@tangle-network/agent-eval 0.128.2 → 0.130.0

package/dist/reward-hacking-qipEpKvY.js

@@ -0,0 +1,596 @@
+import { s as validateRunRecord } from "./run-record-BuoE80Dq.js";
+import { E as pearsonR } from "./statistics-CnnxdpOg.js";
+import { n as observedScore, s as trainingScore, t as isRealnessGated } from "./reward-nw2xZGZG.js";
+//#region src/campaign/run-record.ts
+/**
+* Project one campaign cell into the canonical run format.
+*
+* A dispatch error establishes terminal execution failure. A judge error only
+* establishes that quality measurement failed after dispatch completed.
+* Failures without a stage remain unknown. No failure becomes a zero-quality
+* label.
+*/
+function campaignCellToRunRecord(cell, options) {
+	const quality = projectCampaignCellQuality(cell);
+	const execution = campaignCellExecutionEvidence(cell);
+	const judgeErrorCount = Math.max(quality.raw.judge_error_count ?? 0, execution.judgeErrorCount ?? 0);
+	const cellCostCaptured = Number.isFinite(cell.costUsd) && cell.costUsd >= 0;
+	const costUsd = cellCostCaptured ? cell.costUsd : options.defaultCostUsd ?? null;
+	const costProvenance = costUsd === null ? {
+		kind: "uncaptured",
+		usd: null
+	} : cellCostCaptured && !cell.costEstimated ? {
+		kind: "observed",
+		usd: costUsd
+	} : {
+		kind: "estimated",
+		usd: costUsd
+	};
+	const raw = {
+		...finiteMetrics(options.raw),
+		...quality.raw,
+		rep: cell.rep,
+		duration_ms: cell.durationMs,
+		...costUsd === null ? {} : { cost_usd: costUsd },
+		cost_estimated: cell.costEstimated ? 1 : 0,
+		tokens_input: cell.tokenUsage.input,
+		tokens_output: cell.tokenUsage.output,
+		latency_ms: cell.durationMs,
+		...execution.executionErrorCount === void 0 ? {} : { execution_error_count: execution.executionErrorCount },
+		...judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {},
+		...execution.unclassifiedErrorCount === void 0 ? {} : { unclassified_error_count: execution.unclassifiedErrorCount }
+	};
+	if (typeof cell.generation === "number") raw.generation = cell.generation;
+	if (cell.tokenUsage.reasoning !== void 0) raw.tokens_reasoning = cell.tokenUsage.reasoning;
+	if (cell.tokenUsage.cached !== void 0) raw.tokens_cached = cell.tokenUsage.cached;
+	if (cell.tokenUsage.cacheWrite !== void 0) raw.tokens_cache_write = cell.tokenUsage.cacheWrite;
+	if (costUsd !== null && costUsd > 0) raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd;
+	if (costUsd !== null && quality.score !== void 0 && quality.score > .01) raw.cost_per_quality = costUsd / quality.score;
+	const outcome = {
+		raw,
+		...quality.judgeScores ? { judgeScores: quality.judgeScores } : {}
+	};
+	if (quality.score !== void 0) if (options.splitTag === "holdout") outcome.holdoutScore = quality.score;
+	else outcome.searchScore = quality.score;
+	return validateRunRecord({
+		runId: options.runId,
+		experimentId: options.experimentId,
+		candidateId: options.candidateId,
+		seed: options.seed ?? cell.seed,
+		model: options.model,
+		promptHash: options.promptHash,
+		configHash: options.configHash,
+		commitSha: options.commitSha,
+		wallMs: cell.durationMs,
+		costUsd,
+		costProvenance,
+		tokenUsage: { ...cell.tokenUsage },
+		terminalOutcome: execution.terminalOutcome,
+		...execution.terminalFailureReason ? { terminalFailureReason: execution.terminalFailureReason } : {},
+		outcome,
+		splitTag: options.splitTag,
+		scenarioId: options.scenarioId ?? cell.scenarioId,
+		...options.agentProfile ? { agentProfile: options.agentProfile } : {}
+	});
+}
+function campaignCellExecutionEvidence(cell) {
+	if (cell.errorStage === "dispatch") return {
+		terminalOutcome: "failed",
+		executionErrorCount: 1,
+		...cell.error ? { terminalFailureReason: cell.error } : {}
+	};
+	if (cell.errorStage === "judge") return {
+		terminalOutcome: "succeeded",
+		executionErrorCount: 0,
+		judgeErrorCount: 1
+	};
+	if (!cell.error) return {
+		terminalOutcome: "succeeded",
+		executionErrorCount: 0
+	};
+	return {
+		terminalOutcome: "unknown",
+		unclassifiedErrorCount: 1
+	};
+}
+/**
+* Produce the only task-quality view used by campaign aggregates and exports.
