@tangle-network/agent-eval 0.132.0 → 0.133.1

package/dist/metrics-C9YY1OcL.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"metrics-C9YY1OcL.js","names":[],"sources":["../src/metrics.ts"],"sourcesContent":["import type { ProductClient } from './client'\nimport type { DriverState, TurnMetrics } from './types'\n\ninterface TokenPrice {\n  input: number\n  output: number\n}\n\n/** Per-1K token pricing for exact model ids. */\nexport const MODEL_PRICING: Record<string, TokenPrice> = {\n  'gpt-4o': { input: 0.0025, output: 0.01 },\n  'gpt-4o-mini': { input: 0.00015, output: 0.0006 },\n  'gpt-4-turbo': { input: 0.01, output: 0.03 },\n  'claude-sonnet-4-20250514': { input: 0.003, output: 0.015 },\n  'claude-opus-4-20250514': { input: 0.015, output: 0.075 },\n  'claude-3-haiku-20240307': { input: 0.00025, output: 0.00125 },\n}\n\n/** Family-level pricing fallbacks (per-1K), matched against a normalized id\n *  after exact lookup misses. Ordered — first match wins. Covers the model\n *  ids actually used through the Tangle router + cli-bridge harnesses\n *  (`claude-code/sonnet`, `opencode/zai-coding-plan/glm-5.1`,\n *  `kimi-code/kimi-k2.6`, `deepseek-v4-pro`, `anthropic/claude-sonnet-4-6`, …),\n *  none of which appear in the exact table above — without this they priced\n *  to a silent $0, blanking every cost/Pareto axis downstream. */\nconst FAMILY_PRICING: Array<[RegExp, TokenPrice]> = [\n  [/claude.*opus/, { input: 0.015, output: 0.075 }],\n  [/claude.*haiku/, { input: 0.0008, output: 0.004 }],\n  [/claude.*sonnet|claude-code|claude-sonnet/, { input: 0.003, output: 0.015 }],\n  [/gpt-4o-mini/, { input: 0.00015, output: 0.0006 }],\n  [/gpt-5|gpt-4\\.1|o[134]\\b/, { input: 0.00125, output: 0.01 }],\n  [/gpt-4o|gpt-4/, { input: 0.0025, output: 0.01 }],\n  [/deepseek/, { input: 0.0003, output: 0.0011 }],\n  [/glm|zhipu|zai/, { input: 0.0006, output: 0.0022 }],\n  [/kimi|moonshot/, { input: 0.0006, output: 0.0025 }],\n  [/qwen/, { input: 0.0004, output: 0.0012 }],\n  [/gemini.*flash/, { input: 0.0001, output: 0.0004 }],\n  [/gemini/, { input: 0.00125, output: 0.005 }],\n  [/llama/, { input: 0.0002, output: 0.0006 }],\n]\n\n/** Normalize a model id for pricing: drop a `@snapshot` suffix, lowercase,\n *  and keep the final harness/provider-prefixed segment so family regexes\n *  match (`opencode/zai-coding-plan/glm-5.1` → `glm-5.1`). */\nfunction normalizeModelId(model: string): string {\n  return (model.split('@')[0] ?? model).trim().toLowerCase()\n}\n\n/** Resolve pricing for a model id: exact table, then family fallback.\n *  Returns null when the id matches nothing (caller decides — never a\n *  silent-zero masquerading as a real $0 cost). */\nexport function resolveModelPricing(model: string): TokenPrice | null {\n  if (MODEL_PRICING[model]) return MODEL_PRICING[model]\n  const id = normalizeModelId(model)\n  if (MODEL_PRICING[id]) return MODEL_PRICING[id]\n  for (const [pattern, price] of FAMILY_PRICING) {\n    if (pattern.test(id)) return price\n  }\n  return null\n}\n\n/** True when `model` has known pricing (exact or family). Lets cost-aware\n *  callers distinguish a real $0 from an unpriced model. */\nexport function isModelPriced(model: string): boolean {\n  return resolveModelPricing(model) !== null\n}\n\nconst warnedUnpricedModels = new Set<string>()\n\n/** Estimate token count from string length (chars / 4 approximation) */\nexport function estimateTokens(text: string): number {\n  return Math.ceil(text.length / 4)\n}\n\n/** Calculate cost in USD from token counts and model. Unknown models warn\n *  once (not a silent zero) and return 0 so callers that ignore pricing keep\n *  working; cost-sensitive callers should gate on {@link isModelPriced}. */\nexport function estimateCost(inputTokens: number, outputTokens: number, model: string): number {\n  const pricing = resolveModelPricing(model)\n  if (!pricing) {\n    if (!warnedUnpricedModels.has(model)) {\n      warnedUnpricedModels.add(model)\n      console.warn(\n        `estimateCost: no pricing for model \"${model}\" — returning 0; add it to ` +\n          'MODEL_PRICING/FAMILY_PRICING (cost/Pareto axes will be blank until then)',\n      )\n    }\n    return 0\n  }\n  return (inputTokens / 1000) * pricing.input + (outputTokens / 1000) * pricing.output\n}\n\n/**\n * TokenCounter — accumulates token usage and cost across turns.\n */\nexport class TokenCounter {\n  private totalInput = 0\n  private totalOutput = 0\n  private totalCost = 0\n  private model: string\n\n  constructor(model = 'gpt-4o') {\n    this.model = model\n  }\n\n  /** Record tokens for a turn, returns per-turn cost */\n  record(inputTokens: number, outputTokens: number): number {\n    this.totalInput += inputTokens\n    this.totalOutput += outputTokens\n    const cost = estimateCost(inputTokens, outputTokens, this.model)\n    this.totalCost += cost\n    return cost\n  }\n\n  /** Estimate and record from raw text */\n  recordFromText(\n    inputText: string,\n    outputText: string,\n  ): { inputTokens: number; outputTokens: number; cost: number } {\n    const inputTokens = estimateTokens(inputText)\n    const outputTokens = estimateTokens(outputText)\n    const cost = this.record(inputTokens, outputTokens)\n    return { inputTokens, outputTokens, cost }\n  }\n\n  getTotalInput(): number {\n    return this.totalInput\n  }\n  getTotalOutput(): number {\n    return this.totalOutput\n  }\n  getTotalCost(): number {\n    return this.totalCost\n  }\n}\n\n/**\n * MetricsCollector — collects per-turn metrics from the product.\n *\n * After each turn, queries the product's APIs to measure state changes.\n */\nexport class MetricsCollector {\n  private client: ProductClient\n  private workspaceId: string\n  private metrics: TurnMetrics[] = []\n  constructor(client: ProductClient, workspaceId: string) {\n    this.client = client\n    this.workspaceId = workspaceId\n  }\n\n  /** Collect metrics after a turn completes */\n  async collect(\n    turn: number,\n    responseLatencyMs: number,\n    responseChars: number,\n    codeBlocksProduced: number,\n    blocksExtracted: number,\n    completionCriteriaMet: number,\n    completionCriteriaTotal: number,\n    qualityScore?: number,\n    inputTokens = 0,\n    outputTokens = 0,\n    estimatedCostUsd = 0,\n  ): Promise<TurnMetrics> {\n    const state = await this.getState()\n\n    const m: TurnMetrics = {\n      turn,\n      timestamp: new Date().toISOString(),\n      tasks: state.tasks,\n      events: state.events,\n      proposals: state.proposals,\n      vaultFiles: state.vaultFiles.length,\n      responseLatencyMs,\n      responseChars,\n      codeBlocksProduced,\n      blocksExtracted,\n      qualityScore,\n      inputTokens,\n      outputTokens,\n      estimatedCostUsd,\n      totalCostUsd: estimatedCostUsd,\n      completionPercent:\n        completionCriteriaTotal > 0 ? (completionCriteriaMet / completionCriteriaTotal) * 100 : 0,\n    }\n\n    this.metrics.push(m)\n    return m\n  }\n\n  /** Get current product state */\n  async getState(): Promise<DriverState> {\n    const [tasks, events, approvals, vaultFiles] = await Promise.all([\n      this.client.getTasks(this.workspaceId),\n      this.client.getEvents(this.workspaceId),\n      this.client.getApprovals(this.workspaceId),\n      this.client.getVaultTree(this.workspaceId),\n    ])\n\n    return {\n      tasks: tasks.length,\n      events: events.length,\n      proposals: {\n        pending: approvals.filter((a) => a.status === 'pending').length,\n        approved: approvals.filter((a) => a.status === 'approved').length,\n        rejected: approvals.filter((a) => a.status === 'rejected').length,\n      },\n      vaultFiles,\n      codeBlocks: 0,\n      generations: 0,\n    }\n  }\n\n  /** Get all collected metrics */\n  getMetrics(): TurnMetrics[] {\n    return [...this.metrics]\n  }\n\n  /** Get convergence curve (completion% over turns) */\n  getConvergenceCurve(): number[] {\n    return this.metrics.map((m) => m.completionPercent)\n  }\n}\n"],"mappings":";;AASA,MAAa,gBAA4C;CACvD,UAAU;EAAE,OAAO;EAAQ,QAAQ;CAAK;CACxC,eAAe;EAAE,OAAO;EAAS,QAAQ;CAAO;CAChD,eAAe;EAAE,OAAO;EAAM,QAAQ;CAAK;CAC3C,4BAA4B;EAAE,OAAO;EAAO,QAAQ;CAAM;CAC1D,0BAA0B;EAAE,OAAO;EAAO,QAAQ;CAAM;CACxD,2BAA2B;EAAE,OAAO;EAAS,QAAQ;CAAQ;AAC/D;;;;;;;;AASA,MAAM,iBAA8C;CAClD,CAAC,gBAAgB;EAAE,OAAO;EAAO,QAAQ;CAAM,CAAC;CAChD,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAM,CAAC;CAClD,CAAC,4CAA4C;EAAE,OAAO;EAAO,QAAQ;CAAM,CAAC;CAC5E,CAAC,eAAe;EAAE,OAAO;EAAS,QAAQ;CAAO,CAAC;CAClD,CAAC,2BAA2B;EAAE,OAAO;EAAS,QAAQ;CAAK,CAAC;CAC5D,CAAC,gBAAgB;EAAE,OAAO;EAAQ,QAAQ;CAAK,CAAC;CAChD,CAAC,YAAY;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CAC9C,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CACnD,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CACnD,CAAC,QAAQ;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CAC1C,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CACnD,CAAC,UAAU;EAAE,OAAO;EAAS,QAAQ;CAAM,CAAC;CAC5C,CAAC,SAAS;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;AAC7C;;;;AAKA,SAAS,iBAAiB,OAAuB;CAC/C,QAAQ,MAAM,MAAM,GAAG,CAAC,CAAC,MAAM,MAAA,CAAO,KAAK,CAAC,CAAC,YAAY;AAC3D;;;;AAKA,SAAgB,oBAAoB,OAAkC;CACpE,IAAI,cAAc,QAAQ,OAAO,cAAc;CAC/C,MAAM,KAAK,iBAAiB,KAAK;CACjC,IAAI,cAAc,KAAK,OAAO,cAAc;CAC5C,KAAK,MAAM,CAAC,SAAS,UAAU,gBAC7B,IAAI,QAAQ,KAAK,EAAE,GAAG,OAAO;CAE/B,OAAO;AACT;;;AAIA,SAAgB,cAAc,OAAwB;CACpD,OAAO,oBAAoB,KAAK,MAAM;AACxC;AAEA,MAAM,uCAAuB,IAAI,IAAY;;AAG7C,SAAgB,eAAe,MAAsB;CACnD,OAAO,KAAK,KAAK,KAAK,SAAS,CAAC;AAClC;;;;AAKA,SAAgB,aAAa,aAAqB,cAAsB,OAAuB;CAC7F,MAAM,UAAU,oBAAoB,KAAK;CACzC,IAAI,CAAC,SAAS;EACZ,IAAI,CAAC,qBAAqB,IAAI,KAAK,GAAG;GACpC,qBAAqB,IAAI,KAAK;