@tangle-network/agent-eval 0.193.2 → 0.195.0

package/dist/index.js

@@ -1,7 +1,7 @@
 import { t as __exportAll } from "./rolldown-runtime-8H4AJuhK.js";
 import { i as JudgeError, n as CaptureIntegrityError, o as NotFoundError, r as ConfigError, s as ValidationError, t as AgentEvalError } from "./errors-Dngq5h35.js";
 import { a as hashCanonical, r as canonicalString } from "./canonical-BsZVgBhW.js";
-import { a as CODING_HARNESSES, c as agentProfileId, d as harnessAxisOf, i as verifyCompletion, l as agentProfileModelId, n as completionVerdict, o as HARNESS_NATIVE_MODEL, r as createLlmCorrectnessChecker, s as agentProfileHash, t as extractProducedState, u as expandProfileAxes } from "./produced-state-B2AhDRnq.js";
+import { a as CODING_HARNESSES, c as agentProfileId, d as harnessAxisOf, i as verifyCompletion, l as agentProfileModelId, n as completionVerdict, o as HARNESS_NATIVE_MODEL, r as createLlmCorrectnessChecker, s as agentProfileHash, t as extractProducedState, u as expandProfileAxes } from "./produced-state-C8rk-GCK.js";
 import { c as groupRunsByAgentProfileCell, d as validateAgentProfileCell, f as verifyAgentProfileCell, h as manifestContentDigest, i as agentProfileCellKey, m as hashJson, r as agentProfileCellHashMaterial, s as buildAgentProfileCell, t as AGENT_PROFILE_KINDS, u as toAgentProfileJson } from "./agent-profile-cell-DcFW4vOc.js";
 import { t as mulberry32 } from "./random-Dn5fPWkt.js";
 import { a as spearmanR, c as weightedMean, i as ranks, n as partialCredit, o as summarizeNumberSeries, r as pearsonR, s as weightedComposite, t as confidenceInterval } from "./descriptive-1V17A-qa.js";
@@ -22,12 +22,12 @@ import { r as observedSplitScore, t as isRealnessGated } from "./reward-nw2xZGZG
 import { a as parseRunRecordSafe, c as runTaskScore, i as modelHasSnapshot, l as searchCellRunId, n as UNKNOWN_MODEL, o as roundTripRunRecord, r as isRunRecord, s as runCostFloorUsd, t as RunRecordValidationError, u as validateRunRecord } from "./run-record-BRlM99rX.js";
 import { a as summaryTable, n as gainHistogram, r as paretoChart } from "./summary-report-BMvou2oZ.js";
 import { n as contentHash, r as fileVerdictCache, t as canonicalJson } from "./verdict-cache-BQEsVy9n.js";
-import { $ as redTeamDataset, K as dominates, Q as DEFAULT_RED_TEAM_CORPUS, U as pairedPromotionPower, W as requiredPairsForPairedPromotion, a as transientDispatchFailure, ct as runCampaign, et as redTeamReport, i as quotaExhaustedUntil, nt as runCanaries, o as aggregateRunScore, q as paretoFrontier, s as clamp01, t as llmJudge, tt as scoreRedTeamOutput } from "./llm-judge-Cififl_V.js";
+import { G as requiredPairsForPairedPromotion, J as dominates, W as pairedPromotionPower, Y as paretoFrontier, a as transientDispatchFailure, et as DEFAULT_RED_TEAM_CORPUS, i as quotaExhaustedUntil, it as runCanaries, nt as redTeamReport, o as aggregateRunScore, rt as scoreRedTeamOutput, s as clamp01, t as llmJudge, tt as redTeamDataset, ut as runCampaign } from "./llm-judge-CvKpaRKY.js";
 import { a as resolveModelPricing, i as isModelPriced, n as estimateCost, r as estimateTokens, t as MODEL_PRICING } from "./metrics-DOMpkoeb.js";
 import { a as CostLedgerPersistenceError, c as costForTokenPricing, i as CostLedger, l as costForUsage, n as CostCallConflictError, o as CostReceiptCaptureError, r as CostCeilingReachedError, s as CostReservationExceededError, t as CostAccountingIncompleteError, u as modelPriceKey } from "./cost-ledger-_IbkejCF.js";
 import { B as assertRealBackend, F as surfaceContentHash, T as parseReflectionResponse, V as summarizeBackendIntegrity, c as ATTESTATION_ALGORITHM, l as attest, u as verifyAttestation, w as buildReflectionPrompt, z as BackendIntegrityError } from "./campaign-evidence-DJc60F9C.js";
 import { d as redactForShare, f as redactText, n as REDACTION_VERSION, p as shareAllowed, r as assessShareSafety, u as redact } from "./redact-Dk1ZdJEm.js";
-import { r as selfImprove, t as defineAgentEval } from "./define-agent-eval-qxAPjMSL.js";
+import { r as selfImprove, t as defineAgentEval } from "./define-agent-eval-Ddut4mkS.js";
 import { n as equivalenceVerdict, t as certificationEvidenceDigest } from "./verdict-DhPoXApk.js";
 import { t as TraceEmitter } from "./emitter-DbMXUIVP.js";
 import { t as buildTrajectory } from "./trajectory-D_7rLrvE.js";

package/dist/jev.d.ts

@@ -1,6 +1,6 @@
 import { b as CustomTokenPricing, g as CostReceiptInput } from "./cost-ledger-DbQdN3nO.js";
 import { i as AnalystFinding, n as AnalystContext, t as Analyst } from "./types-CeDfCRXp.js";
-import { R as Scenario, S as JudgeConfig, w as JudgeScore } from "./types-BWwRXofp.js";
+import { R as Scenario, S as JudgeConfig, w as JudgeScore } from "./types-DLtLe21j.js";
 import { EvaluationAnalystOptions, EvaluationContext, EvaluationJudgeOptions, EvaluationResult, Evaluator, EvaluatorOptions, asAnalyst, asJudge, createEvaluator } from "./evaluation.js";
 import { JevAnswer, JevAnswerFor, JevJson, JevQuestion, JevQuestions, JevRequest, JevResponseError, JevResult, JevScoreCriteria, JevState, jevUsage, parseJevRequest, parseJevResult } from "./jev-protocol.js";
 import { z } from "zod";