@tangle-network/agent-eval 0.222.0 → 0.223.0

package/dist/cli.js

@@ -1,14 +1,14 @@
 #!/usr/bin/env node
 import { c as errorMessage } from "./errors-DeEEV9Vx.js";
-import { l as operatorYield, u as INSUFFICIENT_FROM } from "./pareto-B9udAt2i.js";
+import { l as operatorYield, u as INSUFFICIENT_FROM } from "./pareto-7_LEmYGG.js";
 import { i as replaySearchLedgerText } from "./search-ledger-BsVi-Ngv.js";
-import { t as renderSearchSummary } from "./search-summary-Ct7SrmVc.js";
+import { t as renderSearchSummary } from "./search-summary-Bt3uuK53.js";
 import { l as SEARCH_LEDGER_BATCH_MAX_BYTES, m as SearchRunKindSchema, r as shipSearchLedger, s as hostedTenantFromEnv, t as SearchShipConflictError } from "./search-shipper-pdg4viK7.js";
 import { a as privacyHit, f as harnessSessionRolloutLines, m as appendRolloutLines, n as runRolloutReleaseCli, p as gateReceiptRolloutLines, r as PRIVACY_RULE_NAMES, u as harnessSessionRefs } from "./hf-dataset-DPVQXW5n.js";
 import { t as createChatClient } from "./chat-client-CO3uB33i.js";
-import { t as runAnalystBenchmarkCommand } from "./benchmark-command-jYWR2gya.js";
+import { t as runAnalystBenchmarkCommand } from "./benchmark-command-CH86twZj.js";
 import { i as buildOpenApi, l as handleVersion, o as runRpcBatch, r as startServerAsync, s as runRpcOnce } from "./server-DA6h3co5.js";
-import { b as editCreditText, f as landscape, g as surfaceTextEdits, l as metaSearch, m as formatLandscape, n as taskMatrix, o as skillManifold, r as formatSkillManifold, s as renderMetaSearchText, t as tree, v as front, y as editCredit } from "./tree-DrIChF4a.js";
+import { b as editCreditText, f as landscape, g as surfaceTextEdits, l as metaSearch, m as formatLandscape, n as taskMatrix, o as skillManifold, r as formatSkillManifold, s as renderMetaSearchText, t as tree, v as front, y as editCredit } from "./tree-DnBkWXun.js";
 import { c as discoveryRunRolloutLines, i as analyzeSupervisorRun, n as renderSupervisorRunHeadline, r as renderSupervisorRunMarkdown } from "./render-pfAZ7IR4.js";
 import { createHash } from "node:crypto";
 import { existsSync, readFileSync, writeFileSync } from "node:fs";

package/dist/contract/index.js

@@ -1,24 +1,25 @@
 import { a as hashCanonical } from "../canonical-CicJlINA.js";
 import "../run-record-DEjx43cI.js";
-import { r as summarizeExecution, t as analyzeRuns } from "../analyze-runs-CTheASYV.js";
-import { i as pairedBootstrap } from "../mean-CdDYG-us.js";
-import { t as decidePairedPromotion } from "../paired-promotion-decision-FGvHA4p_.js";
-import { a as heldoutSignificance } from "../statistical-heldout-TP6ndL0B.js";
+import { r as summarizeExecution, t as analyzeRuns } from "../analyze-runs-Ck68mbgj.js";
+import { r as pairedBootstrap } from "../paired-tests-C2zUqgUW.js";
+import { t as mean$1 } from "../mean-BPuUNznQ.js";
+import { t as decidePairedPromotion } from "../paired-promotion-decision-BiVumgvM.js";
+import { a as heldoutSignificance } from "../statistical-heldout-C9hz_AF9.js";
 import { f as mapConcurrentRange } from "../journal-C1uLJi0w.js";
-import { X as runEval, Z as runCampaign, a as clamp01, f as runImprovementLoop, r as transientDispatchFailure, t as llmJudge, z as defaultProductionGate } from "../llm-judge-cgbhTKQp.js";
+import { X as runEval, Z as runCampaign, a as clamp01, f as runImprovementLoop, r as transientDispatchFailure, t as llmJudge, z as defaultProductionGate } from "../llm-judge-D6JtIqiU.js";
 import { i as isModelPriced, n as estimateCost } from "../metrics-DNCAUtuB.js";
 import { t as CostLedger } from "../cost-ledger-CS8R3IZj.js";
-import { A as campaignSplitDigest } from "../campaign-evidence-DN8rWuBj.js";
+import { A as campaignSplitDigest } from "../campaign-evidence-MjCjkb5a.js";
 import { M as fsCampaignStorage, N as inMemoryCampaignStorage } from "../external-optimizer-subprocess-U616hV_l.js";
 import { a as makeProposalFinding } from "../types-BK0blWhe.js";
 import { n as isRecord } from "../json-DtvZNAU9.js";
-import { n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-BLoP4Lf9.js";
+import { n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-Cc15qGik.js";
 import { t as buildDefaultAnalystRegistry } from "../default-registry-C1tPEjZ4.js";
 import { t as extractUsage } from "../extract-usage-ISFqhfDz.js";
 import { t as createChatClient } from "../chat-client-CO3uB33i.js";
 import { n as InMemoryOutcomeStore, t as FileSystemOutcomeStore } from "../outcome-store-TarrcfHG.js";
-import { c as composeGate, l as externalTextOptimizationMethod, n as compareOptimizationMethods, r as gepaOptimizationMethod, s as heldOutGate, t as skillOptOptimizationMethod } from "../skillopt-optimization-method-qB1M554K.js";
-import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-CBYeD4zi.js";
+import { c as composeGate, l as externalTextOptimizationMethod, n as compareOptimizationMethods, r as gepaOptimizationMethod, s as heldOutGate, t as skillOptOptimizationMethod } from "../skillopt-optimization-method-BKFbJEyz.js";
+import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-D7B4_jEZ.js";
 import { t as fromOtelSpans } from "../otel-spans-BsGzTWIK.js";
 import { createHash } from "node:crypto";
 import { agentCandidateBenchmarkSuiteSchema, agentCandidateBenchmarkTaskSchema, agentCandidateBundleSchema, agentCandidateEvaluationPolicySchema, agentCandidateExperimentSchema, agentImprovementMeasuredComparisonSchema, agentProfileImprovementExperimentSchema, agentProfileImprovementMeasuredComparisonSchema, agentProfileImprovementRunCellSchema, agentProfileImprovementRunReceiptSchema, agentProfileImprovementSuiteInputsSchema, agentProfileImprovementSuiteSchema, agentProfileImprovementTaskSchema, candidateExecutionEvidenceSchema, canonicalCandidateDigest, canonicalCandidateJson, numbersApproximatelyEqual, omitTopLevelDigest } from "@tangle-network/agent-interface";
@@ -726,7 +727,7 @@ function observedPairedPrecision(options) {
 }
 function mean(values) {
 	if (values.length === 0) throw new Error("candidate experiment requires measured values");
-	return values.reduce((sum, value) => sum + value, 0) / values.length;
+	return mean$1(values);
 }
 // src/contract/candidate-experiment.ts
 /** Seal held-out experiments, execute their signed cells, and publish verified comparisons. */