GAC9B,QAAQ,KACN,uCAAuC,MAAM,oGAE/C;EACF;EACA,OAAO;CACT;CACA,OAAQ,cAAc,MAAQ,QAAQ,QAAS,eAAe,MAAQ,QAAQ;AAChF;;;;AAKA,IAAa,eAAb,MAA0B;CACxB,aAAqB;CACrB,cAAsB;CACtB,YAAoB;CACpB;CAEA,YAAY,QAAQ,UAAU;EAC5B,KAAK,QAAQ;CACf;;CAGA,OAAO,aAAqB,cAA8B;EACxD,KAAK,cAAc;EACnB,KAAK,eAAe;EACpB,MAAM,OAAO,aAAa,aAAa,cAAc,KAAK,KAAK;EAC/D,KAAK,aAAa;EAClB,OAAO;CACT;;CAGA,eACE,WACA,YAC6D;EAC7D,MAAM,cAAc,eAAe,SAAS;EAC5C,MAAM,eAAe,eAAe,UAAU;EAE9C,OAAO;GAAE;GAAa;GAAc,MADvB,KAAK,OAAO,aAAa,YACC;EAAE;CAC3C;CAEA,gBAAwB;EACtB,OAAO,KAAK;CACd;CACA,iBAAyB;EACvB,OAAO,KAAK;CACd;CACA,eAAuB;EACrB,OAAO,KAAK;CACd;AACF;;;;;;AAOA,IAAa,mBAAb,MAA8B;CAC5B;CACA;CACA,UAAiC,CAAC;CAClC,YAAY,QAAuB,aAAqB;EACtD,KAAK,SAAS;EACd,KAAK,cAAc;CACrB;;CAGA,MAAM,QACJ,MACA,mBACA,eACA,oBACA,iBACA,uBACA,yBACA,cACA,cAAc,GACd,eAAe,GACf,mBAAmB,GACG;EACtB,MAAM,QAAQ,MAAM,KAAK,SAAS;EAElC,MAAM,IAAiB;GACrB;GACA,4BAAW,IAAI,KAAK,EAAA,CAAE,YAAY;GAClC,OAAO,MAAM;GACb,QAAQ,MAAM;GACd,WAAW,MAAM;GACjB,YAAY,MAAM,WAAW;GAC7B;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA,cAAc;GACd,mBACE,0BAA0B,IAAK,wBAAwB,0BAA2B,MAAM;EAC5F;EAEA,KAAK,QAAQ,KAAK,CAAC;EACnB,OAAO;CACT;;CAGA,MAAM,WAAiC;EACrC,MAAM,CAAC,OAAO,QAAQ,WAAW,cAAc,MAAM,QAAQ,IAAI;GAC/D,KAAK,OAAO,SAAS,KAAK,WAAW;GACrC,KAAK,OAAO,UAAU,KAAK,WAAW;GACtC,KAAK,OAAO,aAAa,KAAK,WAAW;GACzC,KAAK,OAAO,aAAa,KAAK,WAAW;EAC3C,CAAC;EAED,OAAO;GACL,OAAO,MAAM;GACb,QAAQ,OAAO;GACf,WAAW;IACT,SAAS,UAAU,QAAQ,MAAM,EAAE,WAAW,SAAS,CAAC,CAAC;IACzD,UAAU,UAAU,QAAQ,MAAM,EAAE,WAAW,UAAU,CAAC,CAAC;IAC3D,UAAU,UAAU,QAAQ,MAAM,EAAE,WAAW,UAAU,CAAC,CAAC;GAC7D;GACA;GACA,YAAY;GACZ,aAAa;EACf;CACF;;CAGA,aAA4B;EAC1B,OAAO,CAAC,GAAG,KAAK,OAAO;CACzB;;CAGA,sBAAgC;EAC9B,OAAO,KAAK,QAAQ,KAAK,MAAM,EAAE,iBAAiB;CACpD;AACF"}

package/dist/off-policy-DvgzvtIx.js

@@ -1,220 +0,0 @@
-import { s as ValidationError } from "./errors-8YnH8WlF.js";
-//#region src/rl/off-policy.ts
-/**
-* Off-policy evaluation primitives.
-*
-* Standard inverse-probability-weighted (IPS), self-normalized
-* importance-weighted (SNIPS), and doubly-robust (DR) estimators for the
-* value of a *target* policy given trajectories collected under a
-* *behavior* policy. This is the canonical RL eval task: "we have last
-* week's runs, we changed the policy — how would the new one do without
-* re-running?"
-*
-* The math here is textbook (Dudík, Langford, Li 2011 for DR; Swaminathan
-* & Joachims 2015 for SNIPS) but the *application* to LLM-agent
-* evaluation needs care:
-*
-*   - The "policy" is the (prompt, tool config, model snapshot) triple.
-*     Two policies have the same probability over an action *iff* their
-*     LLM call would emit the same token with the same probability —
-*     which is generally unknowable without the model log-probs.
-*   - For LLM agents, propensity scores must be supplied by the caller
-*     (logged in the trace, recovered from token log-probs, or estimated
-*     via a learned propensity model). We do NOT estimate propensity here.
-*   - Doubly-robust requires two outputs from a Q-function: its prediction
-*     for the logged action and its expectation under the target policy.
-*     Consumers compute these with a tabular estimate, regression fit, or
-*     learned reward model before constructing the trajectories.
-*
-* Bias / variance tradeoffs:
-*   - IPS: unbiased; high variance for small overlap, infinite variance
-*     when target has support outside behavior.
-*   - SNIPS: lower variance, slight bias; usually preferred in practice.
-*   - DR: doubly-robust — unbiased if either propensity OR Q-function is
-*     correct. Lowest practical variance when Q is decent. Use this.
-*
-* Caveat the panel will land: on the LLM-agent setting, propensity scores
-* recovered from token log-probs are noisy, the action space is enormous,
-* and overlap is often poor. These estimators are useful but not magic;
-* complement with `replayCampaign` (exact replay where the request hashes
-* match) for high-confidence answers and OPE for the gap.
-*/
-/**
-* Inverse Probability Weighting (Horvitz-Thompson). Unbiased estimator
-* of E[reward under target policy]. Variance scales with the spread of
-* target/behavior ratios.
-*/
-function inverseProbabilityWeighting(trajectories, opts = {}) {
-	const cap = opts.weightCap ?? Infinity;
-	const clip = opts.rewardClip ?? {
-		low: 0,
-		high: 1
-	};
-	if (trajectories.length === 0) return zeroEstimate();
-	const weights = [];
-	const weightedRewards = [];
-	let maxW = 0;
-	for (const t of trajectories) {
-		if (t.behaviorProb <= 0) throw new ValidationError(`inverseProbabilityWeighting: behaviorProb must be > 0 (runId=${t.runId})`);
-		const w = Math.min(cap, t.targetProb / t.behaviorProb);
-		const r = clamp(t.reward, clip.low, clip.high);
-		weights.push(w);
-		weightedRewards.push(w * r);
-		if (w > maxW) maxW = w;
-	}
-	const n = weights.length;
-	const value = weightedRewards.reduce((s, x) => s + x, 0) / n;
-	const variance = weightedRewards.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1);
-	const sumW = weights.reduce((s, w) => s + w, 0);
-	const sumW2 = weights.reduce((s, w) => s + w * w, 0);
-	const effN = sumW === 0 ? 0 : sumW * sumW / sumW2;
-	return {
-		value,
-		standardError: Math.sqrt(variance / n),
-		effectiveSampleSize: effN,
-		n,
-		maxImportanceWeight: maxW
-	};
-}
-/**
-* Self-Normalized Importance Sampling. Lower variance than vanilla IPS at
-* the cost of small bias (vanishing as N grows). The right default for
-* LLM-agent evaluation where overlap is often poor.
-*/
-function selfNormalizedImportanceWeighting(trajectories, opts = {}) {
-	const cap = opts.weightCap ?? Infinity;
-	const clip = opts.rewardClip ?? {
-		low: 0,
-		high: 1
-	};
-	if (trajectories.length === 0) return zeroEstimate();
-	const weights = [];
-	const rewards = [];
-	let maxW = 0;
-	for (const t of trajectories) {
-		if (t.behaviorProb <= 0) throw new ValidationError(`selfNormalizedImportanceWeighting: behaviorProb must be > 0 (runId=${t.runId})`);
-		const w = Math.min(cap, t.targetProb / t.behaviorProb);
-		weights.push(w);
-		rewards.push(clamp(t.reward, clip.low, clip.high));
-		if (w > maxW) maxW = w;
-	}
-	const sumW = weights.reduce((s, w) => s + w, 0);
-	const sumWR = weights.reduce((s, w, i) => s + w * rewards[i], 0);
-	const value = sumW === 0 ? 0 : sumWR / sumW;
-	const sumW2 = weights.reduce((s, w) => s + w * w, 0);
-	const effN = sumW === 0 ? 0 : sumW * sumW / sumW2;
-	const variance = weights.map((w, i) => w * (rewards[i] - value)).reduce((s, x) => s + x * x, 0) / Math.max(1, sumW * sumW);
-	return {
-		value,
-		standardError: Math.sqrt(variance),
-		effectiveSampleSize: effN,
-		n: trajectories.length,
-		maxImportanceWeight: maxW
-	};
-}
-/**
-* Doubly-robust off-policy estimator (Dudík, Langford, Li 2011).
-*
-*     V_DR = (1/N) * sum_i [ v_hat_target_i
-*              + (target_prob_i / behavior_prob_i) * (r_i - q_hat_chosen_i) ]
-*
-* Unbiased if EITHER:
-*   - the importance ratios are correct (IPS-style validity), OR
-*   - the Q-hat function is correct (model-based validity).
-*
-* In practice both are imperfect, but the residual bias is the *product*
-* of both errors — much smaller than either alone. This is why DR is the
-* default in production OPE pipelines.
-*
-* `qHatChosen` and `vHatTarget` must be supplied together. Rows with neither
-* use the exact IPS contribution. `contributionCounts` makes the mix explicit
-* in the result.
-* Callers must cross-fit the Q-function or train it on independent rows;
-* fitting and evaluating Q on the same outcomes leaks the answer.
-*/
-function doublyRobust(trajectories, opts = {}) {
-	const cap = opts.weightCap ?? Infinity;
-	const clip = opts.rewardClip ?? {
-		low: 0,
-		high: 1
-	};
-	if (trajectories.length === 0) return {
-		...zeroEstimate(),
-		contributionCounts: {
-			dr: 0,
-			ipsFallback: 0
-		}
-	};
-	const contributions = [];
-	const contributionCounts = {
-		dr: 0,
-		ipsFallback: 0
-	};
-	let maxW = 0;
-	let sumW = 0;
-	let sumW2 = 0;
-	for (const t of trajectories) {
-		if (t.behaviorProb <= 0) throw new ValidationError(`doublyRobust: behaviorProb must be > 0 (runId=${t.runId})`);
-		const w = Math.min(cap, t.targetProb / t.behaviorProb);
-		const r = clamp(t.reward, clip.low, clip.high);
-		const rawQHatChosen = t.qHatChosen;
-		const rawVHatTarget = t.vHatTarget;
-		const hasQHatChosen = rawQHatChosen !== null && rawQHatChosen !== void 0;
-		const hasVHatTarget = rawVHatTarget !== null && rawVHatTarget !== void 0;
-		if (hasQHatChosen !== hasVHatTarget) throw new ValidationError(`doublyRobust: qHatChosen and vHatTarget must be supplied together (runId=${t.runId})`);
-		if (hasQHatChosen && hasVHatTarget) {
-			if (!Number.isFinite(rawQHatChosen) || !Number.isFinite(rawVHatTarget)) throw new ValidationError(`doublyRobust: qHatChosen and vHatTarget must be finite (runId=${t.runId})`);
-			const qHatChosen = clamp(rawQHatChosen, clip.low, clip.high);
-			const vHatTarget = clamp(rawVHatTarget, clip.low, clip.high);
-			contributions.push(vHatTarget + w * (r - qHatChosen));
-			contributionCounts.dr += 1;
-		} else {
-			contributions.push(w * r);
-			contributionCounts.ipsFallback += 1;
-		}
-		if (w > maxW) maxW = w;
-		sumW += w;
-		sumW2 += w * w;
-	}
-	const n = contributions.length;
-	const value = contributions.reduce((s, x) => s + x, 0) / n;
-	const variance = contributions.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1);
-	const effN = sumW === 0 ? 0 : sumW * sumW / sumW2;
-	return {
-		value,
-		standardError: Math.sqrt(variance / n),
-		effectiveSampleSize: effN,
-		n,
-		maxImportanceWeight: maxW,
-		contributionCounts
-	};
-}
-/**
-* Convenience: run all three estimators and return them side-by-side.
