@tangle-network/agent-eval 0.216.0 → 0.217.1

package/dist/contract/index.js

@@ -1,29 +1,30 @@
-import { a as hashCanonical } from "../canonical-BsZVgBhW.js";
-import { r as pairedBootstrap } from "../paired-tests-jbwPdSz5.js";
-import { t as decidePairedPromotion } from "../paired-promotion-decision-BN6IQTiu.js";
-import { a as heldoutSignificance } from "../statistical-heldout-1Gbmex7v.js";
-import { n as summarizeExecution, t as analyzeRuns } from "../analyze-runs-3xpby3p2.js";
-import { a as parseRunRecordSafe } from "../run-record-DlTZ_lqe.js";
-import { K as defaultProductionGate, a as transientDispatchFailure, c as compareOptimizationMethods, h as runImprovementLoop, nt as runEval, rt as runCampaign, t as llmJudge } from "../llm-judge-PXV7Xljl.js";
+import { a as hashCanonical } from "../canonical-Dq2hOQ8L.js";
+import "../run-record-DhEJgoPz.js";
+import { r as summarizeExecution, t as analyzeRuns } from "../analyze-runs-DqPJaJ4X.js";
+import { i as pairedBootstrap } from "../mean-BbA_JCzW.js";
+import { t as decidePairedPromotion } from "../paired-promotion-decision-D0SHb-ZL.js";
+import { a as heldoutSignificance } from "../statistical-heldout-Bksv92Th.js";
+import { B as defaultProductionGate, Q as runCampaign, Z as runEval, i as transientDispatchFailure, o as clamp01, p as runImprovementLoop, t as llmJudge } from "../llm-judge-C8jMTNNG.js";
 import { i as isModelPriced, n as estimateCost } from "../metrics-BKaPBtIG.js";
-import { i as CostLedger } from "../cost-ledger-D2FnJPBF.js";
-import { G as campaignSplitDigest } from "../campaign-evidence-OcUciD-N.js";
-import { h as mapConcurrentRange } from "../journal-DvKCvjty.js";
-import { F as inMemoryCampaignStorage, P as fsCampaignStorage } from "../external-optimizer-subprocess-DcerlZSG.js";
-import { o as makeProposalFinding } from "../types-IXD2HJzu.js";
-import { n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-Ch3MWX6k.js";
-import { t as buildDefaultAnalystRegistry } from "../default-registry-Cn1LQ2Ve.js";
-import { t as extractUsage } from "../extract-usage-BrnzRJ6H.js";
-import { t as createChatClient } from "../chat-client-DthPOYGd.js";
+import { t as CostLedger } from "../cost-ledger-D0352Z4d.js";
+import { z as campaignSplitDigest } from "../campaign-evidence-CQYw8_0H.js";
+import { o as mapConcurrentRange } from "../journal-CR4QikYp.js";
+import { M as fsCampaignStorage, N as inMemoryCampaignStorage } from "../external-optimizer-subprocess-XIHBOSKt.js";
+import { a as makeProposalFinding } from "../types-DqFMNTgK.js";
+import { n as isRecord } from "../json-BhdsJaOQ.js";
+import { n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-BDxNlP3D.js";
+import { t as buildDefaultAnalystRegistry } from "../default-registry-ClYjq3QE.js";
+import { t as extractUsage } from "../extract-usage-BZLU01kR.js";
+import { t as createChatClient } from "../chat-client-ldqxYYTN.js";
 import { n as InMemoryOutcomeStore, t as FileSystemOutcomeStore } from "../outcome-store-DUcMj-3D.js";
-import { c as heldOutGate, g as runReferenceEquivalenceJudge, h as createReferenceEquivalenceJudge, l as composeGate, m as REFERENCE_EQUIVALENCE_JUDGE_VERSION, n as gepaOptimizationMethod, p as REFERENCE_EQUIVALENCE_INPUT_LIMITS, t as skillOptOptimizationMethod, u as externalTextOptimizationMethod } from "../skillopt-optimization-method-CCtzralq.js";
-import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-CWW35UWq.js";
-import { t as fromOtelSpans } from "../otel-spans-C8YIJWy1.js";
+import { c as composeGate, l as externalTextOptimizationMethod, n as compareOptimizationMethods, r as gepaOptimizationMethod, s as heldOutGate, t as skillOptOptimizationMethod } from "../skillopt-optimization-method-kyL94-ri.js";
+import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-6A8osGr_.js";
+import { t as fromOtelSpans } from "../otel-spans-CdN909k1.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";
 import { createReadStream } from "node:fs";
-import { dirname, isAbsolute, join, relative, resolve, sep } from "node:path";
-import { mkdir, readFile, readdir, stat, writeFile } from "node:fs/promises";
+import "node:path";
+import "node:fs/promises";
 import { readSessionInput, summarizeSessionInput } from "@tangle-network/harness-sessions";
 //#region src/contract/fixed-spend.ts
 function addFixedSpend(left, right) {
@@ -37,6 +38,189 @@ function addFixedSpend(left, right) {
 		costProvenance: left.costProvenance === "observed" && right.costProvenance === "observed" ? "observed" : "estimated"
 	};
 }
+/** Sum both arms in receipt order while retaining the observed/estimated provenance. */
+function pairedFixedSpendReceipt(measurements, model, usageOf) {
+	const usage = measurements.reduce((sum, measurement) => addFixedSpend(addFixedSpend(sum, usageOf(measurement.baseline)), usageOf(measurement.candidate)), {
+		inputTokens: 0,
+		outputTokens: 0,
+		cachedInputTokens: 0,
+		reasoningTokens: 0,
+		modelCalls: 0,
+		costUsdNanos: 0,
+		costProvenance: "observed"
+	});
+	const costUsd = usage.costUsdNanos / 1e9;
+	return {
+		model,
+		inputTokens: usage.inputTokens,
+		outputTokens: usage.outputTokens,
+		cachedTokens: usage.cachedInputTokens,
+		reasoningTokens: usage.reasoningTokens,
+		...usage.costProvenance === "observed" ? { actualCostUsd: costUsd } : { estimatedCostUsd: costUsd }
+	};
+}
+//#endregion
+//#region src/contract/candidate-experiment-evidence.ts
+/** Verify signed candidate execution receipts and read their measured usage. */
+const candidateExecutionEvidenceAdapter = {
+	score: (evidence) => evidence.receipt.benchmarkResult.material.score,
+	dimensions: (evidence) => evidence.receipt.benchmarkResult.material.dimensions,
+	costUsd: costFromEvidence,
+	costProvenance: (evidence) => combinedUsage(evidence).costProvenance,
+	latencyMs: latencyFromEvidence,
+	completed: completedSuccessfully,
+	passed: (evidence) => evidence.receipt.benchmarkResult.material.passed
+};
+function verifyMeasurement(experiment, input, index) {
+	const suite = experiment.benchmark.suite;
+	const taskIndex = Math.floor(index / suite.reps);
+	const repetition = index % suite.reps;
+	const task = experiment.benchmark.tasks[taskIndex];
+	const seed = suite.seeds[index];
+	if (!task || seed === void 0) throw new Error(`candidate experiment measurement ${index} is outside the signed suite`);
+	const baseline = verifyExecutionEvidence(input.baseline);
+	const candidate = verifyExecutionEvidence(input.candidate);
+	for (const [arm, evidence] of [["baseline", baseline], ["candidate", candidate]]) {
+		const bundle = experiment[arm];
+		const materialization = evidence.materializationReceipt;
+		const runCell = materialization.executionPlan.material.runCell;
+		verifySelfAddressed(runCell, "candidate run cell");
+		if (runCell.experimentDigest !== experiment.digest || runCell.arm !== arm || runCell.bundleDigest !== bundle.digest || runCell.suiteDigest !== suite.digest || runCell.taskDigest !== task.digest || runCell.taskIndex !== taskIndex || runCell.repetition !== repetition || runCell.seed !== seed || runCell.attempt > task.attempt.maxAttempts || materialization.bundleDigest !== bundle.digest || materialization.benchmark.suite.digest !== suite.digest || materialization.benchmark.task.digest !== task.digest || materialization.codeKind !== bundle.code.kind || materialization.profileActivation.profilePlan.material.sourceProfileDigest !== canonicalCandidateDigest(bundle.profile) || evidence.receipt.runCellDigest !== runCell.digest || JSON.stringify(materialization.resolvedModel) !== JSON.stringify(task.model)) throw new Error(`candidate experiment measurement ${index} substituted its ${arm} arm`);
+		verifyTaskOutcome(task, evidence, index, arm);
+	}
+	const baselinePlan = baseline.materializationReceipt.executionPlan.material;
+	const candidatePlan = candidate.materializationReceipt.executionPlan.material;
+	if (baselinePlan.executionId === candidatePlan.executionId || baselinePlan.runCell.digest === candidatePlan.runCell.digest || baseline.materializationReceipt.digest === candidate.materializationReceipt.digest || baseline.receipt.digest === candidate.receipt.digest || baseline.digest === candidate.digest) throw new Error(`candidate experiment measurement ${index} reused one execution across arms`);
+	return {
+		baseline,
+		candidate
+	};
+}
+function verifyExecutionEvidence(input) {
+	const evidence = candidateExecutionEvidenceSchema.parse(input);
+	verifySelfAddressed(evidence, "candidate execution evidence");
+	verifySelfAddressed(evidence.materializationReceipt, "candidate materialization receipt");
+	verifySelfAddressed(evidence.materializationReceipt.profileActivation, "candidate profile activation");
+	verifyMaterialAddressed(evidence.materializationReceipt.profileActivation.profilePlan, "candidate profile plan");
+	verifyMaterialAddressed(evidence.materializationReceipt.executionPlan, "candidate execution plan");
+	verifySelfAddressed(evidence.receipt, "candidate run receipt");
+	verifyMaterialAddressed(evidence.receipt.modelSettlement, "candidate model settlement");
+	verifyMaterialAddressed(evidence.receipt.taskOutcome, "candidate task outcome");
+	verifyMaterialAddressed(evidence.receipt.benchmarkResult, "candidate benchmark result");
+	return evidence;
+}
+function verifySelfAddressed(document, label) {
+	if (canonicalCandidateDigest(omitTopLevelDigest(document)) !== document.digest) throw new Error(`${label} digest is invalid`);
+}
+function verifyTaskOutcome(task, evidence, index, arm) {
+	const outcome = evidence.receipt.taskOutcome.material.outcome;
+	const result = evidence.receipt.benchmarkResult.material;
+	const prefix = `candidate experiment measurement ${index} ${arm}`;
+	if (result.evidence.sha256 === task.grader.artifact.sha256) throw new Error(`${prefix} reused grader bytes as grading evidence`);
+	const usage = combinedUsage(evidence);
+	const usageChecks = [
+		[
+			usage.modelCalls,
+			task.limits.maxModelCalls,
+			"model calls"
+		],
+		[
+			usage.inputTokens,
+			task.limits.maxInputTokens,
+			"input tokens"
+		],
+		[
+			usage.outputTokens,
+			task.limits.maxOutputTokens,
+			"output tokens"
+		],
+		[
+			usage.costUsdNanos,
+			Math.round(task.limits.maxCostUsd * 1e9),
+			"cost"
+		]
+	];
+	for (const [actual, maximum, label] of usageChecks) if (actual > maximum) throw new Error(`${prefix} ${label} ${actual} exceeds the signed limit ${maximum}`);
+	if (outcome.kind !== task.outcome.kind) throw new Error(`${prefix} returned an outcome outside the signed task contract`);
+	if (task.outcome.kind === "output") {
+		if (outcome.kind !== "output" || outcome.spec.mediaType !== task.outcome.mediaType || outcome.spec.maxBytes !== task.outcome.maxBytes) throw new Error(`${prefix} changed the signed output contract`);
+		return;
+	}
+	const repository = task.repository;
+	if (outcome.kind !== "workspace" || repository === void 0 || outcome.baseRepository.identity !== repository.identity || outcome.baseRepository.rootIdentity !== repository.rootIdentity || outcome.baseRepository.commit !== repository.baseCommit || outcome.baseRepository.tree !== repository.baseTree) throw new Error(`${prefix} did not start from the signed repository state`);
+}
+function verifyStableProfileMaterialization(measurements) {
+	for (const arm of ["baseline", "candidate"]) {
+		const expected = measurements[0]?.[arm].materializationReceipt.profileActivation;
+		if (!expected) throw new Error("candidate experiment contains no profile materialization");
+		const expectedDigest = canonicalCandidateDigest({
+			profilePlanDigest: expected.profilePlan.digest,
+			files: expected.files
+		});
+		for (const [index, measurement] of measurements.entries()) {
+			const activation = measurement[arm].materializationReceipt.profileActivation;
+			if (canonicalCandidateDigest({
+				profilePlanDigest: activation.profilePlan.digest,
+				files: activation.files
+			}) !== expectedDigest) throw new Error(`candidate experiment measurement ${index} ${arm} materialized a different profile`);
+		}
+	}
+}
+function completedSuccessfully(evidence) {
+	const termination = evidence.receipt.termination;
+	return termination.kind === "exit" && termination.exitCode === 0;
+}
+function verifyMaterialAddressed(evidence, label) {
+	if (canonicalCandidateDigest(evidence.material) !== evidence.digest) throw new Error(`${label} digest is invalid`);
+}
+function costFromEvidence(evidence) {
+	return combinedUsage(evidence).costUsdNanos / 1e9;
+}
+function latencyFromEvidence(evidence) {
+	return evidence.receipt.timing.durationMs + evidence.receipt.benchmarkResult.material.grading.timing.durationMs;
+}
+function combinedUsage(evidence) {
+	const candidate = evidence.receipt.modelSettlement.material.usage;
+	const grader = evidence.receipt.benchmarkResult.material.grading.usage;
+	return addFixedSpend(candidate, grader);
+}
+function candidateMeasurementCostReceipt(measurements, model) {
+	return pairedFixedSpendReceipt(measurements, model, combinedUsage);
+}
+//#endregion
+//#region src/contract/measured-comparison-record.ts
+/** Keep published timing and cost accounting identical across experiment domains. */
+function comparisonEvaluation(options, evaluation) {
+	return {
+		generationsExplored: options.generationsExplored ?? 0,
+		preparation: options.preparation,
+		measurement: {
+			wallDurationMs: options.measurement.wallDurationMs,
+			workDurationMs: evaluation.measurementWorkDurationMs,
+			cost: evaluation.measurementCost
+		},
+		total: {
+			wallDurationMs: options.preparation.wallDurationMs + options.measurement.wallDurationMs,
+			cost: evaluation.totalCost
+		}
+	};
+}
+/** Replay only the original inputs; derived timing and statistics are recomputed. */
+function comparisonRecomputationInput(comparison) {
+	return {
+		experiment: comparison.experiment,
+		measurements: comparison.measurements,
+		runId: comparison.provenance.runId,
+		...comparison.candidate ? { candidate: comparison.candidate } : {},
+		generationsExplored: comparison.evaluation.generationsExplored,
+		preparation: comparison.evaluation.preparation,
+		measurement: {
+			wallDurationMs: comparison.evaluation.measurement.wallDurationMs,
+			cost: comparison.evaluation.measurement.cost
+		},
+		...comparison.metadata ? { metadata: comparison.metadata } : {}
+	};
+}
 //#endregion
 //#region src/contract/concurrent-map.ts
 /** Map both arms while bounding the actual executions, not the number of pairs. */
@@ -116,104 +300,67 @@ function assertNonNegativeFinite(value, label) {
 	if (!Number.isFinite(value) || value < 0) throw new Error(`${label} must be a non-negative finite number`);
 }
 //#endregion
-//#region src/contract/measured-comparison.ts
-/** Content-address one task before any measured execution can see it. */
-function sealCandidateBenchmarkTask(material) {
-	return agentCandidateBenchmarkTaskSchema.parse({
-		...material,
-		digest: canonicalCandidateDigest(material)
-	});
-}
-/** Freeze task order, repetitions, and every seed before either arm runs. */
-function sealCandidateBenchmarkSuite(options) {
-	for (const task of options.tasks) verifyCandidateBenchmarkTask(task);
-	const material = {
-		kind: "agent-candidate-benchmark-suite",
-		digestAlgorithm: "rfc8785-sha256",
-		taskDigests: options.tasks.map((task) => task.digest),
-		reps: options.reps,
-		seeds: options.seeds
+//#region src/contract/paired-measurement-input.ts
+function projectPairedMeasurement(measurement, index, adapter) {
+	if (typeof measurement.cellId !== "string" || !measurement.cellId.trim()) throw new Error(`paired measurement ${index} requires a cell id`);
+	return {
+		cellId: measurement.cellId,
+		baseline: projectRun(measurement.baseline, adapter, `paired measurement ${index} baseline`),
+		candidate: projectRun(measurement.candidate, adapter, `paired measurement ${index} candidate`)
 	};
+}
+function projectRun(run, adapter, label) {
+	const suppliedDimensions = adapter.dimensions(run);
+	if (!Array.isArray(suppliedDimensions)) throw new Error(`${label} dimensions must be an array`);
+	const dimensions = /* @__PURE__ */ new Map();
+	for (const dimension of suppliedDimensions) {
+		if (typeof dimension.name !== "string" || !dimension.name.trim()) throw new Error(`${label} contains an unnamed dimension`);
+		if (dimensions.has(dimension.name)) throw new Error(`${label} repeats dimension '${dimension.name}'`);
+		dimensions.set(dimension.name, finiteMeasurement(dimension.score, `${label} ${dimension.name}`));
+	}
+	const completed = adapter.completed(run);
+	const passed = adapter.passed(run);
+	if (typeof completed !== "boolean" || typeof passed !== "boolean") throw new Error(`${label} completion and pass values must be booleans`);
+	const costProvenance = adapter.costProvenance(run);
+	if (costProvenance !== "observed" && costProvenance !== "estimated") throw new Error(`${label} cost provenance must be observed or estimated`);
 	return {
-		suite: agentCandidateBenchmarkSuiteSchema.parse({
-			...material,
-			digest: canonicalCandidateDigest(material)
-		}),
-		tasks: options.tasks
+		score: finiteMeasurement(adapter.score(run), `${label} score`),
+		dimensions,
+		costUsd: nonNegativeMeasurement(adapter.costUsd(run), `${label} cost`),
+		costProvenance,
+		latencyMs: nonNegativeMeasurement(adapter.latencyMs(run), `${label} latency`),
+		completed,
+		passed
 	};
 }
-/** Freeze both complete agent states and their exact held-out work. */
-function sealCandidateExperiment(material) {
-	return verifyCandidateExperiment(agentCandidateExperimentSchema.parse({
-		...material,
-		digest: canonicalCandidateDigest(material)
-	}));
+function sharedProjectedDimensions(measurements) {
+	const expected = [...measurements[0].baseline.dimensions.keys()];
+	for (const [index, measurement] of measurements.entries()) for (const [arm, run] of [["baseline", measurement.baseline], ["candidate", measurement.candidate]]) {
+		const actual = [...run.dimensions.keys()];
+		if (JSON.stringify(actual) !== JSON.stringify(expected)) throw new Error(`paired measurement ${index} ${arm} dimensions do not match the suite`);
+	}
+	return expected;
 }
-function verifyCandidateExperiment(input) {
-	const experiment = agentCandidateExperimentSchema.parse(input);
-	verifySelfAddressed(experiment, "candidate experiment");
-	verifyBundle(experiment.baseline, "baseline bundle");
-	verifyBundle(experiment.candidate, "candidate bundle");
-	if (experiment.baseline.digest === experiment.candidate.digest) throw new Error("candidate experiment baseline and candidate bundles are identical");
-	verifyCandidateBenchmarkSuiteInputs(experiment.benchmark);
-	return experiment;
+function dimensionScore(run, name) {
+	const value = run.dimensions.get(name);
+	if (value === void 0) throw new Error(`paired measurement is missing dimension '${name}'`);
+	return value;
 }
-/** Execute each signed cell for both arms. The callback is Runtime's one executor. */
-async function runCandidateExperiment(options) {
-	const experiment = verifyCandidateExperiment(options.experiment);
-	const { suite, tasks } = experiment.benchmark;
-	const run = await runPaidPairedMeasurement({
-		count: suite.taskDigests.length * suite.reps,
-		maxConcurrency: options.maxConcurrency ?? 2,
-		label: "candidate experiment",
-		budgetUsd: experiment.policy.budgetUsd,
-		...options.costLedger ? { costLedger: options.costLedger } : {},
-		maximumCostUsd: candidateMeasurementMaximumCostUsd(experiment),
-		call: {
-			callId: `candidate-measurement:${experiment.digest}`,
-			channel: "measurement",
-			phase: "heldout",
-			actor: experiment.digest,
-			model: candidateMeasurementModel(experiment),
-			tags: { experimentDigest: experiment.digest }
-		},
-		...options.signal ? { signal: options.signal } : {},
-		async execute(index, arm, signal) {
-			const taskIndex = Math.floor(index / suite.reps);
-			const repetition = index % suite.reps;
-			const task = tasks[taskIndex];
-			const seed = suite.seeds[index];
-			if (!task || seed === void 0) throw new Error(`candidate experiment cell ${index} has no signed task or seed`);
-			const benchmarkCell = {
-				suiteDigest: suite.digest,
-				taskIndex,
-				repetition
-			};
-			return verifyExecutionEvidence(await options.execute({
-				experiment,
-				arm,
-				bundle: experiment[arm],
-				task,
-				benchmarkCell,
-				seed,
-				signal
-			}));
-		},
-		receipt(measurements) {
-			return candidateMeasurementCostReceipt(measurements, candidateMeasurementModel(experiment));
-		}
-	});
-	return {
-		measurements: run.measurements.map((measurement, index) => verifyMeasurement(experiment, measurement, index)),
-		measurement: {
-			wallDurationMs: run.wallDurationMs,
-			cost: {
-				usd: run.cost.usd,
-				provenance: run.cost.kind
-			}
-		}
-	};
+function finiteMeasurement(value, label) {
+	if (!Number.isFinite(value)) throw new Error(`${label} must be finite`);
+	return value;
+}
+function nonNegativeMeasurement(value, label) {
+	if (!Number.isFinite(value) || value < 0) throw new Error(`${label} must be non-negative`);
+	return value;
+}
+function assertKnownCost(cost, label) {
+	if (!Number.isFinite(cost.usd) || cost.usd < 0) throw new Error(`${label} must be a non-negative number`);
+	if (cost.provenance !== "observed" && cost.provenance !== "estimated") throw new Error(`${label} provenance must be observed or estimated`);
 }
+//#endregion
+//#region src/contract/paired-measurement-evaluation.ts
+/** Shared significance, regression, precision, and budget decisions on paired receipts. */
 /**
 * Calculate the shared paired decision from any complete receipt shape.
