@tangle-network/agent-eval 0.144.9 → 0.144.11

package/dist/cli.js

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
 import { o as runRolloutReleaseCli } from "./hf-dataset-XggBupCr.js";
-import { n as runAnalystBenchmarkCommand } from "./benchmark-command-Dtym4pA-.js";
+import { n as runAnalystBenchmarkCommand } from "./benchmark-command-BteMFN62.js";
 import { a as runRpcBatch, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi } from "./server-iu0ede49.js";
 import { writeFileSync } from "node:fs";
 //#region src/cli-config.ts

package/dist/contract/index.d.ts

@@ -6,7 +6,7 @@ import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions }
 import { C as JudgeDimension, H as SurfaceProposer, M as OptimizerConfig, R as Scenario, S as JudgeConfig, V as SessionScript, _ as GateDecision, a as CampaignResult, b as GenerationRecord, c as CampaignTraceWriter, d as DispatchContext, f as DispatchFn, g as GateContribution, h as GateContext, i as CampaignCostMeter, j as MutableSurface, k as LabeledScenarioStore, l as CodeSurface, m as GateCheckStatus, n as CampaignArtifactWriter, p as Gate, r as CampaignCellResult, t as CampaignAggregates, v as GateResult, w as JudgeScore, y as GenerationCandidate } from "../types-BnjdJ70P.js";
 import { a as FailureClusterInsight, c as JudgeInsight, d as Recommendation, f as ReleaseSummary, i as FailureClassTally, l as LiftInsight, m as TokenUsageInsight, n as ExecutionErrorOutcomeCell, o as InsightReport, p as ScalarDistribution, r as ExecutionInsight, s as InterRaterInsight, t as CostProvenanceSummary, u as OutcomeCorrelationInsight } from "../insight-report-CgX_s0Ez.js";
 import { D as analyzeRuns, E as SummarizeExecutionOptions, O as summarizeExecution, T as ExecutionReport, w as AnalyzeRunsOptions } from "../engine-3hL-XqwJ.js";
-import { A as RunEvalOptions, At as compareOptimizationMethods, B as OptimizerModelBudget, C as RunImprovementLoopOptions, Ct as OptimizationMethodProvenance, D as RunOptimizationOptions, Dt as OptimizationPackageSource, E as PremeasuredOptimizationBaseline, F as GepaOptimizationMethodConfig, G as DefaultProductionRewardHackingOptions, Gt as inMemoryCampaignStorage, H as heldOutGate, Ht as CampaignStorage, I as GepaOptimizationRecipe, K as defaultProductionGate, L as GepaRunnerCommand, Lt as RunCampaignOptions, M as GepaAdaptiveEngineRun, N as GepaEngineOptions, Nt as CampaignCellFailureReceipt, Ot as OptimizationTokenUsage, P as GepaEngineRun, R as gepaOptimizationMethod, T as runImprovementLoop, U as DefaultProductionGateCheck, V as HeldOutGateOptions, W as DefaultProductionGateOptions, Wt as fsCampaignStorage, Yt as campaignSplitDigest, _t as CompareOptimizationMethodsOptions, an as ReferenceEquivalenceScenario, bt as OptimizationMethodComparison, cn as LlmJudgeDimension, dt as externalTextOptimizationMethod, en as REFERENCE_EQUIVALENCE_INPUT_LIMITS, ft as ExternalTextOptimizationMethodConfig, gt as ExternalTextEvaluationResponse, ht as ExternalOptimizationExample, i as skillOptOptimizationMethod, in as ReferenceEquivalenceJudgeResult, j as runEval, ln as LlmJudgeOptions, mt as ExternalTextOptimizerResult, n as SkillOptRunnerCommand, nn as ReferenceEquivalenceJudgeInput, on as createReferenceEquivalenceJudge, p as LoopProvenanceRecord, pt as ExternalTextOptimizerContext, r as SkillOptTrainerConfig, rn as ReferenceEquivalenceJudgeOptions, sn as runReferenceEquivalenceJudge, t as SkillOptOptimizationMethodConfig, tn as REFERENCE_EQUIVALENCE_JUDGE_VERSION, un as llmJudge, ut as composeGate, vt as ComparisonCost, w as RunImprovementLoopResult, wt as OptimizationMethodResult, xt as OptimizationMethodInput, yt as OptimizationMethod, z as OpenAICompatibleOptimizerModel, zt as runCampaign } from "../skillopt-optimization-method-D8wysvU1.js";
+import { $t as fsCampaignStorage, A as RunEvalOptions, B as OptimizerModelBudget, Bt as runCampaign, C as RunImprovementLoopOptions, D as RunOptimizationOptions, E as PremeasuredOptimizationBaseline, F as GepaOptimizationMethodConfig, G as DefaultProductionRewardHackingOptions, H as heldOutGate, I as GepaOptimizationRecipe, K as defaultProductionGate, L as GepaRunnerCommand, M as GepaAdaptiveEngineRun, N as GepaEngineOptions, Ot as OptimizationPackageSource, P as GepaEngineRun, Pt as CampaignCellFailureReceipt, R as gepaOptimizationMethod, Rt as RunCampaignOptions, St as OptimizationMethodInput, T as runImprovementLoop, Tt as OptimizationMethodResult, U as DefaultProductionGateCheck, V as HeldOutGateOptions, W as DefaultProductionGateOptions, Zt as CampaignStorage, _n as LlmJudgeDimension, bt as OptimizationMethod, dn as ReferenceEquivalenceJudgeInput, dt as externalTextOptimizationMethod, en as inMemoryCampaignStorage, fn as ReferenceEquivalenceJudgeOptions, ft as ExternalTextOptimizationMethodConfig, gn as runReferenceEquivalenceJudge, gt as ExternalTextEvaluationResponse, hn as createReferenceEquivalenceJudge, ht as ExternalOptimizationExample, i as skillOptOptimizationMethod, in as campaignSplitDigest, j as runEval, jt as compareOptimizationMethods, kt as OptimizationTokenUsage, ln as REFERENCE_EQUIVALENCE_INPUT_LIMITS, mn as ReferenceEquivalenceScenario, mt as ExternalTextOptimizerResult, n as SkillOptRunnerCommand, p as LoopProvenanceRecord, pn as ReferenceEquivalenceJudgeResult, pt as ExternalTextOptimizerContext, r as SkillOptTrainerConfig, t as SkillOptOptimizationMethodConfig, un as REFERENCE_EQUIVALENCE_JUDGE_VERSION, ut as composeGate, vn as LlmJudgeOptions, vt as CompareOptimizationMethodsOptions, w as RunImprovementLoopResult, wt as OptimizationMethodProvenance, xt as OptimizationMethodComparison, yn as llmJudge, yt as ComparisonCost, z as OpenAICompatibleOptimizerModel } from "../skillopt-optimization-method-USDKhxSA.js";
 import { a as ObjectiveSource, b as PowerPreflight, c as PromotionPolicy, d as paretoSignificanceGate, i as EvidenceVector, l as buildEvidenceVector, n as AxisVerdict, o as ParetoSignificanceGateOptions, r as BuildEvidenceVectorOptions, s as PromotionObjective, t as AxisEvidence, u as paretoPolicy } from "../promotion-policy-Ckjhzg_4.js";
 import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, n as HostedTenant, o as EvalRunCellScore, s as EvalRunEvent } from "../client-C9gzZE59.js";
 import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "../outcome-store-BYHIuO0e.js";