-* The recommended diagnostic — agreement across estimators is a much
-* stronger signal than any single one.
-*/
-function offPolicyEstimateAll(trajectories, opts = {}) {
-	return {
-		ips: inverseProbabilityWeighting(trajectories, opts),
-		snips: selfNormalizedImportanceWeighting(trajectories, opts),
-		dr: doublyRobust(trajectories, opts)
-	};
-}
-function zeroEstimate() {
-	return {
-		value: 0,
-		standardError: 0,
-		effectiveSampleSize: 0,
-		n: 0,
-		maxImportanceWeight: 0
-	};
-}
-function clamp(x, lo, hi) {
-	if (!Number.isFinite(x)) return lo;
-	return Math.max(lo, Math.min(hi, x));
-}
-//#endregion
-export { selfNormalizedImportanceWeighting as i, inverseProbabilityWeighting as n, offPolicyEstimateAll as r, doublyRobust as t };
-
-//# sourceMappingURL=off-policy-DvgzvtIx.js.map

package/dist/off-policy-DvgzvtIx.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"off-policy-DvgzvtIx.js","names":[],"sources":["../src/rl/off-policy.ts"],"sourcesContent":["/**\n * Off-policy evaluation primitives.\n *\n * Standard inverse-probability-weighted (IPS), self-normalized\n * importance-weighted (SNIPS), and doubly-robust (DR) estimators for the\n * value of a *target* policy given trajectories collected under a\n * *behavior* policy. This is the canonical RL eval task: \"we have last\n * week's runs, we changed the policy — how would the new one do without\n * re-running?\"\n *\n * The math here is textbook (Dudík, Langford, Li 2011 for DR; Swaminathan\n * & Joachims 2015 for SNIPS) but the *application* to LLM-agent\n * evaluation needs care:\n *\n *   - The \"policy\" is the (prompt, tool config, model snapshot) triple.\n *     Two policies have the same probability over an action *iff* their\n *     LLM call would emit the same token with the same probability —\n *     which is generally unknowable without the model log-probs.\n *   - For LLM agents, propensity scores must be supplied by the caller\n *     (logged in the trace, recovered from token log-probs, or estimated\n *     via a learned propensity model). We do NOT estimate propensity here.\n *   - Doubly-robust requires two outputs from a Q-function: its prediction\n *     for the logged action and its expectation under the target policy.\n *     Consumers compute these with a tabular estimate, regression fit, or\n *     learned reward model before constructing the trajectories.\n *\n * Bias / variance tradeoffs:\n *   - IPS: unbiased; high variance for small overlap, infinite variance\n *     when target has support outside behavior.\n *   - SNIPS: lower variance, slight bias; usually preferred in practice.\n *   - DR: doubly-robust — unbiased if either propensity OR Q-function is\n *     correct. Lowest practical variance when Q is decent. Use this.\n *\n * Caveat the panel will land: on the LLM-agent setting, propensity scores\n * recovered from token log-probs are noisy, the action space is enormous,\n * and overlap is often poor. These estimators are useful but not magic;\n * complement with `replayCampaign` (exact replay where the request hashes\n * match) for high-confidence answers and OPE for the gap.\n */\n\nimport { ValidationError } from '../errors'\n\nexport interface OffPolicyTrajectory {\n  /** Stable id, for traceability through the dataset. */\n  runId: string\n  /** Reward observed under the behavior policy (the realized outcome). */\n  reward: number\n  /**\n   * Behavior-policy probability of the action that was taken. For LLM\n   * agents this is typically `exp(sum(token_log_probs))` over the chosen\n   * trajectory. Must be in (0, 1].\n   */\n  behaviorProb: number\n  /**\n   * Target-policy probability of the same action. For replay-style\n   * counterfactual evaluation this is what the *new* policy would have\n   * assigned to the *old* trajectory. Must be in [0, 1].\n   */\n  targetProb: number\n  /**\n   * Model-based reward prediction for the action selected by the behavior\n   * policy: `Q_hat(context, loggedAction)`. Supply this together with\n   * `vHatTarget` for contextual-bandit doubly-robust estimation.\n   */\n  qHatChosen?: number | null\n  /**\n   * Expected model-based reward under the target policy:\n   * `sum_action targetPolicy(action | context) * Q_hat(context, action)`.\n   * Supply this together with `qHatChosen`. For an honest evaluation, both\n   * values must come from a model cross-fitted or trained outside this row.\n   */\n  vHatTarget?: number | null\n}\n\nexport interface OffPolicyContributionCounts {\n  /** Contributions using the contextual-bandit doubly-robust formula. */\n  dr: number\n  /** Contributions using exact IPS because no reward-model estimate was supplied. */\n  ipsFallback: number\n}\n\nexport interface OffPolicyEstimate {\n  /** Estimated value of the target policy. */\n  value: number\n  /** Standard error of the estimate. */\n  standardError: number\n  /** Effective sample size (Kong 1992). Lower = more reliance on a few high-weight samples. */\n  effectiveSampleSize: number\n  /** Number of trajectories used. */\n  n: number\n  /**\n   * Diagnostic: maximum importance weight observed. Large values (>>10x\n   * mean) are a red flag — variance is dominated by a few outliers.\n   */\n  maxImportanceWeight: number\n  /** Populated by `doublyRobust` to expose which formula each row used. */\n  contributionCounts?: OffPolicyContributionCounts\n}\n\nexport interface OffPolicyOptions {\n  /**\n   * Cap importance weights at this value (Ionides 2008 truncated IS) to\n   * trade unbiasedness for variance reduction. Default `Infinity` (no cap).\n   * Set e.g. `10` for stable estimates when the policies are close.\n   */\n  weightCap?: number\n  /** Reward clipping range. Default `[0, 1]`. */\n  rewardClip?: { low: number; high: number }\n}\n\n/**\n * Inverse Probability Weighting (Horvitz-Thompson). Unbiased estimator\n * of E[reward under target policy]. Variance scales with the spread of\n * target/behavior ratios.\n */\nexport function inverseProbabilityWeighting(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n\n  if (trajectories.length === 0) {\n    return zeroEstimate()\n  }\n\n  const weights: number[] = []\n  const weightedRewards: number[] = []\n  let maxW = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(\n        `inverseProbabilityWeighting: behaviorProb must be > 0 (runId=${t.runId})`,\n      )\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    const r = clamp(t.reward, clip.low, clip.high)\n    weights.push(w)\n    weightedRewards.push(w * r)\n    if (w > maxW) maxW = w\n  }\n  const n = weights.length\n  const value = weightedRewards.reduce((s, x) => s + x, 0) / n\n  const variance = weightedRewards.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1)\n  const sumW = weights.reduce((s, w) => s + w, 0)\n  const sumW2 = weights.reduce((s, w) => s + w * w, 0)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n\n  return {\n    value,\n    standardError: Math.sqrt(variance / n),\n    effectiveSampleSize: effN,\n    n,\n    maxImportanceWeight: maxW,\n  }\n}\n\n/**\n * Self-Normalized Importance Sampling. Lower variance than vanilla IPS at\n * the cost of small bias (vanishing as N grows). The right default for\n * LLM-agent evaluation where overlap is often poor.\n */\nexport function selfNormalizedImportanceWeighting(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n  if (trajectories.length === 0) return zeroEstimate()\n\n  const weights: number[] = []\n  const rewards: number[] = []\n  let maxW = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(\n        `selfNormalizedImportanceWeighting: behaviorProb must be > 0 (runId=${t.runId})`,\n      )\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    weights.push(w)\n    rewards.push(clamp(t.reward, clip.low, clip.high))\n    if (w > maxW) maxW = w\n  }\n  const sumW = weights.reduce((s, w) => s + w, 0)\n  const sumWR = weights.reduce((s, w, i) => s + w * rewards[i]!, 0)\n  const value = sumW === 0 ? 0 : sumWR / sumW\n  const sumW2 = weights.reduce((s, w) => s + w * w, 0)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n  // Influence-function-based SE for SNIPS (Owen 2013, Ch. 9).\n  const phi = weights.map((w, i) => w * (rewards[i]! - value))\n  const variance = phi.reduce((s, x) => s + x * x, 0) / Math.max(1, sumW * sumW)\n  return {\n    value,\n    standardError: Math.sqrt(variance),\n    effectiveSampleSize: effN,\n    n: trajectories.length,\n    maxImportanceWeight: maxW,\n  }\n}\n\n/**\n * Doubly-robust off-policy estimator (Dudík, Langford, Li 2011).\n *\n *     V_DR = (1/N) * sum_i [ v_hat_target_i\n *              + (target_prob_i / behavior_prob_i) * (r_i - q_hat_chosen_i) ]\n *\n * Unbiased if EITHER:\n *   - the importance ratios are correct (IPS-style validity), OR\n *   - the Q-hat function is correct (model-based validity).\n *\n * In practice both are imperfect, but the residual bias is the *product*\n * of both errors — much smaller than either alone. This is why DR is the\n * default in production OPE pipelines.\n *\n * `qHatChosen` and `vHatTarget` must be supplied together. Rows with neither\n * use the exact IPS contribution. `contributionCounts` makes the mix explicit\n * in the result.\n * Callers must cross-fit the Q-function or train it on independent rows;\n * fitting and evaluating Q on the same outcomes leaks the answer.\n */\nexport function doublyRobust(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n  if (trajectories.length === 0) {\n    return {\n      ...zeroEstimate(),\n      contributionCounts: { dr: 0, ipsFallback: 0 },\n    }\n  }\n\n  const contributions: number[] = []\n  const contributionCounts: OffPolicyContributionCounts = {\n    dr: 0,\n    ipsFallback: 0,\n  }\n  let maxW = 0\n  let sumW = 0\n  let sumW2 = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(`doublyRobust: behaviorProb must be > 0 (runId=${t.runId})`)\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    const r = clamp(t.reward, clip.low, clip.high)\n    const rawQHatChosen = t.qHatChosen\n    const rawVHatTarget = t.vHatTarget\n    const hasQHatChosen = rawQHatChosen !== null && rawQHatChosen !== undefined\n    const hasVHatTarget = rawVHatTarget !== null && rawVHatTarget !== undefined\n    if (hasQHatChosen !== hasVHatTarget) {\n      throw new ValidationError(\n        `doublyRobust: qHatChosen and vHatTarget must be supplied together (runId=${t.runId})`,\n      )\n    }\n\n    if (hasQHatChosen && hasVHatTarget) {\n      if (!Number.isFinite(rawQHatChosen) || !Number.isFinite(rawVHatTarget)) {\n        throw new ValidationError(\n          `doublyRobust: qHatChosen and vHatTarget must be finite (runId=${t.runId})`,\n        )\n      }\n      const qHatChosen = clamp(rawQHatChosen, clip.low, clip.high)\n      const vHatTarget = clamp(rawVHatTarget, clip.low, clip.high)\n      contributions.push(vHatTarget + w * (r - qHatChosen))\n      contributionCounts.dr += 1\n    } else {\n      contributions.push(w * r)\n      contributionCounts.ipsFallback += 1\n    }\n    if (w > maxW) maxW = w\n    sumW += w\n    sumW2 += w * w\n  }\n  const n = contributions.length\n  const value = contributions.reduce((s, x) => s + x, 0) / n\n  const variance = contributions.