 *
@@ -449,6 +596,164 @@ function significanceReason(significance, deltaThreshold) {
 	if (decision.exactTestVetoes && decision.mcnemar !== null) return `McNemar's exact test refuses at threshold ${deltaThreshold}: b=${decision.mcnemar.b}, c=${decision.mcnemar.c}, p=${decision.mcnemar.pValue}`;
 	return `paired ${decision.statistic} interval lower bound ${decision.low} did not clear ${deltaThreshold}.${decision.methodDetail}`;
 }
+function measuredEstimate(baseline, candidate, options) {
+	const bootstrap = pairedBootstrap(baseline, candidate, {
+		confidence: options.confidence,
+		resamples: options.resamples,
+		statistic: "mean",
+		seed: options.seed
+	});
+	const baselineMean = mean(baseline);
+	const candidateMean = mean(candidate);
+	const delta = candidateMean - baselineMean;
+	return {
+		baseline: baselineMean,
+		candidate: candidateMean,
+		delta,
+		confidenceInterval: {
+			level: bootstrap.confidence,
+			lower: Math.min(bootstrap.low, delta),
+			upper: Math.max(bootstrap.high, delta),
+			method: "paired-bootstrap",
+			statistic: "mean",
+			resamples: bootstrap.resamples
+		},
+		n: bootstrap.n
+	};
+}
+function observedPairedPrecision(options) {
+	const n = options.baselineScores.length;
+	const lowerMargin = Math.max(0, options.delta - options.lowerBound);
+	const minimumDetectableDelta = options.deltaThreshold + lowerMargin;
+	const scaleAssumed = [...options.baselineScores, ...options.candidateScores].every((score) => score >= -.001 && score <= 1.001);
+	const headroom = Math.max(0, 1 - mean(options.baselineScores));
+	const enoughRuns = n >= options.minProductiveRuns;
+	const attainable = !scaleAssumed || minimumDetectableDelta <= headroom;
+	const sufficient = enoughRuns && attainable;
+	const scorerNote = options.sharedScorerChannel ? " The same scorer evaluated both arms, so scorer bias is not measured." : "";
+	if (!enoughRuns) return {
+		sufficient,
+		minimumDetectableDelta,
+		scaleAssumed,
+		reason: `only ${n} paired runs; ${options.minProductiveRuns} required.${scorerNote}`
+	};
+	if (!attainable) return {
+		sufficient,
+		minimumDetectableDelta,
+		scaleAssumed,
+		reason: `observed paired uncertainty gives a minimum detectable delta of ${minimumDetectableDelta.toFixed(3)}, above the remaining score headroom ${headroom.toFixed(3)}.${scorerNote}`
+	};
+	return {
+		sufficient,
+		minimumDetectableDelta,
+		scaleAssumed,
+		reason: `observed paired uncertainty gives a minimum detectable delta of ${minimumDetectableDelta.toFixed(3)} at ${options.confidence.toFixed(2)} confidence from ${n} runs.${scorerNote}`
+	};
+}
+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;
+}
+//#endregion
+//#region src/contract/candidate-experiment.ts
+/** Seal held-out experiments, execute their signed cells, and publish verified comparisons. */
+/** Content-address one task before any measured execution can see it. */
+function sealCandidateBenchmarkTask(material) {
+	return agentCandidateBenchmarkTaskSchema.parse({
+		...material,
+		digest: canonicalCandidateDigest(material)
+	});
+}
+/** Freeze task order, repetitions, and every seed before either arm runs. */
+function sealCandidateBenchmarkSuite(options) {
+	for (const task of options.tasks) verifyCandidateBenchmarkTask(task);
+	const material = {
+		kind: "agent-candidate-benchmark-suite",
+		digestAlgorithm: "rfc8785-sha256",
+		taskDigests: options.tasks.map((task) => task.digest),
+		reps: options.reps,
+		seeds: options.seeds
+	};
+	return {
+		suite: agentCandidateBenchmarkSuiteSchema.parse({
+			...material,
+			digest: canonicalCandidateDigest(material)
+		}),
+		tasks: options.tasks
+	};
+}
+/** Freeze both complete agent states and their exact held-out work. */
+function sealCandidateExperiment(material) {
+	return verifyCandidateExperiment(agentCandidateExperimentSchema.parse({
+		...material,
+		digest: canonicalCandidateDigest(material)
+	}));
+}
+function verifyCandidateExperiment(input) {
+	const experiment = agentCandidateExperimentSchema.parse(input);
+	verifySelfAddressed(experiment, "candidate experiment");
+	verifyBundle(experiment.baseline, "baseline bundle");
+	verifyBundle(experiment.candidate, "candidate bundle");
+	if (experiment.baseline.digest === experiment.candidate.digest) throw new Error("candidate experiment baseline and candidate bundles are identical");
+	verifyCandidateBenchmarkSuiteInputs(experiment.benchmark);
+	return experiment;
+}
+/** Execute each signed cell for both arms. The callback is Runtime's one executor. */
+async function runCandidateExperiment(options) {
+	const experiment = verifyCandidateExperiment(options.experiment);
+	const { suite, tasks } = experiment.benchmark;
+	const run = await runPaidPairedMeasurement({
+		count: suite.taskDigests.length * suite.reps,
+		maxConcurrency: options.maxConcurrency ?? 2,
+		label: "candidate experiment",
+		budgetUsd: experiment.policy.budgetUsd,
+		...options.costLedger ? { costLedger: options.costLedger } : {},
+		maximumCostUsd: candidateMeasurementMaximumCostUsd(experiment),
+		call: {
+			callId: `candidate-measurement:${experiment.digest}`,
+			channel: "measurement",
+			phase: "heldout",
+			actor: experiment.digest,
+			model: candidateMeasurementModel(experiment),
+			tags: { experimentDigest: experiment.digest }
+		},
+		...options.signal ? { signal: options.signal } : {},
+		async execute(index, arm, signal) {
+			const taskIndex = Math.floor(index / suite.reps);
+			const repetition = index % suite.reps;
+			const task = tasks[taskIndex];
+			const seed = suite.seeds[index];
+			if (!task || seed === void 0) throw new Error(`candidate experiment cell ${index} has no signed task or seed`);
+			const benchmarkCell = {
+				suiteDigest: suite.digest,
+				taskIndex,
+				repetition
+			};
+			return verifyExecutionEvidence(await options.execute({
+				experiment,
+				arm,
+				bundle: experiment[arm],
+				task,
+				benchmarkCell,
+				seed,
+				signal
+			}));
+		},
+		receipt(measurements) {
+			return candidateMeasurementCostReceipt(measurements, candidateMeasurementModel(experiment));
+		}
+	});
+	return {
+		measurements: run.measurements.map((measurement, index) => verifyMeasurement(experiment, measurement, index)),
+		measurement: {
+			wallDurationMs: run.wallDurationMs,
+			cost: {
+				usd: run.cost.usd,
+				provenance: run.cost.kind
+			}
+		}
+	};
+}
 /** Build the only publishable comparison: paired statistics over Runtime receipts. */
 function measuredComparisonFromCandidateExperiment(options) {
 	const experiment = verifyCandidateExperiment(options.experiment);
@@ -489,19 +794,7 @@ function measuredComparisonFromCandidateExperiment(options) {
 			candidateContentHash: experiment.candidate.digest
 		},
 		diff,
-		evaluation: {
-			generationsExplored: options.generationsExplored ?? 0,
-			preparation: options.preparation,
-			measurement: {
-				wallDurationMs: options.measurement.wallDurationMs,
-				workDurationMs: evaluation.measurementWorkDurationMs,
-				cost: evaluation.measurementCost
-			},
-			total: {
-				wallDurationMs: options.preparation.wallDurationMs + options.measurement.wallDurationMs,
-				cost: evaluation.totalCost
-			}
-		},
+		evaluation: comparisonEvaluation(options, evaluation),
 		...options.metadata ? { metadata: options.metadata } : {}
 	});
 	const { recordDigest: _recordDigest, ...provenance } = provisional.provenance;
@@ -519,19 +812,7 @@ function measuredComparisonFromCandidateExperiment(options) {
 /** Recompute every statistic and decision from the signed experiment receipts. */
 function verifyCandidateExperimentComparison(input) {
 	const comparison = agentImprovementMeasuredComparisonSchema.parse(input);
-	const recomputed = measuredComparisonFromCandidateExperiment({
-		experiment: comparison.experiment,
-		measurements: comparison.measurements,
-		runId: comparison.provenance.runId,
-		...comparison.candidate ? { candidate: comparison.candidate } : {},
-		generationsExplored: comparison.evaluation.generationsExplored,
-		preparation: comparison.evaluation.preparation,
-		measurement: {
-			wallDurationMs: comparison.evaluation.measurement.wallDurationMs,
-			cost: comparison.evaluation.measurement.cost
-		},
-		...comparison.metadata ? { metadata: comparison.metadata } : {}
-	});
+	const recomputed = measuredComparisonFromCandidateExperiment(comparisonRecomputationInput(comparison));
 	if (canonicalCandidateDigest(recomputed) !== canonicalCandidateDigest(comparison)) throw new Error("candidate experiment comparison does not match its Runtime receipts");
 	return comparison;
 }
@@ -544,34 +825,10 @@ function candidateMeasurementModel(experiment) {
 	const models = [...new Set(experiment.benchmark.tasks.map((task) => task.model.model))];
 	return models.length === 1 ? models[0] : "multiple-models";
 }
-function candidateMeasurementCostReceipt(measurements, model) {
-	const usage = measurements.reduce((sum, measurement) => addFixedSpend(addFixedSpend(sum, combinedUsage(measurement.baseline)), combinedUsage(measurement.candidate)), {
-		inputTokens: 0,
-		outputTokens: 0,
-		cachedInputTokens: 0,
-		reasoningTokens: 0,
-		modelCalls: 0,
-		costUsdNanos: 0,
-		costProvenance: "observed"
-	});
-	const costUsd = usage.costUsdNanos / 1e9;
-	return {
-		model,
-		inputTokens: usage.inputTokens,
-		outputTokens: usage.outputTokens,
-		cachedTokens: usage.cachedInputTokens,
-		reasoningTokens: usage.reasoningTokens,
-		...usage.costProvenance === "observed" ? { actualCostUsd: costUsd } : { estimatedCostUsd: costUsd }
-	};
-}
 function assertPhaseAccounting(accounting, label) {
 	if (!Number.isFinite(accounting.wallDurationMs) || accounting.wallDurationMs < 0) throw new Error(`${label} wall duration must be a non-negative number`);
 	assertKnownCost(accounting.cost, `${label} cost`);
 }
-function assertKnownCost(cost, label) {
-	if (!Number.isFinite(cost.usd) || cost.usd < 0) throw new Error(`${label} must be a non-negative number`);
-	if (cost.provenance !== "observed" && cost.provenance !== "estimated") throw new Error(`${label} provenance must be observed or estimated`);
-}
 function deriveCandidateBundleDiff(experiment) {
 	const changed = [
 		"profile",
@@ -625,235 +882,6 @@ function verifyBundle(input, label) {
 	verifySelfAddressed(bundle, label);
 	return bundle;
 }
-function verifyMeasurement(experiment, input, index) {
-	const suite = experiment.benchmark.suite;
-	const taskIndex = Math.floor(index / suite.reps);
-	const repetition = index % suite.reps;
-	const task = experiment.benchmark.tasks[taskIndex];
-	const seed = suite.seeds[index];
-	if (!task || seed === void 0) throw new Error(`candidate experiment measurement ${index} is outside the signed suite`);
-	const baseline = verifyExecutionEvidence(input.baseline);
-	const candidate = verifyExecutionEvidence(input.candidate);
-	for (const [arm, evidence] of [["baseline", baseline], ["candidate", candidate]]) {
-		const bundle = experiment[arm];
-		const materialization = evidence.materializationReceipt;
-		const runCell = materialization.executionPlan.material.runCell;
-		verifySelfAddressed(runCell, "candidate run cell");
-		if (runCell.experimentDigest !== experiment.digest || runCell.arm !== arm || runCell.bundleDigest !== bundle.digest || runCell.suiteDigest !== suite.digest || runCell.taskDigest !== task.digest || runCell.taskIndex !== taskIndex || runCell.repetition !== repetition || runCell.seed !== seed || runCell.attempt > task.attempt.maxAttempts || materialization.bundleDigest !== bundle.digest || materialization.benchmark.suite.digest !== suite.digest || materialization.benchmark.task.digest !== task.digest || materialization.codeKind !== bundle.code.kind || materialization.profileActivation.profilePlan.material.sourceProfileDigest !== canonicalCandidateDigest(bundle.profile) || evidence.receipt.runCellDigest !== runCell.digest || JSON.stringify(materialization.resolvedModel) !== JSON.stringify(task.model)) throw new Error(`candidate experiment measurement ${index} substituted its ${arm} arm`);
-		verifyTaskOutcome(task, evidence, index, arm);
-	}
-	const baselinePlan = baseline.materializationReceipt.executionPlan.material;
-	const candidatePlan = candidate.materializationReceipt.executionPlan.material;
-	if (baselinePlan.executionId === candidatePlan.executionId || baselinePlan.runCell.digest === candidatePlan.runCell.digest || baseline.materializationReceipt.digest === candidate.materializationReceipt.digest || baseline.receipt.digest === candidate.receipt.digest || baseline.digest === candidate.digest) throw new Error(`candidate experiment measurement ${index} reused one execution across arms`);
-	return {
-		baseline,
-		candidate
-	};
-}
-function verifyExecutionEvidence(input) {
-	const evidence = candidateExecutionEvidenceSchema.parse(input);
-	verifySelfAddressed(evidence, "candidate execution evidence");
-	verifySelfAddressed(evidence.materializationReceipt, "candidate materialization receipt");
-	verifySelfAddressed(evidence.materializationReceipt.profileActivation, "candidate profile activation");
-	verifyMaterialAddressed(evidence.materializationReceipt.profileActivation.profilePlan, "candidate profile plan");
-	verifyMaterialAddressed(evidence.materializationReceipt.executionPlan, "candidate execution plan");
-	verifySelfAddressed(evidence.receipt, "candidate run receipt");
-	verifyMaterialAddressed(evidence.receipt.modelSettlement, "candidate model settlement");
-	verifyMaterialAddressed(evidence.receipt.taskOutcome, "candidate task outcome");
-	verifyMaterialAddressed(evidence.receipt.benchmarkResult, "candidate benchmark result");
-	return evidence;
-}
-function verifySelfAddressed(document, label) {
-	if (canonicalCandidateDigest(omitTopLevelDigest(document)) !== document.digest) throw new Error(`${label} digest is invalid`);
-}
-function verifyTaskOutcome(task, evidence, index, arm) {
-	const outcome = evidence.receipt.taskOutcome.material.outcome;
-	const result = evidence.receipt.benchmarkResult.material;
-	const prefix = `candidate experiment measurement ${index} ${arm}`;
-	if (result.evidence.sha256 === task.grader.artifact.sha256) throw new Error(`${prefix} reused grader bytes as grading evidence`);
-	const usage = combinedUsage(evidence);
-	const usageChecks = [
-		[
-			usage.modelCalls,
-			task.limits.maxModelCalls,
-			"model calls"
-		],
-		[
-			usage.inputTokens,
-			task.limits.maxInputTokens,
-			"input tokens"
-		],
-		[
-			usage.outputTokens,
-			task.limits.maxOutputTokens,
-			"output tokens"
-		],
-		[
-			usage.costUsdNanos,
-			Math.round(task.limits.maxCostUsd * 1e9),
-			"cost"
-		]
-	];
-	for (const [actual, maximum, label] of usageChecks) if (actual > maximum) throw new Error(`${prefix} ${label} ${actual} exceeds the signed limit ${maximum}`);
-	if (outcome.kind !== task.outcome.kind) throw new Error(`${prefix} returned an outcome outside the signed task contract`);
-	if (task.outcome.kind === "output") {
-		if (outcome.kind !== "output" || outcome.spec.mediaType !== task.outcome.mediaType || outcome.spec.maxBytes !== task.outcome.maxBytes) throw new Error(`${prefix} changed the signed output contract`);
-		return;
-	}
-	const repository = task.repository;
