@tangle-network/agent-eval 0.180.0 → 0.182.0

package/dist/opencode-sqlite-CNw3vubS.js

@@ -1,145 +0,0 @@
-import { createRequire } from "node:module";
-import { join } from "node:path";
-import { homedir } from "node:os";
-//#region src/rollout/readers/opencode-sqlite.ts
-/**
-* Read-only backfill reader over the opencode sqlite store
-* (~/.local/share/opencode/opencode.db) → canonical chat-with-tools messages.
-*
-* Schema consumed (observed, 2026-07): `session` rows carry directory /
-* parent_id / agent / model / cost / tokens_*; `message` rows carry a JSON
-* `data` blob ({role, modelID, providerID, tokens, cost, finish}); `part`
-* rows carry the actual content ({type: text|reasoning|tool|step-start|
-* step-finish|snapshot…}). Tool parts hold {callID, state:{input, output,
-* status}} — both the call and its result, which we split into an assistant
-* tool_call plus a role:"tool" result message.
-*
-* The store is mutable and can be corrupt (a `.corrupt-bak` sibling ships
-* next to it in the wild), so `openOpencodeDb` returns null instead of
-* throwing — callers record a gap line, never crash the backfill.
-*/
-const DEFAULT_OPENCODE_DB = join(homedir(), ".local", "share", "opencode", "opencode.db");
-/** Open the store read-only; null = unavailable/corrupt (caller records a gap). */
-async function openOpencodeDb(path = DEFAULT_OPENCODE_DB) {
-	try {
-		const { DatabaseSync } = createRequire(import.meta.url)("node:sqlite");
-		const db = new DatabaseSync(path, { readOnly: true });
-		db.prepare("SELECT id FROM session LIMIT 1").get();
-		return db;
-	} catch {
-		return null;
-	}
-}
-const isRecord = (v) => typeof v === "object" && v !== null && !Array.isArray(v);
-function parseSessionRow(row) {
-	let model = null;
-	if (typeof row.model === "string" && row.model.length > 0) try {
-		const parsed = JSON.parse(row.model);
-		if (isRecord(parsed)) model = parsed;
-	} catch {
-		model = null;
-	}
-	return {
-		id: String(row.id),
-		parentId: row.parent_id === null || row.parent_id === void 0 ? null : String(row.parent_id),
-		directory: String(row.directory),
-		agent: row.agent === null || row.agent === void 0 ? null : String(row.agent),
-		model,
-		costUsd: Number(row.cost ?? 0),
-		tokensInput: Number(row.tokens_input ?? 0),
-		tokensOutput: Number(row.tokens_output ?? 0),
-		tokensReasoning: Number(row.tokens_reasoning ?? 0),
-		tokensCacheRead: Number(row.tokens_cache_read ?? 0),
-		tokensCacheWrite: Number(row.tokens_cache_write ?? 0),
-		timeCreated: Number(row.time_created ?? 0),
-		timeUpdated: Number(row.time_updated ?? 0)
-	};
-}
-const SESSION_COLUMNS = "id, parent_id, directory, agent, model, cost, tokens_input, tokens_output, tokens_reasoning, tokens_cache_read, tokens_cache_write, time_created, time_updated";
-/** Sessions whose cwd is `directory` (the worker-clone join key). */
-function findOpencodeSessionsByDirectory(db, directory) {
-	return db.prepare(`SELECT ${SESSION_COLUMNS} FROM session WHERE directory = ? ORDER BY time_created`).all(directory).map(parseSessionRow);
-}
-function toolResultContent(output) {
-	if (typeof output === "string") return output;
-	if (output === null || output === void 0) return "";
-	return JSON.stringify(output);
-}
-/**
-* Convert one session's message+part rows into canonical messages.
-* An opencode assistant message row spans several model steps; each step's
-* parts (reasoning → text → tool …) become one assistant message followed by
-* the role:"tool" results of its calls, preserving order.
-*/
-function readOpencodeSessionMessages(db, sessionId) {
-	const messageRows = db.prepare("SELECT id, data FROM message WHERE session_id = ? ORDER BY time_created, id").all(sessionId);
-	const partsStmt = db.prepare("SELECT data FROM part WHERE message_id = ? ORDER BY id");
-	const messages = [];
-	for (const messageRow of messageRows) {
-		let data;
-		try {
-			const parsed = JSON.parse(messageRow.data);
-			if (!isRecord(parsed)) continue;
-			data = parsed;
-		} catch {
-			continue;
-		}
-		const parts = [];
-		for (const row of partsStmt.all(messageRow.id)) try {
-			const parsed = JSON.parse(row.data);
-			if (isRecord(parsed)) parts.push(parsed);
-		} catch {}
-		if (data.role === "user") {
-			const text = parts.filter((p) => p.type === "text" && typeof p.text === "string").map((p) => p.text).join("\n");
-			messages.push({
-				role: "user",
-				content: text
-			});
-			continue;
-		}
-		if (data.role !== "assistant") continue;
-		const steps = [];
-		let current = [];
-		for (const part of parts) {
-			if (part.type === "step-start") {
-				if (current.length > 0) steps.push(current);
-				current = [];
-				continue;
-			}
-			if (part.type === "step-finish" || part.type === "snapshot" || part.type === "patch") continue;
-			current.push(part);
-		}
-		if (current.length > 0) steps.push(current);
-		for (const step of steps) {
-			const reasoning = step.filter((p) => p.type === "reasoning" && typeof p.text === "string" && p.text.length > 0).map((p) => p.text).join("\n");
-			const text = step.filter((p) => p.type === "text" && typeof p.text === "string").map((p) => p.text).join("\n");
-			const toolParts = step.filter((p) => p.type === "tool" && typeof p.callID === "string");
-			const toolCalls = toolParts.map((p) => ({
-				id: p.callID,
-				type: "function",
-				function: {
-					name: p.tool ?? "unknown",
-					arguments: JSON.stringify(p.state?.input ?? {})
-				}
-			}));
-			if (reasoning.length === 0 && text.length === 0 && toolCalls.length === 0) continue;
-			messages.push({
-				role: "assistant",
-				content: text.length > 0 ? text : null,
-				...reasoning.length > 0 ? { reasoning_content: reasoning } : {},
-				...toolCalls.length > 0 ? { tool_calls: toolCalls } : {}
-			});
-			for (const p of toolParts) messages.push({
-				role: "tool",
-				tool_call_id: p.callID,
-				name: p.tool ?? "unknown",
-				content: toolResultContent(p.state?.output)
-			});
-		}
-	}
-	return messages;
-}
-//#endregion
-export { readOpencodeSessionMessages as i, findOpencodeSessionsByDirectory as n, openOpencodeDb as r, DEFAULT_OPENCODE_DB as t };
-
-//# sourceMappingURL=opencode-sqlite-CNw3vubS.js.map

package/dist/opencode-sqlite-CNw3vubS.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"opencode-sqlite-CNw3vubS.js","names":[],"sources":["../src/rollout/readers/opencode-sqlite.ts"],"sourcesContent":["/**\n * Read-only backfill reader over the opencode sqlite store\n * (~/.local/share/opencode/opencode.db) → canonical chat-with-tools messages.\n *\n * Schema consumed (observed, 2026-07): `session` rows carry directory /\n * parent_id / agent / model / cost / tokens_*; `message` rows carry a JSON\n * `data` blob ({role, modelID, providerID, tokens, cost, finish}); `part`\n * rows carry the actual content ({type: text|reasoning|tool|step-start|\n * step-finish|snapshot…}). Tool parts hold {callID, state:{input, output,\n * status}} — both the call and its result, which we split into an assistant\n * tool_call plus a role:\"tool\" result message.\n *\n * The store is mutable and can be corrupt (a `.corrupt-bak` sibling ships\n * next to it in the wild), so `openOpencodeDb` returns null instead of\n * throwing — callers record a gap line, never crash the backfill.\n */\n\nimport { createRequire } from 'node:module'\nimport { homedir } from 'node:os'\nimport { join } from 'node:path'\nimport type { DatabaseSync } from 'node:sqlite'\nimport type { ChatMessage, ChatToolCall } from '../schema'\n\nexport const DEFAULT_OPENCODE_DB = join(homedir(), '.local', 'share', 'opencode', 'opencode.db')\n\nexport interface OpencodeSessionRow {\n  id: string\n  parentId: string | null\n  directory: string\n  agent: string | null\n  /** Raw session.model JSON: {id, providerID, variant} where present. */\n  model: { id?: string; providerID?: string } | null\n  costUsd: number\n  tokensInput: number\n  tokensOutput: number\n  tokensReasoning: number\n  tokensCacheRead: number\n  tokensCacheWrite: number\n  timeCreated: number\n  timeUpdated: number\n}\n\n// `node:sqlite` is loaded through CommonJS `require`, not `import()`. esbuild\n// (bundling) and Vite (tests) both rewrite a dynamic import and strip the\n// `node:` prefix under an es20xx target, turning this builtin into a bogus\n// \"sqlite\" package lookup; composing the specifier at runtime does not reliably\n// defeat that (it still resolved through Vite's transform in some workers, so\n// the failure moved around as test files were added). A require obtained from\n// `createRequire` is not an analyzable module reference in either tool, so\n// neither can rewrite it.\n/** Open the store read-only; null = unavailable/corrupt (caller records a gap). */\nexport async function openOpencodeDb(\n  path: string = DEFAULT_OPENCODE_DB,\n): Promise<DatabaseSync | null> {\n  try {\n    // Keep Node-only module initialization inside the Node-only operation.\n    // Root imports are shared with edge consumers that do not define import.meta.url.\n    const nodeRequire = createRequire(import.meta.url)\n    const { DatabaseSync } = nodeRequire('node:sqlite') as typeof import('node:sqlite')\n    const db = new DatabaseSync(path, { readOnly: true })\n    // Probe: a corrupt store can open() fine and fail on first page read.\n    db.prepare('SELECT id FROM session LIMIT 1').get()\n    return db\n  } catch {\n    return null\n  }\n}\n\nconst isRecord = (v: unknown): v is Record<string, unknown> =>\n  typeof v === 'object' && v !== null && !Array.isArray(v)\n\nfunction parseSessionRow(row: Record<string, unknown>): OpencodeSessionRow {\n  let model: OpencodeSessionRow['model'] = null\n  if (typeof row.model === 'string' && row.model.length > 0) {\n    try {\n      const parsed: unknown = JSON.parse(row.model)\n      if (isRecord(parsed)) model = parsed as { id?: string; providerID?: string }\n    } catch {\n      model = null\n    }\n  }\n  return {\n    id: String(row.id),\n    parentId: row.parent_id === null || row.parent_id === undefined ? null : String(row.parent_id),\n    directory: String(row.directory),\n    agent: row.agent === null || row.agent === undefined ? null : String(row.agent),\n    model,\n    costUsd: Number(row.cost ?? 0),\n    tokensInput: Number(row.tokens_input ?? 0),\n    tokensOutput: Number(row.tokens_output ?? 0),\n    tokensReasoning: Number(row.tokens_reasoning ?? 0),\n    tokensCacheRead: Number(row.tokens_cache_read ?? 0),\n    tokensCacheWrite: Number(row.tokens_cache_write ?? 0),\n    timeCreated: Number(row.time_created ?? 0),\n    timeUpdated: Number(row.time_updated ?? 0),\n  }\n}\n\nconst SESSION_COLUMNS =\n  'id, parent_id, directory, agent, model, cost, tokens_input, tokens_output, tokens_reasoning, tokens_cache_read, tokens_cache_write, time_created, time_updated'\n\n/** Sessions whose cwd is `directory` (the worker-clone join key). */\nexport function findOpencodeSessionsByDirectory(\n  db: DatabaseSync,\n  directory: string,\n): OpencodeSessionRow[] {\n  const rows = db\n    .prepare(`SELECT ${SESSION_COLUMNS} FROM session WHERE directory = ? ORDER BY time_created`)\n    .all(directory) as Array<Record<string, unknown>>\n  return rows.map(parseSessionRow)\n}\n\ninterface OpencodePart {\n  type?: string\n  text?: string\n  tool?: string\n  callID?: string\n  state?: { status?: string; input?: unknown; output?: unknown }\n}\n\nfunction toolResultContent(output: unknown): string {\n  if (typeof output === 'string') return output\n  if (output === null || output === undefined) return ''\n  return JSON.stringify(output)\n}\n\n/**\n * Convert one session's message+part rows into canonical messages.