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n  return {\n    value,\n    standardError: Math.sqrt(variance / n),\n    effectiveSampleSize: effN,\n    n,\n    maxImportanceWeight: maxW,\n    contributionCounts,\n  }\n}\n\n/**\n * Convenience: run all three estimators and return them side-by-side.\n * The recommended diagnostic — agreement across estimators is a much\n * stronger signal than any single one.\n */\nexport function offPolicyEstimateAll(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): { ips: OffPolicyEstimate; snips: OffPolicyEstimate; dr: OffPolicyEstimate } {\n  return {\n    ips: inverseProbabilityWeighting(trajectories, opts),\n    snips: selfNormalizedImportanceWeighting(trajectories, opts),\n    dr: doublyRobust(trajectories, opts),\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction zeroEstimate(): OffPolicyEstimate {\n  return { value: 0, standardError: 0, effectiveSampleSize: 0, n: 0, maxImportanceWeight: 0 }\n}\n\nfunction clamp(x: number, lo: number, hi: number): number {\n  if (!Number.isFinite(x)) return lo\n  return Math.max(lo, Math.min(hi, x))\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmHA,SAAgB,4BACd,cACA,OAAyB,CAAC,GACP;CACnB,MAAM,MAAM,KAAK,aAAa;CAC9B,MAAM,OAAO,KAAK,cAAc;EAAE,KAAK;EAAG,MAAM;CAAE;CAElD,IAAI,aAAa,WAAW,GAC1B,OAAO,aAAa;CAGtB,MAAM,UAAoB,CAAC;CAC3B,MAAM,kBAA4B,CAAC;CACnC,IAAI,OAAO;CACX,KAAK,MAAM,KAAK,cAAc;EAC5B,IAAI,EAAE,gBAAgB,GACpB,MAAM,IAAI,gBACR,gEAAgE,EAAE,MAAM,EAC1E;EAEF,MAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;EACrD,MAAM,IAAI,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI;EAC7C,QAAQ,KAAK,CAAC;EACd,gBAAgB,KAAK,IAAI,CAAC;EAC1B,IAAI,IAAI,MAAM,OAAO;CACvB;CACA,MAAM,IAAI,QAAQ;CAClB,MAAM,QAAQ,gBAAgB,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;CAC3D,MAAM,WAAW,gBAAgB,QAAQ,GAAG,MAAM,KAAK,IAAI,UAAU,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,IAAI,CAAC;CAC9F,MAAM,OAAO,QAAQ,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC;CAC9C,MAAM,QAAQ,QAAQ,QAAQ,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC;CACnD,MAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;CAE9C,OAAO;EACL;EACA,eAAe,KAAK,KAAK,WAAW,CAAC;EACrC,qBAAqB;EACrB;EACA,qBAAqB;CACvB;AACF;;;;;;AAOA,SAAgB,kCACd,cACA,OAAyB,CAAC,GACP;CACnB,MAAM,MAAM,KAAK,aAAa;CAC9B,MAAM,OAAO,KAAK,cAAc;EAAE,KAAK;EAAG,MAAM;CAAE;CAClD,IAAI,aAAa,WAAW,GAAG,OAAO,aAAa;CAEnD,MAAM,UAAoB,CAAC;CAC3B,MAAM,UAAoB,CAAC;CAC3B,IAAI,OAAO;CACX,KAAK,MAAM,KAAK,cAAc;EAC5B,IAAI,EAAE,gBAAgB,GACpB,MAAM,IAAI,gBACR,sEAAsE,EAAE,MAAM,EAChF;EAEF,MAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;EACrD,QAAQ,KAAK,CAAC;EACd,QAAQ,KAAK,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI,CAAC;EACjD,IAAI,IAAI,MAAM,OAAO;CACvB;CACA,MAAM,OAAO,QAAQ,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC;CAC9C,MAAM,QAAQ,QAAQ,QAAQ,GAAG,GAAG,MAAM,IAAI,IAAI,QAAQ,IAAK,CAAC;CAChE,MAAM,QAAQ,SAAS,IAAI,IAAI,QAAQ;CACvC,MAAM,QAAQ,QAAQ,QAAQ,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC;CACnD,MAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;CAG9C,MAAM,WADM,QAAQ,KAAK,GAAG,MAAM,KAAK,QAAQ,KAAM,MAClC,CAAC,CAAC,QAAQ,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,OAAO,IAAI;CAC7E,OAAO;EACL;EACA,eAAe,KAAK,KAAK,QAAQ;EACjC,qBAAqB;EACrB,GAAG,aAAa;EAChB,qBAAqB;CACvB;AACF;;;;;;;;;;;;;;;;;;;;;AAsBA,SAAgB,aACd,cACA,OAAyB,CAAC,GACP;CACnB,MAAM,MAAM,KAAK,aAAa;CAC9B,MAAM,OAAO,KAAK,cAAc;EAAE,KAAK;EAAG,MAAM;CAAE;CAClD,IAAI,aAAa,WAAW,GAC1B,OAAO;EACL,GAAG,aAAa;EAChB,oBAAoB;GAAE,IAAI;GAAG,aAAa;EAAE;CAC9C;CAGF,MAAM,gBAA0B,CAAC;CACjC,MAAM,qBAAkD;EACtD,IAAI;EACJ,aAAa;CACf;CACA,IAAI,OAAO;CACX,IAAI,OAAO;CACX,IAAI,QAAQ;CACZ,KAAK,MAAM,KAAK,cAAc;EAC5B,IAAI,EAAE,gBAAgB,GACpB,MAAM,IAAI,gBAAgB,iDAAiD,EAAE,MAAM,EAAE;EAEvF,MAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;EACrD,MAAM,IAAI,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI;EAC7C,MAAM,gBAAgB,EAAE;EACxB,MAAM,gBAAgB,EAAE;EACxB,MAAM,gBAAgB,kBAAkB,QAAQ,kBAAkB,KAAA;EAClE,MAAM,gBAAgB,kBAAkB,QAAQ,kBAAkB,KAAA;EAClE,IAAI,kBAAkB,eACpB,MAAM,IAAI,gBACR,4EAA4E,EAAE,MAAM,EACtF;EAGF,IAAI,iBAAiB,eAAe;GAClC,IAAI,CAAC,OAAO,SAAS,aAAa,KAAK,CAAC,OAAO,SAAS,aAAa,GACnE,MAAM,IAAI,gBACR,iEAAiE,EAAE,MAAM,EAC3E;GAEF,MAAM,aAAa,MAAM,eAAe,KAAK,KAAK,KAAK,IAAI;GAC3D,MAAM,aAAa,MAAM,eAAe,KAAK,KAAK,KAAK,IAAI;GAC3D,cAAc,KAAK,aAAa,KAAK,IAAI,WAAW;GACpD,mBAAmB,MAAM;EAC3B,OAAO;GACL,cAAc,KAAK,IAAI,CAAC;GACxB,mBAAmB,eAAe;EACpC;EACA,IAAI,IAAI,MAAM,OAAO;EACrB,QAAQ;EACR,SAAS,IAAI;CACf;CACA,MAAM,IAAI,cAAc;CACxB,MAAM,QAAQ,cAAc,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;CACzD,MAAM,WAAW,cAAc,QAAQ,GAAG,MAAM,KAAK,IAAI,UAAU,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,IAAI,CAAC;CAC5F,MAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;CAC9C,OAAO;EACL;EACA,eAAe,KAAK,KAAK,WAAW,CAAC;EACrC,qBAAqB;EACrB;EACA,qBAAqB;EACrB;CACF;AACF;;;;;;AAOA,SAAgB,qBACd,cACA,OAAyB,CAAC,GACmD;CAC7E,OAAO;EACL,KAAK,4BAA4B,cAAc,IAAI;EACnD,OAAO,kCAAkC,cAAc,IAAI;EAC3D,IAAI,aAAa,cAAc,IAAI;CACrC;AACF;AAIA,SAAS,eAAkC;CACzC,OAAO;EAAE,OAAO;EAAG,eAAe;EAAG,qBAAqB;EAAG,GAAG;EAAG,qBAAqB;CAAE;AAC5F;AAEA,SAAS,MAAM,GAAW,IAAY,IAAoB;CACxD,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;CAChC,OAAO,KAAK,IAAI,IAAI,KAAK,IAAI,IAAI,CAAC,CAAC;AACrC"}

package/dist/off-policy-mskQw8Mb.d.ts

@@ -1,153 +0,0 @@
-//#region src/rl/off-policy.d.ts
-/**
- * Off-policy evaluation primitives.
- *
- * Standard inverse-probability-weighted (IPS), self-normalized
- * importance-weighted (SNIPS), and doubly-robust (DR) estimators for the
- * value of a *target* policy given trajectories collected under a
- * *behavior* policy. This is the canonical RL eval task: "we have last
- * week's runs, we changed the policy — how would the new one do without
- * re-running?"
- *
- * The math here is textbook (Dudík, Langford, Li 2011 for DR; Swaminathan
- * & Joachims 2015 for SNIPS) but the *application* to LLM-agent
- * evaluation needs care:
- *
- *   - The "policy" is the (prompt, tool config, model snapshot) triple.
- *     Two policies have the same probability over an action *iff* their
- *     LLM call would emit the same token with the same probability —
- *     which is generally unknowable without the model log-probs.
- *   - For LLM agents, propensity scores must be supplied by the caller
- *     (logged in the trace, recovered from token log-probs, or estimated
- *     via a learned propensity model). We do NOT estimate propensity here.
- *   - Doubly-robust requires two outputs from a Q-function: its prediction
- *     for the logged action and its expectation under the target policy.
- *     Consumers compute these with a tabular estimate, regression fit, or
- *     learned reward model before constructing the trajectories.
- *
- * Bias / variance tradeoffs:
- *   - IPS: unbiased; high variance for small overlap, infinite variance
- *     when target has support outside behavior.
- *   - SNIPS: lower variance, slight bias; usually preferred in practice.
- *   - DR: doubly-robust — unbiased if either propensity OR Q-function is
- *     correct. Lowest practical variance when Q is decent. Use this.
- *
- * Caveat the panel will land: on the LLM-agent setting, propensity scores
- * recovered from token log-probs are noisy, the action space is enormous,
- * and overlap is often poor. These estimators are useful but not magic;
- * complement with `replayCampaign` (exact replay where the request hashes
- * match) for high-confidence answers and OPE for the gap.
- */
-interface OffPolicyTrajectory {
-  /** Stable id, for traceability through the dataset. */
-  runId: string;
-  /** Reward observed under the behavior policy (the realized outcome). */
-  reward: number;
-  /**
-   * Behavior-policy probability of the action that was taken. For LLM
-   * agents this is typically `exp(sum(token_log_probs))` over the chosen
-   * trajectory. Must be in (0, 1].
-   */
-  behaviorProb: number;
-  /**
-   * Target-policy probability of the same action. For replay-style
-   * counterfactual evaluation this is what the *new* policy would have
-   * assigned to the *old* trajectory. Must be in [0, 1].
-   */
-  targetProb: number;
-  /**
-   * Model-based reward prediction for the action selected by the behavior
-   * policy: `Q_hat(context, loggedAction)`. Supply this together with
-   * `vHatTarget` for contextual-bandit doubly-robust estimation.
-   */
-  qHatChosen?: number | null;
-  /**
-   * Expected model-based reward under the target policy:
-   * `sum_action targetPolicy(action | context) * Q_hat(context, action)`.
-   * Supply this together with `qHatChosen`. For an honest evaluation, both
-   * values must come from a model cross-fitted or trained outside this row.
-   */
-  vHatTarget?: number | null;
-}
-interface OffPolicyContributionCounts {
-  /** Contributions using the contextual-bandit doubly-robust formula. */
-  dr: number;
-  /** Contributions using exact IPS because no reward-model estimate was supplied. */
-  ipsFallback: number;
-}
-interface OffPolicyEstimate {
-  /** Estimated value of the target policy. */
-  value: number;
-  /** Standard error of the estimate. */
-  standardError: number;
-  /** Effective sample size (Kong 1992). Lower = more reliance on a few high-weight samples. */
-  effectiveSampleSize: number;
-  /** Number of trajectories used. */
-  n: number;
-  /**
-   * Diagnostic: maximum importance weight observed. Large values (>>10x
-   * mean) are a red flag — variance is dominated by a few outliers.
-   */
-  maxImportanceWeight: number;
-  /** Populated by `doublyRobust` to expose which formula each row used. */
-  contributionCounts?: OffPolicyContributionCounts;
-}
-interface OffPolicyOptions {
-  /**
-   * Cap importance weights at this value (Ionides 2008 truncated IS) to
-   * trade unbiasedness for variance reduction. Default `Infinity` (no cap).
-   * Set e.g. `10` for stable estimates when the policies are close.
-   */
-  weightCap?: number;
-  /** Reward clipping range. Default `[0, 1]`. */
-  rewardClip?: {
-    low: number;
-    high: number;
-  };
-}
-/**
- * Inverse Probability Weighting (Horvitz-Thompson). Unbiased estimator
- * of E[reward under target policy]. Variance scales with the spread of
- * target/behavior ratios.
- */
-declare function inverseProbabilityWeighting(trajectories: OffPolicyTrajectory[], opts?: OffPolicyOptions): OffPolicyEstimate;
-/**
- * Self-Normalized Importance Sampling. Lower variance than vanilla IPS at
- * the cost of small bias (vanishing as N grows). The right default for
- * LLM-agent evaluation where overlap is often poor.
