@tangle-network/agent-eval 0.193.0 → 0.194.0

package/dist/{llm-judge-ChbWfeWl.js → llm-judge-D6NIvyrM.js}

@@ -12,14 +12,14 @@ import { i as CostLedger, t as CostAccountingIncompleteError } from "./cost-ledg
 import { F as surfaceContentHash, G as campaignCoverage, H as assertCampaignDesign, I as surfaceDispatchRef, K as campaignScenarioIdentity, L as surfaceHash, O as FinalEvidenceConflictError, P as renderSurfaceDiff, R as surfaceHashMatches, S as campaignMeanCompositeOrNull, U as assertCampaignSplitIdentity, W as assertCompleteCampaign, Y as formatCoverageFailures, b as campaignBreakdown, k as FinalEvidenceError, q as campaignSplitDigest, t as createCampaignEvidenceReceipt, x as campaignMeanComposite, z as BackendIntegrityError } from "./campaign-evidence-DJc60F9C.js";
 import { m as mapConcurrent } from "./journal-DvKCvjty.js";
 import { t as deepFreezeCanonicalJson } from "./deep-freeze-CurcOQg3.js";
-import { C as fsCampaignStorage, S as createRunCostLedger, g as isRecord, h as isExternalTextCandidate, x as acquireSingleRunLock } from "./external-optimizer-subprocess-kbqC-iWf.js";
+import { C as fsCampaignStorage, S as createRunCostLedger, g as isRecord, h as isExternalTextCandidate, x as acquireSingleRunLock } from "./external-optimizer-subprocess-BrghTvEp.js";
 import { r as summarizeEvaluationUnits, t as defineEvaluationClaim } from "./claim-V9Zqbq0J.js";
-import { f as redactText, n as REDACTION_VERSION, o as detectCredential, s as detectPersonalData } from "./redact-B7rt9Zot.js";
+import { f as redactText, n as REDACTION_VERSION, o as detectCredential, s as detectPersonalData } from "./redact-Dk1ZdJEm.js";
 import { t as detectRewardHacking } from "./reward-hacking-D0XwhVWE.js";
 import { n as assertProposalFindings, t as combineAbortSignals } from "./abort-signal-CtzAM_sJ.js";
-import { a as validateSearchLedgerEvent, d as searchTaskSetDigest, i as replaySearchLedgerText, l as searchCellId, n as openSearchLedger, t as FileSearchLedger, u as searchNodeId } from "./search-ledger-CJiAEU3j.js";
-import { c as seedFromDigest, i as estimateNode, n as searchProposerView, s as searchPosterior, t as renderSearchSummary } from "./search-summary-BhY8p0Av.js";
-import { d as stripFencedJson, o as costReceiptFromLlm, s as costReceiptFromLlmError, u as maximumChargeForLlmRequest } from "./llm-client-KYQBh9C8.js";
+import { a as validateSearchLedgerEvent, d as searchTaskSetDigest, i as replaySearchLedgerText, l as searchCellId, n as openSearchLedger, t as FileSearchLedger, u as searchNodeId } from "./search-ledger-BbKtG_WF.js";
+import { c as seedFromDigest, i as estimateNode, n as searchProposerView, s as searchPosterior, t as renderSearchSummary } from "./search-summary-DnDGR702.js";
+import { d as stripFencedJson, o as costReceiptFromLlm, s as costReceiptFromLlmError, u as maximumChargeForLlmRequest } from "./llm-client-c_bKHTXz.js";
 import { createHash } from "node:crypto";
 import { writeFileSync } from "node:fs";
 import { z } from "zod";
@@ -3064,6 +3064,76 @@ function decideSearchClaim(state, plan) {
 		decisions
 	};
 }
+/**
+* Make a closed search's claim again from its ledger alone and compare it with
+* the recorded one, byte for byte: the power check, the finalists, each
+* finalist's test estimate and deciding interval, the selection, the decision
+* and the reason. A producer's claim is not evidence until this holds; a store
+* keeps what this function derives, not what the producer wrote.
+*
+* It needs no blob, so it works on a store that keeps digests only. The
+* finalists are the nodes the ledger decided `finalist`, in decision order.
+* The one input the ledger cannot replay is whether the budget covered k
+* finalists' test cells; the claim cells that ran answer it (none ran: the
+* budget covered none). Test repeats come from the claim cells too; without
+* test cells, the recorded estimator stands for them (binary needs one repeat,
+* so any other count reproduces continuous).
+*
+* Returns null for a search that made no claim.