+*
+* Successful judge results remain available for diagnosis after another judge
+* fails, but a task score exists only for an error-free cell whose reported
+* judge values are all finite.
+*/
+function projectCampaignCellQuality(cell) {
+	if (cell.errorStage === "dispatch") return {
+		successfulJudgeScores: {},
+		failedJudges: [],
+		raw: {}
+	};
+	const perJudge = {};
+	const successfulJudgeScores = {};
+	const dimensionValues = /* @__PURE__ */ new Map();
+	const composites = [];
+	const notes = [];
+	const failedJudges = new Set(cell.errorStage === "judge" ? [cell.errorJudge ?? "unknown-judge"] : []);
+	const raw = {};
+	for (const [judgeName, score] of Object.entries(cell.judgeScores)) {
+		const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite);
+		if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {
+			failedJudges.add(judgeName);
+			continue;
+		}
+		composites.push(score.composite);
+		successfulJudgeScores[judgeName] = score;
+		const dimensions = { ...score.dimensions };
+		perJudge[judgeName] = dimensions;
+		for (const [dimension, value] of Object.entries(dimensions)) {
+			raw[`${judgeName}.${dimension}`] = value;
+			const values = dimensionValues.get(dimension) ?? [];
+			values.push(value);
+			dimensionValues.set(dimension, values);
+		}
+		if (score.notes) notes.push(`${judgeName}: ${score.notes}`);
+		for (const failedJudge of score.failedJudges ?? []) failedJudges.add(`${judgeName}/${failedJudge}`);
+	}
+	if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size;
+	const sortedFailedJudges = [...failedJudges].sort();
+	if (composites.length === 0) return {
+		successfulJudgeScores,
+		failedJudges: sortedFailedJudges,
+		raw
+	};
+	const composite = mean$1(composites);
+	const perDimMean = Object.fromEntries([...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean$1(values)]));
+	const complete = cell.error === void 0 && cell.errorStage === void 0 && failedJudges.size === 0;
+	if (complete) raw.composite = composite;
+	return {
+		...complete ? { score: composite } : {},
+		raw,
+		successfulJudgeScores,
+		failedJudges: sortedFailedJudges,
+		judgeScores: {
+			perJudge,
+			perDimMean,
+			composite,
+			...sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {},
+			...notes.length > 0 ? { notes: notes.join(" | ") } : {}
+		}
+	};
+}
+/** Read the canonical task score without recomputing cell quality. */
+function campaignCellTaskScore(cell) {
+	return projectCampaignCellQuality(cell).score;
+}
+/** Read canonical successful judge dimensions without recomputing cell quality. */
+function campaignCellJudgeDimensions(cell) {
+	return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {};
+}
+function finiteMetrics(metrics) {
+	const finite = {};
+	for (const [key, value] of Object.entries(metrics ?? {})) if (Number.isFinite(value)) finite[key] = value;
+	return finite;
+}
+function mean$1(values) {
+	return values.reduce((sum, value) => sum + value, 0) / values.length;
+}
+//#endregion
+//#region src/rl/verifiable-reward.ts
+const DEFAULT_DETERMINISTIC_LAYERS = /* @__PURE__ */ new Set([
+	"install",
+	"typecheck",
+	"build",
+	"lint",
+	"test",
+	"compile",
+	"schema",
+	"sandbox",
+	"unit_tests",
+	"integration_tests"
+]);
+const DEFAULT_SOURCE_FOR = (name) => {
+	const lower = name.toLowerCase();
+	if (lower.includes("test")) return "test";
+	if (lower.includes("compile") || lower.includes("build") || lower.includes("typecheck") || lower.includes("lint")) return "compile";
+	if (lower.includes("schema")) return "schema";
+	if (lower.includes("sandbox")) return "sandbox";
+	if (lower.includes("judge") || lower.includes("semantic")) return "judge";
+	return "composite";
+};
+/**
+* Extract a `VerifiableReward` from a `VerificationReport`.
