@tangle-network/agent-eval 0.218.1 → 0.219.0

package/dist/{analyze-runs-DjlZOVXx.d.ts → analyze-runs-BWAHBDAw.d.ts}

@@ -1,9 +1,9 @@
 import { o as FailureClass } from "./schema-D6EVKzTB.js";
-import { a as RunRecord, u as RunTerminalOutcome } from "./run-record-CaxbLq6Y.js";
+import { a as RunRecord, u as RunTerminalOutcome } from "./run-record-C3awiXPf.js";
 import { r as ContinuousAgreement } from "./judge-calibration-DFtEMlde.js";
 import { r as DatasetScenario } from "./dataset-DQqhOCPt.js";
-import { f as ParetoFigureSpec, l as GainDistributionBin } from "./summary-report-CkbRcSuu.js";
-import { t as AnalystRegistry } from "./registry-ssNoHzwD.js";
+import { f as ParetoFigureSpec, l as GainDistributionBin } from "./summary-report-C8V49Jt3.js";
+import { t as AnalystRegistry } from "./registry-B_yG-1jh.js";
 //#region src/contract/insight-report.d.ts
 interface InsightReport {
   /** Number of runs analyzed. */
@@ -486,4 +486,4 @@ interface AnalyzeRunsOptions {
 declare function analyzeRuns(opts: AnalyzeRunsOptions): Promise<InsightReport>;
 //#endregion
 export { ReleaseSummary as _, summarizeExecution as a, ExecutionInsight as c, InsightReport as d, InterRaterInsight as f, Recommendation as g, OutcomeCorrelationInsight as h, SummarizeExecutionOptions as i, FailureClassTally as l, LiftInsight as m, analyzeRuns as n, CostProvenanceSummary as o, JudgeInsight as p, ExecutionReport as r, ExecutionErrorOutcomeCell as s, AnalyzeRunsOptions as t, FailureClusterInsight as u, ScalarDistribution as v, TokenUsageInsight as y };
-//# sourceMappingURL=analyze-runs-DjlZOVXx.d.ts.map
+//# sourceMappingURL=analyze-runs-BWAHBDAw.d.ts.map

package/dist/{analyze-runs-DjlZOVXx.d.ts.map → analyze-runs-BWAHBDAw.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"analyze-runs-DjlZOVXx.d.ts","names":[],"sources":["../src/contract/insight-report.ts","../src/contract/execution-insight.ts","../src/contract/analyze-runs.ts"],"mappings":";;;;;;;UAqCiB;;EAEf;;;;EAKA,WAAW;;EAGX,WAAW;;;EAIX,cAAc,eAAe;;EAG7B;IACE,MAAM;IACN,QAAQ;;;;;IAKR,aAAa;;;;;IAKb;MAAa;MAAe;;;;;;EAM9B,QAAQ,eAAe;;;;EAKvB,aAAa;;;;EAKb,OAAO;;;EAIP,kBAAkB;;;EAIlB,gBAAgB;;;;;EAMhB,qBAAqB;;;;;EAMrB,SAAS;;;;;;EAOT,wBAAwB;;;;EAKxB,iBAAiB;;;EAIjB,iBAAiB;;UAGF;EACf;IAAY;IAAW;;EACvB;IAAa;IAAW;;;;EAGxB;IAAc;IAAW;;EACzB;IAAc;;EACd;;UAGe;;EAEf,YAAY;;EAEZ,SAAS;;;EAGT,YAAY;;;EAGZ;IACE;IACA,YAAY;IACZ,SAAS;IACT;;;EAGF,QAAQ;IAAQ;IAAe;;;;;EAI/B;IACE;IACA;IACA;;;;;EAKF;IACE;;;IAGA;;IAEA;;IAEA;;IAEA;;IAEA;;;;;IAKA,mBAAmB,OAAO,oBAAoB;;;;EAIhD;IACE;IACA;IACA;IACA;IACA;;;UAIa;;EAEf;;EAEA;;EAEA;;UAGe;EACf,OAAO;EACP,QAAQ;EACR,WAAW;EACX,QAAQ;EACR,YAAY;EACZ;IACE;IACA;IACA;IACA;IACA;;;;UAOa;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,WAAW;;;;;EAKX,WAAW;IAAQ;IAAe;;;UAGnB;;EAEf;;EAEA;;;EAGA,cAAc;;;EAGd;;;EAGA;;;EAGA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,SAAS;;EAET,mBAAmB;IACjB;IACA,SAAS;MAAQ;MAAe;;IAChC;;;UAIa;EACf;EACA;;EAEA;;EAEA;