationale" in value;
 }
-//#endregion
-//#region src/campaign/presets/optimization-surface-nodes.ts
+// src/campaign/presets/optimization-surface-nodes.ts
 /** Surface-node campaigns, cached cell adoption, and proposer history from the search ledger. */
 /** A mutable surface as a search artifact. */
 const surfaceCodec = {
@@ -4532,8 +4502,7 @@ function toScoredSurfaceOutcome(hash, campaign, coverage, generation) {
 		}
 	};
 }
-//#endregion
-//#region src/campaign/presets/run-optimization.ts
+// src/campaign/presets/run-optimization.ts
 /**
 * `runOptimization` improves a mutable surface with a caller-owned proposer,
 * as a search on the one kernel (`runSearch`).
@@ -4934,8 +4903,7 @@ function validatedPremeasuredBaseline(args) {
 	if (!coverage.complete) throw new Error(`runOptimization: premeasured baseline is incomplete (${coverage.scorableCellIds.length}/${coverage.expectedCellIds.length} designed cells scorable) — ${formatCoverageFailures(coverage)}. Refusing to optimize against an incomplete incumbent.`);
 	return campaign;
 }
-//#endregion
-//#region src/campaign/presets/run-improvement-loop.ts
+// src/campaign/presets/run-improvement-loop.ts
 /**
 * Run a caller-owned candidate generator, compare its winner with the starting
 * surface on separate cases, apply a release rule, and optionally open a pull
@@ -5009,8 +4977,7 @@ async function runImprovementLoop(opts) {
 		cost: costLedger.summary()
 	};
 }
-//#endregion
-//#region src/campaign/optimization-cost.ts
+// src/campaign/optimization-cost.ts
 /** Attribute method calls while retaining the shared account's admission and read behavior. */
 function createMethodCostScope(account, methodName) {
 	const tags = { [`optimizationAttempt:${crypto.randomUUID()}`]: methodName };
@@ -5125,8 +5092,7 @@ function optimizationTokenUsageFromSummary(summary, receipts) {
 		calls: summary.totalCalls
 	};
 }
-//#endregion
-//#region src/campaign/gepa-candidate-population.ts
+// src/campaign/gepa-candidate-population.ts
 /**
 * Read GEPA's exact candidate graph from the artifact addressed by method provenance.
 *
@@ -5310,8 +5276,7 @@ function assertExactKeys(value, expected, label) {
 function assertPositiveSafeInteger(value, label) {
 	if (!Number.isSafeInteger(value) || value <= 0) throw new Error(`GEPA candidate population ${label} must be a positive safe integer`);
 }
-//#endregion
-//#region src/campaign/optimization-method.ts
+// src/campaign/optimization-method.ts
 /** Both complete-method workflows use the same detached inputs, accounting, and evidence admission. */
 async function executeOptimizationMethod(options) {
 	assertSearchHistoryAdmissionOptions(options);
@@ -5458,8 +5423,7 @@ function assertOptimizationProvenance(methodName, value) {
 	}
 	if (value.optimizerModel === void 0 !== (value.optimizerCallRef === void 0) || value.optimizerModel === void 0 !== (value.modelExecutions === void 0)) fail("optimizerModel execution provenance");
 }
-//#endregion
-//#region src/run-score.ts
+// src/run-score.ts
 const DEFAULT_RUN_SCORE_WEIGHTS = {
 	success: 4,
 	goalProgress: 2,
@@ -5487,8 +5451,7 @@ function clamp01(value) {
 function finiteOrZero(value) {
 	return Number.isFinite(value) ? value : 0;
 }
-//#endregion
-//#region src/campaign/transient-failure.ts
+// src/campaign/transient-failure.ts
 const BASE_TRANSIENT = /\b50[234]\b|no stream output|produced no stream|admission timed out|admission_rejected|queue_timeout|fetch failed|ECONNRESET|This operation was aborted/i;
 const TIMEOUT_PATTERN = /timeout after \d+ ?ms|cli-bridge timeout/i;
 /**
@@ -5588,25 +5551,7 @@ function isTransientTransportFailure(message, opts = {}) {
 function transientDispatchFailure(opts = {}) {
 	return (failure) => failure.stage === "dispatch" && isTransientTransportFailure(failure.error.message, opts);
 }
-//#endregion
-//#region src/judges.ts
-/**
-* A judge's model response could not be parsed into scored dimensions.
