@tangle-network/agent-eval 0.190.1 → 0.192.0

package/dist/{campaign-evidence-D40-LHIL.js → campaign-evidence-DcQiYhJI.js}

@@ -1,10 +1,10 @@
-import { s as ValidationError, t as AgentEvalError } from "./errors-Dngq5h35.js";
+import { t as AgentEvalError } from "./errors-Dngq5h35.js";
 import { a as hashCanonical, i as compareCodeUnits, r as canonicalString } from "./canonical-BsZVgBhW.js";
 import { n as aggregatePairedHoldout, o as pairHoldout } from "./statistical-heldout-GcLIPxk4.js";
 import { n as contentHash } from "./verdict-cache-BQEsVy9n.js";
-import { n as campaignCellExecutionEvidence, o as projectCampaignCellQuality } from "./run-record-DnMFyg_g.js";
-import { t as FileLedgerJournal } from "./journal-c3qY0SDP.js";
-import { g as readField } from "./ast-CP9ae9B0.js";
+import { r as projectCampaignCellQuality } from "./run-record-tNnlQIrD.js";
+import { t as FileLedgerJournal } from "./journal-DvKCvjty.js";
+import { t as defineEvaluationClaim } from "./claim-V9Zqbq0J.js";
 import { createHash } from "node:crypto";
 import { z } from "zod";
 import { join } from "node:path";
@@ -403,53 +403,6 @@ function surfaceDispatchRef(surface, executionRef = "anonymous") {
 	return `surface:${executionRef}:${surfaceContentHash(surface)}`;
 }
 //#endregion
-//#region src/experiment/claim.ts
-const nonEmpty = z.string().min(1).refine((value) => value.trim() === value);
-/** The claim's schema, shared by every record that binds a claim. */
-const evaluationClaimSchema = z.object({
-	use: z.enum([
-		"development",
-		"comparison",
-		"certification"
-	]),
-	population: z.object({
-		id: nonEmpty,
-		description: nonEmpty
-	}).strict(),
-	samplingFrame: nonEmpty,
-	independentUnit: nonEmpty,
-	generalization: z.enum(["fixed-roster", "new-units"]),
-	minimumEffect: z.number().finite().positive().optional()
-}).strict();
-/** Validate a claim before binding it to a sealed experiment or final evidence. */
-function defineEvaluationClaim(input) {
-	const parsed = evaluationClaimSchema.safeParse(input);
-	if (!parsed.success) throw new ValidationError(`invalid evaluation claim: ${parsed.error.message}`);
-	return Object.freeze({
-		...parsed.data,
-		population: Object.freeze(parsed.data.population)
-	});
-}
-/** Repetitions retain their denominator without becoming additional independent units. */
-function summarizeEvaluationUnits(claim, rows) {
-	const validated = defineEvaluationClaim(claim);
-	const counts = /* @__PURE__ */ new Map();
-	for (const [index, row] of rows.entries()) {
-		if (row === null || typeof row !== "object" || Array.isArray(row)) throw new ValidationError(`evaluation claim: row ${index} must be an object`);
-		const value = readField(row, validated.independentUnit);
-		if (typeof value !== "string" || !value.trim() || value.trim() !== value) throw new ValidationError(`evaluation claim: row ${index} needs a nonempty string at '${validated.independentUnit}'`);
-		counts.set(value, (counts.get(value) ?? 0) + 1);
-	}
-	return {
-		observations: rows.length,
-		independentUnits: counts.size,
-		units: [...counts].sort(([left], [right]) => compareCodeUnits(left, right)).map(([id, observations]) => ({
-			id,
-			observations
-		}))
-	};
-}
-//#endregion
 //#region src/experiment/final-evidence.ts
 const identity = z.string().min(1).refine((value) => value.trim() === value);
 const digest = z.string().regex(/^sha256:[a-f0-9]{64}$/);
@@ -1017,25 +970,6 @@ function campaignMeanCompositeOrNull(campaign) {
 	});
 	return scores.length === 0 ? null : scores.reduce((sum, score) => sum + score, 0) / scores.length;
 }
-/** Reject rank keys that cannot produce deterministic lexicographic ordering. */
-function assertFiniteRankKey(key, label, expectedLength) {
-	if (!Array.isArray(key) || key.length === 0) throw new Error(`${label} must return a non-empty array`);
-	if (expectedLength !== void 0 && key.length !== expectedLength) throw new Error(`${label} returned ${key.length} elements; expected ${expectedLength}`);
-	for (let index = 0; index < key.length; index++) if (!Number.isFinite(key[index])) throw new Error(`${label}[${index}] must be finite`);
-}
-/** Compare fixed-length lexicographic rank keys where each element is higher-is-better.