-	if (outcome.kind !== "workspace" || repository === void 0 || outcome.baseRepository.identity !== repository.identity || outcome.baseRepository.rootIdentity !== repository.rootIdentity || outcome.baseRepository.commit !== repository.baseCommit || outcome.baseRepository.tree !== repository.baseTree) throw new Error(`${prefix} did not start from the signed repository state`);
-}
-function verifyStableProfileMaterialization(measurements) {
-	for (const arm of ["baseline", "candidate"]) {
-		const expected = measurements[0]?.[arm].materializationReceipt.profileActivation;
-		if (!expected) throw new Error("candidate experiment contains no profile materialization");
-		const expectedDigest = canonicalCandidateDigest({
-			profilePlanDigest: expected.profilePlan.digest,
-			files: expected.files
-		});
-		for (const [index, measurement] of measurements.entries()) {
-			const activation = measurement[arm].materializationReceipt.profileActivation;
-			if (canonicalCandidateDigest({
-				profilePlanDigest: activation.profilePlan.digest,
-				files: activation.files
-			}) !== expectedDigest) throw new Error(`candidate experiment measurement ${index} ${arm} materialized a different profile`);
-		}
-	}
-}
-function completedSuccessfully(evidence) {
-	const termination = evidence.receipt.termination;
-	return termination.kind === "exit" && termination.exitCode === 0;
-}
-function verifyMaterialAddressed(evidence, label) {
-	if (canonicalCandidateDigest(evidence.material) !== evidence.digest) throw new Error(`${label} digest is invalid`);
-}
-function projectPairedMeasurement(measurement, index, adapter) {
-	if (typeof measurement.cellId !== "string" || !measurement.cellId.trim()) throw new Error(`paired measurement ${index} requires a cell id`);
-	return {
-		cellId: measurement.cellId,
-		baseline: projectRun(measurement.baseline, adapter, `paired measurement ${index} baseline`),
-		candidate: projectRun(measurement.candidate, adapter, `paired measurement ${index} candidate`)
-	};
-}
-function projectRun(run, adapter, label) {
-	const suppliedDimensions = adapter.dimensions(run);
-	if (!Array.isArray(suppliedDimensions)) throw new Error(`${label} dimensions must be an array`);
-	const dimensions = /* @__PURE__ */ new Map();
-	for (const dimension of suppliedDimensions) {
-		if (typeof dimension.name !== "string" || !dimension.name.trim()) throw new Error(`${label} contains an unnamed dimension`);
-		if (dimensions.has(dimension.name)) throw new Error(`${label} repeats dimension '${dimension.name}'`);
-		dimensions.set(dimension.name, finiteMeasurement(dimension.score, `${label} ${dimension.name}`));
-	}
-	const completed = adapter.completed(run);
-	const passed = adapter.passed(run);
-	if (typeof completed !== "boolean" || typeof passed !== "boolean") throw new Error(`${label} completion and pass values must be booleans`);
-	const costProvenance = adapter.costProvenance(run);
-	if (costProvenance !== "observed" && costProvenance !== "estimated") throw new Error(`${label} cost provenance must be observed or estimated`);
-	return {
-		score: finiteMeasurement(adapter.score(run), `${label} score`),
-		dimensions,
-		costUsd: nonNegativeMeasurement(adapter.costUsd(run), `${label} cost`),
-		costProvenance,
-		latencyMs: nonNegativeMeasurement(adapter.latencyMs(run), `${label} latency`),
-		completed,
-		passed
-	};
-}
-function sharedProjectedDimensions(measurements) {
-	const expected = [...measurements[0].baseline.dimensions.keys()];
-	for (const [index, measurement] of measurements.entries()) for (const [arm, run] of [["baseline", measurement.baseline], ["candidate", measurement.candidate]]) {
-		const actual = [...run.dimensions.keys()];
-		if (JSON.stringify(actual) !== JSON.stringify(expected)) throw new Error(`paired measurement ${index} ${arm} dimensions do not match the suite`);
-	}
-	return expected;
-}
-function dimensionScore(run, name) {
-	const value = run.dimensions.get(name);
-	if (value === void 0) throw new Error(`paired measurement is missing dimension '${name}'`);
-	return value;
-}
-function measuredEstimate(baseline, candidate, options) {
-	const bootstrap = pairedBootstrap(baseline, candidate, {
-		confidence: options.confidence,
-		resamples: options.resamples,
-		statistic: "mean",
-		seed: options.seed
-	});
-	const baselineMean = mean(baseline);
-	const candidateMean = mean(candidate);
-	const delta = candidateMean - baselineMean;
-	return {
-		baseline: baselineMean,
-		candidate: candidateMean,
-		delta,
-		confidenceInterval: {
-			level: bootstrap.confidence,
-			lower: Math.min(bootstrap.low, delta),
-			upper: Math.max(bootstrap.high, delta),
-			method: "paired-bootstrap",
-			statistic: "mean",
-			resamples: bootstrap.resamples
-		},
-		n: bootstrap.n
-	};
-}
-function observedPairedPrecision(options) {
-	const n = options.baselineScores.length;
-	const lowerMargin = Math.max(0, options.delta - options.lowerBound);
-	const minimumDetectableDelta = options.deltaThreshold + lowerMargin;
-	const scaleAssumed = [...options.baselineScores, ...options.candidateScores].every((score) => score >= -.001 && score <= 1.001);
-	const headroom = Math.max(0, 1 - mean(options.baselineScores));
-	const enoughRuns = n >= options.minProductiveRuns;
-	const attainable = !scaleAssumed || minimumDetectableDelta <= headroom;
-	const sufficient = enoughRuns && attainable;
-	const scorerNote = options.sharedScorerChannel ? " The same scorer evaluated both arms, so scorer bias is not measured." : "";
-	if (!enoughRuns) return {
-		sufficient,
-		minimumDetectableDelta,
-		scaleAssumed,
-		reason: `only ${n} paired runs; ${options.minProductiveRuns} required.${scorerNote}`
-	};
-	if (!attainable) return {
-		sufficient,
-		minimumDetectableDelta,
-		scaleAssumed,
-		reason: `observed paired uncertainty gives a minimum detectable delta of ${minimumDetectableDelta.toFixed(3)}, above the remaining score headroom ${headroom.toFixed(3)}.${scorerNote}`
-	};
-	return {
-		sufficient,
-		minimumDetectableDelta,
-		scaleAssumed,
-		reason: `observed paired uncertainty gives a minimum detectable delta of ${minimumDetectableDelta.toFixed(3)} at ${options.confidence.toFixed(2)} confidence from ${n} runs.${scorerNote}`
-	};
-}
-function finiteMeasurement(value, label) {
-	if (!Number.isFinite(value)) throw new Error(`${label} must be finite`);
-	return value;
-}
-function nonNegativeMeasurement(value, label) {
-	if (!Number.isFinite(value) || value < 0) throw new Error(`${label} must be non-negative`);
-	return value;
-}
-const candidateExecutionEvidenceAdapter = {
-	score: (evidence) => evidence.receipt.benchmarkResult.material.score,
-	dimensions: (evidence) => evidence.receipt.benchmarkResult.material.dimensions,
-	costUsd: costFromEvidence,
-	costProvenance: (evidence) => combinedUsage(evidence).costProvenance,
-	latencyMs: latencyFromEvidence,
-	completed: completedSuccessfully,
-	passed: (evidence) => evidence.receipt.benchmarkResult.material.passed
-};
-function costFromEvidence(evidence) {
-	return combinedUsage(evidence).costUsdNanos / 1e9;
-}
-function latencyFromEvidence(evidence) {
-	return evidence.receipt.timing.durationMs + evidence.receipt.benchmarkResult.material.grading.timing.durationMs;
-}
-function combinedUsage(evidence) {
-	const candidate = evidence.receipt.modelSettlement.material.usage;
-	const grader = evidence.receipt.benchmarkResult.material.grading.usage;
-	return addFixedSpend(candidate, grader);
-}
 function cellIds(experiment) {
 	const { suite, tasks } = experiment.benchmark;
 	return suite.seeds.map((_, index) => {
@@ -862,10 +890,6 @@ function cellIds(experiment) {
 		return `${tasks[taskIndex]?.scenario.id ?? taskIndex}:${repetition}`;
 	});
 }
-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;
-}
 //#endregion
 //#region src/contract/profile-measured-comparison.ts
 /** Content-address one held-out profile task before either state can execute it. */
@@ -991,19 +1015,7 @@ function measuredComparisonFromAgentProfileImprovementExperiment(options) {
 			candidateContentHash: experiment.candidate.stateDigest
 		},
 		diff: canonicalCandidateJson(experiment.change),
-		evaluation: {
-			generationsExplored: options.generationsExplored ?? 0,
-			preparation: options.preparation,
-			measurement: {
-				wallDurationMs: options.measurement.wallDurationMs,
-				workDurationMs: evaluation.measurementWorkDurationMs,
-				cost: evaluation.measurementCost
-			},
-			total: {
-				wallDurationMs: options.preparation.wallDurationMs + options.measurement.wallDurationMs,
-				cost: evaluation.totalCost
-			}
-		},
+		evaluation: comparisonEvaluation(options, evaluation),
 		...options.metadata ? { metadata: options.metadata } : {}
 	});
 	const { recordDigest: _recordDigest, ...provenance } = provisional.provenance;
@@ -1021,19 +1033,7 @@ function measuredComparisonFromAgentProfileImprovementExperiment(options) {
 /** Recompute a profile comparison from the exact sealed experiment and receipts. */
 function verifyAgentProfileImprovementExperimentComparison(input) {
 	const comparison = agentProfileImprovementMeasuredComparisonSchema.parse(input);
-	const recomputed = measuredComparisonFromAgentProfileImprovementExperiment({
-		experiment: comparison.experiment,
-		measurements: comparison.measurements,
-		runId: comparison.provenance.runId,
-		...comparison.candidate ? { candidate: comparison.candidate } : {},
-		generationsExplored: comparison.evaluation.generationsExplored,
-		preparation: comparison.evaluation.preparation,
-		measurement: {
-			wallDurationMs: comparison.evaluation.measurement.wallDurationMs,
-			cost: comparison.evaluation.measurement.cost
-		},
-		...comparison.metadata ? { metadata: comparison.metadata } : {}
-	});
+	const recomputed = measuredComparisonFromAgentProfileImprovementExperiment(comparisonRecomputationInput(comparison));
 	if (canonicalCandidateDigest(recomputed) !== canonicalCandidateDigest(comparison)) throw new Error("profile improvement comparison does not match its Runtime receipts");
 	return comparison;
 }
@@ -1048,24 +1048,7 @@ function profileMeasurementModel(experiment) {
 	return models.length === 1 ? models[0] : "multiple-models";
 }
 function profileMeasurementCostReceipt(measurements, model) {
-	const usage = measurements.reduce((sum, measurement) => addFixedSpend(addFixedSpend(sum, combinedProfileUsage(measurement.baseline)), combinedProfileUsage(measurement.candidate)), {
-		inputTokens: 0,
-		outputTokens: 0,
-		cachedInputTokens: 0,
-		reasoningTokens: 0,
-		modelCalls: 0,
-		costUsdNanos: 0,
-		costProvenance: "observed"
-	});
-	const costUsd = usage.costUsdNanos / 1e9;
-	return {
-		model,
-		inputTokens: usage.inputTokens,
-		outputTokens: usage.outputTokens,
-		cachedTokens: usage.cachedInputTokens,
-		reasoningTokens: usage.reasoningTokens,
-		...usage.costProvenance === "observed" ? { actualCostUsd: costUsd } : { estimatedCostUsd: costUsd }
-	};
+	return pairedFixedSpendReceipt(measurements, model, combinedProfileUsage);
 }
 function combinedProfileUsage(receipt) {
 	return addFixedSpend(receipt.usage, receipt.grading.usage);
@@ -1146,186 +1129,10 @@ const profileReceiptAdapter = {
 	passed: (receipt) => receipt.grading.passed
 };
 //#endregion
-//#region src/contract/intake/run-record-dir.ts
-/**
-* # `intake/run-record-dir` — load a directory or file of `RunRecord`s.
-*
-* The on-disk counterpart to the in-memory intake adapters: point it at a
-* single `.json` (array) / `.jsonl` (one record per line) file or at a
-* directory of such files, and it returns the substrate-canonical
-* `RunRecord[]` ready for `analyzeRuns({ runs })`.
-*
-* Validation is at the boundary: each parsed object goes through
-* `parseRunRecordSafe`. By default an invalid record fails loud with its
-* file + index; pass `onInvalid: 'collect'` to keep the valid records and
-* receive the rejects as structured diagnostics instead.
-*/
-const ANALYSIS_ARTIFACT$1 = "analysis.json";
-function defaultInclude(fileName) {
-	if (fileName === ANALYSIS_ARTIFACT$1) return false;
-	return fileName.endsWith(".json") || fileName.endsWith(".jsonl");
-}
-/**
-* Resolve a file or directory path into validated `RunRecord[]`.
-*
-* A `.json` file must parse to a top-level array; a `.jsonl` file is one
-* record per non-empty line. Directories are read shallowly by default
-* (set `recursive` to descend); the `analysis.json` output artifact is
-* always excluded.
-*/
-async function fromRunRecordDir(path, options = {}) {
-	const onInvalid = options.onInvalid ?? "throw";
-	const include = options.include ?? defaultInclude;
-	const filePaths = (await stat(path)).isDirectory() ? await collectFiles(path, include, options.recursive ?? false) : [path];
-	const runs = [];
-	const rejected = [];
-	for (const file of filePaths) {
-		const raw = await parseRecordFile(file);
-		for (const { index, value } of raw) {
-			const parsed = parseRunRecordSafe(value);
-			if (parsed.ok) {
-				runs.push(parsed.value);
-				continue;
-			}
-			const rejection = {
-				file,
-				index,
-				reason: parsed.error.message
-			};
-			if (onInvalid === "throw") throw new Error(`fromRunRecordDir: invalid RunRecord in '${file}' at index ${index}: ${parsed.error.message}`);
-			rejected.push(rejection);
-		}
-	}
-	return {
-		runs,
-		rejected,
-		files: filePaths
-	};
-}
-/** Read a single `.json` / `.jsonl` file into `{ index, value }` pairs. A
-*  malformed JSONL line throws with its line number rather than being skipped —
-*  silent line-dropping is how corpora quietly shrink. */
-async function parseRecordFile(file) {
-	const trimmed = (await readFile(file, "utf8")).trim();
-	if (trimmed.length === 0) return [];
-	if (trimmed.startsWith("[")) {
-		const parsed = JSON.parse(trimmed);
-		if (!Array.isArray(parsed)) throw new Error(`fromRunRecordDir: file '${file}' did not parse to an array`);
-		return parsed.map((value, index) => ({
-			index,
-			value
-		}));
-	}
-	const out = [];
-	const lines = trimmed.split("\n");
-	for (let i = 0; i < lines.length; i++) {
-		const line = lines[i].trim();
-		if (line.length === 0) continue;
-		try {
-			out.push({
-				index: i,
-				value: JSON.parse(line)
-			});
-		} catch (err) {
-			throw new Error(`fromRunRecordDir: file '${file}' line ${i + 1} is not valid JSON: ${err instanceof Error ? err.message : String(err)}`);
-		}
-	}
-	return out;
-}
-/** Sorted file list under a directory, filtered by `include`. Sorted so the
-*  resulting `RunRecord` order — and any downstream fingerprint — is stable
-*  across filesystems. */
-async function collectFiles(dir, include, recursive) {
-	const entries = await readdir(dir, { withFileTypes: true });
-	const files = [];
-	const subdirs = [];
-	for (const entry of entries) {
-		if (entry.isDirectory()) {
-			if (recursive) subdirs.push(join(dir, entry.name));
-			continue;
-		}
-		if (include(entry.name)) files.push(join(dir, entry.name));
-	}
-	files.sort();
-	subdirs.sort();
-	for (const sub of subdirs) files.push(...await collectFiles(sub, include, recursive));
-	return files;
-}
-//#endregion
-//#region src/contract/eval-reporting-suite.ts
-/**
-* # `evalReportingSuite` — one call from runs (or a run dir) to `analysis.json`.