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;EAWA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;UAGe;;EAEf,UAAU;IACR;IACA;;IAEA;;IAEA;;IAEA;;EAEF;;;;;UAMe;;EAEf,cAAc;;EAEd;;EAEA;;UAGe;;EAEf;;;EAGA;EACA,UAAU;IAAQ;IAAe;IAAgB;;;UAGlC;;;EAGf;;EAEA;;EAEA;;EAEA;;EAEA;IACE;IACA;IACA;;;UAIa;;;EAGf;EACA,MAAM;IACJ;IACA;IACA;;;;EAIF;;UAGQ;;EAER;;EAEA;;;;EAIA;;EAEA;EACA;;KAGU,eACP;EACC;;EAEA;;EAEA;;EAEA;;EAEA;MAED;EACC;;EAEA;EACA;EACA;EACA;;UAGW;;EAEf;EACA;;EAEA;;;;EAIA,SAAS,eAAe;;;EAGxB;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;;EAEA;;;;UCzbe;EACf,MAAM;EACN;;UAGe;EACf,WAAW;EACX,gBAAgB;;;iBAIF,mBAAmB,MAAM,4BAA4B;;;UC4BpD;;EAEf,MAAM;;;EAGN;;;;;EAKA;EACA;;;EAGA,8BAA8B;;;EAG9B,kBAAkB;;;EAGlB,UAAU;;;;EAIV;IACE;IACA,cAAc;;;;;EAKhB,cAAc;IAAQ;IAAe;IAAe;;;EAEpD;;;;EAIA;;;;;;;;;EASA,eAAe;;;EAGf;;iBAGoB,YAAY,MAAM,qBAAqB,QAAQ"}
+{"version":3,"file":"analyze-runs-BWAHBDAw.d.ts","names":[],"sources":["../src/contract/insight-report.ts","../src/contract/execution-insight.ts","../src/contract/analyze-runs.ts"],"mappings":";;;;;;;UAqCiB;;EAEf;;;;EAKA,WAAW;;EAGX,WAAW;;;EAIX,cAAc,eAAe;;EAG7B;IACE,MAAM;IACN,QAAQ;;;;;IAKR,aAAa;;;;;IAKb;MAAa;MAAe;;;;;;EAM9B,QAAQ,eAAe;;;;EAKvB,aAAa;;;;EAKb,OAAO;;;EAIP,kBAAkB;;;EAIlB,gBAAgB;;;;;EAMhB,qBAAqB;;;;;EAMrB,SAAS;;;;;;EAOT,wBAAwB;;;;EAKxB,iBAAiB;;;EAIjB,iBAAiB;;UAGF;EACf;IAAY;IAAW;;EACvB;IAAa;IAAW;;;;EAGxB;IAAc;IAAW;;EACzB;IAAc;;EACd;;UAGe;;EAEf,YAAY;;EAEZ,SAAS;;;EAGT,YAAY;;;EAGZ;IACE;IACA,YAAY;IACZ,SAAS;IACT;;;EAGF,QAAQ;IAAQ;IAAe;;;;;EAI/B;IACE;IACA;IACA;;;;;EAKF;IACE;;;IAGA;;IAEA;;IAEA;;IAEA;;IAEA;;;;;IAKA,mBAAmB,OAAO,oBAAoB;;;;EAIhD;IACE;IACA;IACA;IACA;IACA;;;UAIa;;EAEf;;EAEA;;EAEA;;UAGe;EACf,OAAO;EACP,QAAQ;EACR,WAAW;EACX,QAAQ;EACR,YAAY;EACZ;IACE;IACA;IACA;IACA;IACA;;;;UAOa;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,WAAW;;;;;EAKX,WAAW;IAAQ;IAAe;;;UAGnB;;EAEf;;EAEA;;;EAGA,cAAc;;;EAGd;;;EAGA;;;EAGA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,SAAS;;EAET,mBAAmB;IACjB;IACA,SAAS;MAAQ;MAAe;;IAChC;;;UAIa;EACf;EACA;;EAEA;;EAEA;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;EAWA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;UAGe;;EAEf,UAAU;IACR;IACA;;IAEA;;IAEA;;IAEA;;EAEF;;;;;UAMe;;EAEf,cAAc;;EAEd;;EAEA;;UAGe;;EAEf;;;EAGA;EACA,UAAU;IAAQ;IAAe;IAAgB;;;UAGlC;;;EAGf;;EAEA;;EAEA;;EAEA;;EAEA;IACE;IACA;IACA;;;UAIa;;;EAGf;EACA,MAAM;IACJ;IACA;IACA;;;;EAIF;;UAGQ;;EAER;;EAEA;;;;EAIA;;EAEA;EACA;;KAGU,eACP;EACC;;EAEA;;EAEA;;EAEA;;EAEA;MAED;EACC;;EAEA;EACA;EACA;EACA;;UAGW;;EAEf;EACA;;EAEA;;;;EAIA,SAAS,eAAe;;;EAGxB;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;;EAEA;;;;UCzbe;EACf,MAAM;EACN;;UAGe;EACf,WAAW;EACX,gBAAgB;;;iBAIF,mBAAmB,MAAM,4BAA4B;;;UC4BpD;;EAEf,MAAM;;;EAGN;;;;;EAKA;EACA;;;EAGA,8BAA8B;;;EAG9B,kBAAkB;;;EAGlB,UAAU;;;;EAIV;IACE;IACA,cAAc;;;;;EAKhB,cAAc;IAAQ;IAAe;IAAe;;;EAEpD;;;;EAIA;;;;;;;;;EASA,eAAe;;;EAGf;;iBAGoB,YAAY,MAAM,qBAAqB,QAAQ"}

package/dist/{benchmark-command-BupdHRyf.js → benchmark-command-BzmMTnvC.js}

