@tangle-network/agent-eval 0.221.7 → 0.222.0

package/dist/{skillopt-optimization-method-sB_-yi1G.js → skillopt-optimization-method-qB1M554K.js}

@@ -1,23 +1,24 @@
-import { a as hashCanonical } from "./canonical-Dq2hOQ8L.js";
-import { t as mean$1 } from "./mean-BbA_JCzW.js";
-import { t as decidePairedPromotion } from "./paired-promotion-decision-D0SHb-ZL.js";
-import { a as heldoutSignificance, o as pairHoldout } from "./statistical-heldout-Bksv92Th.js";
-import { n as contentHash, t as canonicalJson } from "./verdict-cache-CL_7tePD.js";
-import { E as assertSearchHistoryAdmissionOptions, J as evaluationUnitMap, K as assertIndependentEvaluationSplit, Q as runCampaign, S as developmentClaim, T as surfaceNode, X as reserveFinalEvidence, Y as exposeFinalEvidence, c as assertGepaCandidatePopulationSummary, d as costFromLedgerSummary, et as resolveRunDir, f as optimizationTokenUsageFromSummary, l as readGepaCandidatePopulationArtifact, q as captureFinalEvidencePolicy, s as executeOptimizationMethod, u as combineComparisonCosts, w as surfaceDiff, x as SearchRecorder, z as captureJudge } from "./llm-judge-Cybq29bL.js";
-import { D as surfaceContentHash, I as assertCompleteCampaign, P as assertCampaignDesign, f as campaignBreakdown, t as createCampaignEvidenceReceipt } from "./campaign-evidence-DunzcJGv.js";
-import { a as mapConcurrent } from "./journal-C7om8Gum.js";
-import { t as deepFreezeCanonicalJson } from "./deep-freeze-CurcOQg3.js";
-import { M as fsCampaignStorage, _ as resolveExternalOptimizerCallbackLimits, d as assertExternalOptimizerModelBudget, f as assertJsonValue, g as removeCredentialEnvironment, h as isExternalTextCandidate, i as startExternalOptimizerModelProxy, j as createRunCostLedger, m as isCandidateText, n as closeExternalOptimizerResources, p as assertNoCredentialValues, r as runWithCleanup, t as runExternalOptimizerProcess, v as resolveExternalOptimizerProcessLimits, y as safePathComponent } from "./external-optimizer-subprocess-B_ukPtRD.js";
-import { r as summarizeEvaluationUnits, t as defineEvaluationClaim } from "./claim-BPVuyHuy.js";
-import { n as openSearchLedger } from "./search-ledger-CZRqtFbW.js";
-import { t as acquireSingleRunLock } from "./single-run-lock-Dekd8_Wv.js";
-import { n as isRecord } from "./json-BhdsJaOQ.js";
-import { d as assertServedModelPolicy, v as combineAbortSignals } from "./llm-client-b1ysZVIU.js";
-import { t as startExternalOptimizerCallback } from "./external-optimizer-process-DEzJ2CGX.js";
+import { a as hashCanonical } from "./canonical-CicJlINA.js";
+import { t as mean$1 } from "./mean-CdDYG-us.js";
+import { t as decidePairedPromotion } from "./paired-promotion-decision-FGvHA4p_.js";
+import { a as heldoutSignificance, o as pairHoldout } from "./statistical-heldout-TP6ndL0B.js";
+import { d as mapConcurrent } from "./journal-C1uLJi0w.js";
+import { n as contentHash, t as canonicalJson } from "./verdict-cache-BVH-fl18.js";
+import { $ as resolveRunDir, C as surfaceDiff, G as assertIndependentEvaluationSplit, J as exposeFinalEvidence, K as captureFinalEvidencePolicy, R as captureJudge, T as assertSearchHistoryAdmissionOptions, Y as reserveFinalEvidence, Z as runCampaign, b as SearchRecorder, c as readGepaCandidatePopulationArtifact, d as optimizationTokenUsageFromSummary, l as combineComparisonCosts, o as executeOptimizationMethod, q as evaluationUnitMap, s as assertGepaCandidatePopulationSummary, u as costFromLedgerSummary, w as surfaceNode, x as developmentClaim } from "./llm-judge-cgbhTKQp.js";
+import { D as assertCompleteCampaign, T as assertCampaignDesign, f as campaignBreakdown, t as createCampaignEvidenceReceipt } from "./campaign-evidence-DN8rWuBj.js";
+import { t as deepFreezeCanonicalJson } from "./deep-freeze-CK8-2W7p.js";
+import { M as fsCampaignStorage, _ as resolveExternalOptimizerCallbackLimits, d as assertExternalOptimizerModelBudget, f as assertJsonValue, g as removeCredentialEnvironment, h as isExternalTextCandidate, i as startExternalOptimizerModelProxy, j as createRunCostLedger, m as isCandidateText, n as closeExternalOptimizerResources, p as assertNoCredentialValues, r as runWithCleanup, t as runExternalOptimizerProcess, v as resolveExternalOptimizerProcessLimits, y as safePathComponent } from "./external-optimizer-subprocess-U616hV_l.js";