- */
-declare function selfNormalizedImportanceWeighting(trajectories: OffPolicyTrajectory[], opts?: OffPolicyOptions): OffPolicyEstimate;
-/**
- * Doubly-robust off-policy estimator (Dudík, Langford, Li 2011).
- *
- *     V_DR = (1/N) * sum_i [ v_hat_target_i
- *              + (target_prob_i / behavior_prob_i) * (r_i - q_hat_chosen_i) ]
- *
- * Unbiased if EITHER:
- *   - the importance ratios are correct (IPS-style validity), OR
- *   - the Q-hat function is correct (model-based validity).
- *
- * In practice both are imperfect, but the residual bias is the *product*
- * of both errors — much smaller than either alone. This is why DR is the
- * default in production OPE pipelines.
- *
- * `qHatChosen` and `vHatTarget` must be supplied together. Rows with neither
- * use the exact IPS contribution. `contributionCounts` makes the mix explicit
- * in the result.
- * Callers must cross-fit the Q-function or train it on independent rows;
- * fitting and evaluating Q on the same outcomes leaks the answer.
- */
-declare function doublyRobust(trajectories: OffPolicyTrajectory[], opts?: OffPolicyOptions): OffPolicyEstimate;
-/**
- * Convenience: run all three estimators and return them side-by-side.
- * The recommended diagnostic — agreement across estimators is a much
- * stronger signal than any single one.
- */
-declare function offPolicyEstimateAll(trajectories: OffPolicyTrajectory[], opts?: OffPolicyOptions): {
-  ips: OffPolicyEstimate;
-  snips: OffPolicyEstimate;
-  dr: OffPolicyEstimate;
-};
-//#endregion
-export { doublyRobust as a, selfNormalizedImportanceWeighting as c, OffPolicyTrajectory as i, OffPolicyEstimate as n, inverseProbabilityWeighting as o, OffPolicyOptions as r, offPolicyEstimateAll as s, OffPolicyContributionCounts as t };
-//# sourceMappingURL=off-policy-mskQw8Mb.d.ts.map

package/dist/off-policy-mskQw8Mb.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"off-policy-mskQw8Mb.d.ts","names":[],"sources":["../src/rl/off-policy.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UA0CiB;;EAEf;;EAEA;;;;;;EAMA;;;;;;EAMA;;;;;;EAMA;;;;;;;EAOA;;UAGe;;EAEf;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;EAKA;;EAEA,qBAAqB;;UAGN;;;;;;EAMf;;EAEA;IAAe;IAAa;;;;;;;;iBAQd,4BACd,cAAc,uBACd,OAAM,mBACL;;;;;;iBA4Ca,kCACd,cAAc,uBACd,OAAM,mBACL;;;;;;;;;;;;;;;;;;;;;iBAwDa,aACd,cAAc,uBACd,OAAM,mBACL;;;;;;iBAuEa,qBACd,cAAc,uBACd,OAAM;EACH,KAAK;EAAmB,OAAO;EAAmB,IAAI"}

package/dist/pre-registration-DakwTRXk.js

@@ -1,96 +0,0 @@
-//#region src/pre-registration.ts
-/**
-* Deterministic JSON canonicalization — sort object keys recursively.
-*
-* Two semantically-equal objects produce byte-identical canonicalized output;
-* this is what makes a content-hash stable across encoders, key insertion
-* orders, and runtime versions. Exported for any consumer that needs the same
-* canonicalization guarantee outside the manifest-signing path (e.g., signing
-* an artifact bundle, hashing a dataset version, etc.).
-*/
-function canonicalize(v) {
-	if (v === null || typeof v !== "object") return v;
-	if (Array.isArray(v)) return v.map(canonicalize);
-	const keys = Object.keys(v).sort();
-	const out = {};
-	for (const k of keys) out[k] = canonicalize(v[k]);
-	return out;
-}
-/**
-* SHA-256 hex (full 64 chars) over the canonicalized JSON encoding of `obj`.
-*
-* The same primitive `signManifest` and `verifyManifest` are built on, exposed
-* directly so consumers signing arbitrary structured content (artifact bundles,
-* production packets, dataset manifests, etc.) don't have to re-derive
-* canonicalize+sha256 from scratch.
-*
-* Stable across:
-*   - object key insertion order (canonicalization sorts keys recursively)
-*   - encoder choice (UTF-8 via TextEncoder, fixed)
-*   - runtime (uses the Web Crypto subtle digest, present in Node ≥18 and browsers)
-*
-* Named `hashJson` to disambiguate from `prompt-registry.ts`'s `hashContent`,
-* which takes a string input and returns a truncated 12-char prompt id.
-* Use `hashJson` when you mean "canonicalize then hash."
-*
-* @example
-*   const hash = await hashJson({ id: '1', kind: 'spec' })
-*   // 'a3f1...' (64 hex chars)
-*/
-async function hashJson(obj) {
-	const canonical = canonicalize(obj);
-	const bytes = new TextEncoder().encode(JSON.stringify(canonical));
-	const digest = await globalThis.crypto.subtle.digest("SHA-256", bytes);
-	return Array.from(new Uint8Array(digest)).map((b) => b.toString(16).padStart(2, "0")).join("");
-}
-/**
-* Sign a manifest with a SHA-256 content hash.
-*
-* The hash covers the canonicalized manifest with the `contentHash`
-* and `algo` fields stripped; this lets verifiers re-sign the rest and
-* compare. Returned manifest always carries `algo: 'sha256-content'`
-* so downstream consumers can identify the scheme; manifests without
-* `algo` still verify because it is stripped before hashing on both sides.
-*/
-async function signManifest(m) {
-	const hash = await hashJson(m);
-	return {
-		...m,
-		contentHash: hash,
-		algo: "sha256-content"
-	};
-}
-/**
-* Verify that a signed manifest has not been tampered with.
-*
-* Strips `contentHash` and `algo` before re-signing so manifests without
-* `algo` verify identically to ones that carry it.
-*/
-async function verifyManifest(m) {
-	const { contentHash, algo: _algo, ...rest } = m;
-	return (await signManifest(rest)).contentHash === contentHash;
-}
-/**
-* Evaluate a pre-registered hypothesis against observed results.
-* Mechanical — no re-interpretation permitted.
-*/
-async function evaluateHypothesis(manifest, observed) {
-	if (!await verifyManifest(manifest)) throw new Error("evaluateHypothesis: manifest content hash mismatch (tampered)");
-	const reasons = [];
-	if (!(manifest.direction === "increase" ? observed.effect > 0 : observed.effect < 0)) reasons.push("wrong_direction");
-	if (Math.abs(observed.effect) < manifest.minEffect) reasons.push("effect_too_small");
-	if (observed.pValue >= manifest.alpha) reasons.push("not_significant");
-	if (observed.n < manifest.preRegisteredN) reasons.push("undersampled");
-	return {
-		manifest,
-		observedN: observed.n,
-		observedEffect: observed.effect,
-		observedPValue: observed.pValue,
-		confirmed: reasons.length === 0,
-		rejectionReasons: reasons
-	};
-}
-//#endregion
-export { verifyManifest as a, signManifest as i, evaluateHypothesis as n, hashJson as r, canonicalize as t };
-
-//# sourceMappingURL=pre-registration-DakwTRXk.js.map

package/dist/pre-registration-DakwTRXk.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"pre-registration-DakwTRXk.js","names":[],"sources":["../src/pre-registration.ts"],"sourcesContent":["/**\n * Pre-registered hypotheses — declare what you're testing BEFORE the\n * run, check it AFTER. Prevents p-hacking, optional stopping, and the\n * \"we ran until it looked good\" failure mode.\n *\n * Manifest is a plain JSON-friendly object. Sign it with a content hash\n * + timestamp; the registered record becomes immutable. Post-run,\n * evaluate the manifest against observed results — the library refuses\n * to let you re-interpret a different metric as the declared one.\n */\n\nexport interface HypothesisManifest {\n  id: string\n  /** Human prose — goes into the audit trail. */\n  hypothesis: string\n  /** Metric the hypothesis claims to move. */\n  metric: string\n  /** 'increase' = candidate should score higher than baseline; 'decrease' = lower. */\n  direction: 'increase' | 'decrease'\n  /** Minimum effect size to count (same units as the metric). */\n  minEffect: number\n  /** Alpha threshold. */\n  alpha: number\n  /** Target statistical power at which sample size was pre-computed. */\n  power: number\n  /** Declared N per arm before running. */\n  preRegisteredN: number\n  /** ISO8601 timestamp the manifest was registered. */\n  registeredAt: string\n  /** Optional identifiers to tie into the trace corpus. */\n  baselineLabel?: string\n  candidateLabel?: string\n}\n\n/**\n * Identifier for the hashing scheme used to produce `contentHash`.\n *\n * `'sha256-content'` — sha256 hex over the canonicalized manifest with\n * the `contentHash` and `algo` fields stripped. Held as a string union\n * so future schemes can be added without breaking parsers; SignedManifest\n * values without `algo` deserialize cleanly because the field is optional.\n */\nexport type SignedManifestAlgo = 'sha256-content'\n\nexport interface SignedManifest extends HypothesisManifest {\n  /** sha256 hex of canonicalized manifest (everything except contentHash and algo). */\n  contentHash: string\n  /**\n   * Algorithm string describing how `contentHash` was produced.\n   *\n   * Optional on the type so serialized manifests without it still parse,\n   * but ALWAYS populated by {@link signManifest}. Consumers that want to\n   * enforce a known algorithm should reject manifests where this field\n   * is missing or unrecognized.\n   */\n  algo?: SignedManifestAlgo\n}\n\nexport interface HypothesisResult {\n  manifest: SignedManifest\n  observedN: number\n  observedEffect: number\n  observedPValue: number\n  /** True iff the observed effect hits the pre-declared direction with\n   *  magnitude ≥ minEffect AND p < alpha. */\n  confirmed: boolean\n  /** Enumerated reasons the hypothesis was rejected (each a machine-tag). */\n  rejectionReasons: Array<\n    'wrong_direction' | 'effect_too_small' | 'not_significant' | 'undersampled'\n  >\n  notes?: string\n}\n\n/**\n * Deterministic JSON canonicalization — sort object keys recursively.