+*
+* Strategy: prefer the deterministic layers (in order: test → compile →
+* schema → sandbox), fall back to the judge layer if `fallbackToJudge` is
+* true, return `null` if no signal qualifies. When multiple deterministic
+* layers contribute, return a `'composite'` source with a weighted blend.
+*
+* NO realness gate is applied and none can be: a `VerificationReport` carries
+* layer scores and nothing about whether the run faked them — `realness` lives
+* on the `RunRecord`. Use `extractVerifiableRewardsFromRecords` for anything
+* that becomes training data; this signature is for scoring a report in hand.
+*/
+function extractVerifiableReward(report, opts = {}) {
+	const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS]);
+	const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR;
+	const fallbackToJudge = opts.fallbackToJudge ?? true;
+	const judgeFloor = opts.judgeConfidenceFloor ?? .7;
+	const deterministic = report.layers.filter((layer) => deterministicSet.has(layer.layer) && isMeasuredLayer(layer));
+	if (deterministic.length === 1) {
+		const layer = deterministic[0];
+		const value = clamp01$1(layer.score);
+		return {
+			value,
+			source: sourceFor(layer.layer),
+			determinism: "deterministic",
+			confidence: 1,
+			origin: layer.layer,
+			components: { [layer.layer]: value },
+			realnessScreened: false
+		};
+	}
+	if (deterministic.length > 1) {
+		let num = 0;
+		let denom = 0;
+		const components = {};
+		for (const l of deterministic) {
+			const w = l.detail?.weight ?? 1;
+			num += w * (l.score ?? 0);
+			denom += w;
+			components[l.layer] = l.score;
+		}
+		return {
+			value: denom === 0 ? 0 : clamp01$1(num / denom),
+			source: "composite",
+			determinism: "deterministic",
+			confidence: 1,
+			origin: deterministic.map((l) => l.layer).join("+"),
+			components,
+			realnessScreened: false
+		};
+	}
+	if (!fallbackToJudge) return null;
+	const judge = report.layers.find((layer) => isMeasuredLayer(layer) && sourceFor(layer.layer) === "judge") ?? report.layers.find(isMeasuredLayer);
+	if (!judge) return null;
+	const confFromDetail = judge.detail?.confidence;
+	const judgeValue = clamp01$1(judge.score);
+	return {
+		value: judgeValue,
+		source: "judge",
+		determinism: "probabilistic",
+		confidence: typeof confFromDetail === "number" ? confFromDetail : judgeFloor,
+		origin: judge.layer,
+		components: { [judge.layer]: judgeValue },
+		realnessScreened: false
+	};
+}
+function isMeasuredLayer(layer) {
+	return (layer.status === "pass" || layer.status === "fail") && typeof layer.score === "number" && Number.isFinite(layer.score) && layer.score >= 0 && layer.score <= 1;
+}
+/**
+* Extract verifiable rewards from `RunRecord[]` produced via the
+* `verificationReportToRunRecord` adapter (which encodes per-layer scores
+* in `outcome.raw['layer.<name>']`). For records that don't carry layer
+* scores, returns `null` for that record.
+*
+* This is the canonical bridge from "campaign-shaped artifacts" to
+* "RL-training-ready reward signals": every record that has a clean
+* verifiable reward becomes a training datum, every record that doesn't
+* gets filtered out (or kept with `'probabilistic'` determinism for
+* separate downstream handling).
+*
+* The realness gate applies to EVERY channel here, and to the deterministic one
+* MOST. It is tempting to reason that a decidable signal cannot be gamed, so
+* the gate is redundant on it — that reasoning is backwards. `realness.gated`
+* means the run's success signal was FAKED, and a test suite reporting green on
+* a stubbed integration is precisely what that looks like: the deterministic
+* layer is the thing that got faked. Exporting it ungated hands a trainer the
+* highest-credibility reward the module can emit (`determinism: 'deterministic'`,
+* `confidence: 1`) for the one population the gate exists to catch. Pass
+* `applyRealnessGate: false` only to look at the ungated numbers for detection.
