@tangle-network/agent-eval 0.187.1 → 0.188.0

package/dist/experiment/index.js

@@ -10,7 +10,6 @@ import { t as eProcess } from "../sequential-eprocess-D1jKoihe.js";
 import { a as heldoutSignificance, o as pairHoldout } from "../statistical-heldout-GcLIPxk4.js";
 import { A as FinalEvidenceConflictError, M as openFinalEvidenceLedger, N as defineEvaluationClaim, P as summarizeEvaluationUnits, a as createEvidenceReceipt, i as INDEPENDENT_EVIDENCE_AUTHORITY_KINDS, j as FinalEvidenceError, k as powerPreflight, n as EVIDENCE_AUTHORITY_KINDS, o as isIndependentEvidence, r as EVIDENCE_RECEIPT_VERSION, s as verifyEvidenceReceipt, t as createCampaignEvidenceReceipt } from "../campaign-evidence-IF1p0oDa.js";
 import { _ as runSelectionRule, a as evaluateHaltRule, c as evaluatePopulationReproducibilityGate, d as evaluateProvenanceGate, f as executeDecisionRule, g as readField, h as projectNLadderBudget, i as evaluateCondition, l as evaluatePowerFloorGate, m as powerFloorProblems, n as computeEstimand, o as evaluateIdentityGate, p as intervalSpecProblems, r as computeInterval, s as evaluateOracleDeterminismGate, t as classifyReissue, u as evaluatePredicate, v as runUniformPassBudget } from "../ast-CP9ae9B0.js";
-import { a as inMemoryExperimentStore, r as fileExperimentStore, t as ExperimentTracker } from "../experiment-tracker-BKEumQug.js";
 import { n as pairedEvalueSequence, r as sequentialCrossingHorizon } from "../sequential-CzK5DarL.js";
 import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-CfMNUro2.js";
 import { n as sequentialPairedGate, t as sequentialDecide } from "../sequential-lwJOTRON.js";
@@ -851,6 +850,124 @@ function mixSeed(seed, effect) {
 	return (seed ^ Math.round(effect * 1000003) * 2654435769) >>> 0 | 0;
 }
 //#endregion
+//#region src/experiment/prospective-canary-join.ts
+function nonEmpty$1(value, label) {
+	if (typeof value !== "string" || value.length === 0 || value.trim() !== value) throw new ValidationError(`prospective canary: ${label} must be a nonempty trimmed string`);
+}
+function utc(value, label) {
+	nonEmpty$1(value, label);
+	const parsed = Date.parse(value);
+	const canonical = Number.isFinite(parsed) ? new Date(parsed).toISOString() : "";
+	if (!value.endsWith("Z") || !Number.isFinite(parsed) || value !== canonical && value !== canonical.replace(".000Z", "Z")) throw new ValidationError(`prospective canary: ${label} must be a canonical UTC timestamp`);
+	return parsed;
+}
+function indexRows(rows, label, assignments) {
+	const indexed = /* @__PURE__ */ new Map();
+	for (const row of rows) {
+		nonEmpty$1(row.assignmentId, `${label} assignmentId`);
+		if (!assignments.has(row.assignmentId)) throw new ValidationError(`prospective canary: orphan ${label} for '${row.assignmentId}'`);
+		if (indexed.has(row.assignmentId)) throw new ValidationError(`prospective canary: duplicate ${label} for '${row.assignmentId}'`);
+		indexed.set(row.assignmentId, row);
+	}
+	return indexed;
+}
+function uniqueIds(ids, label) {
+	if (!Array.isArray(ids)) throw new ValidationError(`prospective canary: ${label} must be an array`);
+	const seen = /* @__PURE__ */ new Set();
+	for (const id of ids) {
+		nonEmpty$1(id, label);
+		if (seen.has(id)) throw new ValidationError(`prospective canary: duplicate ${label} '${id}'`);
+		seen.add(id);
+	}
+	return [...seen].sort();
+}
+/**
+* Join independently sourced served receipts without inventing missing costs.
+* The host still verifies ledger completeness, source authority, isolation,
+* randomization and the pre-traffic witness in `decideRandomizedCanary`.