-*
-* A thin wrapper over the analysis primitive (`analyzeRuns`) and the on-disk
-* intake adapter (`fromRunRecordDir`). It does NOT reimplement any statistics,
-* distributions, or clustering — it resolves the input into validated
-* `RunRecord[]`, calls `analyzeRuns` with the options you'd pass it directly,
-* wraps the result in a small provenance envelope, and (optionally) writes a
-* single `analysis.json` artifact.
-*
-* ```ts
-* // From a directory of run files, write ./runs/analysis.json:
-* const suite = await evalReportingSuite('./runs', { write: true })
-* // From records already in memory, no write:
-* const suite = await evalReportingSuite(records, { analyze: { decisionThreshold: 0.03 } })
-* suite.report // the InsightReport — distributions, paired lift, findings rollup
-* ```
-*/
-const ANALYSIS_ARTIFACT = "analysis.json";
-/**
-* Resolve runs (or a run dir/file), run `analyzeRuns`, and optionally persist a
-* single `analysis.json`. The only analysis logic lives in `analyzeRuns`; this
-* function is composition + I/O.
-*/
-async function evalReportingSuite(input, options = {}) {
-	const fromPath = typeof input === "string";
-	let runs;
-	let files = [];
-	let rejected = [];
-	if (fromPath) {
-		const loaded = await fromRunRecordDir(input, options.load);
-		runs = loaded.runs;
-		files = loaded.files;
-		rejected = loaded.rejected;
-	} else runs = input;
-	if (runs.length === 0) throw new Error(fromPath ? `evalReportingSuite: no RunRecords found at '${input}'` : "evalReportingSuite: no RunRecords to analyze");
-	const result = {
-		report: await analyzeRuns({
-			...options.analyze,
-			runs
-		}),
-		provenance: {
-			generatedAt: (/* @__PURE__ */ new Date()).toISOString(),
-			runCount: runs.length,
-			sourcePath: fromPath ? input : null,
-			files,
-			rejected
-		},
-		writtenTo: null
-	};
-	const target = resolveWriteTarget(options.write, fromPath ? input : null);
-	if (target) {
-		await mkdir(dirname(target), { recursive: true });
-		await writeFile(target, `${JSON.stringify(result, null, 2)}\n`, "utf8");
-		result.writtenTo = target;
-	}
-	return result;
-}
-/** Resolve where (if anywhere) to write `analysis.json`. Returns null when
-*  writing is disabled. Throws on `write: true` with in-memory input — there is
-*  no directory to anchor the artifact to, and silently inventing `cwd` would
-*  scatter files. */
-function resolveWriteTarget(write, sourcePath) {
-	if (!write) return null;
-	if (typeof write === "string") return write.endsWith("/") || !write.endsWith(".json") && !write.endsWith(".jsonl") ? join(write, ANALYSIS_ARTIFACT) : write;
-	if (sourcePath === null) throw new Error("evalReportingSuite: write:true needs a source path to anchor analysis.json — pass an explicit output path when analyzing in-memory records");
-	return sourcePath.endsWith(".json") || sourcePath.endsWith(".jsonl") ? join(dirname(sourcePath), ANALYSIS_ARTIFACT) : join(sourcePath, ANALYSIS_ARTIFACT);
-}
-//#endregion
-//#region src/contract/intake/agent-trace.ts
-function rangeLines(r) {
-	return Math.max(0, r.end_line - r.start_line + 1);
-}
+//#region src/contract/intake/agent-trace.ts
+function rangeLines(r) {
+	return Math.max(0, r.end_line - r.start_line + 1);
+}
 /**
 * Build a commit → provenance index from Agent Trace records. Multiple records
 * for the same revision are merged. Records without `vcs.revision` are skipped
@@ -1402,6 +1209,94 @@ function partitionRunsByAuthoringModel(runs, index) {
 	};
 }
 //#endregion
+//#region src/contract/intake/code-agent-fields.ts
+function record(value) {
+	return isRecord(value) ? value : null;
+}
+function stringField(obj, key) {
+	const value = obj[key];
+	return typeof value === "string" && value.length > 0 ? value : void 0;
+}
+function numberField(obj, key) {
+	const value = obj[key];
+	return typeof value === "number" && Number.isFinite(value) ? value : void 0;
+}
+function timestampMs(value) {
+	if (typeof value === "number" && Number.isFinite(value)) return value > 0xe8d4a51000 ? value : value * 1e3;
+	if (typeof value === "string" && value.length > 0) {
+		const parsed = Date.parse(value);
+		return Number.isFinite(parsed) ? parsed : void 0;
+	}
+}
+//#endregion
+//#region src/contract/intake/code-agent-identity.ts
+function sessionIdFromEntries(source, entries) {
+	for (const entry of entries) {
+		if (source === "pi" && entry.type === "session") {
+			const id = stringField(entry, "id");
+			if (id) return id;
+		}
+		if (source === "codex") {
+			const threadId = stringField(entry, "thread_id");
+			if (threadId) return threadId;
+			const payload = record(entry.payload);
+			const id = payload ? stringField(payload, "id") : void 0;
+			if (id) return id;
+		}
+		const payload = record(record(entry.message)?.payload);
+		const sessionId = stringField(entry, "sessionID") ?? stringField(entry, "sessionId") ?? stringField(entry, "session_id") ?? stringField(payload ?? {}, "session_id");
+		if (sessionId) return sessionId;
+	}
+}
+function modelFromEntries(source, entries) {
+	if (source === "pi") return summarizeSessionInput("pi", { records: entries }).lastServedModel ?? void 0;
+	let providerModel;
+	for (const entry of entries) {
+		if (source === "codex") {
+			const payload = record(entry.payload);
+			const model = payload ? stringField(payload, "model") : void 0;
+			if (model) return model;
+			providerModel = providerModel ?? (payload ? stringField(payload, "model_provider") : void 0);
+		}
+		const openCodeModel = stringField(entry, "modelID");
+		if (openCodeModel) {
+			const provider = stringField(entry, "providerID");
+			return provider ? `${provider}/${openCodeModel}` : openCodeModel;
+		}
+		const modelObject = record(entry.model);
+		const nestedModel = stringField(modelObject ?? {}, "modelID");
+		if (nestedModel) {
+			const provider = stringField(modelObject ?? {}, "providerID");
+			return provider ? `${provider}/${nestedModel}` : nestedModel;
+		}
+		const message = record(entry.message);
+		const model = message ? stringField(message, "model") : void 0;
+		if (model) return model;
+	}
+	return providerModel;
+}
+function cwdFromEntries(entries) {
+	for (const entry of entries) {
+		const path = record(entry.path);
+		const cwd = stringField(entry, "cwd") ?? stringField(path ?? {}, "cwd") ?? stringField(path ?? {}, "root") ?? (record(entry.payload) ? stringField(record(entry.payload), "cwd") : void 0);
+		if (cwd) return cwd;
+	}
+}
+function fallbackSessionId$1(source, sourcePath) {
+	return hashString(`${source}:${sourcePath ?? "unknown-session"}`).slice(7, 27);
+}
+function withSnapshot(model) {
+	const trimmed = model.trim() || "unknown";
+	if (trimmed.includes("@") || /-\d{8}$/.test(trimmed) || /-\d{4}-\d{2}-\d{2}$/.test(trimmed)) return trimmed;
+	return `${trimmed}@observed-local`;
+}
+function hashString(value) {
+	return `sha256:${createHash("sha256").update(value).digest("hex")}`;
+}
+function stableSeed(value) {
+	return createHash("sha256").update(value).digest().readUInt32BE(0);
+}
+//#endregion
 //#region src/contract/intake/code-agent-observation.ts
 /**
 * Where a session came from. `pi` is a native Pi session (`~/.pi/agent/sessions/*.jsonl`), read
@@ -1414,12 +1309,12 @@ function partitionRunsByAuthoringModel(runs, index) {
 * stay out of this projection; callers retain the source JSONL as evidence.
 */
 function observeCodeAgentSession(options) {
-	const entries = options.entries.filter(isRecord$1);
+	const entries = options.entries.filter(isRecord);
 	const projection = projectionFor(options.source, entries);
 	const terminal = terminalWithExecution(projection, options.execution);
 	return {
 		source: options.source,
-		sessionId: sessionIdFromEntries$1(options.source, entries) ?? fallbackSessionId$1(options.source, options.sourcePath),
+		sessionId: sessionIdFromEntries(options.source, entries) ?? fallbackSessionId(options.source, options.sourcePath),
 		finalText: projection.finalText,
 		terminal,
 		actions: projection.actions
@@ -1443,16 +1338,16 @@ function codexProjection(entries) {
 	let terminal = "unknown";
 	let explicitTerminal = false;
 	for (const entry of entries) {
-		const entryType = stringField$1(entry, "type");
-		const payload = record$1(entry.payload) ?? {};
-		const payloadType = stringField$1(payload, "type");
-		const timestampMs = timestamp(entry.timestamp);
+		const entryType = stringField(entry, "type");
+		const payload = record(entry.payload) ?? {};
+		const payloadType = stringField(payload, "type");
+		const timestampMs$1 = timestampMs(entry.timestamp);
 		const itemEvent = admittedCodexItem(entry, transcriptStream);
 		const item = itemEvent?.item ?? null;
 		if (itemEvent?.eventType === "item.started" && item) {
-			const itemType = stringField$1(item, "type");
+			const itemType = stringField(item, "type");
 			if (isCodexActionItem(itemType)) {
-				const id = stringField$1(item, "id") ?? `item-${actions.length}`;
+				const id = stringField(item, "id") ?? `item-${actions.length}`;
 				const action = actionFor({
 					id,
 					stepIndex: actions.length,
@@ -1460,7 +1355,7 @@ function codexProjection(entries) {
 					surface: codexItemSurface(itemType, item),
 					name: codexItemName(itemType, item),
 					status: "started",
-					timestampMs,
+					timestampMs: timestampMs$1,
 					metadata: compactMetadata({
 						sourceEventType: itemEvent.eventType,
 						itemType
@@ -1471,10 +1366,10 @@ function codexProjection(entries) {
 			}
 		}
 		if (itemEvent?.eventType === "item.completed" && item) {
-			const itemType = stringField$1(item, "type");
-			if (itemType === "agent_message" || itemType === "message") finalText = nonEmpty(stringField$1(item, "text")) ?? finalText;
+			const itemType = stringField(item, "type");
+			if (itemType === "agent_message" || itemType === "message") finalText = nonEmpty(stringField(item, "text")) ?? finalText;
 			if (isCodexActionItem(itemType)) {
-				const id = stringField$1(item, "id") ?? `item-${actions.length}`;
+				const id = stringField(item, "id") ?? `item-${actions.length}`;
 				const status = codexItemStatus(itemType, item);
 				const existing = calls.get(id);
 				if (existing) existing.status = status;
@@ -1486,7 +1381,7 @@ function codexProjection(entries) {
 						surface: codexItemSurface(itemType, item),
 						name: codexItemName(itemType, item),
 						status,
-						timestampMs,
+						timestampMs: timestampMs$1,
 						metadata: compactMetadata({
 							sourceEventType: itemEvent.eventType,
 							itemType
@@ -1500,8 +1395,8 @@ function codexProjection(entries) {
 		if (entryType === "response_item") {
 			if (payloadType === "message" || payloadType === "agent_message") finalText = textFromMessagePayload(payload) ?? finalText;
 			if (payloadType === "function_call" || payloadType === "custom_tool_call") {
-				const id = stringField$1(payload, "call_id") ?? stringField$1(payload, "id") ?? `call-${actions.length}`;
-				const name = stringField$1(payload, "name") ?? payloadType;
+				const id = stringField(payload, "call_id") ?? stringField(payload, "id") ?? `call-${actions.length}`;
+				const name = stringField(payload, "name") ?? payloadType;
 				const action = actionFor({
 					id,
 					stepIndex: actions.length,
@@ -1509,7 +1404,7 @@ function codexProjection(entries) {
 					surface: surfaceForTool(name),
 					name,
 					status: "started",
-					timestampMs,
+					timestampMs: timestampMs$1,
 					metadata: compactMetadata({
 						sourceEventType: entryType,
 						payloadType
@@ -1519,23 +1414,23 @@ function codexProjection(entries) {
 				actions.push(action);
 			}
 			if (payloadType === "function_call_output" || payloadType === "custom_tool_call_output") {
-				const id = stringField$1(payload, "call_id") ?? stringField$1(payload, "id");
+				const id = stringField(payload, "call_id") ?? stringField(payload, "id");
 				const existing = id ? calls.get(id) : void 0;
 				if (existing) existing.status = looksLikeError(payload.output) ? "failed" : "completed";
 			}
 		}
 		if (entryType === "event_msg") {
-			if (payloadType === "agent_message") finalText = nonEmpty(stringField$1(payload, "message")) ?? nonEmpty(stringField$1(payload, "text")) ?? finalText;
+			if (payloadType === "agent_message") finalText = nonEmpty(stringField(payload, "message")) ?? nonEmpty(stringField(payload, "text")) ?? finalText;
 			if (payloadType === "patch_apply_end") {
 				const patchStatus = typeof payload.success === "boolean" ? payload.success ? "completed" : "failed" : "unknown";
 				actions.push(actionFor({
-					id: stringField$1(payload, "call_id") ?? `patch-${actions.length}`,
+					id: stringField(payload, "call_id") ?? `patch-${actions.length}`,
 					stepIndex: actions.length,
 					kind: "patch",
 					surface: "code",
 					name: "patch",
 					status: patchStatus,
-					timestampMs,
+					timestampMs: timestampMs$1,
 					metadata: compactMetadata({
 						sourceEventType: entryType,
 						payloadType
@@ -1543,15 +1438,15 @@ function codexProjection(entries) {
 				}));
 			}
 			if (payloadType === "sub_agent_activity") {
-				const name = stringField$1(payload, "kind") ?? "subagent";
+				const name = stringField(payload, "kind") ?? "subagent";
 				actions.push(actionFor({
-					id: stringField$1(payload, "event_id") ?? `subagent-${actions.length}`,
+					id: stringField(payload, "event_id") ?? `subagent-${actions.length}`,
 					stepIndex: actions.length,
 					kind: "tool",
 					surface: "subagent",
 					name,
 					status: statusFrom(payload),
-					timestampMs,
+					timestampMs: timestampMs$1,
 					metadata: compactMetadata({
 						sourceEventType: entryType,
 						payloadType
@@ -1562,20 +1457,20 @@ function codexProjection(entries) {
 				const completed = payloadType === "task_complete";
 				terminal = completed ? "completed" : "failed";
 				explicitTerminal = true;
-				actions.push(terminalAction(actions.length, payloadType, completed, timestampMs));
+				actions.push(terminalAction(actions.length, payloadType, completed, timestampMs$1));
 			}
 		}
 		if (entryType === "turn.completed" || entryType === "exec_done") {
 			const completed = entry.success !== false;
 			terminal = completed ? "completed" : "failed";
 			explicitTerminal = true;
-			finalText = nonEmpty(stringField$1(entry, "message")) ?? finalText;
-			actions.push(terminalAction(actions.length, entryType, completed, timestampMs));
+			finalText = nonEmpty(stringField(entry, "message")) ?? finalText;
+			actions.push(terminalAction(actions.length, entryType, completed, timestampMs$1));
 		}
 		if (entryType === "turn.failed" || entryType === "error") {