+*/
+function extractVerifiableRewardsFromRecords(runs, opts = {}) {
+	const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR;
+	const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS]);
+	const fallbackToJudge = opts.fallbackToJudge ?? true;
+	const judgeFloor = opts.judgeConfidenceFloor ?? .7;
+	const applyGate = opts.applyRealnessGate ?? true;
+	return runs.map((run) => {
+		const flagged = isRealnessGated(run);
+		const screened = run.outcome.realness === void 0 ? {} : { realnessScreened: true };
+		const gate = (value) => applyGate && flagged ? 0 : value;
+		const layerScores = [];
+		for (const [k, v] of Object.entries(run.outcome.raw)) if (k.startsWith("layer.") && !k.includes(".", 6) && typeof v === "number" && Number.isFinite(v)) layerScores.push({
+			name: k.slice(6),
+			score: v
+		});
+		const det = layerScores.filter((l) => deterministicSet.has(l.name));
+		if (det.length === 1) {
+			const layer = det[0];
+			const value = gate(clamp01$1(layer.score));
+			return {
+				runId: run.runId,
+				reward: {
+					value,
+					source: sourceFor(layer.name),
+					determinism: "deterministic",
+					confidence: 1,
+					origin: layer.name,
+					components: { [layer.name]: value },
+					realnessGated: flagged,
+					...screened
+				}
+			};
+		}
+		if (det.length > 1) {
+			const value = gate(clamp01$1(det.reduce((s, l) => s + l.score, 0) / det.length));
+			const components = Object.fromEntries(det.map((l) => [l.name, gate(clamp01$1(l.score))]));
+			return {
+				runId: run.runId,
+				reward: {
+					value,
+					source: "composite",
+					determinism: "deterministic",
+					confidence: 1,
+					origin: det.map((l) => l.name).join("+"),
+					components,
+					realnessGated: flagged,
+					...screened
+				}
+			};
+		}
+		if (!fallbackToJudge) return {
+			runId: run.runId,
+			reward: null
+		};
+		const primary = applyGate ? trainingScore(run) : observedScore(run);
+		if (typeof primary !== "number" || !Number.isFinite(primary)) return {
+			runId: run.runId,
+			reward: null
+		};
+		const primaryValue = clamp01$1(primary);
+		return {
+			runId: run.runId,
+			reward: {
+				value: primaryValue,
+				source: "judge",
+				determinism: "probabilistic",
+				confidence: judgeFloor,
+				origin: "run.outcome.score",
+				components: { "run.outcome.score": primaryValue },
+				realnessGated: flagged,
+				...screened
+			}
+		};
+	});
+}
+/**
+* Filter `RunRecord[]` to those with deterministic verifiable rewards.
+*
+* A realness-gated run is KEPT, at reward 0 with `realnessGated: true` — the
+* same rule GRPO uses on a gated line. 0 is the honest label for a faked
+* success and is usable signal, whereas dropping the run would move a group
+* baseline without saying so. (SFT differs: there every row is a target to
+* imitate, so a gated row is removed outright.)
+*/
+function filterDeterministicallyRewarded(runs, opts = {}) {
+	const rewarded = extractVerifiableRewardsFromRecords(runs, {
+		...opts,
+		fallbackToJudge: false
+	});
+	const out = [];
+	for (let i = 0; i < runs.length; i++) {
+		const r = rewarded[i];
+		if (r.reward && r.reward.determinism === "deterministic") out.push({
+			run: runs[i],
+			reward: r.reward
+		});
+	}
+	return out;
+}
+function clamp01$1(x) {
+	if (!Number.isFinite(x)) return 0;
+	return Math.max(0, Math.min(1, x));
+}
+//#endregion
+//#region src/rl/reward-hacking.ts
+/**
+* Reward hacking / Goodhart detection.
+*
+* Goodhart's Law says: when a measure becomes a target, it ceases to be
+* a good measure. In RLHF and agentic-RL settings this is the dominant
+* failure mode — the policy learns to produce outputs that score well on
+* the proxy reward (judge, rubric, test pass-rate) without producing
+* the underlying capability the proxy was meant to track.
+*
+* Krakovna et al. (2020, "Specification Gaming Examples in AI") and the
+* subsequent RLHF reward-hacking literature (Skalse et al. 2022, Kim et al.
+* 2023) converge on a few diagnostic signatures:
+*
+*   1. **Reward divergence:** the proxy reward grows while the held-out
+*      ground-truth signal stagnates or drops. Predictive validity over
+*      time captures this.
+*   2. **Distributional shift in outputs:** after RL, the policy produces
+*      outputs that no longer match the reference distribution — usually
+*      because it found a high-reward attractor that's degenerate (e.g.