+*/
+function verifySearchClaim(state) {
+	const closed = state.closed;
+	if (!closed) throw new Error(`search ${state.searchId} is open; its claim is made when it closes`);
+	const recorded = closed.claim;
+	if (recorded === null) return null;
+	if (recorded.rule.revision !== SEARCH_CLAIM_RULE.revision) return {
+		status: "unknown",
+		reason: `the claim names rule revision ${recorded.rule.revision}; this code implements ${SEARCH_CLAIM_RULE.revision}`
+	};
+	const claimCells = state.cells().filter((cell) => cell.stage === "claim");
+	const decided = state.nodes().flatMap((node) => {
+		const first = node.decisions.find((record) => record.decision.status === "finalist");
+		return first ? [{
+			nodeId: node.nodeId,
+			sequence: first.sequence
+		}] : [];
+	}).sort((a, b) => a.sequence - b.sequence).map((entry) => entry.nodeId);
+	const reps = claimCells.length > 0 ? 1 + Math.max(...claimCells.map((cell) => cell.rep)) : "adequate" in recorded.power && recorded.power.estimator.kind === "binary" ? 1 : 2;
+	const plan = planSearchClaim(state, {
+		stopReason: closed.reason,
+		reps,
+		affordable: (finalists) => claimCells.length > 0 && finalists <= decided.length
+	});
+	const differences = [];
+	const planned = plan.finalists.map((finalist) => finalist.nodeId);
+	if (canonicalString(planned) !== canonicalString(decided)) differences.push(`finalists: the rule picks [${planned.join(", ")}], the ledger decided [${decided.join(", ")}]`);
+	if (plan.test === "run" !== claimCells.length > 0) differences.push(`test: the plan ${plan.test === "run" ? "runs" : "skips"} the test cells, and ${claimCells.length} ran`);
+	const expected = decideSearchClaim(state, plan).claim;
+	for (const key of [
+		"decision",
+		"selected",
+		"confidence",
+		"power",
+		"reason"
+	]) if (canonicalString(expected[key]) !== canonicalString(recorded[key])) differences.push(`${key}: the ledger supports ${canonicalString(expected[key])}, the claim records ${canonicalString(recorded[key])}`);
+	const count = Math.max(expected.finalists.length, recorded.finalists.length);
+	for (let index = 0; index < count; index++) {
+		const want = expected.finalists[index];
+		const have = recorded.finalists[index];
+		if (canonicalString(want ?? null) === canonicalString(have ?? null)) continue;
+		const fields = [
+			"nodeId",
+			"promote",
+			"test",
+			"estimate"
+		].filter((field) => canonicalString(want?.[field] ?? null) !== canonicalString(have?.[field] ?? null));
+		differences.push(`finalist ${index + 1} (${have?.nodeId ?? want?.nodeId}): ${fields.join(", ")} differ from what the ledger supports`);
+	}
+	return differences.length === 0 ? { status: "verified" } : {
+		status: "mismatch",
+		differences
+	};
+}
 /** Per-finalist confidence for `k` finalists. */
 function bonferroni(k) {
 	return 1 - (1 - FAMILY_CONFIDENCE) / k;
@@ -3321,7 +3391,7 @@ const SEARCH_KERNEL_SOURCE = {
 	revision: hashCanonical({
 		name: "tangle.search-kernel.2026-09",
 		dispatch: "claim, then rung and root, then screen, then train; first allocated first within a stage",
-		expansion: "when no cell waits, fewer than twice the lanes capacity run, and the cap admits one proposal and its expected screens",
+		expansion: "when no cell waits, fewer than twice the lanes capacity run, the cap admits one proposal and its expected screens, and overspend has not taken the search past its cap",
 		reservation: {
 			hard: "the lane per-cell maximum",
 			estimate: "1.5 times the p99 of the lane settled cells once 20 settled, else the lane prior"
@@ -3942,7 +4012,7 @@ var SearchKernel = class {
 		const children = this.options.proposer.childrenPerProposal ?? 1;
 		const need = operation.usd + children * screen * this.maxCellReservation();
 		const { headroomUsd } = this.state.budget;
-		if (headroomUsd !== null && need > headroomUsd + USD_TOLERANCE) return "budget";
+		if (headroomUsd !== null && (need > headroomUsd + USD_TOLERANCE || this.overCap())) return "budget";
 		if (maxCells !== null && this.state.audit.cells.allocated + children * screen > maxCells) return "budget";
 		const expansion = policy.expand(view);
 		if (expansion === null) return "converged";
@@ -4151,13 +4221,19 @@ var SearchKernel = class {
 		});
 		await this.refresh();
 	}
+	/** The closed search, with its claim made again from the ledger alone. A
+	* claim the ledger does not support throws: a tampered ledger, or a
+	* producer and verifier that disagree. */
 	closedResult() {
 		const state = this.state;
+		const claimVerification = verifySearchClaim(state);
+		if (claimVerification?.status === "mismatch") throw new Error(`runSearch: search ${state.searchId} closed with a claim its ledger does not support: ${claimVerification.differences.join("; ")}`);
 		return {
 			state,