 			terminal = "failed";
 			explicitTerminal = true;
-			actions.push(terminalAction(actions.length, entryType, false, timestampMs));
+			actions.push(terminalAction(actions.length, entryType, false, timestampMs$1));
 		}
 	}
 	return {
@@ -1593,20 +1488,20 @@ function claudeProjection(entries) {
 	let terminal = "unknown";
 	let explicitTerminal = false;
 	for (const entry of entries) {
-		const entryType = stringField$1(entry, "type");
-		const timestampMs = timestamp(entry.timestamp);
-		const message = record$1(entry.message);
+		const entryType = stringField(entry, "type");
+		const timestampMs$2 = timestampMs(entry.timestamp);
+		const message = record(entry.message);
 		const content = Array.isArray(message?.content) ? message.content : [];
 		if (entryType === "assistant") {
 			const text = textBlocks(content);
 			if (text !== null) finalText = text;
 			for (const value of content) {
-				const part = record$1(value);
+				const part = record(value);
 				if (!part) continue;
-				const partType = stringField$1(part, "type");
+				const partType = stringField(part, "type");
 				if (partType === "tool_use") {
-					const id = stringField$1(part, "id") ?? `tool-${actions.length}`;
-					const name = stringField$1(part, "name") ?? "tool";
+					const id = stringField(part, "id") ?? `tool-${actions.length}`;
+					const name = stringField(part, "name") ?? "tool";
 					const action = actionFor({
 						id,
 						stepIndex: actions.length,
@@ -1614,7 +1509,7 @@ function claudeProjection(entries) {
 						surface: surfaceForTool(name),
 						name,
 						status: "started",
-						timestampMs,
+						timestampMs: timestampMs$2,
 						metadata: compactMetadata({
 							sourceEventType: entryType,
 							partType
@@ -1625,17 +1520,17 @@ function claudeProjection(entries) {
 				}
 			}
 			completeClaudeTools(content, calls);
-			const stopReason = stringField$1(message ?? {}, "stop_reason");
+			const stopReason = stringField(message ?? {}, "stop_reason");
 			if (stopReason === "end_turn") {
 				terminal = "completed";
 				explicitTerminal = true;
-				actions.push(terminalAction(actions.length, stopReason, true, timestampMs));
+				actions.push(terminalAction(actions.length, stopReason, true, timestampMs$2));
 			}
 		}
 		if (entryType === "user") completeClaudeTools(content, calls);
 		if (entryType === "tool_use") {
-			const id = stringField$1(entry, "id") ?? `tool-${actions.length}`;
-			const name = stringField$1(entry, "name") ?? "tool";
+			const id = stringField(entry, "id") ?? `tool-${actions.length}`;
+			const name = stringField(entry, "name") ?? "tool";
 			const action = actionFor({
 				id,
 				stepIndex: actions.length,
@@ -1643,47 +1538,47 @@ function claudeProjection(entries) {
 				surface: surfaceForTool(name),
 				name,
 				status: "started",
-				timestampMs,
+				timestampMs: timestampMs$2,
 				metadata: { sourceEventType: entryType }
 			});
 			calls.set(id, action);
 			actions.push(action);
 		}
 		if (entryType === "tool_result") {
-			const id = stringField$1(entry, "tool_use_id");
+			const id = stringField(entry, "tool_use_id");
 			const action = id ? calls.get(id) : void 0;
 			if (action) action.status = entry.is_error === true ? "failed" : "completed";
 		}
 		if (entryType === "system") {
-			const subtype = stringField$1(entry, "subtype");
-			const hookName = stringField$1(entry, "hook_name") ?? stringField$1(entry, "hookName");
+			const subtype = stringField(entry, "subtype");
+			const hookName = stringField(entry, "hook_name") ?? stringField(entry, "hookName");
 			if (subtype?.toLowerCase().includes("hook") || hookName) actions.push(actionFor({
-				id: stringField$1(entry, "uuid") ?? `hook-${actions.length}`,
+				id: stringField(entry, "uuid") ?? `hook-${actions.length}`,
 				stepIndex: actions.length,
 				kind: "tool",
 				surface: "hook",
 				name: hookName ?? subtype ?? "hook",
 				status: statusFrom(entry),
-				timestampMs,
+				timestampMs: timestampMs$2,
 				metadata: compactMetadata({
 					sourceEventType: entryType,
 					subtype
 				})
 			}));
 		}
-		if (entryType === "pr-link") actions.push(terminalAction(actions.length, entryType, true, timestampMs));
+		if (entryType === "pr-link") actions.push(terminalAction(actions.length, entryType, true, timestampMs$2));
 		if (entryType === "result") {
-			resultText = nonEmpty(stringField$1(entry, "result")) ?? resultText;
-			const subtype = stringField$1(entry, "subtype");
+			resultText = nonEmpty(stringField(entry, "result")) ?? resultText;
+			const subtype = stringField(entry, "subtype");
 			const completed = entry.is_error !== true && !subtype?.toLowerCase().includes("error");
 			terminal = completed ? "completed" : "failed";
 			explicitTerminal = true;
-			actions.push(terminalAction(actions.length, subtype ?? entryType, completed, timestampMs));
+			actions.push(terminalAction(actions.length, subtype ?? entryType, completed, timestampMs$2));
 		}
 		if (entryType === "error") {
 			terminal = "failed";
 			explicitTerminal = true;
-			actions.push(terminalAction(actions.length, entryType, false, timestampMs));
+			actions.push(terminalAction(actions.length, entryType, false, timestampMs$2));
 		}
 	}
 	return {
@@ -1695,28 +1590,28 @@ function claudeProjection(entries) {
 }
 function openCodeProjection(entries) {
 	const actions = [];
-	const assistantMessageIds = new Set(entries.filter((entry) => stringField$1(entry, "role") === "assistant").map((entry) => stringField$1(entry, "id")).filter((id) => id !== void 0));
+	const assistantMessageIds = new Set(entries.filter((entry) => stringField(entry, "role") === "assistant").map((entry) => stringField(entry, "id")).filter((id) => id !== void 0));
 	let finalText = null;
 	let terminal = "unknown";
 	let explicitTerminal = false;
 	for (const entry of entries) {
-		const entryType = stringField$1(entry, "type");
-		const part = record$1(entry.part) ?? entry;
-		const partType = stringField$1(part, "type");
-		const timestampMs = timestamp(entry.timestamp) ?? timestamp(record$1(entry.time)?.created);
-		if (entryType === "text" && record$1(entry.part)) finalText = nonEmpty(stringField$1(part, "text")) ?? finalText;
-		else if (partType === "text" && (assistantMessageIds.size === 0 || assistantMessageIds.has(stringField$1(part, "messageID") ?? ""))) finalText = nonEmpty(stringField$1(part, "text")) ?? finalText;
+		const entryType = stringField(entry, "type");
+		const part = record(entry.part) ?? entry;
+		const partType = stringField(part, "type");
+		const timestampMs$3 = timestampMs(entry.timestamp) ?? timestampMs(record(entry.time)?.created);
+		if (entryType === "text" && record(entry.part)) finalText = nonEmpty(stringField(part, "text")) ?? finalText;
+		else if (partType === "text" && (assistantMessageIds.size === 0 || assistantMessageIds.has(stringField(part, "messageID") ?? ""))) finalText = nonEmpty(stringField(part, "text")) ?? finalText;
 		if (entryType === "tool_use" || partType === "tool") {
-			const state = record$1(part.state) ?? {};
-			const name = stringField$1(part, "tool") ?? "tool";
+			const state = record(part.state) ?? {};
+			const name = stringField(part, "tool") ?? "tool";
 			actions.push(actionFor({
-				id: stringField$1(part, "callID") ?? stringField$1(part, "id") ?? `tool-${actions.length}`,
+				id: stringField(part, "callID") ?? stringField(part, "id") ?? `tool-${actions.length}`,
 				stepIndex: actions.length,
 				kind: "tool",
 				surface: surfaceForTool(name),
 				name,
 				status: statusFrom(state),
-				timestampMs,
+				timestampMs: timestampMs$3,
 				metadata: compactMetadata({
 					sourceEventType: entryType,
 					partType
@@ -1724,31 +1619,31 @@ function openCodeProjection(entries) {
 			}));
 		}
 		if (partType === "patch") actions.push(actionFor({
-			id: stringField$1(part, "id") ?? `patch-${actions.length}`,
+			id: stringField(part, "id") ?? `patch-${actions.length}`,
 			stepIndex: actions.length,
 			kind: "patch",
 			surface: "code",
 			name: "patch",
 			status: "completed",
-			timestampMs,
+			timestampMs: timestampMs$3,
 			metadata: compactMetadata({
 				sourceEventType: entryType,
 				partType
 			})
 		}));
-		if (stringField$1(entry, "role") === "assistant") {
-			const finish = stringField$1(entry, "finish");
+		if (stringField(entry, "role") === "assistant") {
+			const finish = stringField(entry, "finish");
 			if (finish === "stop" || finish === "error") {
 				const completed = finish === "stop";
 				terminal = completed ? "completed" : "failed";
 				explicitTerminal = true;
-				actions.push(terminalAction(actions.length, finish, completed, timestampMs));
+				actions.push(terminalAction(actions.length, finish, completed, timestampMs$3));
 			}
 		}
 		if (entryType === "error") {
 			terminal = "failed";
 			explicitTerminal = true;
-			actions.push(terminalAction(actions.length, entryType, false, timestampMs));
+			actions.push(terminalAction(actions.length, entryType, false, timestampMs$3));
 		}
 	}
 	return {
@@ -1765,15 +1660,15 @@ function kimiProjection(entries) {
 	let terminal = "unknown";
 	let explicitTerminal = false;
 	for (const entry of entries) {
-		const timestampMs = timestamp(entry.timestamp);
-		const message = record$1(entry.message);
-		const messageType = stringField$1(message ?? {}, "type");
-		const payload = record$1(message?.payload) ?? {};
-		if (messageType === "ContentPart" && stringField$1(payload, "type") === "text") finalText = nonEmpty(stringField$1(payload, "text")) ?? nonEmpty(stringField$1(payload, "content")) ?? finalText;
+		const timestampMs$4 = timestampMs(entry.timestamp);
+		const message = record(entry.message);
+		const messageType = stringField(message ?? {}, "type");
+		const payload = record(message?.payload) ?? {};
+		if (messageType === "ContentPart" && stringField(payload, "type") === "text") finalText = nonEmpty(stringField(payload, "text")) ?? nonEmpty(stringField(payload, "content")) ?? finalText;
 		if (messageType === "ToolCall") {
-			const fn = record$1(payload.function) ?? {};
-			const id = stringField$1(payload, "id") ?? `tool-${actions.length}`;
-			const name = stringField$1(fn, "name") ?? "tool";
+			const fn = record(payload.function) ?? {};
+			const id = stringField(payload, "id") ?? `tool-${actions.length}`;
+			const name = stringField(fn, "name") ?? "tool";
 			const action = actionFor({
 				id,
 				stepIndex: actions.length,
@@ -1781,22 +1676,22 @@ function kimiProjection(entries) {
 				surface: surfaceForTool(name),
 				name,
 				status: "started",
-				timestampMs,
+				timestampMs: timestampMs$4,
 				metadata: { sourceEventType: messageType }
 			});
 			calls.set(id, action);
 			actions.push(action);
 		}
 		if (messageType === "ToolResult") {
-			const id = stringField$1(payload, "tool_call_id");
+			const id = stringField(payload, "tool_call_id");
 			const action = id ? calls.get(id) : void 0;
-			if (action) action.status = record$1(payload.return_value)?.is_error === true ? "failed" : "completed";
+			if (action) action.status = record(payload.return_value)?.is_error === true ? "failed" : "completed";
 		}
 		if (messageType === "TurnEnd" || messageType === "StepInterrupted") {
 			const completed = messageType === "TurnEnd";
 			terminal = completed ? "completed" : "failed";
 			explicitTerminal = true;
-			actions.push(terminalAction(actions.length, messageType, completed, timestampMs));
+			actions.push(terminalAction(actions.length, messageType, completed, timestampMs$4));
 		}
 	}
 	return {
@@ -1816,7 +1711,7 @@ function piSessionProjection(entries) {
 		surface: surfaceForTool(call.name),
 		name: call.name,
 		status: call.status === "completed" ? "completed" : call.status === "error" ? "failed" : call.status === "pending" ? "started" : "unknown",
-		timestampMs: timestamp(call.startedAt),
+		timestampMs: timestampMs(call.startedAt),
 		metadata: { sourceEventType: "pi-tool-call" }
 	}));
 	const answer = [...session.messages].reverse().find((m) => m.role === "assistant" && m.actor === "agent" && m.parts.some((p) => p.type === "text"));
@@ -1836,17 +1731,17 @@ function piGraphProjection(entries) {
 	for (const entry of entries) {
 		const nodes = Array.isArray(entry.nodes) ? entry.nodes : [];
 		for (const value of nodes) {
-			const node = record$1(value);
-			const ir = record$1(node?.ir) ?? node;
-			const kind = stringField$1(ir ?? {}, "kind");
+			const node = record(value);
+			const ir = record(node?.ir) ?? node;
+			const kind = stringField(ir ?? {}, "kind");
 			if (kind === "ToolInvocation") actions.push(actionFor({
-				id: stringField$1(ir ?? {}, "id") ?? stringField$1(node ?? {}, "id") ?? `tool-${actions.length}`,
+				id: stringField(ir ?? {}, "id") ?? stringField(node ?? {}, "id") ?? `tool-${actions.length}`,
 				stepIndex: actions.length,
 				kind: "tool",
 				surface: "tool",
-				name: stringField$1(ir ?? {}, "name") ?? "graph-tool",
+				name: stringField(ir ?? {}, "name") ?? "graph-tool",
 				status: "started",
-				timestampMs: timestamp(ir?.createdAt),
+				timestampMs: timestampMs(ir?.createdAt),
 				metadata: {
 					sourceEventType: "graph-node",
 					graphKind: kind
@@ -1860,13 +1755,13 @@ function piGraphProjection(entries) {
 				terminal = "completed";
 				explicitTerminal = true;
 				actions.push(actionFor({
-					id: stringField$1(ir ?? {}, "id") ?? stringField$1(node ?? {}, "id") ?? `completion-${actions.length}`,
+					id: stringField(ir ?? {}, "id") ?? stringField(node ?? {}, "id") ?? `completion-${actions.length}`,
 					stepIndex: actions.length,
 					kind: "graph-completion",
 					surface: "tool",
 					name: "complete",
 					status: "completed",
-					timestampMs: timestamp(ir?.createdAt),
+					timestampMs: timestampMs(ir?.createdAt),
 					metadata: {
 						sourceEventType: "graph-node",
 						graphKind: kind
@@ -1884,9 +1779,9 @@ function piGraphProjection(entries) {
 }
 function completeClaudeTools(content, calls) {
 	for (const value of content) {
-		const part = record$1(value);
-		if (!part || stringField$1(part, "type") !== "tool_result") continue;
-		const id = stringField$1(part, "tool_use_id");
+		const part = record(value);
+		if (!part || stringField(part, "type") !== "tool_result") continue;
+		const id = stringField(part, "tool_use_id");
 		const action = id ? calls.get(id) : void 0;
 		if (action) action.status = part.is_error === true ? "failed" : "completed";
 	}
@@ -1900,21 +1795,21 @@ function completeClaudeTools(content, calls) {
 * PascalCase. Read both here so one branch serves both transports.