\n * An opencode assistant message row spans several model steps; each step's\n * parts (reasoning → text → tool …) become one assistant message followed by\n * the role:\"tool\" results of its calls, preserving order.\n */\nexport function readOpencodeSessionMessages(db: DatabaseSync, sessionId: string): ChatMessage[] {\n  const messageRows = db\n    .prepare('SELECT id, data FROM message WHERE session_id = ? ORDER BY time_created, id')\n    .all(sessionId) as Array<{ id: string; data: string }>\n  const partsStmt = db.prepare('SELECT data FROM part WHERE message_id = ? ORDER BY id')\n\n  const messages: ChatMessage[] = []\n  for (const messageRow of messageRows) {\n    let data: Record<string, unknown>\n    try {\n      const parsed: unknown = JSON.parse(messageRow.data)\n      if (!isRecord(parsed)) continue\n      data = parsed\n    } catch {\n      continue\n    }\n    const parts: OpencodePart[] = []\n    for (const row of partsStmt.all(messageRow.id) as Array<{ data: string }>) {\n      try {\n        const parsed: unknown = JSON.parse(row.data)\n        if (isRecord(parsed)) parts.push(parsed as OpencodePart)\n      } catch {\n        // Malformed part payload: skip the part, keep the message.\n      }\n    }\n\n    if (data.role === 'user') {\n      const text = parts\n        .filter((p) => p.type === 'text' && typeof p.text === 'string')\n        .map((p) => p.text as string)\n        .join('\\n')\n      messages.push({ role: 'user', content: text })\n      continue\n    }\n    if (data.role !== 'assistant') continue\n\n    // Split the row into steps at step-start boundaries; parts before the\n    // first step-start (none observed, but tolerated) form an implicit step.\n    const steps: OpencodePart[][] = []\n    let current: OpencodePart[] = []\n    for (const part of parts) {\n      if (part.type === 'step-start') {\n        if (current.length > 0) steps.push(current)\n        current = []\n        continue\n      }\n      if (part.type === 'step-finish' || part.type === 'snapshot' || part.type === 'patch') continue\n      current.push(part)\n    }\n    if (current.length > 0) steps.push(current)\n\n    for (const step of steps) {\n      const reasoning = step\n        .filter((p) => p.type === 'reasoning' && typeof p.text === 'string' && p.text.length > 0)\n        .map((p) => p.text as string)\n        .join('\\n')\n      const text = step\n        .filter((p) => p.type === 'text' && typeof p.text === 'string')\n        .map((p) => p.text as string)\n        .join('\\n')\n      const toolParts = step.filter((p) => p.type === 'tool' && typeof p.callID === 'string')\n      const toolCalls: ChatToolCall[] = toolParts.map((p) => ({\n        id: p.callID as string,\n        type: 'function',\n        function: {\n          name: p.tool ?? 'unknown',\n          arguments: JSON.stringify(p.state?.input ?? {}),\n        },\n      }))\n      if (reasoning.length === 0 && text.length === 0 && toolCalls.length === 0) continue\n      messages.push({\n        role: 'assistant',\n        content: text.length > 0 ? text : null,\n        ...(reasoning.length > 0 ? { reasoning_content: reasoning } : {}),\n        ...(toolCalls.length > 0 ? { tool_calls: toolCalls } : {}),\n      })\n      for (const p of toolParts) {\n        messages.push({\n          role: 'tool',\n          tool_call_id: p.callID as string,\n          name: p.tool ?? 'unknown',\n          content: toolResultContent(p.state?.output),\n        })\n      }\n    }\n  }\n  return messages\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAuBA,MAAa,sBAAsB,KAAK,QAAQ,GAAG,UAAU,SAAS,YAAY,aAAa;;AA4B/F,eAAsB,eACpB,OAAe,qBACe;CAC9B,IAAI;EAIF,MAAM,EAAE,iBADY,cAAc,OAAO,KAAK,GACX,CAAC,CAAC,aAAa;EAClD,MAAM,KAAK,IAAI,aAAa,MAAM,EAAE,UAAU,KAAK,CAAC;EAEpD,GAAG,QAAQ,gCAAgC,CAAC,CAAC,IAAI;EACjD,OAAO;CACT,QAAQ;EACN,OAAO;CACT;AACF;AAEA,MAAM,YAAY,MAChB,OAAO,MAAM,YAAY,MAAM,QAAQ,CAAC,MAAM,QAAQ,CAAC;AAEzD,SAAS,gBAAgB,KAAkD;CACzE,IAAI,QAAqC;CACzC,IAAI,OAAO,IAAI,UAAU,YAAY,IAAI,MAAM,SAAS,GACtD,IAAI;EACF,MAAM,SAAkB,KAAK,MAAM,IAAI,KAAK;EAC5C,IAAI,SAAS,MAAM,GAAG,QAAQ;CAChC,QAAQ;EACN,QAAQ;CACV;CAEF,OAAO;EACL,IAAI,OAAO,IAAI,EAAE;EACjB,UAAU,IAAI,cAAc,QAAQ,IAAI,cAAc,KAAA,IAAY,OAAO,OAAO,IAAI,SAAS;EAC7F,WAAW,OAAO,IAAI,SAAS;EAC/B,OAAO,IAAI,UAAU,QAAQ,IAAI,UAAU,KAAA,IAAY,OAAO,OAAO,IAAI,KAAK;EAC9E;EACA,SAAS,OAAO,IAAI,QAAQ,CAAC;EAC7B,aAAa,OAAO,IAAI,gBAAgB,CAAC;EACzC,cAAc,OAAO,IAAI,iBAAiB,CAAC;EAC3C,iBAAiB,OAAO,IAAI,oBAAoB,CAAC;EACjD,iBAAiB,OAAO,IAAI,qBAAqB,CAAC;EAClD,kBAAkB,OAAO,IAAI,sBAAsB,CAAC;EACpD,aAAa,OAAO,IAAI,gBAAgB,CAAC;EACzC,aAAa,OAAO,IAAI,gBAAgB,CAAC;CAC3C;AACF;AAEA,MAAM,kBACJ;;AAGF,SAAgB,gCACd,IACA,WACsB;CAItB,OAHa,GACV,QAAQ,UAAU,gBAAgB,wDAAwD,CAAC,CAC3F,IAAI,SACG,CAAC,CAAC,IAAI,eAAe;AACjC;AAUA,SAAS,kBAAkB,QAAyB;CAClD,IAAI,OAAO,WAAW,UAAU,OAAO;CACvC,IAAI,WAAW,QAAQ,WAAW,KAAA,GAAW,OAAO;CACpD,OAAO,KAAK,UAAU,MAAM;AAC9B;;;;;;;AAQA,SAAgB,4BAA4B,IAAkB,WAAkC;CAC9F,MAAM,cAAc,GACjB,QAAQ,6EAA6E,CAAC,CACtF,IAAI,SAAS;CAChB,MAAM,YAAY,GAAG,QAAQ,wDAAwD;CAErF,MAAM,WAA0B,CAAC;CACjC,KAAK,MAAM,cAAc,aAAa;EACpC,IAAI;EACJ,IAAI;GACF,MAAM,SAAkB,KAAK,MAAM,WAAW,IAAI;GAClD,IAAI,CAAC,SAAS,MAAM,GAAG;GACvB,OAAO;EACT,QAAQ;GACN;EACF;EACA,MAAM,QAAwB,CAAC;EAC/B,KAAK,MAAM,OAAO,UAAU,IAAI,WAAW,EAAE,GAC3C,IAAI;GACF,MAAM,SAAkB,KAAK,MAAM,IAAI,IAAI;GAC3C,IAAI,SAAS,MAAM,GAAG,MAAM,KAAK,MAAsB;EACzD,QAAQ,CAER;EAGF,IAAI,KAAK,SAAS,QAAQ;GACxB,MAAM,OAAO,MACV,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,SAAS,QAAQ,CAAC,CAC9D,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,SAAS,KAAK;IAAE,MAAM;IAAQ,SAAS;GAAK,CAAC;GAC7C;EACF;EACA,IAAI,KAAK,SAAS,aAAa;EAI/B,MAAM,QAA0B,CAAC;EACjC,IAAI,UAA0B,CAAC;EAC/B,KAAK,MAAM,QAAQ,OAAO;GACxB,IAAI,KAAK,SAAS,cAAc;IAC9B,IAAI,QAAQ,SAAS,GAAG,MAAM,KAAK,OAAO;IAC1C,UAAU,CAAC;IACX;GACF;GACA,IAAI,KAAK,SAAS,iBAAiB,KAAK,SAAS,cAAc,KAAK,SAAS,SAAS;GACtF,QAAQ,KAAK,IAAI;EACnB;EACA,IAAI,QAAQ,SAAS,GAAG,MAAM,KAAK,OAAO;EAE1C,KAAK,MAAM,QAAQ,OAAO;GACxB,MAAM,YAAY,KACf,QAAQ,MAAM,EAAE,SAAS,eAAe,OAAO,EAAE,SAAS,YAAY,EAAE,KAAK,SAAS,CAAC,CAAC,CACxF,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,MAAM,OAAO,KACV,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,SAAS,QAAQ,CAAC,CAC9D,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,MAAM,YAAY,KAAK,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,WAAW,QAAQ;GACtF,MAAM,YAA4B,UAAU,KAAK,OAAO;IACtD,IAAI,EAAE;IACN,MAAM;IACN,UAAU;KACR,MAAM,EAAE,QAAQ;KAChB,WAAW,KAAK,UAAU,EAAE,OAAO,SAAS,CAAC,CAAC;IAChD;GACF,EAAE;GACF,IAAI,UAAU,WAAW,KAAK,KAAK,WAAW,KAAK,UAAU,WAAW,GAAG;GAC3E,SAAS,KAAK;IACZ,MAAM;IACN,SAAS,KAAK,SAAS,IAAI,OAAO;IAClC,GAAI,UAAU,SAAS,IAAI,EAAE,mBAAmB,UAAU,IAAI,CAAC;IAC/D,GAAI,UAAU,SAAS,IAAI,EAAE,YAAY,UAAU,IAAI,CAAC;GAC1D,CAAC;GACD,KAAK,MAAM,KAAK,WACd,SAAS,KAAK;IACZ,MAAM;IACN,cAAc,EAAE;IAChB,MAAM,EAAE,QAAQ;IAChB,SAAS,kBAAkB,EAAE,OAAO,MAAM;GAC5C,CAAC;EAEL;CACF;CACA,OAAO;AACT"}

package/dist/outcome-store-BYHIuO0e.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"outcome-store-BYHIuO0e.d.ts","names":[],"sources":["../src/meta-eval/outcome-store.ts"],"mappings":";;;;;;;;;;;;;UAaiB;EACf;EACA;;EAEA,SAAS;;EAET,SAAS;;EAET;;UAGe;EACf;EACA;EACA;EACA;IAAU;IAAa;;EACvB;;UAGe;EACf,OAAO,SAAS,oBAAoB;;;EAGpC,OAAO,gBAAgB,QAAQ;EAC/B,KAAK,SAAS,gBAAgB,QAAQ;;cAG3B,gCAAgC;UACnC;EAEF,OAAO,SAAS,oBAAoB;EAIpC,OAAO,gBAAgB,QAAQ;EAI/B,KAAK,SAAQ,gBAAqB,QAAQ;;UAKjC;EACf;EACA;;cAGW,kCAAkC;UACrC;UACA;UACA;UACA;EAER,YAAY,SAAS;UAKP;EAKR,OAAO,SAAS,oBAAoB;UAiB5B;EAuBR,OAAO,gBAAgB,QAAQ;EAI/B,KAAK,SAAS,gBAAgB,QAAQ"}

package/dist/outcome-store-ChBKlTd_.js

@@ -1,75 +0,0 @@
-//#region src/meta-eval/outcome-store.ts
-var InMemoryOutcomeStore = class {
-	items = [];
-	async append(outcome) {
-		this.items.push({ ...outcome });
-	}
-	async forRun(runId) {
-		return this.items.filter((o) => o.runId === runId).map((o) => ({ ...o }));
-	}
-	async list(filter = {}) {
-		return this.items.filter((o) => matches(o, filter)).map((o) => ({ ...o }));
-	}
-};
-var FileSystemOutcomeStore = class {
-	dir;
-	maxBytes;
-	memo;
-	loaded = false;
-	constructor(options) {
-		this.dir = options.dir;
-		this.maxBytes = options.maxBytes ?? 32 * 1024 * 1024;
-	}
-	async ensureDir() {
-		await (await import("node:fs/promises")).mkdir(this.dir, { recursive: true });
-	}
-	async append(outcome) {
-		await this.ensureDir();
-		const fs = await import("node:fs/promises");
-		const path = await import("node:path");
-		const active = path.join(this.dir, "outcomes.ndjson");
-		try {
-			if ((await fs.stat(active)).size >= this.maxBytes) await fs.rename(active, path.join(this.dir, `outcomes.${Date.now()}.ndjson`));
-		} catch {}
-		await fs.appendFile(active, `${JSON.stringify(outcome)}\n`, "utf8");
-		if (this.memo) await this.memo.append(outcome);
-	}
-	async load() {
-		if (this.loaded && this.memo) return this.memo;
-		const fs = await import("node:fs/promises");
-		const path = await import("node:path");
-		const memo = new InMemoryOutcomeStore();
-		try {
-			const entries = await fs.readdir(this.dir);
-			for (const file of entries) {
-				if (!file.endsWith(".ndjson")) continue;
-				const content = await fs.readFile(path.join(this.dir, file), "utf8");
-				for (const line of content.split("\n")) {
-					if (!line.trim()) continue;
-					await memo.append(JSON.parse(line));
-				}
-			}
-		} catch {}
-		this.memo = memo;
-		this.loaded = true;
-		return memo;
-	}
-	async forRun(runId) {
-		return (await this.load()).forRun(runId);
-	}
-	async list(filter) {
-		return (await this.load()).list(filter);
-	}
-};
-function matches(o, f) {
-	if (f.runIds && !f.runIds.includes(o.runId)) return false;
-	if (f.since !== void 0 && o.capturedAt < f.since) return false;
-	if (f.until !== void 0 && o.capturedAt > f.until) return false;
-	if (f.source && o.source !== f.source) return false;
-	if (f.label && o.labels?.[f.label.key] !== f.label.value) return false;
-	return true;
-}
-//#endregion
-export { InMemoryOutcomeStore as n, FileSystemOutcomeStore as t };
-
-//# sourceMappingURL=outcome-store-ChBKlTd_.js.map