+import { s as surfaceContentHash } from "./json-recovery-DK4MDm5r.js";
+import { r as summarizeEvaluationUnits, t as defineEvaluationClaim } from "./claim-Dod_mjk3.js";
+import { n as openSearchLedger } from "./search-ledger-BsVi-Ngv.js";
+import { t as acquireSingleRunLock } from "./single-run-lock-BGlW3OSk.js";
+import { n as isRecord } from "./json-DtvZNAU9.js";
+import { c as combineAbortSignals, f as assertServedModelPolicy } from "./llm-client-kJ9EOqjt.js";
+import { t as startExternalOptimizerCallback } from "./external-optimizer-process-BRVmaDFe.js";
 import { createHash, randomBytes } from "node:crypto";
 import { mkdir } from "node:fs/promises";
 import { isDeepStrictEqual } from "node:util";
-//#region src/campaign/external-optimizer-observations.ts
+// src/campaign/external-optimizer-observations.ts
 /**
 * Read and verify the exact callback observation artifact addressed by method provenance.
 *
@@ -254,8 +255,7 @@ function openExternalOptimizerExecutionLog(input) {
 		}
 	};
 }
-//#endregion
-//#region src/campaign/external-text-evaluation.ts
+// src/campaign/external-text-evaluation.ts
 function createExternalTextEvaluator(args) {
 	const cached = /* @__PURE__ */ new Map();
 	return async ({ candidate, exampleId }, signal) => {
@@ -387,8 +387,7 @@ function deepFreeze(value) {
 	}
 	return value;
 }
-//#endregion
-//#region src/campaign/external-optimizer-run-budget.ts
+// src/campaign/external-optimizer-run-budget.ts
 function externalOptimizerRunKey(input) {
 	if (!input.attemptId.trim() || input.attemptId.trim() !== input.attemptId) throw new Error("external optimizer attemptId must be trimmed and non-empty");
 	if (typeof input.resumeEnabled !== "boolean") throw new Error("external optimizer resumeEnabled must be a boolean");
@@ -457,8 +456,7 @@ function readState(text, maxEvaluations, statePath) {
 		accepted
 	};
 }
-//#endregion
-//#region src/campaign/external-text-optimization-contract.ts
+// src/campaign/external-text-optimization-contract.ts
 const MAX_TIMER_DELAY_MS$3 = 2147483647;
 function assertExternalTextOptimizationConfig(config) {
 	for (const [field, value] of [
@@ -551,8 +549,7 @@ function assertExternalCostAccounting(value, calls, name) {
 	}
 	if (accounting.kind !== "external" || typeof accounting.reason !== "string" || !accounting.reason.trim() || accounting.reason.trim() !== accounting.reason) throw new Error(`${name}: optimizer returned invalid costAccounting`);
 }
-//#endregion
-//#region src/campaign/external-text-optimization.ts
+// src/campaign/external-text-optimization.ts
 const DEFAULT_MAX_CANDIDATE_CHARS = 2e5;
 const DEFAULT_MAX_EVIDENCE_CHARS = 1e5;
 const DEFAULT_TIMEOUT_MS = 18e5;
@@ -971,8 +968,7 @@ async function raceWithAbort(promise, signal) {
 		});
 	});
 }
-//#endregion
-//#region src/campaign/gates/compose.ts
+// src/campaign/gates/compose.ts
 /** Compose gates — all must `ship` for the composite to `ship`. First
 *  non-ship verdict short-circuits the composite verdict, but ALL gates run
 *  (so the result records every gate's reason — useful for diagnostics). */
@@ -1005,8 +1001,7 @@ function composeGate(...gates) {
 		}
 	};
 }
-//#endregion
-//#region src/campaign/gates/heldout-gate.ts
+// src/campaign/gates/heldout-gate.ts
 /**
 * Composable held-out gate: ships only when the lower bound of the DECIDING
 * paired interval on the candidate-minus-baseline composite delta clears
@@ -1071,8 +1066,7 @@ function heldOutGate(options) {
 		}
 	};
 }
-//#endregion
-//#region src/campaign/external-optimizer-run-config.ts
+// src/campaign/external-optimizer-run-config.ts
 function externalOptimizerRunnerIdentity(runner, module) {
 	return {
 		command: runner?.command ?? "python",
@@ -1095,8 +1089,7 @@ function snapshotJson(value, label) {
 	if (serialized === void 0) throw new Error(`${label} must be JSON-serializable`);
 	return JSON.parse(serialized);
 }