@@ -2,19 +2,19 @@ import { c as errorMessage, s as ValidationError, t as AgentEvalError } from "./
 import { a as hashCanonical, i as compareCodeUnits, o as jsonDocument, r as canonicalString } from "./canonical-Dq2hOQ8L.js";
 import { i as pairedBootstrap, t as mean$3 } from "./mean-BbA_JCzW.js";
 import { a as resolveModelPricing } from "./metrics-BKaPBtIG.js";
-import { a as CostLedgerPersistenceError, i as CostCeilingReachedError, n as CostAccountingIncompleteError, o as CostReceiptCaptureError, r as CostCallConflictError, s as CostReservationExceededError, t as CostLedger } from "./cost-ledger-D0352Z4d.js";
+import { a as CostLedgerPersistenceError, i as CostCeilingReachedError, n as CostAccountingIncompleteError, o as CostReceiptCaptureError, r as CostCallConflictError, s as CostReservationExceededError, t as CostLedger } from "./cost-ledger-BGZB6yY2.js";
 import { l as withLedgerFileLock, u as writeLedgerFileAtomically } from "./journal-C7om8Gum.js";
 import "./ledger-core-CAxUdDk0.js";
-import { M as fsCampaignStorage, b as AnalystFindingSchema, i as startExternalOptimizerModelProxy, j as createRunCostLedger, r as runWithCleanup, v as resolveExternalOptimizerProcessLimits } from "./external-optimizer-subprocess-B5gdLRZv.js";
+import { M as fsCampaignStorage, b as AnalystFindingSchema, i as startExternalOptimizerModelProxy, j as createRunCostLedger, r as runWithCleanup, v as resolveExternalOptimizerProcessLimits } from "./external-optimizer-subprocess-D670Oxf2.js";
 import { s as redactText } from "./redact-mtkTiJVk.js";
 import { i as makeFinding } from "./types-DqFMNTgK.js";
 import { t as acquireSingleRunLock } from "./single-run-lock-Dekd8_Wv.js";
 import { a as LlmCallError, n as isRecord, o as LlmResponseError } from "./json-BhdsJaOQ.js";
-import { A as RawAnalystFindingSchema, U as assertValidAnalystUsageReceipt, W as usageReceiptFromCostLedger, j as evidenceRefsFromRawFinding, k as RAW_FINDING_SCHEMA_PROMPT, m as TRACE_ANALYSIS_LIMITS, n as runTraceAnalyst } from "./kind-factory-Gy4fXixd.js";
-import { n as callLlmJson } from "./llm-client-_Y6ic-tf.js";
-import { t as createDspyRlmTraceEngine } from "./dspy-rlm-engine-CSMwudLO.js";
-import { i as otlpTextToTraceAnalysisStore, n as DEFAULT_MAX_TRACE_FILE_BYTES, t as BufferTraceStore } from "./store-otlp-BteDAbHK.js";
-import "./external-optimizer-process-C3Wm7Vhn.js";
+import { A as RawAnalystFindingSchema, U as assertValidAnalystUsageReceipt, W as usageReceiptFromCostLedger, j as evidenceRefsFromRawFinding, k as RAW_FINDING_SCHEMA_PROMPT, m as TRACE_ANALYSIS_LIMITS, n as runTraceAnalyst } from "./kind-factory-BhxbtUsQ.js";
+import { n as callLlmJson } from "./llm-client-DXuOTseU.js";
+import { t as createDspyRlmTraceEngine } from "./dspy-rlm-engine-Bn_Mt_LW.js";
+import { i as otlpTextToTraceAnalysisStore, n as DEFAULT_MAX_TRACE_FILE_BYTES, t as BufferTraceStore } from "./store-otlp-dyWE3wAG.js";
+import "./external-optimizer-process-wOLfzsCC.js";