package/dist/outcome-store-ChBKlTd_.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"outcome-store-ChBKlTd_.js","names":[],"sources":["../src/meta-eval/outcome-store.ts"],"sourcesContent":["/**\n * OutcomeStore — deployment outcomes attached to Run IDs.\n *\n * Outcomes arrive asynchronously from production telemetry after the\n * eval run completed: user ratings, retention flags, conversion events,\n * revenue, support-ticket rate, anything a product team can measure.\n * The store is a peer to TraceStore — separate lifecycle, same runId\n * foreign key.\n *\n * The whole point of this module is to make the meta-eval correlation\n * question computable: `correlate(evalMetric, outcomeMetric) → r, ρ, n, CI`.\n */\n\nexport interface DeploymentOutcome {\n  runId: string\n  capturedAt: number\n  /** Numeric outcomes keyed by name — retention_7d, csat, revenue_usd, etc. */\n  metrics: Record<string, number>\n  /** Dimensions for stratified analysis — cohort, region, user_segment. */\n  labels?: Record<string, string>\n  /** Free-form provenance (source system, pipeline version). */\n  source?: string\n}\n\nexport interface OutcomeFilter {\n  runIds?: string[]\n  since?: number\n  until?: number\n  label?: { key: string; value: string }\n  source?: string\n}\n\nexport interface OutcomeStore {\n  append(outcome: DeploymentOutcome): Promise<void>\n  /** All outcomes attached to this run (a single run can have many — multiple\n   *  capture windows over deployment time). */\n  forRun(runId: string): Promise<DeploymentOutcome[]>\n  list(filter?: OutcomeFilter): Promise<DeploymentOutcome[]>\n}\n\nexport class InMemoryOutcomeStore implements OutcomeStore {\n  private items: DeploymentOutcome[] = []\n\n  async append(outcome: DeploymentOutcome): Promise<void> {\n    this.items.push({ ...outcome })\n  }\n\n  async forRun(runId: string): Promise<DeploymentOutcome[]> {\n    return this.items.filter((o) => o.runId === runId).map((o) => ({ ...o }))\n  }\n\n  async list(filter: OutcomeFilter = {}): Promise<DeploymentOutcome[]> {\n    return this.items.filter((o) => matches(o, filter)).map((o) => ({ ...o }))\n  }\n}\n\nexport interface FileSystemOutcomeStoreOptions {\n  dir: string\n  maxBytes?: number\n}\n\nexport class FileSystemOutcomeStore implements OutcomeStore {\n  private dir: string\n  private maxBytes: number\n  private memo?: InMemoryOutcomeStore\n  private loaded = false\n\n  constructor(options: FileSystemOutcomeStoreOptions) {\n    this.dir = options.dir\n    this.maxBytes = options.maxBytes ?? 32 * 1024 * 1024\n  }\n\n  private async ensureDir(): Promise<void> {\n    const fs = await import('node:fs/promises')\n    await fs.mkdir(this.dir, { recursive: true })\n  }\n\n  async append(outcome: DeploymentOutcome): Promise<void> {\n    await this.ensureDir()\n    const fs = await import('node:fs/promises')\n    const path = await import('node:path')\n    const active = path.join(this.dir, 'outcomes.ndjson')\n    try {\n      const stat = await fs.stat(active)\n      if (stat.size >= this.maxBytes) {\n        await fs.rename(active, path.join(this.dir, `outcomes.${Date.now()}.ndjson`))\n      }\n    } catch {\n      /* first write */\n    }\n    await fs.appendFile(active, `${JSON.stringify(outcome)}\\n`, 'utf8')\n    if (this.memo) await this.memo.append(outcome)\n  }\n\n  private async load(): Promise<InMemoryOutcomeStore> {\n    if (this.loaded && this.memo) return this.memo\n    const fs = await import('node:fs/promises')\n    const path = await import('node:path')\n    const memo = new InMemoryOutcomeStore()\n    try {\n      const entries = await fs.readdir(this.dir)\n      for (const file of entries) {\n        if (!file.endsWith('.ndjson')) continue\n        const content = await fs.readFile(path.join(this.dir, file), 'utf8')\n        for (const line of content.split('\\n')) {\n          if (!line.trim()) continue\n          await memo.append(JSON.parse(line))\n        }\n      }\n    } catch {\n      /* empty */\n    }\n    this.memo = memo\n    this.loaded = true\n    return memo\n  }\n\n  async forRun(runId: string): Promise<DeploymentOutcome[]> {\n    return (await this.load()).forRun(runId)\n  }\n\n  async list(filter?: OutcomeFilter): Promise<DeploymentOutcome[]> {\n    return (await this.load()).list(filter)\n  }\n}\n\nfunction matches(o: DeploymentOutcome, f: OutcomeFilter): boolean {\n  if (f.runIds && !f.runIds.includes(o.runId)) return false\n  if (f.since !== undefined && o.capturedAt < f.since) return false\n  if (f.until !== undefined && o.capturedAt > f.until) return false\n  if (f.source && o.source !== f.source) return false\n  if (f.label && o.labels?.[f.label.key] !== f.label.value) return false\n  return true\n}\n"],"mappings":";AAwCA,IAAa,uBAAb,MAA0D;CACxD,QAAqC,CAAC;CAEtC,MAAM,OAAO,SAA2C;EACtD,KAAK,MAAM,KAAK,EAAE,GAAG,QAAQ,CAAC;CAChC;CAEA,MAAM,OAAO,OAA6C;EACxD,OAAO,KAAK,MAAM,QAAQ,MAAM,EAAE,UAAU,KAAK,CAAC,CAAC,KAAK,OAAO,EAAE,GAAG,EAAE,EAAE;CAC1E;CAEA,MAAM,KAAK,SAAwB,CAAC,GAAiC;EACnE,OAAO,KAAK,MAAM,QAAQ,MAAM,QAAQ,GAAG,MAAM,CAAC,CAAC,CAAC,KAAK,OAAO,EAAE,GAAG,EAAE,EAAE;CAC3E;AACF;AAOA,IAAa,yBAAb,MAA4D;CAC1D;CACA;CACA;CACA,SAAiB;CAEjB,YAAY,SAAwC;EAClD,KAAK,MAAM,QAAQ;EACnB,KAAK,WAAW,QAAQ,YAAY,KAAK,OAAO;CAClD;CAEA,MAAc,YAA2B;EAEvC,OAAM,MADW,OAAO,oBAAA,CACf,MAAM,KAAK,KAAK,EAAE,WAAW,KAAK,CAAC;CAC9C;CAEA,MAAM,OAAO,SAA2C;EACtD,MAAM,KAAK,UAAU;EACrB,MAAM,KAAK,MAAM,OAAO;EACxB,MAAM,OAAO,MAAM,OAAO;EAC1B,MAAM,SAAS,KAAK,KAAK,KAAK,KAAK,iBAAiB;EACpD,IAAI;GAEF,KAAI,MADe,GAAG,KAAK,MAAM,EAAA,CACxB,QAAQ,KAAK,UACpB,MAAM,GAAG,OAAO,QAAQ,KAAK,KAAK,KAAK,KAAK,YAAY,KAAK,IAAI,EAAE,QAAQ,CAAC;EAEhF,QAAQ,CAER;EACA,MAAM,GAAG,WAAW,QAAQ,GAAG,KAAK,UAAU,OAAO,EAAE,KAAK,MAAM;EAClE,IAAI,KAAK,MAAM,MAAM,KAAK,KAAK,OAAO,OAAO;CAC/C;CAEA,MAAc,OAAsC;EAClD,IAAI,KAAK,UAAU,KAAK,MAAM,OAAO,KAAK;EAC1C,MAAM,KAAK,MAAM,OAAO;EACxB,MAAM,OAAO,MAAM,OAAO;EAC1B,MAAM,OAAO,IAAI,qBAAqB;EACtC,IAAI;GACF,MAAM,UAAU,MAAM,GAAG,QAAQ,KAAK,GAAG;GACzC,KAAK,MAAM,QAAQ,SAAS;IAC1B,IAAI,CAAC,KAAK,SAAS,SAAS,GAAG;IAC/B,MAAM,UAAU,MAAM,GAAG,SAAS,KAAK,KAAK,KAAK,KAAK,IAAI,GAAG,MAAM;IACnE,KAAK,MAAM,QAAQ,QAAQ,MAAM,IAAI,GAAG;KACtC,IAAI,CAAC,KAAK,KAAK,GAAG;KAClB,MAAM,KAAK,OAAO,KAAK,MAAM,IAAI,CAAC;IACpC;GACF;EACF,QAAQ,CAER;EACA,KAAK,OAAO;EACZ,KAAK,SAAS;EACd,OAAO;CACT;CAEA,MAAM,OAAO,OAA6C;EACxD,QAAQ,MAAM,KAAK,KAAK,EAAA,CAAG,OAAO,KAAK;CACzC;CAEA,MAAM,KAAK,QAAsD;EAC/D,QAAQ,MAAM,KAAK,KAAK,EAAA,CAAG,KAAK,MAAM;CACxC;AACF;AAEA,SAAS,QAAQ,GAAsB,GAA2B;CAChE,IAAI,EAAE,UAAU,CAAC,EAAE,OAAO,SAAS,EAAE,KAAK,GAAG,OAAO;CACpD,IAAI,EAAE,UAAU,KAAA,KAAa,EAAE,aAAa,EAAE,OAAO,OAAO;CAC5D,IAAI,EAAE,UAAU,KAAA,KAAa,EAAE,aAAa,EAAE,OAAO,OAAO;CAC5D,IAAI,EAAE,UAAU,EAAE,WAAW,EAAE,QAAQ,OAAO;CAC9C,IAAI,EAAE,SAAS,EAAE,SAAS,EAAE,MAAM,SAAS,EAAE,MAAM,OAAO,OAAO;CACjE,OAAO;AACT"}

package/dist/paired-arms-D4aeIHUy.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"paired-arms-D4aeIHUy.js","names":[],"sources":["../src/statistics/paired-binary.ts","../src/math/normal.ts","../src/statistics/rank-tests.ts","../src/paired-arms.ts"],"sourcesContent":["import { regularizedIncompleteBeta } from '../math/special-functions'\nimport { binomialSignTwoSided, zQuantile } from './internal'\n\n// ── Binomial proportion + paired-binary + coding-eval estimators ─────\n//\n// The paired family above (pairedBootstrap/pairedTTest/wilcoxonSignedRank)\n// operates on continuous scores. Pass/fail A/B comparisons — \"does treatment\n// X raise the success RATE vs control\" — are binary and paired, so they need\n// their own correct estimators: McNemar for significance (only the discordant\n// pairs carry signal), the paired risk difference for effect size, Wilson for\n// a single-arm proportion CI, and pass@k for the standard k-sample coding-eval\n// metric. The normal approximation is wrong for proportions near 0/1 and for\n// the small discordant counts typical of eval runs, so these are exact /\n// Wilson-based, not Wald.\n\n/** A binomial proportion estimate with a confidence interval. */\nexport interface ProportionInterval {\n  /** Point estimate successes / n (0 when n = 0). */\n  estimate: number\n  /** Lower bound, clamped to [0, 1]. */\n  lower: number\n  /** Upper bound, clamped to [0, 1]. */\n  upper: number\n}\n\n/**\n * Wilson score interval for a binomial proportion. Correct at small n and near\n * 0/1, where the normal (Wald) approximation produces bounds outside [0, 1] and\n * understates coverage. Use this for any pass-rate / hit-rate / realness-rate\n * CI — the continuous `confidenceInterval` assumes the wrong distribution for a\n * proportion. `n = 0 ⇒ {0, 0, 0}`.\n */\nexport function wilson(successes: number, n: number, confidence = 0.95): ProportionInterval {\n  if (n <= 0) return { estimate: 0, lower: 0, upper: 0 }\n  if (successes < 0 || successes > n) {\n    throw new Error(`wilson: successes (${successes}) must be in [0, ${n}]`)\n  }\n  const z = zQuantile(1 - (1 - confidence) / 2)\n  const p = successes / n\n  const z2 = z * z\n  const denom = 1 + z2 / n\n  const center = (p + z2 / (2 * n)) / denom\n  const half = (z * Math.sqrt((p * (1 - p) + z2 / (4 * n)) / n)) / denom\n  return {\n    estimate: p,\n    lower: Math.max(0, center - half),\n    upper: Math.min(1, center + half),\n  }\n}\n\n/**\n * Are these per-item outcomes binary (every value exactly 0 or 1)?\n *\n * The discriminator a promotion gate needs before choosing a paired statistic.\n * On binary outcomes the paired delta vector lives in {-1, 0, +1} and is\n * normally dominated by zeros (both arms solve, or both arms miss, most items),\n * so its MEDIAN is pinned at exactly 0 no matter how large the real shift in\n * success rate is — and a bootstrap CI on that median collapses to [0, 0].\n * A gate keying on `ci.low > threshold` is then structurally unable to see\n * either a gain or a regression. Detect this shape and switch to the\n * paired-binary estimators ({@link mcnemar}, {@link pairedRiskDifference})\n * instead of silently answering \"no\" forever.\n *\n * Empty input is NOT binary: there is no evidence of the outcome's shape, and\n * defaulting an empty vector into the binary branch would pick a statistic on\n * no data at all.\n *\n * NOT the right discriminator for a gate. It recognises the literal {0, 1}\n * encoding and nothing else, so a pass/fail dimension emitted on 0-100 — which\n * judges in this codebase do routinely — reads as non-binary, and a single\n * partial-credit score in an otherwise pass/fail vector flips it to false while\n * leaving the median just as blind. Gates want {@link pairedBinaryScale} (any\n * two-point encoding). This predicate remains for callers that specifically\n * mean \"literally 0/1\".\n */\nexport function isBinaryOutcomeVector(values: ArrayLike<number>): boolean {\n  if (values.length === 0) return false\n  for (let i = 0; i < values.length; i++) {\n    const v = values[i]!\n    if (v !== 0 && v !== 1) return false\n  }\n  return true\n}\n\n/** Result of a McNemar paired-binary significance test. */\nexport interface McNemarResult {\n  /** Total paired observations. */\n  n: number\n  /** Discordant pairs (b + c) — the only ones that carry signal. */\n  nDiscordant: number\n  /** Pairs where treatment succeeded and control failed (\"newly correct\"). */\n  b: number\n  /** Pairs where control succeeded and treatment failed (\"newly wrong\"). */\n  c: number\n  /** Continuity-corrected chi-square statistic (reference; exact p drives the call). */\n  statistic: number\n  /** Two-sided p-value. Exact (binomial sign test on discordant pairs). */\n  pValue: number\n}\n\n/**\n * McNemar's test for paired binary outcomes — the correct significance test for\n * \"does treatment change the success rate vs control on the SAME items\". Only\n * discordant pairs (one arm right, the other wrong) carry information; concordant\n * pairs are uninformative, so a paired t-test / two-proportion z-test on the raw\n * rates is wrong here. The p-value is exact: under H0 the b \"treatment-wins\" are\n * Binomial(b + c, 0.5), so the two-sided p is the doubled binomial tail — correct\n * at the small discordant counts typical of eval runs (no continuity-corrected\n * chi-square approximation needed, though it is returned as `statistic` for\n * reference). Inputs are paired 0/1 (or boolean) arrays, control first to match\n * the module's (before, after) convention. Throws on unequal lengths.