-* Carries the raw response and paid-call metadata without fabricating a score.
-*/
-var JudgeParseError = class extends JudgeError {
-	judgeName;
-	raw;
-	llmCall;
-	constructor(judgeName, raw, options) {
-		super(`judge '${judgeName}' returned an unparseable response: ${raw.slice(0, 200)}`, options);
-		this.judgeName = judgeName;
-		this.raw = raw;
-		this.llmCall = options?.llmCall;
-	}
-};
-//#endregion
-//#region src/llm-judge.ts
+// src/llm-judge.ts
 /**
 * `llmJudge` — the single-LLM-call bridge that turns a rubric prompt into a
 * canonical campaign `JudgeConfig`.
@@ -5930,7 +5875,6 @@ function parseResponse(name, response, schema, llmCall) {
 function firstString(value) {
 	return typeof value === "string" && value.trim() ? value : void 0;
 }
-//#endregion
-export { planCampaignRun as $, verifySearchHistoryReceipt as A, defaultProductionGate as B, searchEdgeId as C, assertSearchHistoryMatchesState as D, assertSearchHistoryAdmissionOptions as E, asha as F, runCanaries as G, redTeamDataset as H, uniform as I, evaluationUnitMap as J, assertIndependentEvaluationSplit as K, runFinalComparison as L, searchModelIdentity as M, searchProposalExecution as N, createSearchHistoryReceipt as O, searchReceiptAccounting as P, runCampaign as Q, openAutoPr as R, developmentClaim as S, surfaceNode as T, redTeamReport as U, DEFAULT_RED_TEAM_CORPUS as V, scoreRedTeamOutput as W, reserveFinalEvidence as X, exposeFinalEvidence as Y, runEval as Z, beam as _, aggregateRunScore as a, incumbentWithOperatorBandit as b, assertGepaCandidatePopulationSummary as c, costFromLedgerSummary as d, resolveRunDir as et, optimizationTokenUsageFromSummary as f, aide as g, isProposedCandidate as h, transientDispatchFailure as i, campaignCellSearchResult as j, verifySearchHistoryArtifact as k, readGepaCandidatePopulationArtifact as l, runOptimization as m, JudgeParseError as n, cellCachePath as nt, clamp01 as o, runImprovementLoop as p, captureFinalEvidencePolicy as q, isTransientTransportFailure as r, computeManifestHash as rt, executeOptimizationMethod as s, llmJudge as t, buildCellSchedule as tt, combineComparisonCosts as u, crowdedFrontierParent as v, surfaceDiff as w, SearchRecorder as x, incumbent as y, captureJudge as z };
+export { resolveRunDir as $, campaignCellSearchResult as A, DEFAULT_RED_TEAM_CORPUS as B, surfaceDiff as C, createSearchHistoryReceipt as D, assertSearchHistoryMatchesState as E, uniform as F, assertIndependentEvaluationSplit as G, redTeamReport as H, runFinalComparison as I, exposeFinalEvidence as J, captureFinalEvidencePolicy as K, openAutoPr as L, searchProposalExecution as M, searchReceiptAccounting as N, verifySearchHistoryArtifact as O, asha as P, planCampaignRun as Q, captureJudge as R, searchEdgeId as S, assertSearchHistoryAdmissionOptions as T, scoreRedTeamOutput as U, redTeamDataset as V, runCanaries as W, runEval as X, reserveFinalEvidence as Y, runCampaign as Z, crowdedFrontierParent as _, clamp01 as a, SearchRecorder as b, readGepaCandidatePopulationArtifact as c, optimizationTokenUsageFromSummary as d, buildCellSchedule as et, runImprovementLoop as f, beam as g, aide as h, aggregateRunScore as i, searchModelIdentity as j, verifySearchHistoryReceipt as k, combineComparisonCosts as l, isProposedCandidate as m, isTransientTransportFailure as n, computeManifestHash as nt, executeOptimizationMethod as o, runOptimization as p, evaluationUnitMap as q, transientDispatchFailure as r, assertGepaCandidatePopulationSummary as s, llmJudge as t, cellCachePath as tt, costFromLedgerSummary as u, incumbent as v, surfaceNode as w, developmentClaim as x, incumbentWithOperatorBandit as y, defaultProductionGate as z };
 
-//# sourceMappingURL=llm-judge-Cybq29bL.js.map
+//# sourceMappingURL=llm-judge-cgbhTKQp.js.map