 import { t as assertEqualDeclarativeTerms } from "./equal-terms-DbPGcVtI.js";
 import { c as runPrimeExchange, l as decodeReplyRows, n as buildPrimePrompt, o as primeProtocolSha256, s as projectPrimeTrajectory, t as analystUsageReceiptFromPrimeUsage } from "./prime-protocol-BFgGrEMP.js";
 import { t as nodeHttpPrimeBridgeTransport } from "./prime-bridge-transport-Z3cpKGoV.js";
@@ -1877,7 +1877,7 @@ Object.freeze([
 	"package.json",
 	"pnpm-lock.yaml"
 ]);
-const ANALYST_BENCHMARK_DEPENDENCY_LOCK_SHA256 = "79b7d34288f4a5c3fd39cc1d875b98507d822a27ae5adf2ad299c7ba2935fca4";
+const ANALYST_BENCHMARK_DEPENDENCY_LOCK_SHA256 = "2cbded9c738fb0bea1a90ac7da6714c75b6c7aae70aebf6ef4c833df2db652f8";
 const ANALYST_BENCHMARK_EVIDENCE_DEPENDENCY_LOCK_SHA256 = "1e03f2daed356d60316aabefb407ec1e437ac94d408d61eea4ae096e9c6fbb5b";
 const ANALYST_BENCHMARK_EVIDENCE_IMPLEMENTATION_SHA256 = "4dba263b6256a30d56c7fdb2d992d3a953c0035d731f359b704db806f68f75ac";
 const ANALYST_BENCHMARK_IMPLEMENTATION_FILES = Object.freeze([
@@ -2029,7 +2029,7 @@ const ANALYST_BENCHMARK_IMPLEMENTATION_FILES = Object.freeze([
 	"src/trace/redaction/detectors.ts",
 	"src/verdict-cache.ts"
 ]);
-const ANALYST_BENCHMARK_IMPLEMENTATION_SHA256 = "f3ef71f1819dc0f9b35907c01bd2218e2802931b8749b77749127c3054149f3e";
+const ANALYST_BENCHMARK_IMPLEMENTATION_SHA256 = "ae18283837ab8e3eb1feb97e54f9b2574a45737e455a8c6b457c17f044bf5df9";
 //#endregion
 //#region src/analyst/benchmark-evidence-validation.ts
 const MIN_ACTION_EXCERPT_CHARACTERS = 12;
@@ -6890,4 +6890,4 @@ function shellQuote(value) {
 //#endregion
 export { agentRxBenchmarkCase as A, ANALYST_BENCHMARK_COST_LEDGER_FILE as C, AGENT_RX_UPSTREAM_REVISION as D, ANALYST_BENCHMARK_OBSERVATIONS_FILE as E, normalizeAgentRxCategory as M, runAnalystBenchmark as N, codeTraceBenchCase as O, traceStoreEvidenceResolver as P, ANALYST_BENCHMARK_IMPLEMENTATION_FILES as S, ANALYST_BENCHMARK_MANIFEST_FILE as T, readAnalystBenchmarkArtifact as _, runInlineAnalystDefinition as a, ANALYST_BENCHMARK_EVIDENCE_DEPENDENCY_LOCK_SHA256 as b, publicRlmAnalystDefinition as c, publicDirectAnalystDefinition as d, runChunkedAnalystDefinition as f, expandCodeTraceFailureBlocks as g, analystDefinitionProtocolSha256 as h, primeCodeTraceAnalystDefinition as i, agentRxPredictionsToFindings as j, codeTracerPredictionsToFindings as k, runReplVariableAnalystDefinition as l, analystDefinitionAsymmetries as m, createPrimeBenchmarkRunner as n, loadPublicBenchmarkRows as o, AnalystExpressivenessError as p, primeAnalystProtocolSha256 as r, createPublicBenchmarkRlmRunner as s, runAnalystBenchmarkCommand as t, createPublicBenchmarkDirectRunner as u, CODE_TRACE_BENCH_ANALYST_PROMPT as v, ANALYST_BENCHMARK_LOCAL_RECEIPT_FILE as w, ANALYST_BENCHMARK_EVIDENCE_IMPLEMENTATION_SHA256 as x, publicBenchmarkProtocolSha256 as y };
 
-//# sourceMappingURL=benchmark-command-BupdHRyf.js.map
+//# sourceMappingURL=benchmark-command-BzmMTnvC.js.map