\n */\nexport function mcnemar(\n  control: ArrayLike<number | boolean>,\n  treatment: ArrayLike<number | boolean>,\n): McNemarResult {\n  if (control.length !== treatment.length) {\n    throw new Error(`mcnemar: unequal sample sizes (${control.length} vs ${treatment.length})`)\n  }\n  const n = control.length\n  let b = 0 // treatment 1, control 0\n  let c = 0 // treatment 0, control 1\n  for (let i = 0; i < n; i++) {\n    const ctrl = control[i] ? 1 : 0\n    const treat = treatment[i] ? 1 : 0\n    if (treat === 1 && ctrl === 0) b++\n    else if (treat === 0 && ctrl === 1) c++\n  }\n  const nDiscordant = b + c\n  const statistic = nDiscordant === 0 ? 0 : (Math.abs(b - c) - 1) ** 2 / nDiscordant\n  return { n, nDiscordant, b, c, statistic, pValue: binomialSignTwoSided(b, c) }\n}\n\n/** A paired binary effect size (treatment rate − control rate) with a CI. */\nexport interface RiskDifferenceResult {\n  /** Total paired observations. */\n  n: number\n  /** Discordant pairs: treatment-win count. */\n  b: number\n  /** Discordant pairs: control-win count. */\n  c: number\n  /** Paired risk difference p(treatment) − p(control) = (b − c) / n. */\n  riskDifference: number\n  /** Lower bound of the CI, clamped to [-1, 1]. */\n  lower: number\n  /** Upper bound of the CI, clamped to [-1, 1]. */\n  upper: number\n  /** Confidence level used. */\n  confidence: number\n}\n\n/**\n * Paired risk difference (the effect-size companion to {@link mcnemar}): the\n * change in success rate p(treatment) − p(control) on matched items, which for\n * paired binary data equals (b − c) / n. The CI uses the paired variance from\n * the discordant counts, not the independent-samples formula (which overstates\n * the interval by ignoring the pairing). Inputs are paired 0/1 (or boolean)\n * arrays, control first. Throws on unequal lengths.\n *\n * REPORTING ONLY — do NOT decide a promotion on this interval. The CI is a Wald\n * normal approximation, which badly UNDERCOVERS when only a handful of pairs are\n * discordant: at n = 3 with b = 2, c = 0 it returns [0.133, 1.000], excluding 0,\n * while McNemar's exact test on the same data gives p = 0.50. A gate keying on\n * `lower > 0` would promote noise. Use {@link pairedRiskDifferenceExact}, whose\n * interval is dual to the exact test by construction, for any decision.\n */\nexport function pairedRiskDifference(\n  control: ArrayLike<number | boolean>,\n  treatment: ArrayLike<number | boolean>,\n  confidence = 0.95,\n): RiskDifferenceResult {\n  if (control.length !== treatment.length) {\n    throw new Error(\n      `pairedRiskDifference: unequal sample sizes (${control.length} vs ${treatment.length})`,\n    )\n  }\n  const n = control.length\n  if (n === 0) return { n: 0, b: 0, c: 0, riskDifference: 0, lower: 0, upper: 0, confidence }\n  let b = 0\n  let c = 0\n  for (let i = 0; i < n; i++) {\n    const ctrl = control[i] ? 1 : 0\n    const treat = treatment[i] ? 1 : 0\n    if (treat === 1 && ctrl === 0) b++\n    else if (treat === 0 && ctrl === 1) c++\n  }\n  const rd = (b - c) / n\n  const variance = (b + c - (b - c) ** 2 / n) / (n * n)\n  const z = zQuantile(1 - (1 - confidence) / 2)\n  const half = z * Math.sqrt(Math.max(0, variance))\n  return {\n    n,\n    b,\n    c,\n    riskDifference: rd,\n    lower: Math.max(-1, rd - half),\n    upper: Math.min(1, rd + half),\n    confidence,\n  }\n}\n\n/** A paired binary effect size with an EXACT interval and the exact test that\n *  bounds it — one object so a caller cannot read the estimate without the\n *  significance it is entitled to. */\nexport interface ExactRiskDifferenceResult {\n  /** Total paired observations. */\n  n: number\n  /** Discordant pairs: treatment-win count. */\n  b: number\n  /** Discordant pairs: control-win count. */\n  c: number\n  /** Discordant pairs (b + c) — the only ones carrying information. */\n  nDiscordant: number\n  /** Paired risk difference p(treatment) − p(control) = (b − c) / n. */\n  riskDifference: number\n  /** Exact conditional CI lower bound. 0 when there are no discordant pairs. */\n  lower: number\n  /** Exact conditional CI upper bound. 0 when there are no discordant pairs. */\n  upper: number\n  /** Confidence level used. */\n  confidence: number\n  /** McNemar's exact two-sided p-value on the same discordant counts. */\n  pValue: number\n}\n\n/**\n * Paired risk difference with the EXACT CONDITIONAL interval — the estimator a\n * promotion gate may decide on.\n *\n * Conditional on the number of discordant pairs m = b + c, the treatment-win\n * count b is Binomial(m, π) with π = P(treatment wins | discordant), and the\n * risk difference is an exact reparameterisation: RD = (2π − 1)·m/n. So a\n * Clopper-Pearson exact interval for π maps straight onto RD. This buys the\n * property the Wald interval in {@link pairedRiskDifference} does not have:\n *\n *   **`lower > 0` ⟺ McNemar's exact test rejects at α = 1 − confidence.**\n *\n * Clopper-Pearson excludes π = 0.5 exactly when the two-sided exact binomial\n * test of π = 0.5 rejects, and that test IS {@link mcnemar}'s p-value — so the\n * interval and the test can never disagree, and a gate keyed on `lower` cannot\n * promote what the exact test refuses. The exact p is returned in the same\n * object so the two are impossible to compute apart.\n *\n * The interval is conservative (exact intervals over-cover; conditioning on m\n * discards the concordant pairs' information about m itself). That is the\n * correct direction for a promotion gate: it refuses more often, never less.\n *\n * With m = 0 there are no discordant pairs and π is not identified: the result\n * is the degenerate [0, 0] with p = 1. That is NOT evidence of equivalence —\n * callers must treat a zero-width interval as \"cannot decide\", not as \"no\n * difference\". Inputs are paired 0/1 (or boolean) arrays, control first.\n * Throws on unequal lengths.\n */\nexport function pairedRiskDifferenceExact(\n  control: ArrayLike<number | boolean>,\n  treatment: ArrayLike<number | boolean>,\n  confidence = 0.95,\n): ExactRiskDifferenceResult {\n  if (control.length !== treatment.length) {\n    throw new Error(\n      `pairedRiskDifferenceExact: unequal sample sizes (${control.length} vs ${treatment.length})`,\n    )\n  }\n  if (confidence <= 0 || confidence >= 1) {\n    throw new Error(`pairedRiskDifferenceExact: confidence must be in (0,1), got ${confidence}`)\n  }\n  const n = control.length\n  if (n === 0) {\n    return {\n      n: 0,\n      b: 0,\n      c: 0,\n      nDiscordant: 0,\n      riskDifference: 0,\n      lower: 0,\n      upper: 0,\n      confidence,\n      pValue: 1,\n    }\n  }\n  let b = 0\n  let c = 0\n  for (let i = 0; i < n; i++) {\n    const ctrl = control[i] ? 1 : 0\n    const treat = treatment[i] ? 1 : 0\n    if (treat === 1 && ctrl === 0) b++\n    else if (treat === 0 && ctrl === 1) c++\n  }\n  const m = b + c\n  const riskDifference = (b - c) / n\n  const pValue = binomialSignTwoSided(b, c)\n  if (m === 0) {\n    return { n, b, c, nDiscordant: 0, riskDifference: 0, lower: 0, upper: 0, confidence, pValue }\n  }\n  const alpha = 1 - confidence\n  const piLow = b === 0 ? 0 : betaQuantile(alpha / 2, b, m - b + 1)\n  const piHigh = b === m ? 1 : betaQuantile(1 - alpha / 2, b + 1, m - b)\n  const scale = m / n\n  return {\n    n,\n    b,\n    c,\n    nDiscordant: m,\n    riskDifference,\n    lower: Math.max(-1, (2 * piLow - 1) * scale),\n    upper: Math.min(1, (2 * piHigh - 1) * scale),\n    confidence,\n    pValue,\n  }\n}\n\n/** Inverse regularized incomplete beta by bisection on\n *  {@link regularizedIncompleteBeta}, which is monotone increasing in x. 80\n *  halvings of [0,1] resolve to ~8e-25, far past the continued fraction's own\n *  3e-15 tolerance, so the quantile is as exact as the CDF it inverts. */\nfunction betaQuantile(p: number, a: number, b: number): number {\n  if (p <= 0) return 0\n  if (p >= 1) return 1\n  let lo = 0\n  let hi = 1\n  for (let i = 0; i < 80; i++) {\n    const mid = (lo + hi) / 2\n    if (regularizedIncompleteBeta(mid, a, b) < p) lo = mid\n    else hi = mid\n  }\n  return (lo + hi) / 2\n}\n\n/** A paired binary effect size with an interval that is valid at a NONZERO\n *  margin — the estimator a noninferiority decision may be made on. */\nexport interface ScoreRiskDifferenceResult {\n  /** Total paired observations. */\n  n: number\n  /** Discordant pairs: treatment-win count. */\n  b: number\n  /** Discordant pairs: control-win count. */\n  c: number\n  /** Discordant pairs (b + c). */\n  nDiscordant: number\n  /** Paired risk difference p(treatment) − p(control) = (b − c) / n. */\n  riskDifference: number\n  /** Score-interval lower bound on the population risk difference. */\n  lower: number\n  /** Score-interval upper bound on the population risk difference. */\n  upper: number\n  /** Confidence level used. */\n  confidence: number\n}\n\n/**\n * Constrained MLE of q = P(treatment loses) under the hypothesis RD = `delta`.\n *\n * Profiling the two concordant cells out of the multinomial leaves\n * `L(q) = b·log(q+delta) + c·log(q) + e·log(1 − 2q − delta)` with `e = n − b − c`,\n * whose stationary point is the positive root of\n * `2n·q² − [(b + c) − delta·(b + 3c + 2e)]·q − c·delta·(1 − delta) = 0`.\n * At `delta = 0` this returns `(b + c) / 2n`, the familiar null.\n */\nfunction constrainedLossRate(b: number, c: number, n: number, delta: number): number {\n  const e = n - b - c\n  const quadratic = 2 * n\n  const linear = -(b + c - delta * (b + 3 * c + 2 * e))\n  const constant = -c * delta * (1 - delta)\n  const discriminant = linear * linear - 4 * quadratic * constant\n  const root = discriminant > 0 ? Math.sqrt(discriminant) : 0\n  const q = (-linear + root) / (2 * quadratic)\n  // Clamp into the region where all four cell probabilities stay non-negative.\n  return Math.min(Math.max(q, Math.max(0, -delta)), Math.max(0, (1 - delta) / 2))\n}\n\n/** Tango's score statistic for H0: RD = `delta`. `Var(b − c) = n·(2q + delta −\n *  delta²)` under that hypothesis, evaluated at the constrained MLE of q. */\nfunction tangoScore(b: number, c: number, n: number, delta: number): number {\n  const numerator = b - c - n * delta\n  const q = constrainedLossRate(b, c, n, delta)\n  const variance = n * (2 * q + delta * (1 - delta))\n  if (!(variance > 0)) {\n    if (numerator === 0) return 0\n    return numerator > 0 ? Number.POSITIVE_INFINITY : Number.NEGATIVE_INFINITY\n  }\n  return numerator / Math.sqrt(variance)\n}\n\n/**\n * Paired risk difference with TANGO'S (1998) SCORE INTERVAL — the estimator a\n * promotion gate may decide on **at a nonzero margin**.