+*/
+function joinProspectiveCanaryReceipts(input) {
+	if (input.enabled !== true) throw new ValidationError("prospective canary: pilot is disabled");
+	const witnessedAt = utc(input.protocolWitnessedAt, "protocolWitnessedAt");
+	const cutoff = utc(input.sealed.rule.analysisCutoff, "analysisCutoff");
+	if (witnessedAt >= cutoff) throw new ValidationError("prospective canary: protocol witness must precede cutoff");
+	if (!Array.isArray(input.assignments) || input.assignments.length === 0) throw new ValidationError("prospective canary: assignment ledger is empty");
+	const historical = new Set(uniqueIds(input.historicalSourceUnitIds, "historical source unit"));
+	const assignmentIds = /* @__PURE__ */ new Set();
+	const assignmentSources = /* @__PURE__ */ new Set();
+	const sourceUnits = /* @__PURE__ */ new Set();
+	for (const row of input.assignments) {
+		nonEmpty$1(row.assignmentId, "assignmentId");
+		nonEmpty$1(row.assignmentSourceId, "assignmentSourceId");
+		nonEmpty$1(row.clusterId, "clusterId");
+		nonEmpty$1(row.sourceUnitId, "sourceUnitId");
+		if (row.arm !== "control" && row.arm !== "candidate") throw new ValidationError("prospective canary: assignment arm is invalid");
+		if (assignmentIds.has(row.assignmentId) || assignmentSources.has(row.assignmentSourceId)) throw new ValidationError("prospective canary: duplicate assignment identity");
+		if (sourceUnits.has(row.sourceUnitId) || historical.has(row.sourceUnitId)) throw new ValidationError(`prospective canary: source unit '${row.sourceUnitId}' is reused`);
+		const assignedAt = utc(row.assignedAt, `assignment '${row.assignmentId}' assignedAt`);
+		if (assignedAt <= witnessedAt || assignedAt > cutoff) throw new ValidationError("prospective canary: assignment falls outside witnessed window");
+		assignmentIds.add(row.assignmentId);
+		assignmentSources.add(row.assignmentSourceId);
+		sourceUnits.add(row.sourceUnitId);
+	}
+	const outcomes = indexRows(input.outcomes, "outcome", assignmentIds);
+	const executions = indexRows(input.executions, "execution", assignmentIds);
+	const billing = indexRows(input.billing, "billing", assignmentIds);
+	const sourceIds = new Set(assignmentSources);
+	const claimSource = (id, label) => {
+		nonEmpty$1(id, label);
+		if (sourceIds.has(id)) throw new ValidationError(`prospective canary: reused independent source '${id}'`);
+		sourceIds.add(id);
+	};
+	return input.assignments.map((assignment) => {
+		const outcome = outcomes.get(assignment.assignmentId);
+		const execution = executions.get(assignment.assignmentId);
+		const bill = billing.get(assignment.assignmentId);
+		const assignedAt = Date.parse(assignment.assignedAt);
+		if (outcome) {
+			claimSource(outcome.outcomeSourceId, "outcomeSourceId");
+			nonEmpty$1(outcome.outcomeCheckerDigest, "outcomeCheckerDigest");
+			if (utc(outcome.checkedAt, "checkedAt") < assignedAt) throw new ValidationError("prospective canary: outcome precedes assignment");
+			if (![
+				0,
+				1,
+				null
+			].includes(outcome.checkedOutcome)) throw new ValidationError("prospective canary: checked outcome must be 0, 1 or null");
+		}
+		if (execution) {
+			if (execution.traceSourceId === null === (execution.noExecutionSourceId === null)) throw new ValidationError("prospective canary: execution needs one trace or no-execution source");
+			claimSource(execution.traceSourceId ?? execution.noExecutionSourceId, "execution source");
+			if (utc(execution.recordedAt, "execution recordedAt") < assignedAt) throw new ValidationError("prospective canary: execution precedes assignment");
+			if (execution.turns !== null && (!Number.isSafeInteger(execution.turns) || execution.turns < 0)) throw new ValidationError("prospective canary: turns must be a nonnegative integer");
+			if (execution.noExecutionSourceId !== null && (execution.turns !== 0 || execution.providerCallIds.length !== 0)) throw new ValidationError("prospective canary: no-execution receipt contains turns or calls");
+		}
+		if (bill) {
+			claimSource(bill.billingSourceId, "billingSourceId");
+			if (utc(bill.recordedAt, "billing recordedAt") < assignedAt) throw new ValidationError("prospective canary: billing precedes assignment");
+			if (bill.settlement === "settled") {
+				if (bill.billedCostUsd === null || !Number.isFinite(bill.billedCostUsd) || bill.billedCostUsd < 0) throw new ValidationError("prospective canary: settled billing needs nonnegative actual cost");
+			} else if (bill.settlement !== "unknown" || bill.billedCostUsd !== null) throw new ValidationError("prospective canary: unknown billing cost must stay null");
+			if (execution && uniqueIds(bill.providerCallIds, "billing call").join("\0") !== uniqueIds(execution.providerCallIds, "execution call").join("\0")) throw new ValidationError("prospective canary: provider-call billing join is incomplete");
+		}
+		return {