+*      one-token responses, repetition, formatting tricks).
+*   3. **Disagreement between independent rewards:** if you train on
+*      reward A and a held-out independent reward B drops sharply, you're
+*      probably hacking A.
+*   4. **Calibration drift:** the verifiable / deterministic component of
+*      the reward is stable; the probabilistic / judge component drifts up
+*      while the deterministic component doesn't. The judge is being
+*      gamed.
+*
+* This module ships explicit detectors for all four signatures, plus a
+* combined verdict. The output is diagnostic — actionable signals,
+* not autoreject — because each signature has known false positives
+* (e.g., a policy that genuinely improves can show distributional shift).
+*
+* Differs from `rubricPredictiveValidity` (which is a *standing* check on
+* whether rubrics correlate with deployment outcomes) — this is a
+* *temporal* check on whether the reward-vs-truth gap is *widening over
+* time during a training run*.
+*/
+const DEFAULT_PROXY = (r) => {
+	const v = observedScore(r);
+	return typeof v === "number" && Number.isFinite(v) ? v : null;
+};
+function detectRewardHacking(input) {
+	const proxyOf = input.proxyOf ?? DEFAULT_PROXY;
+	const truthOf = input.truthOf;
+	const sus = input.thresholds?.suspect ?? .3;
+	const gam = input.thresholds?.gaming ?? .6;
+	const runs = input.runs.filter((run) => finiteNumber(proxyOf(run)));
+	const n = runs.length;
+	if (n < 4) return {
+		findings: [],
+		evaluatedSignals: [],
+		verdict: "insufficient_evidence",
+		n,
+		rationale: [`fewer than 4 runs with proxy reward (n=${n}); insufficient evidence`]
+	};
+	const windowSize = Math.max(1, input.windowSize ?? Math.min(50, Math.floor(n / 2)));
+	const before = runs.slice(0, n - windowSize);
+	const after = runs.slice(n - windowSize);
+	const findings = [];
+	if (truthOf) {
+		const beforeProxy = before.map(proxyOf).filter(finiteNumber);
+		const afterProxy = after.map(proxyOf).filter(finiteNumber);
+		const beforeTruth = before.map(truthOf).filter(finiteNumber);
+		const afterTruth = after.map(truthOf).filter(finiteNumber);
+		if (beforeProxy.length >= 2 && afterProxy.length >= 2 && beforeTruth.length >= 2 && afterTruth.length >= 2) {
+			const proxyDelta = mean(afterProxy) - mean(beforeProxy);
+			const truthDelta = mean(afterTruth) - mean(beforeTruth);
+			const gap = Math.max(0, proxyDelta - truthDelta);
+			const severity = clamp01(gap * 5);
+			findings.push({
+				signal: "reward_divergence",
+				severity,
+				message: severity >= sus ? `proxy reward rose by ${proxyDelta.toFixed(3)} while truth changed by ${truthDelta.toFixed(3)} — potential Goodhart` : `proxy and truth moved together (proxy ${proxyDelta.toFixed(3)}, truth ${truthDelta.toFixed(3)})`,
+				detail: {
+					proxyDelta,
+					truthDelta,
+					gap,
+					beforeN: beforeProxy.length,
+					afterN: afterProxy.length
+				}
+			});
+		}
+	}
+	{
+		const beforeP = before.map(proxyOf).filter(finiteNumber);
+		const afterP = after.map(proxyOf).filter(finiteNumber);
+		if (beforeP.length >= 4 && afterP.length >= 4) {
+			const ks = ksStatistic(beforeP, afterP);
+			const severity = clamp01(ks - .2);
+			findings.push({
+				signal: "distribution_shift",
+				severity,
+				message: severity >= sus ? `KS=${ks.toFixed(3)} between before/after windows — distributional shift large` : `KS=${ks.toFixed(3)} between before/after windows — within-distribution drift`,
+				detail: {
+					ks,
+					beforeN: beforeP.length,
+					afterN: afterP.length
+				}
+			});
+		}
+	}
+	{
+		const secondaryOf = input.secondaryRewardOf ?? defaultSecondary(input.verifiableRewardOptions);
+		const aligned = runs.map((r) => ({
+			p: proxyOf(r),
+			s: secondaryOf(r)