\n *\n * {@link pairedRiskDifferenceExact} conditions on the observed discordant count\n * `m = b + c`, builds a Clopper-Pearson interval for the win share among those\n * `m` pairs, and multiplies by the observed `m/n`. That is exact for testing\n * RD = 0 — it is dual to McNemar — but it is NOT a confidence interval for the\n * population risk difference at a nonzero margin, because the sampling\n * variability of `m/n` itself is discarded. The gap is not academic: with the\n * production caller's `pairedDeltaThreshold: -0.05`, a process whose true risk\n * difference sits exactly on that margin clears a nominal-95 % `lower > margin`\n * check 24.75 % of the time at n = 40 and 43.95 % at n = 76 (2000 replicates\n * each) when the conditional interval decides.\n *\n * Tango's interval inverts the score test of RD = delta, which estimates the\n * nuisance loss rate under each hypothesised delta instead of fixing it at the\n * observed value, so `m` contributes its own uncertainty. It is the method\n * `ratesci::scorepairci` uses for paired risk-difference noninferiority, and it\n * is not conditional, so it stays valid as the margin moves away from zero.\n *\n * The bounds are found by bisecting `tangoScore(delta) = ±z` — the score is\n * monotone decreasing in delta, so each crossing is unique. Inputs are paired\n * 0/1 (or boolean) arrays, control first. Throws on unequal lengths.\n */\nexport function pairedRiskDifferenceScore(\n  control: ArrayLike<number | boolean>,\n  treatment: ArrayLike<number | boolean>,\n  confidence = 0.95,\n): ScoreRiskDifferenceResult {\n  if (control.length !== treatment.length) {\n    throw new Error(\n      `pairedRiskDifferenceScore: unequal sample sizes (${control.length} vs ${treatment.length})`,\n    )\n  }\n  if (confidence <= 0 || confidence >= 1) {\n    throw new Error(`pairedRiskDifferenceScore: confidence must be in (0,1), got ${confidence}`)\n  }\n  const n = control.length\n  if (n === 0) {\n    return { n: 0, b: 0, c: 0, nDiscordant: 0, riskDifference: 0, lower: -1, upper: 1, confidence }\n  }\n  let b = 0\n  let c = 0\n  for (let i = 0; i < n; i++) {\n    const ctrl = control[i] ? 1 : 0\n    const treat = treatment[i] ? 1 : 0\n    if (treat === 1 && ctrl === 0) b++\n    else if (treat === 0 && ctrl === 1) c++\n  }\n  const riskDifference = (b - c) / n\n  const z = zQuantile(1 - (1 - confidence) / 2)\n\n  // Lower bound: the smallest delta still inside the interval, i.e. the root of\n  // score(delta) = +z on [-1, riskDifference]. score(riskDifference) = 0 < z, so\n  // the right endpoint is always inside and the bisection is well posed.\n  let lo = -1\n  let hi = riskDifference\n  for (let i = 0; i < 200; i++) {\n    const mid = (lo + hi) / 2\n    if (tangoScore(b, c, n, mid) > z) lo = mid\n    else hi = mid\n  }\n  const lower = (lo + hi) / 2\n\n  // Upper bound: root of score(delta) = -z on [riskDifference, 1].\n  let ulo = riskDifference\n  let uhi = 1\n  for (let i = 0; i < 200; i++) {\n    const mid = (ulo + uhi) / 2\n    if (tangoScore(b, c, n, mid) > -z) ulo = mid\n    else uhi = mid\n  }\n  const upper = (ulo + uhi) / 2\n\n  return {\n    n,\n    b,\n    c,\n    nDiscordant: b + c,\n    riskDifference,\n    lower: Math.max(-1, lower),\n    upper: Math.min(1, upper),\n    confidence,\n  }\n}\n\n/**\n * The common positive level `s` such that EVERY value across both paired arms is\n * exactly 0 or `s` — i.e. the outcome is pass/fail, whatever encoding it arrived\n * in. Returns null when the outcomes are not two-point, when the two arms use\n * different levels, or when no positive value was observed at all (all-zero\n * arms: the level is not identified, and there is nothing to decide anyway).\n *\n * This is the scale-aware successor to {@link isBinaryOutcomeVector}, which only\n * recognises literal {0, 1}. Judges in this codebase emit dimensions on 0-100 as\n * well as [0,1] (see `detectScale` in `campaign/gates/statistical-heldout.ts`),\n * so a pass/fail dimension routinely arrives as {0, 100} and a {0,1}-only test\n * silently sends it down the median path that cannot see it. Any positive level\n * is accepted, not just 1 and 100: for a two-point {0, s} outcome the mean paired\n * delta is exactly s·(b − c)/n, so the binary estimators apply after dividing by\n * s and rescaling the result back into the caller's native units.\n *\n * Non-finite values ⇒ null: an unusable outcome must not be classified as a\n * clean pass/fail shape.\n */\nexport function pairedBinaryScale(\n  before: ArrayLike<number>,\n  after: ArrayLike<number>,\n): number | null {\n  let level: number | null = null\n  for (const arm of [before, after]) {\n    for (let i = 0; i < arm.length; i++) {\n      const v = arm[i]!\n      if (!Number.isFinite(v)) return null\n      if (v === 0) continue\n      if (v < 0) return null\n      if (level === null) level = v\n      else if (v !== level) return null\n    }\n  }\n  return level\n}\n\n/**\n * Unbiased pass@k for code generation (Chen et al. 2021, \"Evaluating Large\n * Language Models Trained on Code\"). Given `n` independent samples for one\n * problem of which `c` pass, the probability that at least one of a random k of\n * them passes is 1 − C(n−c, k) / C(n, k). Estimating pass@k as \"did any of the\n * first k pass\" is biased high at small n; this is the variance-reduced estimator\n * averaged implicitly over all k-subsets. Average the per-problem values across\n * the suite for the corpus pass@k. Computed in the numerically stable product\n * form. Requires 1 ≤ k ≤ n and 0 ≤ c ≤ n.\n */\nexport function passAtK(n: number, c: number, k: number): number {\n  if (!Number.isInteger(n) || !Number.isInteger(c) || !Number.isInteger(k)) {\n    throw new Error(`passAtK: n, c, k must be integers (got n=${n}, c=${c}, k=${k})`)\n  }\n  if (k < 1 || k > n || c < 0 || c > n) {\n    throw new Error(`passAtK: require 1 ≤ k ≤ n and 0 ≤ c ≤ n (got n=${n}, c=${c}, k=${k})`)\n  }\n  if (n - c < k) return 1\n  let prob = 1\n  for (let i = n - c + 1; i <= n; i++) prob *= 1 - k / i\n  return 1 - prob\n}\n","/**\n * Standard normal cumulative distribution using Abramowitz and Stegun 7.1.26.\n *\n * The approximation is evaluated as erf(x / sqrt(2)). Computing the negative\n * tail from the complementary term avoids cancellation when x is far below 0.\n * The maximum absolute CDF error is approximately 7.5e-8.\n */\nexport function normalCdf(x: number): number {\n  if (x === 0) return 0.5\n\n  const a1 = 0.254829592\n  const a2 = -0.284496736\n  const a3 = 1.421413741\n  const a4 = -1.453152027\n  const a5 = 1.061405429\n  const p = 0.3275911\n\n  const scaled = Math.abs(x) / Math.SQRT2\n  const t = 1 / (1 + p * scaled)\n  const complement = ((((a5 * t + a4) * t + a3) * t + a2) * t + a1) * t * Math.exp(-scaled * scaled)\n\n  return x < 0 ? complement / 2 : 1 - complement / 2\n}\n","import { ValidationError } from '../errors'\nimport { normalCdf } from '../math/normal'\nimport { lnGamma } from '../math/special-functions'\nimport { assertFiniteSample, makeRng, symmetricTwoSampleSeed } from './internal'\n\n// ── Rank tests: exact by default ─────────────────────────────────────\n//\n// At 3–10 repetitions per arm the binding constraint on a rank test is\n// combinatorial, not numerical. Three versus three admits only 20 splits, so\n// the attainable two-sided p-grid starts at 0.1000 and α = 0.05 is out of\n// reach at that design; a normal approximation reports 0.0495, a p-value that\n// describes no attainable outcome. No better approximation fixes this — adding\n// the tie correction moves `[0,0,0]` versus `[1,1,1]` from 0.0495 to 0.0469,\n// further from its exact 0.1000, not closer.\n//\n// So both null distributions are enumerated exactly inside bounded state and\n// work budgets, by convolution over the observed midranks (identical to\n// enumerating every split / sign pattern, and cheaper), which conditions on the\n// realised tie pattern for free. Above those budgets the default is a seeded\n// Monte Carlo permutation. The asymptotic path is never chosen automatically,\n// and asking for it inside the exact-feasible range throws.\n//\n// Every result carries `method` and `pFloor` so a downstream gate can SEE the\n// discreteness rather than infer it: a gate handed data whose `pFloor` exceeds\n// its alpha is underpowered by construction, which is a true statement about\n// the experiment, not a false one about the effect.\n\n/** How a rank test's p-value was actually computed. */\nexport type RankTestMethod = 'exact' | 'permutation' | 'asymptotic'\n\n/**\n * What the caller asks for. `'auto'` selects `'exact'` inside the enumeration\n * threshold and `'permutation'` above it, and never selects `'asymptotic'`.\n */\nexport type RankTestMethodRequest = 'auto' | 'exact' | 'asymptotic'\n\nexport interface RankTestOptions {\n  /** Default `'auto'`. `'asymptotic'` inside the exact-feasible range throws. */\n  method?: RankTestMethodRequest\n  /** Resamples on the Monte Carlo permutation path. Default 100000. */\n  permutations?: number\n  /** Seed for the permutation path. Omitted ⇒ derived from the data itself, so\n   *  the result is reproducible either way. */\n  seed?: number\n}\n\n/** Maximum dynamic-programming cells used by an exact two-sample rank test. */\nexport const MANN_WHITNEY_EXACT_MAX_STATES = 8_192\n/** Maximum inner-loop transitions used by an exact two-sample rank test. */\nexport const MANN_WHITNEY_EXACT_MAX_WORK = 250_000\n/** Non-zero differences up to which the signed-rank null is enumerated exactly. */\nexport const WILCOXON_EXACT_MAX_N = 20\n/** Resamples used when a rank test falls back to Monte Carlo permutation. */\nexport const DEFAULT_PERMUTATIONS = 100_000\n\nexport interface MannWhitneyResult {\n  /** `min(U_a, U_b)` — the conventional reported statistic. */\n  u: number\n  /** U for sample `a`. Carries the direction of the effect, which `u` discards. */\n  uA: number\n  /** Two-sided p-value. */\n  p: number\n  /** How `p` was computed. */\n  method: RankTestMethod\n  /** Smallest two-sided p this design can produce. `p` can never be below it. */\n  pFloor: number\n}\n\n/**\n * Mann-Whitney U — two independent samples, no distributional assumption.\n *\n * Exact conditional (permutation) p by default when the dynamic program fits\n * {@link MANN_WHITNEY_EXACT_MAX_STATES} cells and\n * {@link MANN_WHITNEY_EXACT_MAX_WORK} transitions, seeded Monte Carlo\n * permutation above those limits. This keeps imbalanced designs such as 1+24\n * exact without admitting expensive balanced designs merely because they have\n * the same total size. Throws on non-finite input and on `method:\n * 'asymptotic'` where an exact answer is available. Empty input yields `p = 1,\n * pFloor = 1` — no design, no attainable evidence.\n */\nexport function mannWhitneyU(\n  a: number[],\n  b: number[],\n  opts: RankTestOptions = {},\n): MannWhitneyResult {\n  assertFiniteSample('mannWhitneyU', 'a', a)\n  assertFiniteSample('mannWhitneyU', 'b', b)\n\n  const n1 = a.length\n  const n2 = b.length\n  if (n1 === 0 || n2 === 0) return { u: 0, uA: 0, p: 1, method: 'exact', pFloor: 1 }\n\n  const total = n1 + n2\n  const combined = [\n    ...a.map((v) => ({ v, fromA: true })),\n    ...b.map((v) => ({ v, fromA: false })),\n  ].sort((x, y) => x.v - y.v)\n\n  const { midranks, tieTerm } = midranksWithTieTerm(combined.map((entry) => entry.v))\n  let rankSumA = 0\n  for (let k = 0; k < total; k++) {\n    if (combined[k]!.fromA) rankSumA += midranks[k]!