 */
 function codexItemEvent(entry) {
-	const entryType = stringField$1(entry, "type");
+	const entryType = stringField(entry, "type");
 	if (entryType === "item.started" || entryType === "item.completed") {
-		const item = record$1(entry.item);
+		const item = record(entry.item);
 		return item ? {
 			eventType: entryType,
 			item
 		} : null;
 	}
 	if (entryType !== "event_msg") return null;
-	const payload = record$1(entry.payload);
-	const payloadType = payload ? stringField$1(payload, "type") : void 0;
+	const payload = record(entry.payload);
+	const payloadType = payload ? stringField(payload, "type") : void 0;
 	if (payloadType !== "item_started" && payloadType !== "item_completed") return null;
-	const item = record$1(payload?.item);
+	const item = record(payload?.item);
 	if (!item) return null;
-	const itemType = codexItemType(stringField$1(item, "type"));
+	const itemType = codexItemType(stringField(item, "type"));
 	return {
 		eventType: payloadType === "item_started" ? "item.started" : "item.completed",
 		item: itemType === void 0 ? item : {
@@ -1934,7 +1829,7 @@ function codexItemType(itemType) {
 	return itemType.replace(/([a-z0-9])([A-Z])/g, "$1_$2").toLowerCase();
 }
 function hasCodexTranscriptStream(entries) {
-	return entries.some((entry) => stringField$1(entry, "type") === "response_item");
+	return entries.some((entry) => stringField(entry, "type") === "response_item");
 }
 /**
 * A rollout file records each turn twice: `response_item` holds the model
@@ -1962,7 +1857,7 @@ function admittedCodexItem(entry, transcriptStream) {
 	const event = codexItemEvent(entry);
 	if (!event) return null;
 	if (!transcriptStream) return event;
-	return isCodexExecutionOnlyItem(stringField$1(event.item, "type")) ? event : null;
+	return isCodexExecutionOnlyItem(stringField(event.item, "type")) ? event : null;
 }
 function isCodexActionItem(itemType) {
 	return itemType === "command_execution" || itemType === "mcp_tool_call" || itemType === "collab_tool_call" || itemType === "web_search" || itemType === "file_change";
@@ -1972,27 +1867,27 @@ function codexItemSurface(itemType, item) {
 	if (itemType === "collab_tool_call") return "subagent";
 	if (itemType === "web_search") return "web";
 	if (itemType === "file_change") return "code";
-	return surfaceForTool(stringField$1(item, "name") ?? itemType ?? "tool");
+	return surfaceForTool(stringField(item, "name") ?? itemType ?? "tool");
 }
 function codexItemName(itemType, item) {
-	if (itemType === "mcp_tool_call") return [stringField$1(item, "server"), stringField$1(item, "tool")].filter(Boolean).join("/") || "mcp";
-	if (itemType === "collab_tool_call") return stringField$1(item, "tool") ?? "subagent";
+	if (itemType === "mcp_tool_call") return [stringField(item, "server"), stringField(item, "tool")].filter(Boolean).join("/") || "mcp";
+	if (itemType === "collab_tool_call") return stringField(item, "tool") ?? "subagent";
 	if (itemType === "web_search") return "web_search";
 	if (itemType === "file_change") return "file_change";
-	return stringField$1(item, "name") ?? itemType ?? "tool";
+	return stringField(item, "name") ?? itemType ?? "tool";
 }
 function codexItemStatus(itemType, item) {
 	const status = statusFrom(item);
 	if (status === "failed") return status;
 	if (itemType === "command_execution") {
-		const exitCode = numberField$1(item, "exit_code");
+		const exitCode = numberField(item, "exit_code");
 		if (exitCode !== void 0) return exitCode === 0 ? "completed" : "failed";
 	}
-	if (itemType === "mcp_tool_call" && record$1(item.error)) return "failed";
+	if (itemType === "mcp_tool_call" && record(item.error)) return "failed";
 	return status === "unknown" || status === "started" ? "completed" : status;
 }
 function textFromMessagePayload(payload) {
-	const direct = nonEmpty(stringField$1(payload, "text"));
+	const direct = nonEmpty(stringField(payload, "text"));
 	if (direct) return direct;
 	const content = payload.content;
 	if (typeof content === "string") return nonEmpty(content);
@@ -2000,7 +1895,7 @@ function textFromMessagePayload(payload) {
 	return null;
 }
 function textBlocks(values) {
-	return nonEmpty(values.map((value) => record$1(value)).filter((value) => value !== null).filter((value) => stringField$1(value, "type") === "text").map((value) => stringField$1(value, "text")).filter((value) => value !== void 0).join("\n"));
+	return nonEmpty(values.map((value) => record(value)).filter((value) => value !== null).filter((value) => stringField(value, "type") === "text").map((value) => stringField(value, "text")).filter((value) => value !== void 0).join("\n"));
 }
 function terminalWithExecution(projection, execution) {
 	if (!execution) return {
@@ -2056,7 +1951,7 @@ function isCodexCollaborationTool(normalized) {
 	return /^[a-z_]+_agents?$/.test(verb) || verb === "send_message" || verb === "send_input" || verb === "followup_task";
 }
 function statusFrom(value) {
-	const status = (stringField$1(value, "status") ?? "").toLowerCase();
+	const status = (stringField(value, "status") ?? "").toLowerCase();
 	if (status === "completed" || status === "success" || status === "succeeded") return "completed";
 	if (status === "error" || status === "failed" || status === "declined") return "failed";
 	if (status === "running" || status === "started" || status === "in_progress") return "started";
@@ -2065,31 +1960,13 @@ function statusFrom(value) {
 function compactMetadata(input) {
 	return Object.fromEntries(Object.entries(input).filter((entry) => entry[1] !== void 0));
 }
-function sessionIdFromEntries$1(source, entries) {
-	for (const entry of entries) {
-		if (source === "pi" && entry.type === "session") {
-			const id = stringField$1(entry, "id");
-			if (id) return id;
-		}
-		if (source === "codex") {
-			const threadId = stringField$1(entry, "thread_id");
-			if (threadId) return threadId;
-			const payload = record$1(entry.payload);
-			const id = payload ? stringField$1(payload, "id") : void 0;
-			if (id) return id;
-		}
-		const payload = record$1(record$1(entry.message)?.payload);
-		const sessionId = stringField$1(entry, "sessionID") ?? stringField$1(entry, "sessionId") ?? stringField$1(entry, "session_id") ?? stringField$1(payload ?? {}, "session_id");
-		if (sessionId) return sessionId;
-	}
-}
-function fallbackSessionId$1(source, sourcePath) {
+function fallbackSessionId(source, sourcePath) {
 	return `${source}:${sourcePath ?? "unknown-session"}`;
 }
 function looksLikeError(value) {
-	if (isRecord$1(value)) {
+	if (isRecord(value)) {
 		if (value.success === false || value.is_error === true) return true;
-		const exitCode = numberField$1(value, "exit_code");
+		const exitCode = numberField(value, "exit_code");
 		if (exitCode !== void 0 && exitCode !== 0) return true;
 	}
 	const text = typeof value === "string" ? value : JSON.stringify(value ?? "");
@@ -2098,298 +1975,147 @@ function looksLikeError(value) {
 function nonEmpty(value) {
 	return value !== void 0 && value.length > 0 ? value : null;
 }
-function isRecord$1(value) {
-	return value !== null && typeof value === "object" && !Array.isArray(value);
+//#endregion
+//#region src/contract/intake/code-agent-session-metrics.ts
+function emptyMetrics(entries) {
+	return {
+		entries,
+		userMessages: 0,
+		assistantMessages: 0,
+		reasoningItems: 0,
+		toolCalls: 0,
+		toolOutputs: 0,
+		toolErrors: 0,
+		unclassifiedErrors: 0,
+		patchAttempts: 0,
+		patchSuccesses: 0,
+		patchFailures: 0,
+		turnsStarted: 0,
+		turnsCompleted: 0,
+		turnsAborted: 0,
+		contextCompactions: 0,
+		mcpCalls: 0,
+		subagentCalls: 0,
+		skillCalls: 0,
+		hookCalls: 0,
+		webCalls: 0,
+		codeActions: 0,
+		prLinks: 0,
+		fileSnapshots: 0,
+		graphNodes: 0,
+		graphEdges: 0,
+		actionCandidates: 0,
+		verificationReports: 0,
+		completionDecisions: 0,
+		reliabilityRows: 0,
+		reliabilityLift: 0,
+		inputTokens: 0,
+		outputTokens: 0,
+		reasoningTokens: 0,
+		cachedTokens: 0,
+		cacheWriteTokens: 0,
+		observedCostUsd: 0,
+		observedCostCaptured: false,
+		wallMs: 0,
+		processScore: 0
+	};
 }
-function record$1(value) {
-	return isRecord$1(value) ? value : null;
+function penalizeErrors(base, metrics) {
+	const operations = Math.max(1, metrics.toolCalls + metrics.patchAttempts);
+	const errorRate = Math.min(1, (metrics.toolErrors + metrics.patchFailures) / operations);
+	return clamp01(base * (1 - .5 * errorRate));
 }
-function stringField$1(value, key) {
-	const field = value[key];
-	return typeof field === "string" && field.length > 0 ? field : void 0;
+function emptyTokenTotals() {
+	return {
+		input: 0,
+		output: 0,
+		reasoning: 0,
+		cached: 0,
+		cacheWrite: 0
+	};
 }
-function numberField$1(value, key) {
-	const field = value[key];
-	return typeof field === "number" && Number.isFinite(field) ? field : void 0;
+function hasTokenTotals(usage) {
+	return usage.input > 0 || usage.output > 0 || usage.reasoning > 0 || usage.cached > 0 || usage.cacheWrite > 0;
 }
-function timestamp(value) {
-	if (typeof value === "number" && Number.isFinite(value)) return value > 0xe8d4a51000 ? value : value * 1e3;
-	if (typeof value === "string" && value.length > 0) {
-		const parsed = Date.parse(value);
-		return Number.isFinite(parsed) ? parsed : void 0;
-	}
+function addUsage(metrics, usage) {
+	const parsed = readUsage(usage);
+	metrics.inputTokens += parsed.input;
+	metrics.outputTokens += parsed.output;
+	metrics.reasoningTokens += parsed.reasoning;
+	metrics.cachedTokens += parsed.cached;
+	metrics.cacheWriteTokens += parsed.cacheWrite;
 }
-//#endregion
-//#region src/contract/intake/code-agent-session.ts
-/** The single per-line rule. Both the string path and the file path call this,
-*  so malformed-line handling and entry validation cannot drift apart. */
-function readCodeAgentJsonlLine(line, lineNumber) {
-	const trimmed = line.trim();
-	if (!trimmed) return void 0;
-	try {
-		return {
-			kind: "entry",
-			lineNumber,
-			entry: JSON.parse(trimmed)
-		};
-	} catch {
-		return {
-			kind: "malformed",
-			lineNumber
-		};
-	}
+function addUsageTo(totals, usage) {
+	const parsed = readUsage(usage);
+	totals.input += parsed.input;
+	totals.output += parsed.output;
+	totals.reasoning += parsed.reasoning;
+	totals.cached += parsed.cached;
+	totals.cacheWrite += parsed.cacheWrite;
 }
-function parseCodeAgentJsonl(jsonl) {
-	const entries = [];
-	let malformedLines = 0;
-	let lineNumber = 0;
-	for (const line of jsonl.split("\n")) {
-		lineNumber += 1;
-		const read = readCodeAgentJsonlLine(line, lineNumber);
-		if (!read) continue;
-		if (read.kind === "malformed") malformedLines += 1;
-		else entries.push(read.entry);
-	}
+function readUsage(usage) {
+	const parsed = extractUsage(usage);
 	return {
-		entries,
-		malformedLines
+		input: parsed?.input ?? 0,
+		output: parsed?.output ?? 0,
+		reasoning: parsed?.reasoning ?? 0,
+		cached: parsed?.cached ?? 0,
+		cacheWrite: parsed?.cacheWrite ?? 0
 	};
 }
-/** Reads a transcript one line at a time and never holds the file as a single
-*  string. `parseCodeAgentJsonl` needs the whole file in one string, so a
-*  session above V8's ~512MB string ceiling throws `ERR_STRING_TOO_LONG` and
-*  cannot be ingested at all; the largest real Codex rollout on record is 695MB.
-*
-*  Lines break on `\n` only, which is what the string path's `split('\n')` does.
-*  `node:readline` also breaks on a bare `\r`, so it is deliberately not used
-*  here: a lone carriage return inside a line must stay inside that line for the
-*  two paths to report the same malformed count. */
-async function* streamCodeAgentJsonlFile(path) {
-	const stream = createReadStream(path, { encoding: "utf8" });
-	let pending = "";
-	let lineNumber = 0;
-	try {
-		for await (const chunk of stream) {
-			pending += chunk;
-			let start = 0;
-			for (let at = pending.indexOf("\n"); at !== -1; at = pending.indexOf("\n", start)) {
-				lineNumber += 1;
-				const read = readCodeAgentJsonlLine(pending.slice(start, at), lineNumber);
-				if (read) yield read;
-				start = at + 1;
-			}
-			pending = pending.slice(start);
+function resolveSessionCost(explicit, metrics, model) {
+	if (explicit) {
+		if (explicit.kind === "uncaptured") {
+			if (explicit.usd !== null) throw new Error("uncaptured cost must have usd: null");
+			return {
+				costUsd: null,
+				costProvenance: explicit
+			};
 		}
-	} finally {
-		stream.destroy();
+		if (explicit.kind === "lower-bound") {
+			if (explicit.usd !== null) throw new Error("lower-bound cost must have usd: null");
+			if (!Number.isFinite(explicit.knownLowerBoundUsd) || explicit.knownLowerBoundUsd <= 0) throw new Error("lower-bound cost must have a finite, positive knownLowerBoundUsd");
+			return {
+				costUsd: null,
+				costProvenance: explicit
+			};
+		}
+		if (!Number.isFinite(explicit.usd) || explicit.usd < 0) throw new Error(`${explicit.kind} cost must be a finite, non-negative USD amount`);
+		return {
+			costUsd: explicit.usd,
+			costProvenance: explicit
+		};
 	}
-	const last = readCodeAgentJsonlLine(pending, lineNumber + 1);
-	if (last) yield last;
-}
-/** Streaming counterpart to `parseCodeAgentJsonl` for a transcript on disk.
-*  It returns the same shape, so a caller that holds every entry keeps working
-*  above the string ceiling. The entry array still grows with the transcript;
-*  consume `streamCodeAgentJsonlFile` directly when memory must stay flat. */
-async function parseCodeAgentJsonlFile(path) {
-	const entries = [];
-	let malformedLines = 0;
-	for await (const read of streamCodeAgentJsonlFile(path)) if (read.kind === "malformed") malformedLines += 1;
-	else entries.push(read.entry);
-	return {
-		entries,
-		malformedLines
-	};
-}
-function fromCodexSession(options) {
-	return fromCodeAgentSession("codex", options);
-}
-function fromClaudeCodeSession(options) {
-	return fromCodeAgentSession("claude-code", options);
-}
-function fromOpenCodeSession(options) {
-	return fromCodeAgentSession("opencode", options);
-}
-function fromKimiCodeSession(options) {
-	return fromCodeAgentSession("kimi-code", options);
-}
-/** A native Pi session (`~/.pi/agent/sessions/<cwd>/<ts>_<id>.jsonl`), parsed entries in order. */
-function fromPiSession(options) {
-	return fromCodeAgentSession("pi", options);
-}
-/** The graph IR a Pi graph run exports (`nodes[].ir`), which is not a Pi session. */
-function fromPiGraphSession(options) {
-	return fromCodeAgentSession("pi-graph", options);
-}
-function fromCodeAgentSession(source, options) {
-	const entries = options.entries.filter(isRecord);
-	if (entries.length === 0) return {
-		runs: [],
-		diagnostics: [{
-			source,
-			sessionId: fallbackSessionId(source, options.sourcePath),
-			sourcePath: options.sourcePath,
-			entries: 0,
-			malformedLines: options.malformedLines ?? 0,
-			hasExplicitTerminalSignal: false,
-			hasFinalOutput: false,
-			hasQualityLabel: false,
-			hasTokenUsage: false,
-			hasCost: false,
-			costKind: "uncaptured",
-			warnings: ["no parseable session entries"]
-		}],
-		metrics: [],
-		observations: []
-	};
-	const metrics = metricsFor(source, entries);
-	const sessionId = sessionIdFromEntries(source, entries) ?? fallbackSessionId(source, options.sourcePath);
-	const observation = observeCodeAgentSession({
-		source,
-		entries,
-		sourcePath: options.sourcePath,
-		execution: options.execution
-	});
-	applyObservedSurfaceMetrics(metrics, observation);
-	const model = withSnapshot(options.model ?? modelFromEntries(source, entries) ?? source);
-	const tokenUsage = {
-		input: metrics.inputTokens,
-		output: metrics.outputTokens,
-		...metrics.reasoningTokens > 0 ? { reasoning: metrics.reasoningTokens } : {},
-		...metrics.cachedTokens > 0 ? { cached: metrics.cachedTokens } : {},
-		...metrics.cacheWriteTokens > 0 ? { cacheWrite: metrics.cacheWriteTokens } : {}
-	};
-	const { costUsd, costProvenance } = resolveSessionCost(options.costProvenance, metrics, model);
-	const qualityScore = finiteScore(options.score);
-	const invalidQualityScore = options.score !== void 0 && qualityScore === void 0;
-	const promptHash = options.promptHash ?? hashString(`prompt:${source}:${firstUserText(entries) ?? ""}`);
-	const configHash = options.configHash ?? hashJson({
-		source,
-		model,
-		sourcePath: options.sourcePath ?? null,
-		cwd: cwdFromEntries(entries) ?? null,
-		entryCount: entries.length
-	});
-	const explicitTerminal = observation.terminal.explicit;
-	const warnings = diagnosticsFor(metrics, {
-		model,
-		explicitTerminal,
-		malformedLines: options.malformedLines ?? 0,
-		hasQualityLabel: qualityScore !== void 0,
-		invalidQualityScore,
-		costKind: costProvenance.kind
-	});
-	const terminalOutcome = terminalOutcomeFromSession(observation.terminal);
-	const processExitCode = options.execution?.exitCode;
-	const splitTag = options.splitTag ?? "holdout";
-	return {
-		runs: [{
-			runId: `${source}:${sessionId}`,
-			experimentId: options.experimentId ?? `${source}-local-sessions`,
-			candidateId: options.candidateId ?? model,
-			seed: options.seed ?? stableSeed(sessionId),
-			model,
-			promptHash,
-			configHash,
-			commitSha: options.commitSha ?? "local-session",
-			wallMs: metrics.wallMs,
+	if (metrics.observedCostCaptured) {
+		const costProvenance = {
+			kind: "observed",
+			usd: metrics.observedCostUsd
+		};
+		return {
+			costUsd: metrics.observedCostUsd,
+			costProvenance
+		};
+	}
+	if ((metrics.inputTokens > 0 || metrics.outputTokens > 0) && isModelPriced(model)) {
+		const costUsd = estimateCost(metrics.inputTokens, metrics.outputTokens, model);
+		return {
 			costUsd,
-			costProvenance,
-			tokenUsage,
-			terminalOutcome,
-			...terminalOutcome === "failed" ? { terminalFailureReason: processExitCode !== void 0 && processExitCode !== 0 ? `process exited with code ${processExitCode}` : metrics.turnsAborted > 0 ? "turn aborted" : "session reported terminal failure" } : {},
-			outcome: {
-				...qualityScore === void 0 ? {} : splitTag === "holdout" ? { holdoutScore: qualityScore } : { searchScore: qualityScore },
-				raw: {
-					entries: metrics.entries,
-					user_messages: metrics.userMessages,