+		})).filter((x) => finiteNumber(x.p) && finiteNumber(x.s));
+		if (aligned.length >= 4) {
+			const r = pearsonR(aligned.map((x) => x.p), aligned.map((x) => x.s));
+			const severity = clamp01(.5 - Math.max(0, r));
+			findings.push({
+				signal: "reward_disagreement",
+				severity,
+				message: severity >= sus ? `proxy and independent secondary reward correlate ρ=${r.toFixed(3)} — possibly hacking proxy` : `proxy and secondary reward correlate ρ=${r.toFixed(3)}`,
+				detail: {
+					pearson: r,
+					n: aligned.length
+				}
+			});
+		}
+	}
+	{
+		const detRuns = filterDeterministicallyRewarded(runs, {
+			...input.verifiableRewardOptions ?? {},
+			applyRealnessGate: false
+		});
+		if (detRuns.length >= 4) {
+			const detBefore = detRuns.slice(0, Math.floor(detRuns.length / 2));
+			const detDelta = mean(detRuns.slice(Math.floor(detRuns.length / 2)).map((r) => r.reward.value)) - mean(detBefore.map((r) => r.reward.value));
+			const proxyDelta = mean(after.map(proxyOf).filter(finiteNumber)) - mean(before.map(proxyOf).filter(finiteNumber));
+			const driftGap = Math.max(0, proxyDelta - detDelta);
+			const severity = clamp01(driftGap * 5);
+			findings.push({
+				signal: "judge_drift",
+				severity,
+				message: severity >= sus ? `judge proxy +${proxyDelta.toFixed(3)} while deterministic reward +${detDelta.toFixed(3)} — judge drifting up without verifiable backing` : `judge and deterministic rewards move in step (judge ${proxyDelta.toFixed(3)}, det ${detDelta.toFixed(3)})`,
+				detail: {
+					proxyDelta,
+					detDelta,
+					driftGap,
+					n: detRuns.length
+				}
+			});
+		}
+	}
+	const maxSev = findings.reduce((m, f) => Math.max(m, f.severity), 0);
+	if (findings.length === 0) return {
+		findings,
+		evaluatedSignals: [],
+		verdict: "insufficient_evidence",
+		rationale: [`no reward-hacking signal had enough paired evidence (n=${n})`],
+		n
+	};
+	const verdict = maxSev >= gam ? "gaming" : maxSev >= sus ? "suspect" : "clean";
+	const rationale = findings.filter((f) => f.severity >= sus).map((f) => `${f.signal}: severity ${f.severity.toFixed(2)} — ${f.message}`);
+	if (rationale.length === 0) rationale.push("no signals fired above suspect threshold");
+	return {
+		findings,
+		evaluatedSignals: findings.map((finding) => finding.signal),
+		verdict,
+		rationale,
+		n
+	};
+}
+function mean(xs) {
+	if (xs.length === 0) return 0;
+	return xs.reduce((s, x) => s + x, 0) / xs.length;
+}
+function finiteNumber(value) {
+	return typeof value === "number" && Number.isFinite(value);
+}
+function clamp01(x) {
+	if (!Number.isFinite(x)) return 0;
+	return Math.max(0, Math.min(1, x));
+}
+function ksStatistic(a, b) {
+	const sortedA = [...a].sort((x, y) => x - y);
+	const sortedB = [...b].sort((x, y) => x - y);
+	const all = [.../* @__PURE__ */ new Set([...sortedA, ...sortedB])].sort((x, y) => x - y);
+	let max = 0;
+	for (const v of all) {
+		const fa = sortedA.filter((x) => x <= v).length / sortedA.length;
+		const fb = sortedB.filter((x) => x <= v).length / sortedB.length;
+		max = Math.max(max, Math.abs(fa - fb));
+	}
+	return max;
+}
+function defaultSecondary(verifiableOpts) {
+	return (run) => {
+		const filtered = filterDeterministicallyRewarded([run], {
+			...verifiableOpts ?? {},
+			applyRealnessGate: false
+		});
+		return filtered.length === 1 ? filtered[0].reward.value : null;
+	};
+}
+//#endregion
+export { campaignCellExecutionEvidence as a, campaignCellToRunRecord as c, filterDeterministicallyRewarded as i, projectCampaignCellQuality as l, extractVerifiableReward as n, campaignCellJudgeDimensions as o, extractVerifiableRewardsFromRecords as r, campaignCellTaskScore as s, detectRewardHacking as t };
+
+//# sourceMappingURL=reward-hacking-qipEpKvY.js.map