 			leader: state.audit.selectedNodeId ?? state.rootNodeId,
 			reason: state.closed.reason,
-			claim: state.closed.claim
+			claim: state.closed.claim,
+			claimVerification
 		};
 	}
 	policyView() {
@@ -4203,6 +4279,14 @@ var SearchKernel = class {
 			usd: estimate(this.operationCosts, this.options.proposer.reservationUsd ?? 0)
 		};
 	}
+	/** Overspend took committed spend, open holds and the unspent claim reserve
+	* past the cap. Nothing new is proposed then, even at a zero-dollar prior. */
+	overCap() {
+		const { maxUsd } = this.state.header.budget;
+		if (maxUsd === null) return false;
+		const { committedUsd, openReservationUsd } = this.state.audit.spend;
+		return committedUsd + openReservationUsd + this.state.budget.claimReserveUsd > maxUsd + USD_TOLERANCE;
+	}
 	pastDeadline() {
 		const deadline = this.state.header.budget.deadline;
 		return deadline !== null && this.now() >= Date.parse(deadline);
@@ -7250,6 +7334,6 @@ function firstString(value) {
 	return typeof value === "string" && value.trim() ? value : void 0;
 }
 //#endregion
-export { redTeamDataset as $, createSearchHistoryReceipt as A, decideSearchClaim as B, searchEdgeId as C, assertCompleteSearchHistory as D, SearchHistoryRequiredError as E, runSearch as F, uniform as G, searchClaimReserveUsd as H, searchDivergence as I, runFinalComparison as J, dominates as K, searchExpansionIndex as L, verifySearchHistoryArtifact as M, verifySearchHistoryReceipt as N, assertSearchHistoryAdmissionOptions as O, SEARCH_KERNEL_SOURCE as P, DEFAULT_RED_TEAM_CORPUS as Q, searchPolicyView as R, developmentClaim as S, surfaceNode as T, pairedPromotionPower as U, planSearchClaim as V, requiredPairsForPairedPromotion as W, captureJudge as X, openAutoPr as Y, defaultProductionGate as Z, isProposedCandidate as _, transientDispatchFailure as a, evaluationUnitMap as at, incumbent as b, compareOptimizationMethods as c, runCampaign as ct, combineComparisonCosts as d, tangleTracesRoot as dt, redTeamReport as et, costFromLedgerSummary as f, readCachedCell as ft, runOptimization as g, runImprovementLoop as h, computeManifestHash as ht, quotaExhaustedUntil as i, captureFinalEvidencePolicy as it, searchHistoryCoverageRow as j, assertSearchHistoryMatchesState as k, optimizationTokenUsageFromSummary as l, planCampaignRun as lt, readGepaCandidatePopulationArtifact as m, cellCachePath as mt, JudgeParseError as n, runCanaries as nt, aggregateRunScore as o, reserveFinalEvidence as ot, assertGepaCandidatePopulationSummary as p, buildCellSchedule as pt, paretoFrontier as q, isTransientTransportFailure as r, assertIndependentEvaluationSplit as rt, clamp01 as s, runEval as st, llmJudge as t, scoreRedTeamOutput as tt, executeOptimizationMethod as u, resolveRunDir as ut, labelTrustRank as v, surfaceDiff as w, SearchRecorder as x, crowdedFrontierParent as y, SEARCH_CLAIM_RULE as z };
+export { DEFAULT_RED_TEAM_CORPUS as $, createSearchHistoryReceipt as A, decideSearchClaim as B, searchEdgeId as C, assertCompleteSearchHistory as D, SearchHistoryRequiredError as E, runSearch as F, requiredPairsForPairedPromotion as G, searchClaimReserveUsd as H, searchDivergence as I, paretoFrontier as J, uniform as K, searchExpansionIndex as L, verifySearchHistoryArtifact as M, verifySearchHistoryReceipt as N, assertSearchHistoryAdmissionOptions as O, SEARCH_KERNEL_SOURCE as P, defaultProductionGate as Q, searchPolicyView as R, developmentClaim as S, surfaceNode as T, verifySearchClaim as U, planSearchClaim as V, pairedPromotionPower as W, openAutoPr as X, runFinalComparison as Y, captureJudge as Z, isProposedCandidate as _, transientDispatchFailure as a, captureFinalEvidencePolicy as at, incumbent as b, compareOptimizationMethods as c, runEval as ct, combineComparisonCosts as d, resolveRunDir as dt, redTeamDataset as et, costFromLedgerSummary as f, tangleTracesRoot as ft, runOptimization as g, computeManifestHash as gt, runImprovementLoop as h, cellCachePath as ht, quotaExhaustedUntil as i, assertIndependentEvaluationSplit as it, searchHistoryCoverageRow as j, assertSearchHistoryMatchesState as k, optimizationTokenUsageFromSummary as l, runCampaign as lt, readGepaCandidatePopulationArtifact as m, buildCellSchedule as mt, JudgeParseError as n, scoreRedTeamOutput as nt, aggregateRunScore as o, evaluationUnitMap as ot, assertGepaCandidatePopulationSummary as p, readCachedCell as pt, dominates as q, isTransientTransportFailure as r, runCanaries as rt, clamp01 as s, reserveFinalEvidence as st, llmJudge as t, redTeamReport as tt, executeOptimizationMethod as u, planCampaignRun as ut, labelTrustRank as v, surfaceDiff as w, SearchRecorder as x, crowdedFrontierParent as y, SEARCH_CLAIM_RULE as z };
 
-//# sourceMappingURL=llm-judge-ChbWfeWl.js.map
+//# sourceMappingURL=llm-judge-D6NIvyrM.js.map