-					assistant_messages: metrics.assistantMessages,
-					reasoning_items: metrics.reasoningItems,
-					tool_calls: metrics.toolCalls,
-					tool_outputs: metrics.toolOutputs,
-					tool_errors: metrics.toolErrors,
-					execution_error_count: metrics.toolErrors,
-					process_error_count: terminalOutcome === "failed" ? 1 : 0,
-					unclassified_error_count: metrics.unclassifiedErrors,
-					...processExitCode !== void 0 ? { process_exit_code: processExitCode } : {},
-					patch_attempts: metrics.patchAttempts,
-					patch_successes: metrics.patchSuccesses,
-					patch_failures: metrics.patchFailures,
-					turns_started: metrics.turnsStarted,
-					turns_completed: metrics.turnsCompleted,
-					turns_aborted: metrics.turnsAborted,
-					context_compactions: metrics.contextCompactions,
-					mcp_calls: metrics.mcpCalls,
-					subagent_calls: metrics.subagentCalls,
-					skill_calls: metrics.skillCalls,
-					hook_calls: metrics.hookCalls,
-					web_calls: metrics.webCalls,
-					code_actions: metrics.codeActions,
-					pr_links: metrics.prLinks,
-					file_snapshots: metrics.fileSnapshots,
-					graph_nodes: metrics.graphNodes,
-					graph_edges: metrics.graphEdges,
-					action_candidates: metrics.actionCandidates,
-					verification_reports: metrics.verificationReports,
-					completion_decisions: metrics.completionDecisions,
-					reliability_rows: metrics.reliabilityRows,
-					reliability_lift: metrics.reliabilityLift,
-					input_tokens: metrics.inputTokens,
-					output_tokens: metrics.outputTokens,
-					reasoning_tokens: metrics.reasoningTokens,
-					cached_tokens: metrics.cachedTokens,
-					cache_write_tokens: metrics.cacheWriteTokens,
-					observed_cost_usd: metrics.observedCostUsd,
-					observed_cost_captured: metrics.observedCostCaptured ? 1 : 0,
-					estimated_cost_usd: costProvenance.kind === "estimated" ? costProvenance.usd : 0,
-					process_score: metrics.processScore,
-					explicit_terminal_signal: explicitTerminal ? 1 : 0,
-					quality_label_present: qualityScore !== void 0 ? 1 : 0,
-					cost_observed: costProvenance.kind === "observed" ? 1 : 0,
-					cost_estimated: costProvenance.kind === "estimated" ? 1 : 0,
-					cost_uncaptured: costProvenance.kind === "uncaptured" ? 1 : 0,
-					cost_unknown: costProvenance.usd === null ? 1 : 0
-				}
-			},
-			splitTag,
-			scenarioId: options.scenarioId ?? sessionId
-		}],
-		diagnostics: [{
-			source,
-			sessionId,
-			sourcePath: options.sourcePath,
-			entries: entries.length,
-			malformedLines: options.malformedLines ?? 0,
-			hasExplicitTerminalSignal: explicitTerminal,
-			hasFinalOutput: observation.finalText !== null,
-			hasQualityLabel: qualityScore !== void 0,
-			hasTokenUsage: metrics.inputTokens > 0 || metrics.outputTokens > 0 || metrics.reasoningTokens > 0 || metrics.cachedTokens > 0 || metrics.cacheWriteTokens > 0,
-			hasCost: costProvenance.usd !== null,
-			costKind: costProvenance.kind,
-			warnings
-		}],
-		metrics: [metrics],
-		observations: [{
-			...observation,
-			sessionId
-		}]
-	};
-}
-function terminalOutcomeFromSession(terminal) {
-	if (!terminal.explicit) return "unknown";
-	if (terminal.status === "completed") return "succeeded";
-	if (terminal.status === "failed") return "failed";
-	return "unknown";
-}
-function metricsFor(source, entries) {
-	switch (source) {
-		case "codex": return codexMetrics(entries);
-		case "claude-code": return claudeCodeMetrics(entries);
-		case "opencode": return openCodeMetrics(entries);
-		case "kimi-code": return kimiCodeMetrics(entries);
-		case "pi": return piSessionMetrics(entries);
-		case "pi-graph": return piGraphMetrics(entries);
+			costProvenance: {
+				kind: "estimated",
+				usd: costUsd
+			}
+		};
 	}
+	return {
+		costUsd: null,
+		costProvenance: {
+			kind: "uncaptured",
+			usd: null
+		}
+	};
 }
+//#endregion
+//#region src/contract/intake/code-agent-codex-metrics.ts
 function codexMetrics(entries) {
 	const metrics = emptyMetrics(entries.length);
 	const startedToolIds = /* @__PURE__ */ new Set();
@@ -2408,7 +2134,7 @@ function codexMetrics(entries) {
 		if (entryType === "turn.started") metrics.turnsStarted += 1;
 		if (entryType === "turn.completed") {
 			metrics.turnsCompleted += 1;
-			addCodexExecUsage(metrics, record(entry.usage));
+			addUsage(metrics, record(entry.usage));
 		}
 		if (entryType === "turn.failed") metrics.turnsAborted += 1;
 		if (entryType === "error") metrics.unclassifiedErrors += 1;
@@ -2496,14 +2222,6 @@ function codexExecToolFailed(itemType, item) {
 	if (itemType === "mcp_tool_call") return status === "failed" || record(item.error) !== null;
 	return itemType === "collab_tool_call" && status === "failed";
 }
-function addCodexExecUsage(metrics, usage) {
-	const parsed = readUsage(usage);
-	metrics.inputTokens += parsed.input;
-	metrics.outputTokens += parsed.output;
-	metrics.reasoningTokens += parsed.reasoning;
-	metrics.cachedTokens += parsed.cached;
-	metrics.cacheWriteTokens += parsed.cacheWrite;
-}
 function setCodexCumulativeUsage(metrics, usage) {
 	const parsed = readUsage(usage);
 	metrics.inputTokens = Math.max(metrics.inputTokens, parsed.input);
@@ -2512,6 +2230,12 @@ function setCodexCumulativeUsage(metrics, usage) {
 	metrics.cachedTokens = Math.max(metrics.cachedTokens, parsed.cached);
 	metrics.cacheWriteTokens = Math.max(metrics.cacheWriteTokens, parsed.cacheWrite);
 }
+function codexProcessScore(metrics) {
+	const terminalTurns = metrics.turnsCompleted + metrics.turnsAborted;
+	return penalizeErrors(terminalTurns > 0 ? metrics.turnsCompleted / terminalTurns : metrics.assistantMessages > 0 ? .5 : 0, metrics);
+}
+//#endregion
+//#region src/contract/intake/code-agent-harness-metrics.ts
 function claudeCodeMetrics(entries) {
 	const metrics = emptyMetrics(entries.length);
 	let startedAt;
@@ -2708,170 +2432,302 @@ function piGraphMetrics(entries) {
 				metrics.turnsCompleted += 1;
 			}
 		}
-		const averageReliability = numberField(entry, "averageNodeReliability");
-		if (averageReliability !== void 0) bestReliabilityScore = maxOptional(bestReliabilityScore, averageReliability);
-		const pessimisticPath = numberField(entry, "pessimisticPathEstimate");
-		if (pessimisticPath !== void 0) bestReliabilityScore = maxOptional(bestReliabilityScore, pessimisticPath);
-		const rows = Array.isArray(entry.rows) ? entry.rows : [];
-		metrics.reliabilityRows += rows.length;
-		for (const row of rows) {
-			const obj = record(row);
-			if (!obj) continue;
-			const lift = numberField(obj, "lift");
-			if (lift !== void 0) metrics.reliabilityLift = Math.max(metrics.reliabilityLift, lift);
-			const validated = numberField(obj, "validatedSuccessEstimate");
-			if (validated !== void 0) bestReliabilityScore = maxOptional(bestReliabilityScore, validated);
-		}
+		const averageReliability = numberField(entry, "averageNodeReliability");
+		if (averageReliability !== void 0) bestReliabilityScore = maxOptional(bestReliabilityScore, averageReliability);
+		const pessimisticPath = numberField(entry, "pessimisticPathEstimate");
+		if (pessimisticPath !== void 0) bestReliabilityScore = maxOptional(bestReliabilityScore, pessimisticPath);
+		const rows = Array.isArray(entry.rows) ? entry.rows : [];
+		metrics.reliabilityRows += rows.length;
+		for (const row of rows) {
+			const obj = record(row);
+			if (!obj) continue;
+			const lift = numberField(obj, "lift");
+			if (lift !== void 0) metrics.reliabilityLift = Math.max(metrics.reliabilityLift, lift);
+			const validated = numberField(obj, "validatedSuccessEstimate");
+			if (validated !== void 0) bestReliabilityScore = maxOptional(bestReliabilityScore, validated);
+		}
+	}
+	metrics.processScore = bestReliabilityScore !== void 0 ? clamp01(bestReliabilityScore) : metrics.completionDecisions > 0 ? 1 : 0;
+	return metrics;
+}
+function claudeProcessScore(metrics) {
+	return penalizeErrors(metrics.prLinks > 0 ? 1 : metrics.assistantMessages > 0 ? .75 : 0, metrics);
+}
+function terminalProcessScore(metrics) {
+	const terminalTurns = metrics.turnsCompleted + metrics.turnsAborted;
+	return penalizeErrors(terminalTurns > 0 ? metrics.turnsCompleted / terminalTurns : metrics.assistantMessages > 0 ? .75 : metrics.toolCalls > 0 ? .5 : 0, metrics);
+}
+function maxOptional(current, next) {
+	return current === void 0 ? next : Math.max(current, next);
+}
+//#endregion
+//#region src/contract/intake/code-agent-jsonl.ts
+/** The single per-line rule. Both the string path and the file path call this,
+*  so malformed-line handling and entry validation cannot drift apart. */
+function readCodeAgentJsonlLine(line, lineNumber) {
+	const trimmed = line.trim();
+	if (!trimmed) return void 0;
+	try {
+		return {
+			kind: "entry",
+			lineNumber,
+			entry: JSON.parse(trimmed)
+		};
+	} catch {
+		return {
+			kind: "malformed",
+			lineNumber
+		};
+	}
+}
+function parseCodeAgentJsonl(jsonl) {
+	const entries = [];
+	let malformedLines = 0;
+	let lineNumber = 0;
+	for (const line of jsonl.split("\n")) {
+		lineNumber += 1;
+		const read = readCodeAgentJsonlLine(line, lineNumber);
+		if (!read) continue;
+		if (read.kind === "malformed") malformedLines += 1;
+		else entries.push(read.entry);
+	}
+	return {
+		entries,
+		malformedLines
+	};
+}
+/** Reads a transcript one line at a time and never holds the file as a single
+*  string. `parseCodeAgentJsonl` needs the whole file in one string, so a
+*  session above V8's ~512MB string ceiling throws `ERR_STRING_TOO_LONG` and
+*  cannot be ingested at all; the largest real Codex rollout on record is 695MB.
+*
+*  Lines break on `\n` only, which is what the string path's `split('\n')` does.
+*  `node:readline` also breaks on a bare `\r`, so it is deliberately not used
+*  here: a lone carriage return inside a line must stay inside that line for the
+*  two paths to report the same malformed count. */
+async function* streamCodeAgentJsonlFile(path) {
+	const stream = createReadStream(path, { encoding: "utf8" });
+	let pending = "";
+	let lineNumber = 0;
+	try {
+		for await (const chunk of stream) {
+			pending += chunk;
+			let start = 0;
+			for (let at = pending.indexOf("\n"); at !== -1; at = pending.indexOf("\n", start)) {
+				lineNumber += 1;
+				const read = readCodeAgentJsonlLine(pending.slice(start, at), lineNumber);
+				if (read) yield read;
+				start = at + 1;
+			}
+			pending = pending.slice(start);
+		}
+	} finally {
+		stream.destroy();
 	}
-	metrics.processScore = bestReliabilityScore !== void 0 ? clamp01(bestReliabilityScore) : metrics.completionDecisions > 0 ? 1 : 0;
-	return metrics;
+	const last = readCodeAgentJsonlLine(pending, lineNumber + 1);
+	if (last) yield last;
 }
-function emptyMetrics(entries) {
+/** Streaming counterpart to `parseCodeAgentJsonl` for a transcript on disk.
+*  It returns the same shape, so a caller that holds every entry keeps working
+*  above the string ceiling. The entry array still grows with the transcript;
+*  consume `streamCodeAgentJsonlFile` directly when memory must stay flat. */
+async function parseCodeAgentJsonlFile(path) {
+	const entries = [];
+	let malformedLines = 0;
+	for await (const read of streamCodeAgentJsonlFile(path)) if (read.kind === "malformed") malformedLines += 1;
+	else entries.push(read.entry);
 	return {
 		entries,
-		userMessages: 0,
-		assistantMessages: 0,
-		reasoningItems: 0,
-		toolCalls: 0,
-		toolOutputs: 0,
-		toolErrors: 0,
-		unclassifiedErrors: 0,
-		patchAttempts: 0,
-		patchSuccesses: 0,
-		patchFailures: 0,
-		turnsStarted: 0,
-		turnsCompleted: 0,
-		turnsAborted: 0,
-		contextCompactions: 0,
-		mcpCalls: 0,
-		subagentCalls: 0,
-		skillCalls: 0,
-		hookCalls: 0,
-		webCalls: 0,
-		codeActions: 0,
-		prLinks: 0,
-		fileSnapshots: 0,
-		graphNodes: 0,
-		graphEdges: 0,
-		actionCandidates: 0,
-		verificationReports: 0,
-		completionDecisions: 0,
-		reliabilityRows: 0,
-		reliabilityLift: 0,
-		inputTokens: 0,
-		outputTokens: 0,
-		reasoningTokens: 0,
-		cachedTokens: 0,
-		cacheWriteTokens: 0,
-		observedCostUsd: 0,
-		observedCostCaptured: false,
-		wallMs: 0,
-		processScore: 0
+		malformedLines
 	};
 }
-function codexProcessScore(metrics) {
-	const terminalTurns = metrics.turnsCompleted + metrics.turnsAborted;
-	return penalizeErrors(terminalTurns > 0 ? metrics.turnsCompleted / terminalTurns : metrics.assistantMessages > 0 ? .5 : 0, metrics);
+//#endregion
+//#region src/contract/intake/code-agent-session.ts
+function fromCodexSession(options) {
+	return fromCodeAgentSession("codex", options);
 }
-function claudeProcessScore(metrics) {
-	return penalizeErrors(metrics.prLinks > 0 ? 1 : metrics.assistantMessages > 0 ? .75 : 0, metrics);
+function fromClaudeCodeSession(options) {
+	return fromCodeAgentSession("claude-code", options);
 }
-function terminalProcessScore(metrics) {
-	const terminalTurns = metrics.turnsCompleted + metrics.turnsAborted;
-	return penalizeErrors(terminalTurns > 0 ? metrics.turnsCompleted / terminalTurns : metrics.assistantMessages > 0 ? .75 : metrics.toolCalls > 0 ? .5 : 0, metrics);
+function fromOpenCodeSession(options) {
+	return fromCodeAgentSession("opencode", options);
 }
-function penalizeErrors(base, metrics) {
-	const operations = Math.max(1, metrics.toolCalls + metrics.patchAttempts);
-	return clamp01(base * (1 - .5 * Math.min(1, (metrics.toolErrors + metrics.patchFailures) / operations)));
+function fromKimiCodeSession(options) {
+	return fromCodeAgentSession("kimi-code", options);
 }
-function emptyTokenTotals() {
+/** A native Pi session (`~/.pi/agent/sessions/<cwd>/<ts>_<id>.jsonl`), parsed entries in order. */
+function fromPiSession(options) {
+	return fromCodeAgentSession("pi", options);
+}
+function fromCodeAgentSession(source, options) {
+	const entries = options.entries.filter(isRecord);
+	if (entries.length === 0) return {
+		runs: [],
+		diagnostics: [{
+			source,
+			sessionId: fallbackSessionId$1(source, options.sourcePath),
+			sourcePath: options.sourcePath,
+			entries: 0,
+			malformedLines: options.malformedLines ?? 0,
+			hasExplicitTerminalSignal: false,
+			hasFinalOutput: false,
+			hasQualityLabel: false,
+			hasTokenUsage: false,
+			hasCost: false,
+			costKind: "uncaptured",
+			warnings: ["no parseable session entries"]
+		}],
+		metrics: [],
+		observations: []
+	};
+	const metrics = metricsFor(source, entries);
+	const sessionId = sessionIdFromEntries(source, entries) ?? fallbackSessionId$1(source, options.sourcePath);
+	const observation = observeCodeAgentSession({
+		source,
+		entries,
+		sourcePath: options.sourcePath,
+		execution: options.execution
+	});
+	applyObservedSurfaceMetrics(metrics, observation);
+	const model = withSnapshot(options.model ?? modelFromEntries(source, entries) ?? source);
+	const tokenUsage = {
+		input: metrics.inputTokens,
+		output: metrics.outputTokens,
+		...metrics.reasoningTokens > 0 ? { reasoning: metrics.reasoningTokens } : {},
+		...metrics.cachedTokens > 0 ? { cached: metrics.cachedTokens } : {},
+		...metrics.cacheWriteTokens > 0 ? { cacheWrite: metrics.cacheWriteTokens } : {}
+	};
+	const { costUsd, costProvenance } = resolveSessionCost(options.costProvenance, metrics, model);
+	const qualityScore = finiteScore(options.score);
+	const invalidQualityScore = options.score !== void 0 && qualityScore === void 0;
+	const promptHash = options.promptHash ?? hashString(`prompt:${source}:${firstUserText(entries) ?? ""}`);
+	const configHash = options.configHash ?? hashCanonical({
+		source,
+		model,
+		sourcePath: options.sourcePath ?? null,
+		cwd: cwdFromEntries(entries) ?? null,
+		entryCount: entries.length
+	});
+	const explicitTerminal = observation.terminal.explicit;
+	const warnings = diagnosticsFor(metrics, {
+		model,
+		explicitTerminal,
+		malformedLines: options.malformedLines ?? 0,
+		hasQualityLabel: qualityScore !== void 0,
+		invalidQualityScore,
+		costKind: costProvenance.kind
+	});
+	const terminalOutcome = terminalOutcomeFromSession(observation.terminal);
+	const processExitCode = options.execution?.exitCode;
+	const splitTag = options.splitTag ?? "holdout";
 	return {
-		input: 0,
-		output: 0,
-		reasoning: 0,
-		cached: 0,
-		cacheWrite: 0
+		runs: [{
+			runId: `${source}:${sessionId}`,
+			experimentId: options.experimentId ?? `${source}-local-sessions`,
+			candidateId: options.candidateId ?? model,
+			seed: options.seed ?? stableSeed(sessionId),
+			model,
+			promptHash,
+			configHash,
+			commitSha: options.commitSha ?? "local-session",
+			wallMs: metrics.wallMs,
+			costUsd,
+			costProvenance,
+			tokenUsage,
+			terminalOutcome,
+			...terminalOutcome === "failed" ? { terminalFailureReason: processExitCode !== void 0 && processExitCode !== 0 ? `process exited with code ${processExitCode}` : metrics.turnsAborted > 0 ? "turn aborted" : "session reported terminal failure" } : {},
+			outcome: {
+				...qualityScore === void 0 ? {} : splitTag === "holdout" ? { holdoutScore: qualityScore } : { searchScore: qualityScore },
+				raw: {
+					entries: metrics.entries,
+					user_messages: metrics.userMessages,
+					assistant_messages: metrics.assistantMessages,
+					reasoning_items: metrics.reasoningItems,
+					tool_calls: metrics.toolCalls,
+					tool_outputs: metrics.toolOutputs,
+					tool_errors: metrics.toolErrors,
+					execution_error_count: metrics.toolErrors,
+					process_error_count: terminalOutcome === "failed" ? 1 : 0,
+					unclassified_error_count: metrics.unclassifiedErrors,
+					...processExitCode !== void 0 ? { process_exit_code: processExitCode } : {},
+					patch_attempts: metrics.patchAttempts,
+					patch_successes: metrics.patchSuccesses,
+					patch_failures: metrics.patchFailures,
+					turns_started: metrics.turnsStarted,
+					turns_completed: metrics.turnsCompleted,
+					turns_aborted: metrics.turnsAborted,
+					context_compactions: metrics.contextCompactions,
+					mcp_calls: metrics.mcpCalls,
+					subagent_calls: metrics.subagentCalls,
+					skill_calls: metrics.skillCalls,
+					hook_calls: metrics.hookCalls,
+					web_calls: metrics.webCalls,
+					code_actions: metrics.codeActions,
+					pr_links: metrics.prLinks,
+					file_snapshots: metrics.fileSnapshots,