\n  }\n\n  const uA = rankSumA - (n1 * (n1 + 1)) / 2\n  const u = Math.min(uA, n1 * n2 - uA)\n  // Midranks are integers or halves, so doubling makes the null convolution\n  // integral. U is centred on n₁n₂/2, hence a doubled centre of n₁n₂.\n  const doubled = midranks.map((rank) => Math.round(rank * 2))\n  const doubledDeviation = Math.abs(2 * uA - n1 * n2)\n  const selectedN = Math.min(n1, n2)\n  const otherN = total - selectedN\n  const exactCost = exactTwoSampleCost(doubled, selectedN)\n\n  const designFloor = exactTwoSampleFloor(doubled, selectedN)\n  const method = selectRankTestMethod(\n    'mannWhitneyU',\n    opts.method ?? 'auto',\n    `n1=${n1}, n2=${n2}`,\n    exactCost.states <= MANN_WHITNEY_EXACT_MAX_STATES &&\n      exactCost.work <= MANN_WHITNEY_EXACT_MAX_WORK,\n    designFloor,\n    `${MANN_WHITNEY_EXACT_MAX_STATES.toLocaleString('en-US')} states and ` +\n      `${MANN_WHITNEY_EXACT_MAX_WORK.toLocaleString('en-US')} transitions`,\n  )\n\n  if (method === 'exact') {\n    const { p, pFloor } = exactTwoSampleP(doubled, selectedN, otherN, doubledDeviation)\n    return { u, uA, p, method, pFloor }\n  }\n\n  if (method === 'asymptotic') {\n    return {\n      u,\n      uA,\n      p: asymptoticTwoSidedP(doubledDeviation / 2, twoSampleSigma(n1, n2, total, tieTerm)),\n      method,\n      pFloor: designFloor,\n    }\n  }\n\n  const permutations = resolvePermutations('mannWhitneyU', opts.permutations)\n  const rng = opts.seed === undefined ? makeRng(symmetricTwoSampleSeed(a, b)) : makeRng(opts.seed)\n  let atLeastAsExtreme = 0\n  const pool = [...doubled]\n  for (let iteration = 0; iteration < permutations; iteration++) {\n    let doubledRankSum = 0\n    for (let k = 0; k < selectedN; k++) {\n      const pick = k + Math.floor(rng() * (total - k))\n      const swapped = pool[pick]!\n      pool[pick] = pool[k]!\n      pool[k] = swapped\n      doubledRankSum += swapped\n    }\n    if (\n      Math.abs(doubledRankSum - selectedN * (selectedN + 1) - selectedN * otherN) >=\n      doubledDeviation\n    ) {\n      atLeastAsExtreme++\n    }\n  }\n  const pFloor = Math.max(1 / (permutations + 1), designFloor)\n  return {\n    u,\n    uA,\n    p: Math.max((1 + atLeastAsExtreme) / (permutations + 1), pFloor),\n    method,\n    pFloor,\n  }\n}\n\nexport interface WilcoxonSignedRankResult {\n  /** W⁺, the rank sum of the positive differences. (scipy reports\n   *  `min(W⁺, W⁻)`; compare statistics only after converting.) */\n  w: number\n  /** Two-sided p-value. */\n  p: number\n  /** How `p` was computed. */\n  method: RankTestMethod\n  /** Smallest two-sided p this design can produce. */\n  pFloor: number\n  /** Non-zero differences — zero differences are dropped and carry no rank. */\n  nNonZero: number\n}\n\n/**\n * Wilcoxon signed-rank — paired, no distributional assumption on the deltas.\n *\n * Exact conditional (sign-flip) p by default at `n ≤\n * {@link WILCOXON_EXACT_MAX_N}` non-zero differences, seeded Monte Carlo\n * permutation above it. Throws on non-finite input and on `method:\n * 'asymptotic'` where an exact answer is available.\n *\n * `n` is the count of NON-ZERO differences: exact ties are dropped before\n * ranking, so a run of tied pairs shrinks the design and raises `pFloor`.\n * All-tied input yields `p = 1, pFloor = 1` — no attainable evidence, which\n * `pFloor` states rather than leaving `p = 1` to be read as a measured null.\n */\nexport function wilcoxonSignedRank(\n  before: number[],\n  after: number[],\n  opts: RankTestOptions = {},\n): WilcoxonSignedRankResult {\n  if (before.length !== after.length) {\n    throw new ValidationError(\n      `wilcoxonSignedRank: unequal sample sizes (${before.length} vs ${after.length})`,\n    )\n  }\n  assertFiniteSample('wilcoxonSignedRank', 'before', before)\n  assertFiniteSample('wilcoxonSignedRank', 'after', after)\n\n  const diffs = before.map((b, i) => after[i]! - b).filter((d) => d !== 0)\n  const n = diffs.length\n  if (n === 0) return { w: 0, p: 1, method: 'exact', pFloor: 1, nNonZero: 0 }\n\n  const order = diffs.map((d, i) => ({ abs: Math.abs(d), i })).sort((x, y) => x.abs - y.abs)\n  const { midranks, tieTerm } = midranksWithTieTerm(order.map((entry) => entry.abs))\n  const ranks: number[] = new Array(n)\n  for (let k = 0; k < n; k++) ranks[order[k]!.i] = midranks[k]!\n\n  let wPlus = 0\n  for (let k = 0; k < n; k++) if (diffs[k]! > 0) wPlus += ranks[k]!\n\n  // Midranks are integers or halves; doubling makes the convolution integral.\n  // Σ midranks = n(n+1)/2 whatever the tie pattern, so E[W⁺] = n(n+1)/4 and\n  // the doubled centre is n(n+1)/2.\n  const doubled = midranks.map((rank) => Math.round(rank * 2))\n  const doubledDeviation = Math.abs(2 * wPlus - (n * (n + 1)) / 2)\n\n  const designFloor = Math.min(1, 2 ** (1 - n))\n  const method = selectRankTestMethod(\n    'wilcoxonSignedRank',\n    opts.method ?? 'auto',\n    `n=${n} non-zero differences`,\n    n <= WILCOXON_EXACT_MAX_N,\n    designFloor,\n    `${WILCOXON_EXACT_MAX_N} non-zero differences`,\n  )\n\n  if (method === 'exact') {\n    const { p, pFloor } = exactSignedRankP(doubled, doubledDeviation)\n    return { w: wPlus, p, method, pFloor, nNonZero: n }\n  }\n\n  if (method === 'asymptotic') {\n    const variance = (n * (n + 1) * (2 * n + 1)) / 24 - tieTerm / 48\n    return {\n      w: wPlus,\n      p: asymptoticTwoSidedP(doubledDeviation / 2, Math.sqrt(variance)),\n      method,\n      pFloor: designFloor,\n      nNonZero: n,\n    }\n  }\n\n  const permutations = resolvePermutations('wilcoxonSignedRank', opts.permutations)\n  const rng = makeRng(opts.seed, before, after)\n  const doubledCentre = (n * (n + 1)) / 2\n  let atLeastAsExtreme = 0\n  for (let iteration = 0; iteration < permutations; iteration++) {\n    let doubledWPlus = 0\n    for (let k = 0; k < n; k++) if (rng() < 0.5) doubledWPlus += doubled[k]!\n    if (Math.abs(doubledWPlus - doubledCentre) >= doubledDeviation) atLeastAsExtreme++\n  }\n  return {\n    w: wPlus,\n    p: (1 + atLeastAsExtreme) / (permutations + 1),\n    method,\n    pFloor: Math.max(1 / (permutations + 1), designFloor),\n    nNonZero: n,\n  }\n}\n\n/**\n * Average ranks over an ASCENDING-sorted array, plus `Σ(t³ − t)` over tie\n * groups of size `t` — the correction term both asymptotic rank-test variances\n * need.\n */\nfunction midranksWithTieTerm(sorted: readonly number[]): { midranks: number[]; tieTerm: number } {\n  const midranks = new Array<number>(sorted.length)\n  let tieTerm = 0\n  let i = 0\n  while (i < sorted.length) {\n    let j = i\n    while (j < sorted.length && sorted[j] === sorted[i]) j++\n    const average = (i + 1 + j) / 2\n    for (let k = i; k < j; k++) midranks[k] = average\n    const groupSize = j - i\n    if (groupSize > 1) tieTerm += groupSize ** 3 - groupSize\n    i = j\n  }\n  return { midranks, tieTerm }\n}\n\nfunction selectRankTestMethod(\n  fn: string,\n  request: RankTestMethodRequest,\n  design: string,\n  exactFeasible: boolean,\n  designFloor: number,\n  threshold: string,\n): RankTestMethod {\n  if (request === 'auto') return exactFeasible ? 'exact' : 'permutation'\n  if (request === 'exact') {\n    if (exactFeasible) return 'exact'\n    throw new ValidationError(\n      `${fn}: method 'exact' is out of range at ${design} — enumeration is bounded by ` +\n        `${threshold}. Use 'auto' for the seeded Monte Carlo permutation, which converges ` +\n        'to the same answer.',\n    )\n  }\n  if (exactFeasible) {\n    throw new ValidationError(\n      `${fn}: method 'asymptotic' is refused at ${design} — the exact p-grid at this design ` +\n        `starts at ${formatProbability(designFloor)}, so an asymptotic p below it describes no ` +\n        `attainable outcome. Use method 'exact' (the default) or add repetitions past ` +\n        `${threshold}.`,\n    )\n  }\n  return 'asymptotic'\n}\n\nfunction resolvePermutations(fn: string, permutations: number | undefined): number {\n  if (permutations === undefined) return DEFAULT_PERMUTATIONS\n  if (!Number.isInteger(permutations) || permutations < 1) {\n    throw new ValidationError(`${fn}: permutations must be a positive integer, got ${permutations}`)\n  }\n  return permutations\n}\n\n/** Two-sided normal-approximation tail with the continuity correction. */\nfunction asymptoticTwoSidedP(deviation: number, sigma: number): number {\n  if (!(sigma > 0)) return 1\n  return Math.min(1, 2 * (1 - normalCdf(Math.max(0, deviation - 0.5) / sigma)))\n}\n\n/** SD of U under the permutation null, corrected for the realised ties. The\n *  tie term reduces (N+1) and reaches it exactly when every value is tied, so\n *  the variance floors at 0 rather than going negative. */\nfunction twoSampleSigma(n1: number, n2: number, total: number, tieTerm: number): number {\n  if (total < 2) return 0\n  const variance = ((n1 * n2) / 12) * (total + 1 - tieTerm / (total * (total - 1)))\n  return Math.sqrt(Math.max(0, variance))\n}\n\nfunction logChoose(n: number, k: number): number {\n  return lnGamma(n + 1) - lnGamma(k + 1) - lnGamma(n - k + 1)\n}\n\nfunction formatProbability(value: number): string {\n  return value >= 1e-4 || value === 0 ? value.toFixed(4) : value.toExponential(3)\n}\n\n/**\n * Exact DP allocation and loop count for this observed rank vector.\n *\n * The smaller arm is sufficient because selecting its complement produces the\n * same two-sided U deviation while using fewer rows in the state table.\n */\nfunction exactTwoSampleCost(\n  doubledRanks: readonly number[],\n  selectedN: number,\n): { states: number; work: number } {\n  const maxSum = doubledRanks.reduce((sum, rank) => sum + rank, 0)\n  let work = 0\n  for (let placed = 0; placed < doubledRanks.length; placed++) {\n    work += Math.min(selectedN, placed + 1) * (maxSum - doubledRanks[placed]! + 1)\n  }\n  return {\n    states: (selectedN + 1) * (maxSum + 1),\n    work,\n  }\n}\n\n/**\n * Smallest attainable two-sided p under the observed ties.\n *\n * Only subsets with the minimum or maximum rank sum can attain the largest\n * deviation. Their multiplicity is the number of ways to choose within the\n * tie group at each boundary, so this calculation is exact without allocating\n * the full null distribution.\n */\nfunction exactTwoSampleFloor(doubledRanks: readonly number[], selectedN: number): number {\n  const total = doubledRanks.length\n  const otherN = total - selectedN\n  const minimumSum = doubledRanks.slice(0, selectedN).reduce((sum, rank) => sum + rank, 0)\n  const maximumSum = doubledRanks.slice(total - selectedN).reduce((sum, rank) => sum + rank, 0)\n  if (minimumSum === maximumSum) return 1\n\n  const centre = selectedN * (selectedN + 1) + selectedN * otherN\n  const minimumDeviation = Math.abs(minimumSum - centre)\n  const maximumDeviation = Math.abs(maximumSum - centre)\n  const totalLogWays = logChoose(total, selectedN)\n  const minimumMass = Math.exp(\n    logExtremeSubsetWays(doubledRanks, selectedN, 'minimum') - totalLogWays,\n  )\n  const maximumMass = Math.exp(\n    logExtremeSubsetWays(doubledRanks, selectedN, 'maximum') - totalLogWays,\n  )\n\n  if (minimumDeviation > maximumDeviation) return minimumMass\n  if (maximumDeviation > minimumDeviation) return maximumMass\n  return Math.min(1, minimumMass + maximumMass)\n}\n\nfunction logExtremeSubsetWays(\n  sortedRanks: readonly number[],\n  selectedN: number,\n  side: 'minimum' | 'maximum',\n): number {\n  const boundaryIndex = side === 'minimum' ? selectedN - 1 : sortedRanks.length - selectedN\n  const boundary = sortedRanks[boundaryIndex]!