+			assignmentId: assignment.assignmentId,
+			assignmentSourceId: assignment.assignmentSourceId,
+			clusterId: assignment.clusterId,
+			arm: assignment.arm,
+			checkedOutcome: outcome?.checkedOutcome ?? null,
+			outcomeSourceId: outcome?.outcomeSourceId ?? null,
+			billedCostUsd: bill?.settlement === "settled" ? bill.billedCostUsd : null,
+			billingSourceId: bill?.billingSourceId ?? null,
+			traceSourceId: execution?.traceSourceId ?? null,
+			noExecutionSourceId: execution?.noExecutionSourceId ?? null,
+			servedProfileDigest: execution?.servedProfileDigest ?? null,
+			servedCodeRevisionDigest: execution?.servedCodeRevisionDigest ?? null,
+			outcomeCheckerDigest: outcome?.outcomeCheckerDigest ?? null,
+			turns: execution?.turns ?? null
+		};
+	});
+}
+//#endregion
 //#region src/experiment/randomized-canary.ts
 /** Fixed-horizon, assignment-level decision for an unpaired served canary. */
 function nonEmpty(value, label) {
@@ -1116,6 +1233,6 @@ function decideRandomizedCanary(sealed, cohort, rows, verifyCohort) {
 	return result;
 }
 //#endregion
-export { BOOTSTRAP_GATE_MIN_N, DesignRefusalError, EVIDENCE_AUTHORITY_KINDS, EVIDENCE_RECEIPT_VERSION, EVIDENCE_STATES, ExperimentTracker, FinalEvidenceConflictError, FinalEvidenceError, FunnelIntegrityError, INDEPENDENT_EVIDENCE_AUTHORITY_KINDS, MatchedBudgetError, SealIntegrityError, amendExperiment, assertDesignAdequate, assertFunnelReconciles, assertMatchedBudgets, benjaminiHochberg, bonferroni, buildEvidenceVector, buildFunnel, classifyReissue, clusteredPower, comparePairedArms, composeFunnels, computeEstimand, computeInterval, createCampaignEvidenceReceipt, createEvidenceReceipt, decideRandomizedCanary, defineEvaluationClaim, defineExperiment, eProcess, evaluateCondition, evaluateHaltRule, evaluateHypothesis, evaluateIdentityGate, evaluateOracleDeterminismGate, evaluatePopulationReproducibilityGate, evaluatePowerFloorGate, evaluatePredicate, evaluateProvenanceGate, evidenceRegistryRecordSchema, executeAdmissionRule, executeDecisionRule, fileExperimentStore, hashJson, heldoutSignificance, holm, inMemoryExperimentStore, isIndependentEvidence, manifestContentDigest, mcnemar, mcnemarPower, mcnemarRequiredN, mulberry32, openFinalEvidenceLedger, openSealedExperiment, pairArms, pairHoldout, pairRunRecords, pairedBootstrap, pairedEvalueSequence, pairedMde, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, paretoPolicy, paretoSignificanceGate, parseEvidenceRegistryRecord, powerPreflight, projectNLadderBudget, randomizedCanaryObservationDigest, randomizedCanaryRosterDigest, readField, renderEvidenceIndex, renderFunnelTable, requiredPairedSampleSize, requiredSampleSize, runSelectionRule, runUniformPassBudget, sealExperiment, sealRandomizedCanaryRule, sequentialCrossingHorizon, sequentialDecide, sequentialPairedGate, signManifest, summarizeEvaluationUnits, validateEvidenceRegistry, verifyEvidenceReceipt, verifyManifest, verifyMatchedBudgets, verifySealedExperiment, wilson };
+export { BOOTSTRAP_GATE_MIN_N, DesignRefusalError, EVIDENCE_AUTHORITY_KINDS, EVIDENCE_RECEIPT_VERSION, EVIDENCE_STATES, FinalEvidenceConflictError, FinalEvidenceError, FunnelIntegrityError, INDEPENDENT_EVIDENCE_AUTHORITY_KINDS, MatchedBudgetError, SealIntegrityError, amendExperiment, assertDesignAdequate, assertFunnelReconciles, assertMatchedBudgets, benjaminiHochberg, bonferroni, buildEvidenceVector, buildFunnel, classifyReissue, clusteredPower, comparePairedArms, composeFunnels, computeEstimand, computeInterval, createCampaignEvidenceReceipt, createEvidenceReceipt, decideRandomizedCanary, defineEvaluationClaim, defineExperiment, eProcess, evaluateCondition, evaluateHaltRule, evaluateHypothesis, evaluateIdentityGate, evaluateOracleDeterminismGate, evaluatePopulationReproducibilityGate, evaluatePowerFloorGate, evaluatePredicate, evaluateProvenanceGate, evidenceRegistryRecordSchema, executeAdmissionRule, executeDecisionRule, hashJson, heldoutSignificance, holm, isIndependentEvidence, joinProspectiveCanaryReceipts, manifestContentDigest, mcnemar, mcnemarPower, mcnemarRequiredN, mulberry32, openFinalEvidenceLedger, openSealedExperiment, pairArms, pairHoldout, pairRunRecords, pairedBootstrap, pairedEvalueSequence, pairedMde, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, paretoPolicy, paretoSignificanceGate, parseEvidenceRegistryRecord, powerPreflight, projectNLadderBudget, randomizedCanaryObservationDigest, randomizedCanaryRosterDigest, readField, renderEvidenceIndex, renderFunnelTable, requiredPairedSampleSize, requiredSampleSize, runSelectionRule, runUniformPassBudget, sealExperiment, sealRandomizedCanaryRule, sequentialCrossingHorizon, sequentialDecide, sequentialPairedGate, signManifest, summarizeEvaluationUnits, validateEvidenceRegistry, verifyEvidenceReceipt, verifyManifest, verifyMatchedBudgets, verifySealedExperiment, wilson };
 
 //# sourceMappingURL=index.js.map