-//#endregion
-//#region src/campaign/optimizer-model.ts
+// src/campaign/optimizer-model.ts
 function assertOptimizerModel(value, label) {
 	if (!value || typeof value !== "object") throw new Error(`${label} is required`);
 	for (const [field, item] of [["model", value.model], ["callRef", value.callRef]]) if (typeof item !== "string" || !item.trim() || item.trim() !== item) throw new Error(`${label}.${field} must be trimmed and non-empty`);
@@ -1292,8 +1285,7 @@ function addEvaluationLimit(total, increment) {
 	if (!Number.isSafeInteger(next)) throw new Error("gepaOptimizationMethod: recipe evaluation limit exceeds safe integer range");
 	return next;
 }
-//#endregion
-//#region src/campaign/external-optimizer-accounting.ts
+// src/campaign/external-optimizer-accounting.ts
 const TOKEN_USAGE_FIELDS = [
 	"inputTokens",
 	"outputTokens",
@@ -1316,8 +1308,7 @@ function assertExternalOptimizerCompletionCount(upstream, requestAttempts, succe
 	if (upstream.calls !== successfulCompletions) throw new Error(`${name}: ${optimizer} reported ${upstream.calls} successful model calls but the proxy completed ${successfulCompletions} across ${requestAttempts} attempts`);
 	if (upstream.requestAttempts !== void 0 && upstream.requestAttempts !== requestAttempts) throw new Error(`${name}: ${optimizer} reported ${upstream.requestAttempts} model attempts but the proxy received ${requestAttempts}`);
 }
-//#endregion
-//#region src/campaign/external-optimizer-source.ts
+// src/campaign/external-optimizer-source.ts
 function assertExternalOptimizerPackageSource(value, expectedPackage, name, optimizer) {
 	assertExternalOptimizerPackageIdentity(value, expectedPackage, name, optimizer);
 	assertExternalOptimizerSourceDetails(value, name, optimizer);
@@ -1341,8 +1332,7 @@ function observedExternalOptimizerPackageSource(value) {
 		...value.sourceSha256 ? { sourceSha256: value.sourceSha256 } : {}
 	};
 }
-//#endregion
-//#region src/campaign/external-optimizer-runtime.ts
+// src/campaign/external-optimizer-runtime.ts
 async function inspectExternalOptimizerRuntime(args) {
 	const result = await runExternalOptimizerProcess({
 		label: `${args.label} source inspection`,
@@ -1382,8 +1372,7 @@ function assertExternalOptimizerRuntimeIdentity(value, expectedPackage, label) {
 	if (!isRecord(value.python) || typeof value.python.implementation !== "string" || value.python.implementation.length === 0 || value.python.implementation !== value.python.implementation.trim() || typeof value.python.version !== "string" || value.python.version.length === 0 || value.python.version !== value.python.version.trim()) throw new Error(`${label}: source inspection returned an invalid Python runtime`);
 	if (!Array.isArray(value.engineModules) || value.engineModules.some((module) => !isRecord(module) || typeof module.module !== "string" || module.module.length === 0 || module.module !== module.module.trim() || typeof module.sourceSha256 !== "string" || !/^[0-9a-f]{64}$/.test(module.sourceSha256))) throw new Error(`${label}: source inspection returned invalid engine modules`);
 }
-//#endregion
-//#region src/campaign/gepa-optimization-result.ts
+// src/campaign/gepa-optimization-result.ts
 function assertGepaBridgeOutput(result, name, maxCandidateChars, recipeKind, maxEvaluations, maxPopulationCandidates, scenarioIds, expectsComponents, requiresCandidatePopulation) {
 	if (result.recipeKind !== recipeKind) throw new Error(`${name}: GEPA bridge reported recipe '${String(result.recipeKind)}'`);
 	if (!isGepaCandidate(result.bestCandidate, maxCandidateChars) || expectsComponents !== (typeof result.bestCandidate !== "string")) throw new Error(`${name}: GEPA bridge returned an invalid candidate`);
@@ -1409,8 +1398,7 @@ function isGepaCandidate(value, maxChars) {
 	if (!isExternalTextCandidate(value)) return false;
 	return (typeof value === "string" ? value.length : JSON.stringify(value).length) <= maxChars;
 }
-//#endregion
-//#region src/campaign/gepa-search-import.ts
+// src/campaign/gepa-search-import.ts
 /**
 * Import an external optimizer's own record into a search ledger.