\n *\n * Two semantically-equal objects produce byte-identical canonicalized output;\n * this is what makes a content-hash stable across encoders, key insertion\n * orders, and runtime versions. Exported for any consumer that needs the same\n * canonicalization guarantee outside the manifest-signing path (e.g., signing\n * an artifact bundle, hashing a dataset version, etc.).\n */\nexport function canonicalize(v: unknown): unknown {\n  if (v === null || typeof v !== 'object') return v\n  if (Array.isArray(v)) return v.map(canonicalize)\n  const keys = Object.keys(v as Record<string, unknown>).sort()\n  const out: Record<string, unknown> = {}\n  for (const k of keys) out[k] = canonicalize((v as Record<string, unknown>)[k])\n  return out\n}\n\n/**\n * SHA-256 hex (full 64 chars) over the canonicalized JSON encoding of `obj`.\n *\n * The same primitive `signManifest` and `verifyManifest` are built on, exposed\n * directly so consumers signing arbitrary structured content (artifact bundles,\n * production packets, dataset manifests, etc.) don't have to re-derive\n * canonicalize+sha256 from scratch.\n *\n * Stable across:\n *   - object key insertion order (canonicalization sorts keys recursively)\n *   - encoder choice (UTF-8 via TextEncoder, fixed)\n *   - runtime (uses the Web Crypto subtle digest, present in Node ≥18 and browsers)\n *\n * Named `hashJson` to disambiguate from `prompt-registry.ts`'s `hashContent`,\n * which takes a string input and returns a truncated 12-char prompt id.\n * Use `hashJson` when you mean \"canonicalize then hash.\"\n *\n * @example\n *   const hash = await hashJson({ id: '1', kind: 'spec' })\n *   // 'a3f1...' (64 hex chars)\n */\nexport async function hashJson<T>(obj: T): Promise<string> {\n  const canonical = canonicalize(obj)\n  const bytes = new TextEncoder().encode(JSON.stringify(canonical))\n  const digest = await globalThis.crypto.subtle.digest('SHA-256', bytes)\n  return Array.from(new Uint8Array(digest))\n    .map((b) => b.toString(16).padStart(2, '0'))\n    .join('')\n}\n\n/**\n * Sign a manifest with a SHA-256 content hash.\n *\n * The hash covers the canonicalized manifest with the `contentHash`\n * and `algo` fields stripped; this lets verifiers re-sign the rest and\n * compare. Returned manifest always carries `algo: 'sha256-content'`\n * so downstream consumers can identify the scheme; manifests without\n * `algo` still verify because it is stripped before hashing on both sides.\n */\nexport async function signManifest(m: HypothesisManifest): Promise<SignedManifest> {\n  const hash = await hashJson(m)\n  return { ...m, contentHash: hash, algo: 'sha256-content' }\n}\n\n/**\n * Verify that a signed manifest has not been tampered with.\n *\n * Strips `contentHash` and `algo` before re-signing so manifests without\n * `algo` verify identically to ones that carry it.\n */\nexport async function verifyManifest(m: SignedManifest): Promise<boolean> {\n  const { contentHash, algo: _algo, ...rest } = m\n  void _algo\n  const resigned = await signManifest(rest)\n  return resigned.contentHash === contentHash\n}\n\n/**\n * Evaluate a pre-registered hypothesis against observed results.\n * Mechanical — no re-interpretation permitted.\n */\nexport async function evaluateHypothesis(\n  manifest: SignedManifest,\n  observed: { n: number; effect: number; pValue: number },\n): Promise<HypothesisResult> {\n  if (!(await verifyManifest(manifest))) {\n    throw new Error('evaluateHypothesis: manifest content hash mismatch (tampered)')\n  }\n  const reasons: HypothesisResult['rejectionReasons'] = []\n  const directionOk = manifest.direction === 'increase' ? observed.effect > 0 : observed.effect < 0\n  if (!directionOk) reasons.push('wrong_direction')\n  if (Math.abs(observed.effect) < manifest.minEffect) reasons.push('effect_too_small')\n  if (observed.pValue >= manifest.alpha) reasons.push('not_significant')\n  if (observed.n < manifest.preRegisteredN) reasons.push('undersampled')\n  return {\n    manifest,\n    observedN: observed.n,\n    observedEffect: observed.effect,\n    observedPValue: observed.pValue,\n    confirmed: reasons.length === 0,\n    rejectionReasons: reasons,\n  }\n}\n"],"mappings":";;;;;;;;;;AAkFA,SAAgB,aAAa,GAAqB;CAChD,IAAI,MAAM,QAAQ,OAAO,MAAM,UAAU,OAAO;CAChD,IAAI,MAAM,QAAQ,CAAC,GAAG,OAAO,EAAE,IAAI,YAAY;CAC/C,MAAM,OAAO,OAAO,KAAK,CAA4B,CAAC,CAAC,KAAK;CAC5D,MAAM,MAA+B,CAAC;CACtC,KAAK,MAAM,KAAK,MAAM,IAAI,KAAK,aAAc,EAA8B,EAAE;CAC7E,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;;AAuBA,eAAsB,SAAY,KAAyB;CACzD,MAAM,YAAY,aAAa,GAAG;CAClC,MAAM,QAAQ,IAAI,YAAY,CAAC,CAAC,OAAO,KAAK,UAAU,SAAS,CAAC;CAChE,MAAM,SAAS,MAAM,WAAW,OAAO,OAAO,OAAO,WAAW,KAAK;CACrE,OAAO,MAAM,KAAK,IAAI,WAAW,MAAM,CAAC,CAAC,CACtC,KAAK,MAAM,EAAE,SAAS,EAAE,CAAC,CAAC,SAAS,GAAG,GAAG,CAAC,CAAC,CAC3C,KAAK,EAAE;AACZ;;;;;;;;;;AAWA,eAAsB,aAAa,GAAgD;CACjF,MAAM,OAAO,MAAM,SAAS,CAAC;CAC7B,OAAO;EAAE,GAAG;EAAG,aAAa;EAAM,MAAM;CAAiB;AAC3D;;;;;;;AAQA,eAAsB,eAAe,GAAqC;CACxE,MAAM,EAAE,aAAa,MAAM,OAAO,GAAG,SAAS;CAG9C,QAAO,MADgB,aAAa,IAAI,EAAA,CACxB,gBAAgB;AAClC;;;;;AAMA,eAAsB,mBACpB,UACA,UAC2B;CAC3B,IAAI,CAAE,MAAM,eAAe,QAAQ,GACjC,MAAM,IAAI,MAAM,+DAA+D;CAEjF,MAAM,UAAgD,CAAC;CAEvD,IAAI,EADgB,SAAS,cAAc,aAAa,SAAS,SAAS,IAAI,SAAS,SAAS,IAC9E,QAAQ,KAAK,iBAAiB;CAChD,IAAI,KAAK,IAAI,SAAS,MAAM,IAAI,SAAS,WAAW,QAAQ,KAAK,kBAAkB;CACnF,IAAI,SAAS,UAAU,SAAS,OAAO,QAAQ,KAAK,iBAAiB;CACrE,IAAI,SAAS,IAAI,SAAS,gBAAgB,QAAQ,KAAK,cAAc;CACrE,OAAO;EACL;EACA,WAAW,SAAS;EACpB,gBAAgB,SAAS;EACzB,gBAAgB,SAAS;EACzB,WAAW,QAAQ,WAAW;EAC9B,kBAAkB;CACpB;AACF"}

package/dist/profile-cell.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"profile-cell.js","names":[],"sources":["../src/agent-profile-cell.ts"],"sourcesContent":["import type { AgentProfile } from '@tangle-network/agent-interface'\nimport { ValidationError } from './errors'\nimport { hashJson } from './pre-registration'\n\nexport type AgentProfileCellSchemaVersion = 'agent-profile-cell/v1'\n\nexport type AgentProfileJsonObject = { [key: string]: AgentProfileJson }\n\nexport type AgentProfileJson =\n  | string\n  | number\n  | boolean\n  | null\n  | AgentProfileJson[]\n  | AgentProfileJsonObject\n\nexport type AgentProfileDimensionValue = string | number | boolean | null\n\nexport interface AgentProfileSource {\n  /** Runtime/profile contract being fingerprinted, e.g. `agent-interface-profile`. */\n  kind: string\n  /** sha256 over the canonical source profile object. */\n  hash: string\n}\n\nexport interface AgentProfileSourceInput {\n  kind: string\n  /** Precomputed sha256 for callers that already sign their profile artifact. */\n  hash?: string\n  /** Full canonical runtime profile; hashed and then discarded from the cell. */\n  profile?: AgentProfileJson\n}\n\nexport interface AgentProfileHarness {\n  id: string\n  version?: string\n  hash?: string\n}\n\nexport interface AgentProfileCellInput {\n  profileId: string\n  sourceProfile: AgentProfileSourceInput\n  harness?: AgentProfileHarness\n  model?: string\n  promptHash?: string\n  dimensions?: Record<string, AgentProfileDimensionValue>\n}\n\nexport interface AgentProfileCell {\n  schemaVersion: AgentProfileCellSchemaVersion\n  cellId: string\n  profileId: string\n  sourceProfile: AgentProfileSource\n  harness?: AgentProfileHarness\n  model?: string\n  promptHash?: string\n  dimensions?: Record<string, AgentProfileDimensionValue>\n}\n\nexport class AgentProfileCellValidationError extends ValidationError {\n  readonly path: string\n  constructor(message: string, path = '') {\n    super(path ? `${message} (at ${path})` : message)\n    this.path = path\n  }\n}\n\nconst SHA256_HEX = /^[0-9a-f]{64}$/\nconst CELL_ID = /^agent-profile-cell:sha256:[0-9a-f]{64}$/\n\nexport async function buildAgentProfileCell(\n  input: AgentProfileCellInput,\n): Promise<AgentProfileCell> {\n  const material = await normalizeAgentProfileCellInput(input)\n  const cellId = `agent-profile-cell:sha256:${await hashJson(material)}`\n  return { ...material, cellId }\n}\n\nexport function agentProfileCellHashMaterial(\n  cell: AgentProfileCell,\n): Omit<AgentProfileCell, 'cellId'> {\n  const { cellId: _cellId, ...material } = cell\n  void _cellId\n  return normalizeAgentProfileCell(material)\n}\n\n/**\n * Verify an `AgentProfileCell`'s `cellId` matches the sha256 of its hash-material fields, confirming the record has not been tampered with.\n */\nexport async function verifyAgentProfileCell(cell: AgentProfileCell): Promise<boolean> {\n  validateAgentProfileCell(cell)\n  return (\n    cell.cellId ===\n    `agent-profile-cell:sha256:${await hashJson(agentProfileCellHashMaterial(cell))}`\n  )\n}\n\nexport function validateAgentProfileCell(input: unknown): AgentProfileCell {\n  if (input === null || typeof input !== 'object') {\n    throw new AgentProfileCellValidationError('expected object')\n  }\n  const obj = input as Record<string, unknown>\n  expectLiteral(obj.schemaVersion, 'agent-profile-cell/v1', 'schemaVersion')\n  if (typeof obj.cellId !== 'string' || !CELL_ID.test(obj.cellId)) {\n    throw new AgentProfileCellValidationError(\n      'cellId must match agent-profile-cell:sha256:<64 lowercase hex chars>',\n      'cellId',\n    )\n  }\n  expectString(obj.profileId, 'profileId')\n  validateSource(obj.sourceProfile, 'sourceProfile')\n  if (obj.harness !== undefined) validateHarness(obj.harness, 'harness')\n  if (obj.model !== undefined) expectString(obj.model, 'model')\n  if (obj.promptHash !== undefined) expectString(obj.promptHash, 'promptHash')\n  if (obj.dimensions !