package/dist/contract/index.js

@@ -1,17 +1,17 @@
 import { c as ValidationError } from "../errors-D-LKuDhb.js";
 import { i as estimateCost, o as isModelPriced } from "../metrics-C9YY1OcL.js";
 import { i as CostLedger } from "../cost-ledger-DMFxsLKr.js";
-import { C as createChatClient, t as buildDefaultAnalystRegistry } from "../default-registry-BaQXW1Ow.js";
+import { C as createChatClient, t as buildDefaultAnalystRegistry } from "../default-registry-BmktKy8r.js";
 import { LLM_MODEL_ATTR_KEYS, SPAN_KIND_ATTR_KEYS } from "../trace-attributes.js";
-import { nt as classifyOtlpSpanRole, rt as isOtlpModelCall } from "../kind-factory-BHIgPmzS.js";
-import { s as makeProposalFinding } from "../usage-receipt-EVI8B8Xu.js";
-import { i as inMemoryCampaignStorage, n as createRunCostLedger, r as fsCampaignStorage } from "../single-run-lock-BMQEv1wG.js";
+import { nt as isOtlpModelCall, tt as classifyOtlpSpanRole } from "../kind-factory-B8-r8-y8.js";
+import { c as makeProposalFinding } from "../usage-receipt-t7vAzCRQ.js";
+import { a as inMemoryCampaignStorage, i as fsCampaignStorage, r as createRunCostLedger } from "../single-run-lock-DFWHEB09.js";
 import { m as mapConcurrentRange } from "../ledger-core-DXZIqu17.js";
-import { $ as runReferenceEquivalenceJudge, H as resolveRunDir, M as surfaceContentHash, N as surfaceHash, Q as createReferenceEquivalenceJudge, S as assertOptimizationResult, V as runCampaign, X as REFERENCE_EQUIVALENCE_INPUT_LIMITS, Z as REFERENCE_EQUIVALENCE_JUDGE_VERSION, _ as heldOutGate, a as emitLoopProvenance, b as composeGate, ct as llmJudge, d as runImprovementLoop, g as gepaOptimizationMethod, o as loopProvenanceArgsFromResult, p as runEval, q as campaignSplitDigest, t as skillOptOptimizationMethod, v as defaultProductionGate, w as compareOptimizationMethods, x as externalTextOptimizationMethod } from "../skillopt-optimization-method-Dtzp7QpP.js";
+import { $ as REFERENCE_EQUIVALENCE_INPUT_LIMITS, F as surfaceHash, P as surfaceContentHash, S as assertOptimizationResult, U as runCampaign, W as resolveRunDir, X as campaignSplitDigest, _ as heldOutGate, a as emitLoopProvenance, b as composeGate, d as runImprovementLoop, dt as llmJudge, et as REFERENCE_EQUIVALENCE_JUDGE_VERSION, g as gepaOptimizationMethod, nt as runReferenceEquivalenceJudge, o as loopProvenanceArgsFromResult, p as runEval, t as skillOptOptimizationMethod, tt as createReferenceEquivalenceJudge, v as defaultProductionGate, w as compareOptimizationMethods, x as externalTextOptimizationMethod } from "../skillopt-optimization-method-CQdVeM8k.js";
 import { E as pairedBootstrap } from "../statistics-ByxzSiOM.js";
 import { i as parseRunRecordSafe, r as modelHasSnapshot } from "../run-record-BmSPWXJR.js";
 import { c as heldoutSignificance, i as powerPreflight, n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector, u as decidePairedPromotion } from "../promotion-policy-CrLrmys8.js";
-import { a as recordAggregateMeasurements, i as readTaskFailureLabels, o as summarizeExecutionMeasurements, s as summarizeTraceErrors, t as extractUsage } from "../extract-usage-BW27f3XW.js";
+import { a as recordAggregateMeasurements, i as readTaskFailureLabels, o as summarizeExecutionMeasurements, s as summarizeTraceErrors, t as extractUsage } from "../extract-usage-CdZdoj1s.js";
 import { n as summarizeExecution, t as analyzeRuns } from "../analyze-runs-C30yljDJ.js";
 import { c as campaignCellTaskScore, l as campaignCellToRunRecord, o as campaignCellExecutionEvidence, s as campaignCellJudgeDimensions } from "../reward-hacking-BDToousL.js";
 import { n as InMemoryOutcomeStore, t as FileSystemOutcomeStore } from "../outcome-store-ChBKlTd_.js";