 *
@@ -1804,8 +1792,7 @@ async function recordExternalSearch(input) {
 		runId: input.searchId
 	});
 }
-//#endregion
-//#region src/campaign/gepa-optimization-run.ts
+// src/campaign/gepa-optimization-run.ts
 function anthropicShimEnvironment(proxy, optimizer) {
 	return {
 		ANTHROPIC_BASE_URL: new URL(proxy.baseUrl).origin,
@@ -2261,8 +2248,7 @@ function buildGepaOutcome(ctx, result, outputDir, modelProxy, searchHistory, cos
 		}
 	};
 }
-//#endregion
-//#region src/campaign/gepa-optimization-method.ts
+// src/campaign/gepa-optimization-method.ts
 /**
 * Turn an optional GEPA installation into an `OptimizationMethod`.
 *
@@ -2293,8 +2279,7 @@ function gepaOptimizationMethod(config) {
 		}
 	};
 }
-//#endregion
-//#region src/campaign/presets/comparison-controls.ts
+// src/campaign/presets/comparison-controls.ts
 /** Admission checks for the shared controls of a method comparison. */
 /** The caller-supplied judges, dispatch function and references, and the baseline surface. */
 function assertCallerInputs(opts) {
@@ -2356,8 +2341,7 @@ function minimumBootstrapResamples(confidence, comparisonCount) {
 	const exact = 2 * comparisonCount / (1 - confidence);
 	return Math.ceil(exact - Number.EPSILON * Math.max(1, exact) * 32);
 }
-//#endregion
-//#region src/campaign/presets/compare-optimization-methods.ts
+// src/campaign/presets/compare-optimization-methods.ts
 /**
 * Compare optimization methods on shared train, selection, and test data.
 * Optimizers receive only train and selection data. After every optimizer
@@ -2883,8 +2867,7 @@ function assertPositiveSafeInteger(value, label) {
 function assertOptionalEnum(value, allowed, label) {
 	if (value !== void 0 && !allowed.includes(value)) throw new Error(`skillOptOptimizationMethod: ${label} must be one of ${allowed.join(", ")}`);
 }
-//#endregion
-//#region src/campaign/skillopt-optimization-result.ts
+// src/campaign/skillopt-optimization-result.ts
 function assertSkillOptBridgeOutput(result, name, maxCandidateChars, maxEvaluations) {
 	if (!isCandidateText(result.bestCandidate, maxCandidateChars)) throw new Error(`${name}: SkillOpt bridge returned an invalid candidate`);
 	if (!Number.isFinite(result.bestScore) || result.bestScore < 0 || result.bestScore > 1) throw new Error(`${name}: SkillOpt bridge returned an invalid bestScore`);
@@ -2896,8 +2879,7 @@ function assertSkillOptBridgeOutput(result, name, maxCandidateChars, maxEvaluati
 	assertExternalOptimizerSourceDetails(result.upstream, name, "SkillOpt");
 	assertExternalOptimizerTokenUsage(result.tokenUsage, name, "SkillOpt");
 }
-//#endregion
-//#region src/campaign/skillopt-optimization-method.ts
+// src/campaign/skillopt-optimization-method.ts
 /** Run Microsoft's SkillOpt trainer as a complete optimization method. */
 function skillOptOptimizationMethod(config) {
 	assertSkillOptOptimizationConfig(config);
@@ -3242,7 +3224,6 @@ async function finishSkillOptRun(context, result, outputDir, modelProxy) {
 		}
 	};
 }
-//#endregion
 export { importExternalEvaluations as a, composeGate as c, readExternalOptimizerObservationArtifact as d, externalSurface as i, externalTextOptimizationMethod as l, compareOptimizationMethods as n, importGepaPopulation as o, gepaOptimizationMethod as r, heldOutGate as s, skillOptOptimizationMethod as t, decodeExternalTextCandidate as u };
 
-//# sourceMappingURL=skillopt-optimization-method-sB_-yi1G.js.map
+//# sourceMappingURL=skillopt-optimization-method-qB1M554K.js.map