== undefined) validateDimensions(obj.dimensions, 'dimensions')\n  return input as AgentProfileCell\n}\n\nexport function requireAgentProfileCell(record: {\n  runId: string\n  agentProfile?: AgentProfileCell\n}): AgentProfileCell {\n  if (!record.agentProfile) {\n    throw new AgentProfileCellValidationError(\n      `run \"${record.runId}\" is missing agentProfile; profile-cell grouping requires explicit profile identity`,\n      'agentProfile',\n    )\n  }\n  return validateAgentProfileCell(record.agentProfile)\n}\n\nexport function agentProfileCellKey(record: {\n  runId: string\n  agentProfile?: AgentProfileCell\n}): string {\n  return requireAgentProfileCell(record).cellId\n}\n\nexport async function assertRunAgentProfileCell(record: {\n  runId: string\n  model: string\n  promptHash: string\n  agentProfile?: AgentProfileCell\n}): Promise<AgentProfileCell> {\n  const profile = requireAgentProfileCell(record)\n  if (!(await verifyAgentProfileCell(profile))) {\n    throw new AgentProfileCellValidationError(\n      `run \"${record.runId}\" has an agentProfile.cellId that does not match its content`,\n      'agentProfile.cellId',\n    )\n  }\n  if (profile.model !== undefined && profile.model !== record.model) {\n    throw new AgentProfileCellValidationError(\n      `run \"${record.runId}\" agentProfile.model \"${profile.model}\" does not match model \"${record.model}\"`,\n      'agentProfile.model',\n    )\n  }\n  if (profile.promptHash !== undefined && profile.promptHash !== record.promptHash) {\n    throw new AgentProfileCellValidationError(\n      `run \"${record.runId}\" agentProfile.promptHash \"${profile.promptHash}\" does not match promptHash \"${record.promptHash}\"`,\n      'agentProfile.promptHash',\n    )\n  }\n  return profile\n}\n\nexport function groupRunsByAgentProfileCell<\n  T extends { runId: string; agentProfile?: AgentProfileCell },\n>(records: readonly T[]): Map<string, T[]> {\n  const groups = new Map<string, T[]>()\n  for (const record of records) {\n    const key = agentProfileCellKey(record)\n    const bucket = groups.get(key)\n    if (bucket) bucket.push(record)\n    else groups.set(key, [record])\n  }\n  return groups\n}\n\nasync function normalizeAgentProfileCellInput(\n  input: AgentProfileCellInput,\n): Promise<Omit<AgentProfileCell, 'cellId'>> {\n  return normalizeAgentProfileCell({\n    schemaVersion: 'agent-profile-cell/v1',\n    profileId: input.profileId,\n    sourceProfile: await normalizeSourceInput(input.sourceProfile),\n    harness: input.harness,\n    model: input.model,\n    promptHash: input.promptHash,\n    dimensions: input.dimensions,\n  })\n}\n\nfunction normalizeAgentProfileCell(\n  input: Omit<AgentProfileCell, 'cellId'>,\n): Omit<AgentProfileCell, 'cellId'> {\n  return compactObject({\n    schemaVersion: 'agent-profile-cell/v1' as const,\n    profileId: requireNonEmpty(input.profileId, 'profileId'),\n    sourceProfile: normalizeSource(input.sourceProfile),\n    harness: input.harness ? normalizeHarness(input.harness, 'harness') : undefined,\n    model: optionalNonEmpty(input.model, 'model'),\n    promptHash: optionalNonEmpty(input.promptHash, 'promptHash'),\n    dimensions: input.dimensions\n      ? nonEmptyRecord(normalizeDimensions(input.dimensions))\n      : undefined,\n  })\n}\n\nasync function normalizeSourceInput(input: AgentProfileSourceInput): Promise<AgentProfileSource> {\n  const kind = requireNonEmpty(input.kind, 'sourceProfile.kind')\n  if (input.hash !== undefined && input.profile !== undefined) {\n    throw new AgentProfileCellValidationError(\n      'sourceProfile must provide either hash or profile, not both',\n      'sourceProfile',\n    )\n  }\n  if (input.hash !== undefined) {\n    return { kind, hash: requireSha256Hex(input.hash, 'sourceProfile.hash') }\n  }\n  if (input.profile === undefined) {\n    throw new AgentProfileCellValidationError(\n      'sourceProfile must provide hash or profile',\n      'sourceProfile',\n    )\n  }\n  assertJson(input.profile, 'sourceProfile.profile')\n  return { kind, hash: await hashJson(input.profile) }\n}\n\nfunction normalizeSource(input: AgentProfileSource): AgentProfileSource {\n  return {\n    kind: requireNonEmpty(input.kind, 'sourceProfile.kind'),\n    hash: requireSha256Hex(input.hash, 'sourceProfile.hash'),\n  }\n}\n\nfunction normalizeHarness(input: AgentProfileHarness, path: string): AgentProfileHarness {\n  return compactObject({\n    id: requireNonEmpty(input.id, `${path}.id`),\n    version: optionalNonEmpty(input.version, `${path}.version`),\n    hash: optionalNonEmpty(input.hash, `${path}.hash`),\n  })\n}\n\nfunction normalizeDimensions(\n  input: Record<string, AgentProfileDimensionValue>,\n): Record<string, AgentProfileDimensionValue> {\n  const out: Record<string, AgentProfileDimensionValue> = {}\n  for (const key of Object.keys(input).sort()) {\n    const value = input[key]\n    requireNonEmpty(key, 'dimensions.<key>')\n    if (\n      value !== null &&\n      typeof value !== 'string' &&\n      typeof value !== 'number' &&\n      typeof value !== 'boolean'\n    ) {\n      throw new AgentProfileCellValidationError(\n        'expected primitive dimension value',\n        `dimensions.${key}`,\n      )\n    }\n    if (typeof value === 'number' && !Number.isFinite(value)) {\n      throw new AgentProfileCellValidationError('expected finite number', `dimensions.${key}`)\n    }\n    out[key] = value\n  }\n  return out\n}\n\nfunction compactObject<T extends Record<string, unknown>>(input: T): T {\n  const out: Record<string, unknown> = {}\n  for (const [key, value] of Object.entries(input)) {\n    if (value !== undefined) out[key] = value\n  }\n  return out as T\n}\n\nfunction nonEmptyRecord<T extends Record<string, unknown>>(input: T): T | undefined {\n  return Object.keys(input).length > 0 ? input : undefined\n}\n\nfunction validateSource(value: unknown, path: string): void {\n  if (value === null || typeof value !== 'object' || Array.isArray(value)) {\n    throw new AgentProfileCellValidationError('expected object', path)\n  }\n  const rec = value as Record<string, unknown>\n  expectString(rec.kind, `${path}.kind`)\n  requireSha256Hex(rec.hash, `${path}.hash`)\n}\n\nfunction validateHarness(value: unknown, path: string): void {\n  if (value === null || typeof value !== 'object' || Array.isArray(value)) {\n    throw new AgentProfileCellValidationError('expected object', path)\n  }\n  const rec = value as Record<string, unknown>\n  expectString(rec.id, `${path}.id`)\n  if (rec.version !== undefined) expectString(rec.version, `${path}.version`)\n  if (rec.hash !== undefined) expectString(rec.hash, `${path}.hash`)\n}\n\nfunction validateDimensions(value: unknown, path: string): void {\n  if (value === null || typeof value !== 'object' || Array.isArray(value)) {\n    throw new AgentProfileCellValidationError('expected object', path)\n  }\n  normalizeDimensions(value as Record<string, AgentProfileDimensionValue>)\n}\n\nfunction assertJson(value: AgentProfileJson, path: string): void {\n  if (value === null) return\n  const type = typeof value\n  if (type === 'string' || type === 'boolean') return\n  if (type === 'number') {\n    if (!Number.isFinite(value)) {\n      throw new AgentProfileCellValidationError('expected finite number', path)\n    }\n    return\n  }\n  if (Array.isArray(value)) {\n    value.forEach((item, index) => {\n      assertJson(item, `${path}[${index}]`)\n    })\n    return\n  }\n  if (type === 'object') {\n    for (const [key, nested] of Object.entries(value)) {\n      requireNonEmpty(key, `${path}.<key>`)\n      assertJson(nested, `${path}.${key}`)\n    }\n    return\n  }\n  throw new AgentProfileCellValidationError('expected JSON-compatible value', path)\n}\n\nfunction expectLiteral(value: unknown, expected: string, path: string): void {\n  if (value !== expected) {\n    throw new AgentProfileCellValidationError(`expected ${expected}`, path)\n  }\n}\n\nfunction expectString(value: unknown, path: string): void {\n  if (typeof value !== 'string' || value.length === 0) {\n    throw new AgentProfileCellValidationError('expected non-empty string', path)\n  }\n}\n\nfunction requireNonEmpty(value: string, path: string): string {\n  if (typeof value !== 'string' || value.length === 0) {\n    throw new AgentProfileCellValidationError('expected non-empty string', path)\n  }\n  return value\n}\n\nfunction optionalNonEmpty(value: string | undefined, path: string): string | undefined {\n  if (value === undefined) return undefined\n  return requireNonEmpty(value, path)\n}\n\nfunction requireSha256Hex(value: unknown, path: string): string {\n  if (typeof value !== 'string' || !SHA256_HEX.test(value)) {\n    throw new AgentProfileCellValidationError('expected 64 lowercase sha256 hex chars', path)\n  }\n  return value\n}\n\n// ── Consumer helpers ─────────────────────────────────────────────────\n//\n// Boilerplate every product consuming `buildAgentProfileCell` used to duplicate:\n//\n//   1. A `JSON.parse(JSON.stringify(value))` helper that canonicalizes an\n//      arbitrary `@tangle-network/agent-interface` `AgentProfile` into the recursive\n//      `AgentProfileJson` shape, with a fail-loud error when the profile\n//      is not JSON-serializable.\n//\n//   2. The magic string `'agent-interface-profile'` for `sourceProfile.kind`.\n//\n// Both belong here so the cross-product cell join (same canonical profile\n// hashes to the same `sourceProfile.hash` across products) is enforced by\n// the type system, not by every consumer remembering to do it right.