\n  let first = boundaryIndex\n  let afterLast = boundaryIndex + 1\n  while (first > 0 && sortedRanks[first - 1] === boundary) first--\n  while (afterLast < sortedRanks.length && sortedRanks[afterLast] === boundary) afterLast++\n\n  const tieSize = afterLast - first\n  const fixed = side === 'minimum' ? first : sortedRanks.length - afterLast\n  return logChoose(tieSize, selectedN - fixed)\n}\n\n/**\n * Exact conditional two-sided p for the two-sample rank test.\n *\n * Convolves the observed doubled midranks into the null distribution of group\n * a's rank sum over every `C(n₁+n₂, n₁)` split — identical to enumerating the\n * splits, but `O(N·n₁·ΣR)` instead of `O(C(N,n₁)·n₁n₂)`. Conditioning on the\n * realised multiset makes the tie handling exact rather than a correction.\n *\n * The null is symmetric about `n₁n₂/2` (negating every value maps `U → n₁n₂ −\n * U` and permutes the split set onto itself), so the two-sided p is the mass\n * at least as far from the centre as the observation.\n */\nfunction exactTwoSampleP(\n  doubledRanks: readonly number[],\n  n1: number,\n  n2: number,\n  doubledDeviation: number,\n): { p: number; pFloor: number } {\n  const maxSum = doubledRanks.reduce((sum, rank) => sum + rank, 0)\n  const width = maxSum + 1\n  // ways[k][s] = number of size-k subsets whose doubled rank sum is s.\n  const ways: Float64Array[] = Array.from({ length: n1 + 1 }, () => new Float64Array(width))\n  ways[0]![0] = 1\n  let placed = 0\n  for (const rank of doubledRanks) {\n    for (let k = Math.min(n1, placed + 1); k >= 1; k--) {\n      const from = ways[k - 1]!\n      const into = ways[k]!\n      for (let sum = maxSum - rank; sum >= 0; sum--) {\n        const count = from[sum]!\n        if (count !== 0) into[sum + rank]! += count\n      }\n    }\n    placed++\n  }\n\n  // U = rankSum − n₁(n₁+1)/2, so doubled U = s − n₁(n₁+1), and the doubled\n  // centre 2·(n₁n₂/2) is n₁n₂.\n  const shift = n1 * (n1 + 1) + n1 * n2\n  const chosen = ways[n1]!\n  let totalWays = 0\n  let extremeWays = 0\n  let tailWays = 0\n  let maxDeviation = -1\n  for (let sum = 0; sum < width; sum++) {\n    const count = chosen[sum]!\n    if (count === 0) continue\n    totalWays += count\n    const deviation = Math.abs(sum - shift)\n    if (deviation >= doubledDeviation) tailWays += count\n    if (deviation > maxDeviation) {\n      maxDeviation = deviation\n      extremeWays = count\n    } else if (deviation === maxDeviation) {\n      extremeWays += count\n    }\n  }\n  return { p: tailWays / totalWays, pFloor: extremeWays / totalWays }\n}\n\n/**\n * Exact conditional two-sided p for the paired signed-rank test.\n *\n * Convolves the observed doubled absolute midranks over all `2ⁿ` sign\n * assignments in `O(n·ΣR)`. Probabilities rather than counts keep `2ⁿ` off the\n * arithmetic. The null is symmetric about `n(n+1)/4`.\n */\nfunction exactSignedRankP(\n  doubledRanks: readonly number[],\n  doubledDeviation: number,\n): { p: number; pFloor: number } {\n  const maxSum = doubledRanks.reduce((sum, rank) => sum + rank, 0)\n  const width = maxSum + 1\n  let mass = new Float64Array(width)\n  mass[0] = 1\n  for (const rank of doubledRanks) {\n    const next = new Float64Array(width)\n    for (let sum = 0; sum < width; sum++) {\n      const probability = mass[sum]!\n      if (probability === 0) continue\n      next[sum]! += probability * 0.5\n      next[sum + rank]! += probability * 0.5\n    }\n    mass = next\n  }\n\n  const centre = maxSum / 2\n  let tail = 0\n  let extreme = 0\n  let maxDeviation = -1\n  for (let sum = 0; sum < width; sum++) {\n    const probability = mass[sum]!\n    if (probability === 0) continue\n    const deviation = Math.abs(sum - centre)\n    if (deviation >= doubledDeviation) tail += probability\n    if (deviation > maxDeviation) {\n      maxDeviation = deviation\n      extreme = probability\n    } else if (deviation === maxDeviation) {\n      extreme += probability\n    }\n  }\n  return { p: Math.min(1, tail), pFloor: Math.min(1, extreme) }\n}\n","/**\n * Matched-pair arm comparison — \"did the treatment arm beat the baseline arm\n * on the SAME work items?\"\n *\n * An arm A/B over run records is only trustworthy when it is PAIRED: the same\n * task/scenario/seed evaluated under both arms, compared item-by-item, so\n * inter-item difficulty variance cancels instead of masquerading as an arm\n * effect. This module owns the two error-prone steps every consumer otherwise\n * hand-rolls:\n *\n *   1. Pairing — matching rows across arms by `pairKey` (and by `repKey`\n *      within multi-rep items), with leftovers REPORTED rather than silently\n *      dropped (a silently unbalanced pairing biases every paired statistic\n *      downstream). Pairing never keys on outcome content: matching reps by\n *      their outcomes deflates discordant-pair counts and makes McNemar\n *      anti-conservative, so reps pair only by (`pairKey`, `repKey`) identity.\n *   2. Composition — feeding the matched pairs to the correct paired\n *      estimators that already live in `statistics`: `mcnemar` +\n *      `pairedRiskDifference` for pass/fail, `pairedBootstrap` +\n *      `wilcoxonSignedRank` for continuous metrics. No statistic is\n *      re-implemented here.\n *\n * The row shape is deliberately structural — callers project a `RunRecord`\n * (or any record) into `{ pairKey, arm, pass?, metrics? }`. Arm names are\n * caller-supplied parameters; the module ships no domain literal.\n */\n\nimport { ValidationError } from './errors'\nimport type { RunRecord } from './run-record'\nimport type { McNemarResult, PairedBootstrapOptions, PairedBootstrapResult } from './statistics'\nimport {\n  mcnemar,\n  pairedBootstrap,\n  pairedRiskDifference,\n  type RiskDifferenceResult,\n  wilcoxonSignedRank,\n} from './statistics'\n\n/** One arm observation of one work item. Structural on purpose: callers\n *  project their own record type (e.g. a `RunRecord`) into this shape. */\nexport interface PairedArmRow {\n  /** Matching key — rows sharing a `pairKey` across both arms form pairs\n   *  (typically the task/scenario/seed identity). */\n  pairKey: string\n  /** Rep identity within a `pairKey` (e.g. a seed or rep number). Required on\n   *  every row of a `pairKey` that has more than one rep in either arm; reps\n   *  then pair only on exact (`pairKey`, `repKey`) match, never on outcome\n   *  content. Optional when each arm has at most one rep of the item. */\n  repKey?: string\n  /** Arm label this row was produced under. */\n  arm: string\n  /** Binary outcome; omit when the comparison has no pass/fail notion. */\n  pass?: boolean\n  /** Named numeric measurements (score, cost, latency, …). */\n  metrics?: Record<string, number>\n}\n\nexport interface PairArmsOptions {\n  /** Arm treated as the control side of every pair. */\n  baselineArm: string\n  /** Arm treated as the treatment side of every pair. */\n  treatmentArm: string\n}\n\n/** One matched (baseline, treatment) observation of the same work item. */\nexport interface MatchedPair {\n  pairKey: string\n  /** 0-based position of this pair within its `pairKey`, ordered by sorted\n   *  `repKey` (always 0 for a single-rep item). The rep identity itself is on\n   *  the rows (`baseline.repKey` / `treatment.repKey`). */\n  repIndex: number\n  baseline: PairedArmRow\n  treatment: PairedArmRow\n}\n\nexport interface PairArmsResult {\n  /** Matched pairs, ordered by (`pairKey`, `repIndex`). */\n  pairs: MatchedPair[]\n  /** Baseline rows left without a treatment counterpart — reported, never\n   *  silently dropped. */\n  unpairedBaseline: PairedArmRow[]\n  /** Treatment rows left without a baseline counterpart. */\n  unpairedTreatment: PairedArmRow[]\n}\n\n/**\n * Match rows across two arms into (baseline, treatment) pairs by `pairKey`.\n *\n * A `pairKey` with at most one row per arm pairs directly, no `repKey`\n * needed. A `pairKey` with multiple reps in either arm requires `repKey` on\n * every one of its rows, and reps pair only on exact (`pairKey`, `repKey`)\n * match — pairing is keyed purely on row identity, never on outcome content\n * (outcome-keyed matching deflates discordant counts and biases McNemar), and\n * is therefore independent of input order. Reps whose `repKey` has no\n * counterpart in the other arm, and items present in only one arm, land in\n * the unpaired lists — reported, never truncated.\n *\n * Fail-loud: throws when either named arm has zero rows (an unknown arm\n * name would otherwise read as \"everything unpaired\"), when the two arm\n * names are equal, when a multi-rep `pairKey` has a row without `repKey`, or\n * when a (`pairKey`, arm) group repeats a `repKey` (the match would be\n * ambiguous).\n */\nexport function pairArms(rows: readonly PairedArmRow[], opts: PairArmsOptions): PairArmsResult {\n  const { baselineArm, treatmentArm } = opts\n  if (baselineArm === treatmentArm) {\n    throw new ValidationError(\n      `pairArms: baselineArm and treatmentArm are both '${baselineArm}' — an arm cannot be compared to itself`,\n    )\n  }\n\n  // arm → pairKey → rows\n  const byArm = new Map<string, Map<string, PairedArmRow[]>>()\n  const armsSeen = new Set<string>()\n  for (const row of rows) {\n    armsSeen.add(row.arm)\n    if (row.arm !== baselineArm && row.arm !== treatmentArm) continue\n    const byKey = byArm.get(row.arm) ?? new Map<string, PairedArmRow[]>()\n    const group = byKey.get(row.pairKey) ?? []\n    group.push(row)\n    byKey.set(row.pairKey, group)\n    byArm.set(row.arm, byKey)\n  }\n\n  for (const arm of [baselineArm, treatmentArm]) {\n    if (!byArm.has(arm)) {\n      const seen = [...armsSeen].sort().join(', ') || '<none>'\n      throw new ValidationError(`pairArms: no rows for arm '${arm}' (arms present: ${seen})`)\n    }\n  }\n\n  const baselineByKey = byArm.get(baselineArm)!\n  const treatmentByKey = byArm.get(treatmentArm)!\n\n  const allKeys = [...new Set([...baselineByKey.keys(), ...treatmentByKey.keys()])].sort()\n  const pairs: MatchedPair[] = []\n  const unpairedBaseline: PairedArmRow[] = []\n  const unpairedTreatment: PairedArmRow[] = []\n  for (const pairKey of allKeys) {\n    const b = baselineByKey.get(pairKey) ?? []\n    const t = treatmentByKey.get(pairKey) ?? []\n\n    if (b.length <= 1 && t.length <= 1) {\n      if (b.length === 1 && t.length === 1) {\n        const baseline = b[0]!\n        const treatment = t[0]!\n        if (baseline.repKey !== undefined || treatment.repKey !