-*  Returns a positive number when `a` ranks above `b`, negative when below, and
-*  zero when equal. */
-function compareRankKeys(a, b) {
-	assertFiniteRankKey(a, "rank key a");
-	assertFiniteRankKey(b, "rank key b", a.length);
-	for (let i = 0; i < a.length; i++) {
-		const av = a[i];
-		const bv = b[i];
-		if (av !== bv) return av - bv;
-	}
-	return 0;
-}
 /** Per-candidate evidence a reflective/patch proposer grounds its next proposal
 *  on: mean score per judge dimension + per-scenario composite. */
 function campaignBreakdown(campaign) {
@@ -1182,7 +1116,7 @@ function buildLoopProvenanceRecord(args) {
 		if (surfaceByHash.size !== candidateByHash.size) throw new Error("buildLoopProvenanceRecord: every measured candidate requires exactly one surface");
 		for (const promotedHash of promotedSet) if (!candidateByHash.has(promotedHash)) throw new Error("buildLoopProvenanceRecord: promoted hash has no measured candidate");
 		for (const c of gen.candidates) {
-			validateCandidateMeasurement(c, incumbentSurfaceHash, incumbentComposite, promotedSet.has(c.surfaceHash));
+			validateCandidateMeasurement(c, incumbentComposite, promotedSet.has(c.surfaceHash));
 			const measured = surfaceByHash.get(c.surfaceHash);
 			if (measured === void 0) throw new Error("buildLoopProvenanceRecord: measured candidate is missing its surface");
 			const { surface, campaign } = measured;
@@ -1193,8 +1127,6 @@ function buildLoopProvenanceRecord(args) {
 				surfaceHash: c.surfaceHash,
 				contentHash: surfaceContentHash(surface),
 				campaignDigest: campaignMeasurementDigest(campaign),
-				parentSurfaceHash: c.parentSurfaceHash,
-				parentComposite: c.parentComposite,
 				eligibleForPromotion: c.eligibleForPromotion,
 				coverage: {
 					expectedCells: c.coverage.expectedCells,
@@ -1207,7 +1139,6 @@ function buildLoopProvenanceRecord(args) {
 			if (c.label) entry.label = c.label;
 			if (c.rationale) entry.rationale = c.rationale;
 			if (c.attribution) entry.attribution = c.attribution;
-			if (c.observedDeltaFromParent !== void 0) entry.observedDeltaFromParent = c.observedDeltaFromParent;
 			candidates.push(entry);
 		}
 		const promotedHash = gen.promoted[0];
@@ -1377,14 +1308,10 @@ function assertGateContributions(value, source) {
 		if ("passed" in item) throw new Error(`${source}: gate contribution '${item.name}' uses obsolete passed; use status instead`);
 	}
 }
-function validateCandidateMeasurement(candidate, expectedParentHash, expectedParentComposite, promoted) {
-	if (!candidate.parentSurfaceHash || !/^[a-f0-9]{16}$/.test(candidate.parentSurfaceHash)) throw new Error("buildLoopProvenanceRecord: parentSurfaceHash must be 16 lowercase hex characters");
-	if (candidate.parentSurfaceHash !== expectedParentHash) throw new Error("buildLoopProvenanceRecord: candidate parent does not match the incumbent");
-	if (candidate.parentComposite === void 0 || !Number.isFinite(candidate.parentComposite) || Math.abs(candidate.parentComposite - expectedParentComposite) > 1e-12) throw new Error("buildLoopProvenanceRecord: candidate parentComposite does not match the incumbent");
-	if (candidate.observedDeltaFromParent !== void 0) {
-		if (!Number.isFinite(candidate.observedDeltaFromParent)) throw new Error("buildLoopProvenanceRecord: observedDeltaFromParent must be finite");