package/dist/{default-registry-BaQXW1Ow.js → default-registry-BmktKy8r.js}

@@ -1,8 +1,8 @@
 import { n as LlmClient } from "./llm-client-DzvMUsS_.js";
 import { LLM_CONTEXT_TOKENS, LLM_INPUT_TOKEN_ATTR_KEYS, LLM_OUTPUT_TOKEN_ATTR_KEYS, TOOL_NAME_ATTR_KEYS } from "./trace-attributes.js";
 import { t as executionTrackByLane } from "./execution-tracks-CpgFPpS5.js";
-import { $ as spanEpochMillis, H as snapshotExactExecutionPlan, R as findingSubjectGrammarPromptFor, U as deepFreezeCanonicalJson, V as snapshotExactExecutionComponentIdentity, t as createTraceAnalyst } from "./kind-factory-BHIgPmzS.js";
-import { i as validateUsageSettlementTimeout, o as makeFinding, t as assertValidAnalystUsageReceipt } from "./usage-receipt-EVI8B8Xu.js";
+import { H as snapshotExactExecutionPlan, Q as spanEpochMillis, R as findingSubjectGrammarPromptFor, V as snapshotExactExecutionComponentIdentity, t as createTraceAnalyst } from "./kind-factory-B8-r8-y8.js";
+import { a as deepFreezeCanonicalJson, i as validateUsageSettlementTimeout, s as makeFinding, t as assertValidAnalystUsageReceipt } from "./usage-receipt-t7vAzCRQ.js";
 import { i as hashCanonical, r as canonicalString } from "./canonical-D011XM8r.js";
 import { t as analyzeSupervisorRunIntegrity } from "./integrity-DY6tIbl0.js";
 import { o as combineAbortSignals } from "./proposal-findings-2GIUo1et.js";
@@ -2449,4 +2449,4 @@ function buildDefaultAnalystRegistry(options = {}) {
 //#endregion
 export { createChatClient as C, computeTraceMetrics as S, CONTROL_INTEGRITY_ANALYST as _, analystFindingDigest as a, behavioralAnalyst as b, completedAnalystReviewQuality as c, validateAnalystReviewDecisions as d, DEFAULT_TRACE_ANALYST_KINDS as f, FAILURE_MODE_KIND_SPEC as g, IMPROVEMENT_KIND_SPEC as h, assertExactRegistryRunOpts as i, readAnalystReview as l, KNOWLEDGE_GAP_KIND_SPEC as m, AnalystRegistry as n, analystRunDigest as o, KNOWLEDGE_POISONING_KIND_SPEC as p, ExactAnalystRunExecutionError as r, assertUniqueFindingIds as s, buildDefaultAnalystRegistry as t, snapshotAnalystRun as u, ControlIntegrityAnalyst as v, deriveEfficiencyFindings as x, emitControlIntegrityFindings as y };
 
-//# sourceMappingURL=default-registry-BaQXW1Ow.js.map
+//# sourceMappingURL=default-registry-BmktKy8r.js.map