+					graph_nodes: metrics.graphNodes,
+					graph_edges: metrics.graphEdges,
+					action_candidates: metrics.actionCandidates,
+					verification_reports: metrics.verificationReports,
+					completion_decisions: metrics.completionDecisions,
+					reliability_rows: metrics.reliabilityRows,
+					reliability_lift: metrics.reliabilityLift,
+					input_tokens: metrics.inputTokens,
+					output_tokens: metrics.outputTokens,
+					reasoning_tokens: metrics.reasoningTokens,
+					cached_tokens: metrics.cachedTokens,
+					cache_write_tokens: metrics.cacheWriteTokens,
+					observed_cost_usd: metrics.observedCostUsd,
+					observed_cost_captured: metrics.observedCostCaptured ? 1 : 0,
+					estimated_cost_usd: costProvenance.kind === "estimated" ? costProvenance.usd : 0,
+					process_score: metrics.processScore,
+					explicit_terminal_signal: explicitTerminal ? 1 : 0,
+					quality_label_present: qualityScore !== void 0 ? 1 : 0,
+					cost_observed: costProvenance.kind === "observed" ? 1 : 0,
+					cost_estimated: costProvenance.kind === "estimated" ? 1 : 0,
+					cost_uncaptured: costProvenance.kind === "uncaptured" ? 1 : 0,
+					cost_unknown: costProvenance.usd === null ? 1 : 0
+				}
+			},
+			splitTag,
+			scenarioId: options.scenarioId ?? sessionId
+		}],
+		diagnostics: [{
+			source,
+			sessionId,
+			sourcePath: options.sourcePath,
+			entries: entries.length,
+			malformedLines: options.malformedLines ?? 0,
+			hasExplicitTerminalSignal: explicitTerminal,
+			hasFinalOutput: observation.finalText !== null,
+			hasQualityLabel: qualityScore !== void 0,
+			hasTokenUsage: metrics.inputTokens > 0 || metrics.outputTokens > 0 || metrics.reasoningTokens > 0 || metrics.cachedTokens > 0 || metrics.cacheWriteTokens > 0,
+			hasCost: costProvenance.usd !== null,
+			costKind: costProvenance.kind,
+			warnings
+		}],
+		metrics: [metrics],
+		observations: [{
+			...observation,
+			sessionId
+		}]
 	};
 }
-function hasTokenTotals(usage) {
-	return usage.input > 0 || usage.output > 0 || usage.reasoning > 0 || usage.cached > 0 || usage.cacheWrite > 0;
-}
-function addUsage(metrics, usage) {
-	const parsed = readUsage(usage);
-	metrics.inputTokens += parsed.input;
-	metrics.outputTokens += parsed.output;
-	metrics.reasoningTokens += parsed.reasoning;
-	metrics.cachedTokens += parsed.cached;
-	metrics.cacheWriteTokens += parsed.cacheWrite;
-}
-function addUsageTo(totals, usage) {
-	const parsed = readUsage(usage);
-	totals.input += parsed.input;
-	totals.output += parsed.output;
-	totals.reasoning += parsed.reasoning;
-	totals.cached += parsed.cached;
-	totals.cacheWrite += parsed.cacheWrite;
-}
-function readUsage(usage) {
-	const parsed = extractUsage(usage);
-	return {
-		input: parsed?.input ?? 0,
-		output: parsed?.output ?? 0,
-		reasoning: parsed?.reasoning ?? 0,
-		cached: parsed?.cached ?? 0,
-		cacheWrite: parsed?.cacheWrite ?? 0
-	};
+function terminalOutcomeFromSession(terminal) {
+	if (!terminal.explicit) return "unknown";
+	if (terminal.status === "completed") return "succeeded";
+	if (terminal.status === "failed") return "failed";
+	return "unknown";
 }
-function resolveSessionCost(explicit, metrics, model) {
-	if (explicit) {
-		if (explicit.kind === "uncaptured") {
-			if (explicit.usd !== null) throw new Error("uncaptured cost must have usd: null");
-			return {
-				costUsd: null,
-				costProvenance: explicit
-			};
-		}
-		if (explicit.kind === "lower-bound") {
-			if (explicit.usd !== null) throw new Error("lower-bound cost must have usd: null");
-			if (!Number.isFinite(explicit.knownLowerBoundUsd) || explicit.knownLowerBoundUsd <= 0) throw new Error("lower-bound cost must have a finite, positive knownLowerBoundUsd");
-			return {
-				costUsd: null,
-				costProvenance: explicit
-			};
-		}
-		if (!Number.isFinite(explicit.usd) || explicit.usd < 0) throw new Error(`${explicit.kind} cost must be a finite, non-negative USD amount`);
-		return {
-			costUsd: explicit.usd,
-			costProvenance: explicit
-		};
-	}
-	if (metrics.observedCostCaptured) {
-		const costProvenance = {
-			kind: "observed",
-			usd: metrics.observedCostUsd
-		};
-		return {
-			costUsd: metrics.observedCostUsd,
-			costProvenance
-		};
-	}
-	if ((metrics.inputTokens > 0 || metrics.outputTokens > 0) && isModelPriced(model)) {
-		const costUsd = estimateCost(metrics.inputTokens, metrics.outputTokens, model);
-		return {
-			costUsd,
-			costProvenance: {
-				kind: "estimated",
-				usd: costUsd
-			}
-		};
+function metricsFor(source, entries) {
+	switch (source) {
+		case "codex": return codexMetrics(entries);
+		case "claude-code": return claudeCodeMetrics(entries);
+		case "opencode": return openCodeMetrics(entries);
+		case "kimi-code": return kimiCodeMetrics(entries);
+		case "pi": return piSessionMetrics(entries);
+		case "pi-graph": return piGraphMetrics(entries);
 	}
-	return {
-		costUsd: null,
-		costProvenance: {
-			kind: "uncaptured",
-			usd: null
-		}
-	};
 }
 function diagnosticsFor(metrics, options) {
 	const warnings = [];
@@ -2895,58 +2751,6 @@ function applyObservedSurfaceMetrics(metrics, observation) {
 	metrics.webCalls = actions.filter((action) => action.surface === "web").length;
 	metrics.codeActions = actions.filter((action) => action.surface === "code").length;
 }
-function sessionIdFromEntries(source, entries) {
-	for (const entry of entries) {
-		if (source === "pi" && entry.type === "session") {
-			const id = stringField(entry, "id");
-			if (id) return id;
-		}
-		if (source === "codex") {
-			const threadId = stringField(entry, "thread_id");
-			if (threadId) return threadId;
-			const payload = record(entry.payload);
-			const id = payload ? stringField(payload, "id") : void 0;
-			if (id) return id;
-		}
-		const payload = record(record(entry.message)?.payload);
-		const sessionId = stringField(entry, "sessionID") ?? stringField(entry, "sessionId") ?? stringField(entry, "session_id") ?? stringField(payload ?? {}, "session_id");
-		if (sessionId) return sessionId;
-	}
-}
-function modelFromEntries(source, entries) {
-	if (source === "pi") return summarizeSessionInput("pi", { records: entries }).lastServedModel ?? void 0;
-	let providerModel;
-	for (const entry of entries) {
-		if (source === "codex") {
-			const payload = record(entry.payload);
-			const model = payload ? stringField(payload, "model") : void 0;
-			if (model) return model;
-			providerModel = providerModel ?? (payload ? stringField(payload, "model_provider") : void 0);
-		}
-		const openCodeModel = stringField(entry, "modelID");
-		if (openCodeModel) {
-			const provider = stringField(entry, "providerID");
-			return provider ? `${provider}/${openCodeModel}` : openCodeModel;
-		}
-		const modelObject = record(entry.model);
-		const nestedModel = stringField(modelObject ?? {}, "modelID");
-		if (nestedModel) {
-			const provider = stringField(modelObject ?? {}, "providerID");
-			return provider ? `${provider}/${nestedModel}` : nestedModel;
-		}
-		const message = record(entry.message);
-		const model = message ? stringField(message, "model") : void 0;
-		if (model) return model;
-	}
-	return providerModel;
-}
-function cwdFromEntries(entries) {
-	for (const entry of entries) {
-		const path = record(entry.path);
-		const cwd = stringField(entry, "cwd") ?? stringField(path ?? {}, "cwd") ?? stringField(path ?? {}, "root") ?? (record(entry.payload) ? stringField(record(entry.payload), "cwd") : void 0);
-		if (cwd) return cwd;
-	}
-}
 /** Join the `text` of every part in a `[{ type: 'text', text }]` content list. */
 function textFromContentParts(content) {
 	if (!Array.isArray(content)) return void 0;
@@ -2981,245 +2785,9 @@ function firstUserText(entries) {
 		}
 	}
 }
-function fallbackSessionId(source, sourcePath) {
-	return hashString(`${source}:${sourcePath ?? "unknown-session"}`).slice(7, 27);
-}
-function withSnapshot(model) {
-	const trimmed = model.trim() || "unknown";
-	if (trimmed.includes("@") || /-\d{8}$/.test(trimmed) || /-\d{4}-\d{2}-\d{2}$/.test(trimmed)) return trimmed;
-	return `${trimmed}@observed-local`;
-}
-function hashJson(value) {
-	return hashCanonical(value);
-}
-function hashString(value) {
-	return `sha256:${createHash("sha256").update(value).digest("hex")}`;
-}
-function stableSeed(value) {
-	return createHash("sha256").update(value).digest().readUInt32BE(0);
-}
-function maxOptional(current, next) {
-	return current === void 0 ? next : Math.max(current, next);
-}
-function clamp01(value) {
-	if (!Number.isFinite(value)) return 0;
-	return Math.max(0, Math.min(1, value));
-}
 function finiteScore(value) {
 	return value !== void 0 && Number.isFinite(value) ? clamp01(value) : void 0;
 }
-function isRecord(value) {
-	return value !== null && typeof value === "object" && !Array.isArray(value);
-}
-function record(value) {
-	return isRecord(value) ? value : null;
-}
-function stringField(obj, key) {
-	const value = obj[key];
-	return typeof value === "string" && value.length > 0 ? value : void 0;
-}
-function numberField(obj, key) {
-	const value = obj[key];
-	return typeof value === "number" && Number.isFinite(value) ? value : void 0;
-}
-function timestampMs(value) {
-	if (typeof value === "number" && Number.isFinite(value)) return value > 0xe8d4a51000 ? value : value * 1e3;
-	if (typeof value === "string" && value.length > 0) {
-		const parsed = Date.parse(value);
-		return Number.isFinite(parsed) ? parsed : void 0;
-	}
-}
-//#endregion
-//#region src/contract/intake/code-agent-store.ts
-const SESSION_SUFFIX = ".jsonl";
-const REJECTION_SAMPLE_LIMIT = 8;
-/**
-* Read every session this run captured under `scope.root` and project them.
-*
-* A store directory that does not exist, or that holds no session the scope attributes, returns
-* `status: 'unavailable'`. That is not an error: a run whose harness store was never synced back
-* has no evidence, and saying so is the correct answer.
-*/
-async function observeCodeAgentStore(scope) {
-	if (!isAbsolute(scope.root)) throw new TypeError(`observeCodeAgentStore: scope.root must be an absolute path to this run's own store, received ${JSON.stringify(scope.root)}`);
-	const workspaceRoot = scope.workspaceRoot === void 0 ? void 0 : resolve(scope.workspaceRoot);
-	const paths = await listSessionFiles(resolve(scope.root));
-	const sessions = [];
-	const byReason = /* @__PURE__ */ new Map();
-	const sample = [];
-	const reject = (path, reason) => {
-		byReason.set(reason, (byReason.get(reason) ?? 0) + 1);
-		if (sample.length < REJECTION_SAMPLE_LIMIT) sample.push({
-			path,
-			reason
-		});
-	};
-	for (const path of paths) {
-		let entries;
-		try {
-			entries = (await parseCodeAgentJsonlFile(path)).entries;
-		} catch {
-			reject(path, "unreadable");
-			continue;
-		}
-		if (entries.length === 0) {
-			reject(path, "no-entries");
-			continue;
-		}
-		const workspace = sessionWorkspace(entries);
-		if (workspaceRoot !== void 0) {
-			if (workspace === void 0) {
-				reject(path, "no-workspace-recorded");
-				continue;
-			}
-			if (!isSameOrBelow(workspaceRoot, workspace)) {
-				reject(path, "outside-workspace");
-				continue;
-			}
-		}
-		const startedAtMs = sessionStartedAtMs(entries);
-		if (scope.window !== void 0) {
-			if (startedAtMs === void 0) {
-				reject(path, "no-start-recorded");
-				continue;
-			}
-			if (startedAtMs < scope.window.startedAtMs) {
-				reject(path, "before-run");
-				continue;
-			}
-			if (scope.window.finishedAtMs !== void 0 && startedAtMs > scope.window.finishedAtMs) {
-				reject(path, "after-run");
-				continue;
-			}
-		}
-		const observation = observeCodeAgentSession({
-			source: scope.source,
-			entries,
-			sourcePath: path
-		});
-		sessions.push({
-			path,
-			sessionId: observation.sessionId,
-			...workspace === void 0 ? {} : { workspace },
-			...startedAtMs === void 0 ? {} : { startedAtMs },
-			observation
-		});
-	}
-	const rejections = {
-		total: [...byReason.values()].reduce((sum, count) => sum + count, 0),
-		byReason: Object.fromEntries(byReason),
-		sample
-	};
-	if (sessions.length === 0) return {
-		status: "unavailable",
-		scope,
-		filesScanned: paths.length,
-		reason: unavailableReason(paths.length, rejections),
-		rejections
-	};
-	const bySurface = /* @__PURE__ */ new Map();
-	let actions = 0;
-	let subagentActions = 0;
-	let sessionsWithSubagentActions = 0;
-	for (const session of sessions) {
-		let sessionSubagents = 0;
-		for (const action of session.observation.actions) {
-			actions += 1;
-			bySurface.set(action.surface, (bySurface.get(action.surface) ?? 0) + 1);
-			if (action.surface === "subagent") sessionSubagents += 1;
-		}
-		subagentActions += sessionSubagents;
-		if (sessionSubagents > 0) sessionsWithSubagentActions += 1;
-	}
-	return {
-		status: "observed",
-		scope,
-		filesScanned: paths.length,
-		sessions,
-		actions,
-		subagentActions,
-		sessionsWithSubagentActions,
-		bySurface: Object.fromEntries(bySurface),
-		rejections
-	};
-}
-function unavailableReason(filesScanned, rejections) {
-	if (filesScanned === 0) return "the run captured no session files in this store";
-	return `no session in this store belongs to the run: ${filesScanned} file(s) scanned, all rejected (${Object.entries(rejections.byReason).map(([reason, count]) => `${count} ${reason}`).sort().join(", ")})`;
-}
-async function listSessionFiles(root) {
-	const found = [];
-	const walk = async (dir) => {
-		let entries;
-		try {
-			entries = await readdir(dir, { withFileTypes: true });
-		} catch {
-			return;
-		}
-		for (const entry of entries) {
-			const path = join(dir, entry.name);
-			if (entry.isDirectory()) await walk(path);
-			else if (entry.isFile() && entry.name.endsWith(SESSION_SUFFIX)) found.push(path);
-		}
-	};
-	await walk(root);
-	if (found.length === 0 && await isSessionFile(root)) found.push(root);
-	return found.sort();
-}
-async function isSessionFile(path) {
-	if (!path.endsWith(SESSION_SUFFIX)) return false;
-	try {
-		return (await stat(path)).isFile();
-	} catch {
-		return false;
-	}
-}
-/**
-* The working directory the session RECORDED, from the first entry that states one.
-*
-* codex writes it on `session_meta.payload.cwd`; claude-code writes `cwd` on every line. Neither is
-* inferred from the file's own location: where a captured file was copied to says nothing about
-* where the agent ran.
-*/
-function sessionWorkspace(entries) {
-	for (const entry of entries) {
-		const record = plainRecord(entry);
-		if (record === void 0) continue;
-		const direct = stringOf(record.cwd);
-		if (direct !== void 0) return direct;
-		const nested = stringOf(plainRecord(record.payload)?.cwd);
-		if (nested !== void 0) return nested;
-	}
-}
-/** When the session itself says it started, from the first entry carrying a parseable timestamp. */
-function sessionStartedAtMs(entries) {
-	for (const entry of entries) {
-		const record = plainRecord(entry);
-		if (record === void 0) continue;
-		const payload = plainRecord(record.payload);
-		for (const candidate of [payload?.timestamp, record.timestamp]) {
-			const text = stringOf(candidate);
-			if (text === void 0) continue;
-			const parsed = Date.parse(text);
-			if (Number.isFinite(parsed)) return parsed;
-		}
-		const epochMs = numberOf(record.timestampMs) ?? numberOf(payload?.timestampMs);
-		if (epochMs !== void 0) return epochMs;
-	}
-}
-function isSameOrBelow(root, candidate) {
-	const rel = relative(root, resolve(candidate));
-	return rel === "" || rel !== ".." && !rel.startsWith(`..${sep}`) && !isAbsolute(rel);
-}
-function plainRecord(value) {
-	return value !== null && typeof value === "object" && !Array.isArray(value) ? value : void 0;
-}
-function stringOf(value) {
-	return typeof value === "string" && value.length > 0 ? value : void 0;
-}
-function numberOf(value) {
-	return typeof value === "number" && Number.isFinite(value) ? value : void 0;
-}
 //#endregion
 //#region src/contract/intake/feedback-table.ts
 function fromFeedbackTable(opts) {
@@ -3300,6 +2868,6 @@ function fromFeedbackTable(opts) {
 	};
 }
 //#endregion
-export { FileSystemOutcomeStore, InMemoryOutcomeStore, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, SelfImproveRunError, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, createReferenceEquivalenceJudge, defaultProductionGate, defineAgentEval, evalReportingSuite, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiGraphSession, fromPiSession, fromRunRecordDir, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, makeProposalFinding, measuredComparisonFromAgentProfileImprovementExperiment, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, observeCodeAgentStore, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, parseCodeAgentJsonlFile, partitionRunsByAuthoringModel, runAgentProfileImprovementExperiment, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, runReferenceEquivalenceJudge, sealAgentProfileImprovementExperiment, sealAgentProfileImprovementSuite, sealAgentProfileImprovementTask, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, streamCodeAgentJsonlFile, summarizeExecution, transientDispatchFailure, verifyAgentProfileImprovementExperimentComparison, verifyCandidateExperiment, verifyCandidateExperimentComparison };
+export { FileSystemOutcomeStore, InMemoryOutcomeStore, SelfImproveRunError, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, defaultProductionGate, defineAgentEval, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiSession, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, makeProposalFinding, measuredComparisonFromAgentProfileImprovementExperiment, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, parseCodeAgentJsonlFile, partitionRunsByAuthoringModel, runAgentProfileImprovementExperiment, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, sealAgentProfileImprovementExperiment, sealAgentProfileImprovementSuite, sealAgentProfileImprovementTask, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, streamCodeAgentJsonlFile, summarizeExecution, transientDispatchFailure, verifyAgentProfileImprovementExperimentComparison, verifyCandidateExperiment, verifyCandidateExperimentComparison };
 
 //# sourceMappingURL=index.js.map