-		if (candidate.eligibleForPromotion !== true) throw new Error("buildLoopProvenanceRecord: observedDeltaFromParent requires a complete eligible candidate and parentSurfaceHash");
-	}
+/** A candidate's coverage receipt and composite agree, and a promoted
+* candidate is complete and beats the incumbent it replaced. Its lineage is
+* the search ledger's to prove, not this record's. */
+function validateCandidateMeasurement(candidate, incumbentComposite, promoted) {
 	const coverage = candidate.coverage;
 	if (!Number.isSafeInteger(coverage.expectedCells) || coverage.expectedCells <= 0 || !Number.isSafeInteger(coverage.scorableCells) || coverage.scorableCells < 0 || coverage.scorableCells > coverage.expectedCells) throw new Error("buildLoopProvenanceRecord: invalid candidate coverage denominator");
 	const unscorableIds = /* @__PURE__ */ new Set();
@@ -1397,14 +1324,8 @@ function validateCandidateMeasurement(candidate, expectedParentHash, expectedPar
 	if (candidate.eligibleForPromotion !== complete) throw new Error("buildLoopProvenanceRecord: candidate eligibility contradicts its coverage receipt");
 	if (complete) {
 		if (candidate.composite === null || !Number.isFinite(candidate.composite)) throw new Error("buildLoopProvenanceRecord: complete candidate composite must be finite");
-		if (candidate.observedDeltaFromParent === void 0) throw new Error("buildLoopProvenanceRecord: complete candidate is missing observedDeltaFromParent");
-		const recomputed = candidate.composite - candidate.parentComposite;
-		if (Math.abs(candidate.observedDeltaFromParent - recomputed) > 1e-12) throw new Error("buildLoopProvenanceRecord: observed delta does not match measured scores");
-	} else {
-		if (candidate.composite !== null && !Number.isFinite(candidate.composite)) throw new Error("buildLoopProvenanceRecord: candidate composite must be finite or null");
-		if (candidate.observedDeltaFromParent !== void 0) throw new Error("buildLoopProvenanceRecord: incomplete candidate cannot carry observed delta");
-	}
-	if (promoted && (!complete || (candidate.observedDeltaFromParent ?? 0) <= 0)) throw new Error("buildLoopProvenanceRecord: promoted candidate must improve the incumbent");
+	} else if (candidate.composite !== null && !Number.isFinite(candidate.composite)) throw new Error("buildLoopProvenanceRecord: candidate composite must be finite or null");
+	if (promoted && (!complete || (candidate.composite ?? Number.NEGATIVE_INFINITY) <= incumbentComposite)) throw new Error("buildLoopProvenanceRecord: promoted candidate must improve the incumbent");
 }
 function hashId(parts) {
 	return createHash("sha256").update(parts.join(":")).digest("hex");
@@ -1496,8 +1417,6 @@ function loopProvenanceSpans(record, opts = {}) {
 				"tangle.generation": generation,
 				"tangle.surfaceHash": c.surfaceHash,
 				"tangle.contentHash": c.contentHash,
-				"tangle.parentSurfaceHash": c.parentSurfaceHash,
-				"tangle.parentComposite": c.parentComposite,
 				"tangle.eligibleForPromotion": c.eligibleForPromotion,
 				"tangle.expectedCells": c.coverage.expectedCells,
 				"tangle.scorableCells": c.coverage.scorableCells,
@@ -1505,7 +1424,6 @@ function loopProvenanceSpans(record, opts = {}) {
 				"tangle.promoted": c.promoted
 			};
 			if (c.composite !== null) attributes["tangle.composite"] = c.composite;