\n// See blueprint-agent issue tangle-network/agent-eval#82.\n\n/** Canonical `sourceProfile.kind` values. Two products fingerprinting the\n *  same canonical profile MUST use the same kind for their cells to share\n *  `sourceProfile.hash`. Extend rather than create new strings — adding a\n *  new kind is a deliberate cross-product schema change. */\nexport const AGENT_PROFILE_KINDS = {\n  /** A profile declared via `defineAgentProfile(...)` from\n   *  `@tangle-network/agent-interface`. The default kind for router-backed\n   *  and sandbox-backed products. */\n  AGENT_INTERFACE_PROFILE: 'agent-interface-profile',\n} as const\n\nexport type AgentProfileKind = (typeof AGENT_PROFILE_KINDS)[keyof typeof AGENT_PROFILE_KINDS]\n\n/** Canonicalize an arbitrary value into `AgentProfileJson` by JSON\n *  round-trip. Throws when the value contains anything not representable\n *  as JSON (functions, BigInt, cycles) — non-portable profiles fail loud\n *  rather than silently dropping fields. */\nexport function toAgentProfileJson(value: unknown): AgentProfileJson {\n  let serialized: string | undefined\n  try {\n    serialized = JSON.stringify(value)\n  } catch (err) {\n    throw new AgentProfileCellValidationError(\n      `agent profile must be JSON-serializable: ${err instanceof Error ? err.message : String(err)}`,\n      'sourceProfile.profile',\n    )\n  }\n  if (serialized === undefined) {\n    throw new AgentProfileCellValidationError(\n      'agent profile must be JSON-serializable (got undefined after JSON.stringify)',\n      'sourceProfile.profile',\n    )\n  }\n  return JSON.parse(serialized) as AgentProfileJson\n}\n\n/** Canonical AgentProfile shape required when deriving a stable cell id. */\nexport type AgentInterfaceProfileLike = AgentProfile & { name: string; version: string }\n\n/** Higher-level helper that hard-codes the canonical\n *  `agent-interface-profile` kind plus the JSON canonicalization. Equivalent\n *  to calling `buildAgentProfileCell` with `profileId = \\`${name}@${version}\\``\n *  and `sourceProfile = { kind: AGENT_INTERFACE_PROFILE, profile: <round-tripped> }`.\n *\n *  Use this from any product consuming an agent-interface `AgentProfile`; the\n *  manual `buildAgentProfileCell` call is reserved for advanced cases\n *  (custom kinds, pre-computed source hashes, alternate profileId\n *  conventions). */\nexport async function buildAgentInterfaceProfileCell(\n  profile: AgentInterfaceProfileLike,\n  input: Omit<AgentProfileCellInput, 'profileId' | 'sourceProfile'>,\n): Promise<AgentProfileCell> {\n  if (!profile || typeof profile !== 'object') {\n    throw new AgentProfileCellValidationError('AgentProfile must be an object', 'profile')\n  }\n  if (typeof profile.name !== 'string' || profile.name.length === 0) {\n    throw new AgentProfileCellValidationError(\n      'AgentProfile must have a non-empty `name`',\n      'profile.name',\n    )\n  }\n  if (typeof profile.version !== 'string' || profile.version.length === 0) {\n    throw new AgentProfileCellValidationError(\n      'AgentProfile must have a non-empty `version`',\n      'profile.version',\n    )\n  }\n  return buildAgentProfileCell({\n    ...input,\n    profileId: `${profile.name}@${profile.version}`,\n    sourceProfile: {\n      kind: AGENT_PROFILE_KINDS.AGENT_INTERFACE_PROFILE,\n      profile: toAgentProfileJson(profile),\n    },\n  })\n}\n"],"mappings":";;;AA2DA,IAAa,kCAAb,cAAqD,gBAAgB;CACnE;CACA,YAAY,SAAiB,OAAO,IAAI;EACtC,MAAM,OAAO,GAAG,QAAQ,OAAO,KAAK,KAAK,OAAO;EAChD,KAAK,OAAO;CACd;AACF;AAEA,MAAM,aAAa;AACnB,MAAM,UAAU;AAEhB,eAAsB,sBACpB,OAC2B;CAC3B,MAAM,WAAW,MAAM,+BAA+B,KAAK;CAC3D,MAAM,SAAS,6BAA6B,MAAM,SAAS,QAAQ;CACnE,OAAO;EAAE,GAAG;EAAU;CAAO;AAC/B;AAEA,SAAgB,6BACd,MACkC;CAClC,MAAM,EAAE,QAAQ,SAAS,GAAG,aAAa;CAEzC,OAAO,0BAA0B,QAAQ;AAC3C;;;;AAKA,eAAsB,uBAAuB,MAA0C;CACrF,yBAAyB,IAAI;CAC7B,OACE,KAAK,WACL,6BAA6B,MAAM,SAAS,6BAA6B,IAAI,CAAC;AAElF;AAEA,SAAgB,yBAAyB,OAAkC;CACzE,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,gCAAgC,iBAAiB;CAE7D,MAAM,MAAM;CACZ,cAAc,IAAI,eAAe,yBAAyB,eAAe;CACzE,IAAI,OAAO,IAAI,WAAW,YAAY,CAAC,QAAQ,KAAK,IAAI,MAAM,GAC5D,MAAM,IAAI,gCACR,wEACA,QACF;CAEF,aAAa,IAAI,WAAW,WAAW;CACvC,eAAe,IAAI,eAAe,eAAe;CACjD,IAAI,IAAI,YAAY,KAAA,GAAW,gBAAgB,IAAI,SAAS,SAAS;CACrE,IAAI,IAAI,UAAU,KAAA,GAAW,aAAa,IAAI,OAAO,OAAO;CAC5D,IAAI,IAAI,eAAe,KAAA,GAAW,aAAa,IAAI,YAAY,YAAY;CAC3E,IAAI,IAAI,eAAe,KAAA,GAAW,mBAAmB,IAAI,YAAY,YAAY;CACjF,OAAO;AACT;AAEA,SAAgB,wBAAwB,QAGnB;CACnB,IAAI,CAAC,OAAO,cACV,MAAM,IAAI,gCACR,QAAQ,OAAO,MAAM,sFACrB,cACF;CAEF,OAAO,yBAAyB,OAAO,YAAY;AACrD;AAEA,SAAgB,oBAAoB,QAGzB;CACT,OAAO,wBAAwB,MAAM,CAAC,CAAC;AACzC;AAEA,eAAsB,0BAA0B,QAKlB;CAC5B,MAAM,UAAU,wBAAwB,MAAM;CAC9C,IAAI,CAAE,MAAM,uBAAuB,OAAO,GACxC,MAAM,IAAI,gCACR,QAAQ,OAAO,MAAM,+DACrB,qBACF;CAEF,IAAI,QAAQ,UAAU,KAAA,KAAa,QAAQ,UAAU,OAAO,OAC1D,MAAM,IAAI,gCACR,QAAQ,OAAO,MAAM,wBAAwB,QAAQ,MAAM,0BAA0B,OAAO,MAAM,IAClG,oBACF;CAEF,IAAI,QAAQ,eAAe,KAAA,KAAa,QAAQ,eAAe,OAAO,YACpE,MAAM,IAAI,gCACR,QAAQ,OAAO,MAAM,6BAA6B,QAAQ,WAAW,+BAA+B,OAAO,WAAW,IACtH,yBACF;CAEF,OAAO;AACT;AAEA,SAAgB,4BAEd,SAAyC;CACzC,MAAM,yBAAS,IAAI,IAAiB;CACpC,KAAK,MAAM,UAAU,SAAS;EAC5B,MAAM,MAAM,oBAAoB,MAAM;EACtC,MAAM,SAAS,OAAO,IAAI,GAAG;EAC7B,IAAI,QAAQ,OAAO,KAAK,MAAM;OACzB,OAAO,IAAI,KAAK,CAAC,MAAM,CAAC;CAC/B;CACA,OAAO;AACT;AAEA,eAAe,+BACb,OAC2C;CAC3C,OAAO,0BAA0B;EAC/B,eAAe;EACf,WAAW,MAAM;EACjB,eAAe,MAAM,qBAAqB,MAAM,aAAa;EAC7D,SAAS,MAAM;EACf,OAAO,MAAM;EACb,YAAY,MAAM;EAClB,YAAY,MAAM;CACpB,CAAC;AACH;AAEA,SAAS,0BACP,OACkC;CAClC,OAAO,cAAc;EACnB,eAAe;EACf,WAAW,gBAAgB,MAAM,WAAW,WAAW;EACvD,eAAe,gBAAgB,MAAM,aAAa;EAClD,SAAS,MAAM,UAAU,iBAAiB,MAAM,SAAS,SAAS,IAAI,KAAA;EACtE,OAAO,iBAAiB,MAAM,OAAO,OAAO;EAC5C,YAAY,iBAAiB,MAAM,YAAY,YAAY;EAC3D,YAAY,MAAM,aACd,eAAe,oBAAoB,MAAM,UAAU,CAAC,IACpD,KAAA;CACN,CAAC;AACH;AAEA,eAAe,qBAAqB,OAA6D;CAC/F,MAAM,OAAO,gBAAgB,MAAM,MAAM,oBAAoB;CAC7D,IAAI,MAAM,SAAS,KAAA,KAAa,MAAM,YAAY,KAAA,GAChD,MAAM,IAAI,gCACR,+DACA,eACF;CAEF,IAAI,MAAM,SAAS,KAAA,GACjB,OAAO;EAAE;EAAM,MAAM,iBAAiB,MAAM,MAAM,oBAAoB;CAAE;CAE1E,IAAI,MAAM,YAAY,KAAA,GACpB,MAAM,IAAI,gCACR,8CACA,eACF;CAEF,WAAW,MAAM,SAAS,uBAAuB;CACjD,OAAO;EAAE;EAAM,MAAM,MAAM,SAAS,MAAM,OAAO;CAAE;AACrD;AAEA,SAAS,gBAAgB,OAA+C;CACtE,OAAO;EACL,MAAM,gBAAgB,MAAM,MAAM,oBAAoB;EACtD,MAAM,iBAAiB,MAAM,MAAM,oBAAoB;CACzD;AACF;AAEA,SAAS,iBAAiB,OAA4B,MAAmC;CACvF,OAAO,cAAc;EACnB,IAAI,gBAAgB,MAAM,IAAI,GAAG,KAAK,IAAI;EAC1C,SAAS,iBAAiB,MAAM,SAAS,GAAG,KAAK,SAAS;EAC1D,MAAM,iBAAiB,MAAM,MAAM,GAAG,KAAK,MAAM;CACnD,CAAC;AACH;AAEA,SAAS,oBACP,OAC4C;CAC5C,MAAM,MAAkD,CAAC;CACzD,KAAK,MAAM,OAAO,OAAO,KAAK,KAAK,CAAC,CAAC,KAAK,GAAG;EAC3C,MAAM,QAAQ,MAAM;EACpB,gBAAgB,KAAK,kBAAkB;EACvC,IACE,UAAU,QACV,OAAO,UAAU,YACjB,OAAO,UAAU,YACjB,OAAO,UAAU,WAEjB,MAAM,IAAI,gCACR,sCACA,cAAc,KAChB;EAEF,IAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,GACrD,MAAM,IAAI,gCAAgC,0BAA0B,cAAc,KAAK;EAEzF,IAAI,OAAO;CACb;CACA,OAAO;AACT;AAEA,SAAS,cAAiD,OAAa;CACrE,MAAM,MAA+B,CAAC;CACtC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,KAAK,GAC7C,IAAI,UAAU,KAAA,GAAW,IAAI,OAAO;CAEtC,OAAO;AACT;AAEA,SAAS,eAAkD,OAAyB;CAClF,OAAO,OAAO,KAAK,KAAK,CAAC,CAAC,SAAS,IAAI,QAAQ,KAAA;AACjD;AAEA,SAAS,eAAe,OAAgB,MAAoB;CAC1D,IAAI,UAAU,QAAQ,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GACpE,MAAM,IAAI,gCAAgC,mBAAmB,IAAI;CAEnE,MAAM,MAAM;CACZ,aAAa,IAAI,MAAM,GAAG,KAAK,MAAM;CACrC,iBAAiB,IAAI,MAAM,GAAG,KAAK,MAAM;AAC3C;AAEA,SAAS,gBAAgB,OAAgB,MAAoB;CAC3D,IAAI,UAAU,QAAQ,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GACpE,MAAM,IAAI,gCAAgC,mBAAmB,IAAI;CAEnE,MAAM,MAAM;CACZ,aAAa,IAAI,IAAI,GAAG,KAAK,IAAI;CACjC,IAAI,IAAI,YAAY,KAAA,GAAW,aAAa,IAAI,SAAS,GAAG,KAAK,SAAS;CAC1E,IAAI,IAAI,SAAS,KAAA,GAAW,aAAa,IAAI,MAAM,GAAG,KAAK,MAAM;AACnE;AAEA,SAAS,mBAAmB,OAAgB,MAAoB;CAC9D,IAAI,UAAU,QAAQ,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GACpE,MAAM,IAAI,gCAAgC,mBAAmB,IAAI;CAEnE,oBAAoB,KAAmD;AACzE;AAEA,SAAS,WAAW,OAAyB,MAAoB;CAC/D,IAAI,UAAU,MAAM;CACpB,MAAM,OAAO,OAAO;CACpB,IAAI,SAAS,YAAY,SAAS,WAAW;CAC7C,IAAI,SAAS,UAAU;EACrB,IAAI,CAAC,OAAO,SAAS,KAAK,GACxB,MAAM,IAAI,gCAAgC,0BAA0B,IAAI;EAE1E;CACF;CACA,IAAI,MAAM,QAAQ,KAAK,GAAG;EACxB,MAAM,SAAS,MAAM,UAAU;GAC7B,WAAW,MAAM,GAAG,KAAK,GAAG,MAAM,EAAE;EACtC,CAAC;EACD;CACF;CACA,IAAI,SAAS,UAAU;EACrB,KAAK,MAAM,CAAC,KAAK,WAAW,OAAO,QAAQ,KAAK,GAAG;GACjD,gBAAgB,KAAK,GAAG,KAAK,OAAO;GACpC,WAAW,QAAQ,GAAG,KAAK,GAAG,KAAK;EACrC;EACA;CACF;CACA,MAAM,IAAI,gCAAgC,kCAAkC,IAAI;AAClF;AAEA,SAAS,cAAc,OAAgB,UAAkB,MAAoB;CAC3E,IAAI,UAAU,UACZ,MAAM,IAAI,gCAAgC,YAAY,YAAY,IAAI;AAE1E;AAEA,SAAS,aAAa,OAAgB,MAAoB;CACxD,IAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAChD,MAAM,IAAI,gCAAgC,6BAA6B,IAAI;AAE/E;AAEA,SAAS,gBAAgB,OAAe,MAAsB;CAC5D,IAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAChD,MAAM,IAAI,gCAAgC,6BAA6B,IAAI;CAE7E,OAAO;AACT;AAEA,SAAS,iBAAiB,OAA2B,MAAkC;CACrF,IAAI,UAAU,KAAA,GAAW,OAAO,KAAA;CAChC,OAAO,gBAAgB,OAAO,IAAI;AACpC;AAEA,SAAS,iBAAiB,OAAgB,MAAsB;CAC9D,IAAI,OAAO,UAAU,YAAY,CAAC,WAAW,KAAK,KAAK,GACrD,MAAM,IAAI,gCAAgC,0CAA0C,IAAI;CAE1F,OAAO;AACT;;;;;AAsBA,MAAa,sBAAsB;;;;AAIjC,yBAAyB,0BAC3B;;;;;AAQA,SAAgB,mBAAmB,OAAkC;CACnE,IAAI;CACJ,IAAI;EACF,aAAa,KAAK,UAAU,KAAK;CACnC,SAAS,KAAK;EACZ,MAAM,IAAI,gCACR,4CAA4C,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,KAC3F,uBACF;CACF;CACA,IAAI,eAAe,KAAA,GACjB,MAAM,IAAI,gCACR,gFACA,uBACF;CAEF,OAAO,KAAK,MAAM,UAAU;AAC9B;;;;;;;;;;AAcA,eAAsB,+BACpB,SACA,OAC2B;CAC3B,IAAI,CAAC,WAAW,OAAO,YAAY,UACjC,MAAM,IAAI,gCAAgC,kCAAkC,SAAS;CAEvF,IAAI,OAAO,QAAQ,SAAS,YAAY,QAAQ,KAAK,WAAW,GAC9D,MAAM,IAAI,gCACR,6CACA,cACF;CAEF,IAAI,OAAO,QAAQ,YAAY,YAAY,QAAQ,QAAQ,WAAW,GACpE,MAAM,IAAI,gCACR,gDACA,iBACF;CAEF,OAAO,sBAAsB;EAC3B,GAAG;EACH,WAAW,GAAG,QAAQ,KAAK,GAAG,QAAQ;EACtC,eAAe;GACb,MAAM,oBAAoB;GAC1B,SAAS,mBAAmB,OAAO;EACrC;CACF,CAAC;AACH"}