-			if (c.observedDeltaFromParent !== void 0) attributes["tangle.observedDeltaFromParent"] = c.observedDeltaFromParent;
 			if (c.label) attributes["tangle.candidateLabel"] = c.label;
 			if (c.rationale) attributes["tangle.candidateRationale"] = c.rationale;
 			spans.push({
@@ -1556,52 +1474,6 @@ function provenanceRecordPath(runDir) {
 function provenanceSpansPath(runDir) {
 	return join(runDir, "loop-provenance-spans.jsonl");
 }
-/** Snapshot a held-out campaign into the hosted `EvalRunGenerationSnapshot`
-*  shape — per-cell composite + per-judge dimensions, aggregate mean, cost,
-*  duration. The dashboard renders these as the baseline → winner comparison. */
-function snapshotFromHoldout(index, surfaceHash, surface, campaign) {
-	return {
-		index,
-		surfaceHash,
-		surface,
-		cells: campaign.cells.map((cell) => {
-			const execution = campaignCellExecutionEvidence(cell);
-			const quality = projectCampaignCellQuality(cell);
-			const score = {
-				scenarioId: cell.scenarioId,
-				rep: cell.rep,
-				compositeMean: quality.score ?? null,
-				dimensions: quality.judgeScores?.perJudge ?? {},
-				terminalOutcome: execution.terminalOutcome,
-				executionErrorCount: execution.executionErrorCount ?? null
-			};
-			if (cell.error) score.errorMessage = cell.error;
-			return score;
-		}),
-		compositeMean: campaignMeanCompositeOrNull(campaign),
-		costUsd: campaign.aggregates.cost.totalCostUsd,
-		durationMs: campaign.durationMs
-	};
-}
-/** Build the hosted `EvalRunEvent` from the loop args + record — baseline +
-*  winner snapshots, gate decision, held-out lift, cost, duration. Shipped to
-*  `/v1/ingest/eval-runs` so the run appears in the dashboard's run list (the
-*  trace spans, shipped separately, back the per-candidate drill-down). */
-function buildEvalRunEvent(args, record) {
-	return {
-		runId: args.runId,
-		runDir: args.runDir,
-		timestamp: args.timestamp,
-		status: "finished",
-		labels: {},
-		baseline: snapshotFromHoldout(0, record.baselineContentHash, args.baselineSurface, args.baselineOnHoldout),
-		generations: [snapshotFromHoldout(1, record.winnerContentHash, args.winnerSurface, args.winnerOnHoldout)],
-		gateDecision: args.gate.decision,
-		...record.heldOutLift !== void 0 ? { holdoutLift: record.heldOutLift } : {},
-		totalCostUsd: args.totalCostUsd,
-		totalDurationMs: args.totalDurationMs
-	};
-}
 /**
 * Build the provenance record + OTel spans and persist them durably under the
 * run dir (and ship spans to a hosted collector when one is wired). Returns
@@ -1620,19 +1492,11 @@ async function emitLoopProvenance(args) {
 	const spansPath = provenanceSpansPath(args.runDir);
 	args.storage.write(recordPath, JSON.stringify(record, null, 2));
 	args.storage.write(spansPath, spans.map((s) => JSON.stringify(s)).join("\n"));
-	if (args.hostedClient) {
-		try {
-			await args.hostedClient.ingestEvalRun(buildEvalRunEvent(args, record));
-		} catch (err) {
-			const msg = err instanceof Error ? err.message : String(err);
-			console.warn(`[agent-eval] hosted eval-run ingest failed (continuing): ${msg}`);
-		}
-		try {
-			await args.hostedClient.ingestTraces(spans);
-		} catch (err) {
-			const msg = err instanceof Error ? err.message : String(err);
-			console.warn(`[agent-eval] provenance span ingest failed (continuing): ${msg}`);
-		}
+	if (args.hostedClient) try {
+		await args.hostedClient.ingestTraces(spans);
+	} catch (err) {
+		const msg = err instanceof Error ? err.message : String(err);
+		console.warn(`[agent-eval] provenance span ingest failed (continuing): ${msg}`);
 	}
 	return {
 		record,
@@ -1872,6 +1736,6 @@ function createCampaignEvidenceReceipt(input) {
 	});
 }
 //#endregion
-export { campaignSplitDigestFromIdentities as $, FinalEvidenceConflictError as A, surfaceContentHash as B, campaignMeanCompositeOrNull as C, parseReflectionResponse as D, buildReflectionPrompt as E, summarizeEvaluationUnits as F, assertRealBackend as G, surfaceHash as H, assertCodeSurfaceIdentity as I, assertCampaignSplitIdentity as J, summarizeBackendIntegrity as K, codeSurfaceIdentityMaterial as L, openFinalEvidenceLedger as M, defineEvaluationClaim as N, recoverTruncatedJson as O, evaluationClaimSchema as P, campaignSplitDigest as Q, componentSurfaceIdentityMaterial as R, campaignMeanComposite as S, pairedCampaignComposites as T, surfaceHashMatches as U, surfaceDispatchRef as V, BackendIntegrityError as W, campaignCoverage as X, assertCompleteCampaign as Y, campaignScenarioIdentity as Z, provenanceRecordPath as _, createEvidenceReceipt as a, assertFiniteRankKey as b, ATTESTATION_ALGORITHM as c, buildLoopProvenanceRecord as d, formatCoverageFailures as et, campaignMeasurementDigest as f, loopProvenanceSpans as g, loopProvenanceArgsFromResult as h, INDEPENDENT_EVIDENCE_AUTHORITY_KINDS as i, FinalEvidenceError as j, powerPreflight as k, attest as l, emitLoopProvenance as m, EVIDENCE_AUTHORITY_KINDS as n, isIndependentEvidence as o, canonicalDigest as p, assertCampaignDesign as q, EVIDENCE_RECEIPT_VERSION as r, verifyEvidenceReceipt as s, createCampaignEvidenceReceipt as t, verifyAttestation as u, provenanceSpansPath as v, compareRankKeys as w, campaignBreakdown as x, verifyLoopProvenanceRecord as y, renderSurfaceDiff as z };
+export { openFinalEvidenceLedger as A, assertRealBackend as B, pairedCampaignComposites as C, powerPreflight as D, recoverTruncatedJson as E, surfaceContentHash as F, campaignCoverage as G, assertCampaignDesign as H, surfaceDispatchRef as I, campaignSplitDigestFromIdentities as J, campaignScenarioIdentity as K, surfaceHash as L, codeSurfaceIdentityMaterial as M, componentSurfaceIdentityMaterial as N, FinalEvidenceConflictError as O, renderSurfaceDiff as P, surfaceHashMatches as R, campaignMeanCompositeOrNull as S, parseReflectionResponse as T, assertCampaignSplitIdentity as U, summarizeBackendIntegrity as V, assertCompleteCampaign as W, formatCoverageFailures as Y, provenanceRecordPath as _, createEvidenceReceipt as a, campaignBreakdown as b, ATTESTATION_ALGORITHM as c, buildLoopProvenanceRecord as d, campaignMeasurementDigest as f, loopProvenanceSpans as g, loopProvenanceArgsFromResult as h, INDEPENDENT_EVIDENCE_AUTHORITY_KINDS as i, assertCodeSurfaceIdentity as j, FinalEvidenceError as k, attest as l, emitLoopProvenance as m, EVIDENCE_AUTHORITY_KINDS as n, isIndependentEvidence as o, canonicalDigest as p, campaignSplitDigest as q, EVIDENCE_RECEIPT_VERSION as r, verifyEvidenceReceipt as s, createCampaignEvidenceReceipt as t, verifyAttestation as u, provenanceSpansPath as v, buildReflectionPrompt as w, campaignMeanComposite as x, verifyLoopProvenanceRecord as y, BackendIntegrityError as z };
 
-//# sourceMappingURL=campaign-evidence-D40-LHIL.js.map
+//# sourceMappingURL=campaign-evidence-DcQiYhJI.js.map