== undefined) {\n          if (\n            baseline.repKey === undefined ||\n            treatment.repKey === undefined ||\n            baseline.repKey !== treatment.repKey\n          ) {\n            unpairedBaseline.push(baseline)\n            unpairedTreatment.push(treatment)\n            continue\n          }\n        }\n        pairs.push({ pairKey, repIndex: 0, baseline, treatment })\n      } else {\n        unpairedBaseline.push(...b)\n        unpairedTreatment.push(...t)\n      }\n      continue\n    }\n\n    const bByRep = indexByRepKey(b, pairKey, baselineArm)\n    const tByRep = indexByRepKey(t, pairKey, treatmentArm)\n    const repKeys = [...new Set([...bByRep.keys(), ...tByRep.keys()])].sort()\n    let repIndex = 0\n    for (const repKey of repKeys) {\n      const baseline = bByRep.get(repKey)\n      const treatment = tByRep.get(repKey)\n      if (baseline !== undefined && treatment !== undefined) {\n        pairs.push({ pairKey, repIndex: repIndex++, baseline, treatment })\n      } else if (baseline !== undefined) {\n        unpairedBaseline.push(baseline)\n      } else if (treatment !== undefined) {\n        unpairedTreatment.push(treatment)\n      }\n    }\n  }\n\n  return { pairs, unpairedBaseline, unpairedTreatment }\n}\n\n/** Index a multi-rep (pairKey, arm) group by `repKey`, enforcing that every\n *  row carries one and that no repKey repeats within the group. */\nfunction indexByRepKey(\n  group: readonly PairedArmRow[],\n  pairKey: string,\n  arm: string,\n): Map<string, PairedArmRow> {\n  const byRep = new Map<string, PairedArmRow>()\n  for (const row of group) {\n    if (row.repKey === undefined) {\n      throw new ValidationError(\n        `pairArms: pairKey '${pairKey}' has multiple reps in an arm, but a row in arm '${arm}' ` +\n          `is missing repKey — multi-rep items require an explicit repKey on every row so reps ` +\n          `pair by identity (pairing reps by outcome or by index would bias the paired statistics)`,\n      )\n    }\n    if (byRep.has(row.repKey)) {\n      throw new ValidationError(\n        `pairArms: duplicate repKey '${row.repKey}' for pairKey '${pairKey}' in arm '${arm}' — ` +\n          `(pairKey, repKey) must uniquely identify a rep within an arm`,\n      )\n    }\n    byRep.set(row.repKey, row)\n  }\n  return byRep\n}\n\n/** Paired pass/fail comparison over the pairs where BOTH sides carry `pass`. */\nexport interface PairedCorrectness {\n  /** Discordant pairs where the treatment passed and the baseline failed. */\n  b10: number\n  /** Discordant pairs where the baseline passed and the treatment failed. */\n  b01: number\n  /** Exact McNemar significance over the paired outcomes (`b === b10`, `c === b01`). */\n  mcnemar: McNemarResult\n  /** Paired effect size: p(treatment) − p(baseline) with a paired-variance CI. */\n  riskDifference: RiskDifferenceResult\n}\n\n/** Paired delta summary for one named metric (delta = treatment − baseline). */\nexport interface PairedMetricDelta {\n  name: string\n  /** Pairs where BOTH sides carry a finite value for this metric. */\n  n: number\n  /** Pairs where at least one side does not carry the metric. */\n  nMissing: number\n  /** Median paired delta, or null when `n === 0`. */\n  medianDelta: number | null\n  /** Mean paired delta, or null when `n === 0`. */\n  meanDelta: number | null\n  /** Bootstrap CI on the paired delta (`pairedBootstrap`); null when\n   *  `n === 0` — a zero-width [0, 0] interval on no data would read as a\n   *  measured tight null. */\n  bootstrapCi: PairedBootstrapResult | null\n  /** Wilcoxon signed-rank test on the paired deltas; null when `n === 0`. */\n  wilcoxon: { w: number; p: number } | null\n}\n\nexport interface ComparePairedArmsOptions extends PairArmsOptions {\n  /** Metrics to compare. Default: every metric name observed on any matched\n   *  pair, sorted. A name that appears on no pair is still reported (with\n   *  `n = 0`) so a misspelled metric is visible instead of vanishing. */\n  metricNames?: string[]\n  /** Passed through to `pairedBootstrap` — set `seed` for reproducible CIs. */\n  bootstrap?: PairedBootstrapOptions\n}\n\nexport interface PairedArmsComparison {\n  nPairs: number\n  nUnpairedBaseline: number\n  nUnpairedTreatment: number\n  /** null when no matched pair carries `pass` on both sides — a pass/fail\n   *  verdict over rows that never measured pass/fail would be fabricated. */\n  correctness: PairedCorrectness | null\n  metricDeltas: PairedMetricDelta[]\n}\n\n/**\n * Full matched-pair arm comparison: pair via {@link pairArms}, then compose\n * the paired estimators from `statistics` over the matched pairs.\n *\n * Correctness uses only the pairs where both sides carry `pass` (`mcnemar.n`\n * is that subset's size); each metric uses only the pairs where both sides\n * carry a finite value for it, with the remainder counted in `nMissing`.\n * Deltas are treatment − baseline throughout.\n *\n * Fail-loud: inherits {@link pairArms}'s unknown-arm throw, and throws on a\n * non-finite metric value — silently treating corrupt telemetry as \"metric\n * absent\" would misreport it as missing coverage.\n */\nexport function comparePairedArms(\n  rows: readonly PairedArmRow[],\n  opts: ComparePairedArmsOptions,\n): PairedArmsComparison {\n  const { pairs, unpairedBaseline, unpairedTreatment } = pairArms(rows, opts)\n\n  let correctness: PairedCorrectness | null = null\n  const baselinePass: number[] = []\n  const treatmentPass: number[] = []\n  for (const pair of pairs) {\n    if (pair.baseline.pass === undefined || pair.treatment.pass === undefined) continue\n    baselinePass.push(pair.baseline.pass ? 1 : 0)\n    treatmentPass.push(pair.treatment.pass ? 1 : 0)\n  }\n  if (baselinePass.length > 0) {\n    const mc = mcnemar(baselinePass, treatmentPass)\n    correctness = {\n      b10: mc.b,\n      b01: mc.c,\n      mcnemar: mc,\n      riskDifference: pairedRiskDifference(baselinePass, treatmentPass),\n    }\n  }\n\n  const metricNames =\n    opts.metricNames ??\n    [\n      ...new Set(\n        pairs.flatMap((p) => [\n          ...Object.keys(p.baseline.metrics ?? {}),\n          ...Object.keys(p.treatment.metrics ?? {}),\n        ]),\n      ),\n    ].sort()\n\n  const metricDeltas: PairedMetricDelta[] = metricNames.map((name) => {\n    const before: number[] = []\n    const after: number[] = []\n    let nMissing = 0\n    for (const pair of pairs) {\n      const b = metricValue(pair.baseline, name)\n      const t = metricValue(pair.treatment, name)\n      if (b === undefined || t === undefined) {\n        nMissing++\n        continue\n      }\n      before.push(b)\n      after.push(t)\n    }\n    const bootstrapCi = before.length === 0 ? null : pairedBootstrap(before, after, opts.bootstrap)\n    return {\n      name,\n      n: before.length,\n      nMissing,\n      medianDelta: bootstrapCi?.median ?? null,\n      meanDelta: bootstrapCi?.mean ?? null,\n      bootstrapCi,\n      wilcoxon: before.length === 0 ? null : wilcoxonSignedRank(before, after),\n    }\n  })\n\n  return {\n    nPairs: pairs.length,\n    nUnpairedBaseline: unpairedBaseline.length,\n    nUnpairedTreatment: unpairedTreatment.length,\n    correctness,\n    metricDeltas,\n  }\n}\n\nexport interface MatchedRunRecordPair {\n  pairKey: string\n  repKey: string\n  baseline: RunRecord\n  treatment: RunRecord\n}\n\nexport interface PairRunRecordsResult {\n  pairs: MatchedRunRecordPair[]\n  unpairedBaseline: RunRecord[]\n  unpairedTreatment: RunRecord[]\n}\n\ninterface RunRecordArmRow extends PairedArmRow {\n  run: RunRecord\n  repKey: string\n}\n\n/**\n * Pair two RunRecord arms by the identity of the evaluated work:\n * `(experimentId, scenarioId, seed)`.\n *\n * Falling back to array order, candidate id, or experiment id can compare\n * different tasks and fabricate lift. Duplicate identities throw.\n */\nexport function pairRunRecords(\n  baselineRuns: readonly RunRecord[],\n  treatmentRuns: readonly RunRecord[],\n): PairRunRecordsResult {\n  const baselineRows = runRecordArmRows(baselineRuns, 'baseline')\n  const treatmentRows = runRecordArmRows(treatmentRuns, 'treatment')\n  validateRunRecordArmRows(baselineRows, 'baseline')\n  validateRunRecordArmRows(treatmentRows, 'treatment')\n  if (baselineRows.length === 0 || treatmentRows.length === 0) {\n    return {\n      pairs: [],\n      unpairedBaseline: baselineRows.map((row) => row.run),\n      unpairedTreatment: treatmentRows.map((row) => row.run),\n    }\n  }\n\n  const result = pairArms([...baselineRows, ...treatmentRows], {\n    baselineArm: 'baseline',\n    treatmentArm: 'treatment',\n  })\n  return {\n    pairs: result.pairs.map((pair) => {\n      const baseline = pair.baseline as RunRecordArmRow\n      const treatment = pair.treatment as RunRecordArmRow\n      return {\n        pairKey: pair.pairKey,\n        repKey: baseline.repKey,\n        baseline: baseline.run,\n        treatment: treatment.run,\n      }\n    }),\n    unpairedBaseline: result.unpairedBaseline.map((row) => (row as RunRecordArmRow).run),\n    unpairedTreatment: result.unpairedTreatment.map((row) => (row as RunRecordArmRow).run),\n  }\n}\n\nfunction runRecordArmRows(runs: readonly RunRecord[], arm: string): RunRecordArmRow[] {\n  return runs.map((run) => {\n    const scenarioId = run.scenarioId.trim()\n    if (!scenarioId) {\n      throw new ValidationError(\n        `pairRunRecords: run '${run.runId}' is missing scenarioId; paired comparisons require explicit scenario identity`,\n      )\n    }\n    return {\n      pairKey: JSON.stringify([run.experimentId, scenarioId]),\n      repKey: String(run.seed),\n      arm,\n      run,\n    }\n  })\n}\n\nfunction validateRunRecordArmRows(rows: readonly RunRecordArmRow[], arm: string): void {\n  const byPairKey = new Map<string, RunRecordArmRow[]>()\n  for (const row of rows) {\n    const group = byPairKey.get(row.pairKey) ?? []\n    group.push(row)\n    byPairKey.set(row.pairKey, group)\n  }\n  for (const [pairKey, group] of byPairKey) {\n    if (group.length > 1) indexByRepKey(group, pairKey, arm)\n  }\n}\n\nfunction metricValue(row: PairedArmRow, name: string): number | undefined {\n  const v = row.metrics?.[name]\n  if (v === undefined) return undefined\n  if (!Number.isFinite(v)) {\n    throw new ValidationError(\n      `comparePairedArms: non-finite value for metric '${name}' on pairKey '${row.pairKey}' (arm '${row.arm}'): ${v}`,\n    )\n  }\n  return v\n}\n"],"mappings":";;;;;;;;;;;AAgCA,SAAgB,OAAO,WAAmB,GAAW,aAAa,KAA0B;CAC1F,IAAI,KAAK,GAAG,OAAO;EAAE,UAAU;EAAG,OAAO;EAAG,OAAO;CAAE;CACrD,IAAI,YAAY,KAAK,YAAY,GAC/B,MAAM,IAAI,MAAM,sBAAsB,UAAU,mBAAmB,EAAE,EAAE;CAEzE,MAAM,IAAI,UAAU,KAAK,IAAI,cAAc,CAAC;CAC5C,MAAM,IAAI,YAAY;CACtB,MAAM,KAAK,IAAI;CACf,MAAM,QAAQ,IAAI,KAAK;CACvB,MAAM,UAAU,IAAI,MAAM,IAAI,MAAM;CACpC,MAAM,OAAQ,IAAI,KAAK,MAAM,KAAK,IAAI,KAAK,MAAM,IAAI,MAAM,CAAC,IAAK;CACjE,OAAO;EACL,UAAU;EACV,OAAO,KAAK,IAAI,GAAG,SAAS,IAAI;EAChC,OAAO,KAAK,IAAI,GAAG,SAAS,IAAI;CAClC;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;AA2BA,SAAgB,sBAAsB,QAAoC;CACxE,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK;EACtC,MAAM,IAAI,OAAO;EACjB,IAAI,MAAM,KAAK,MAAM,GAAG,OAAO;CACjC;CACA,OAAO;AACT;;;;;;;;;;;;;AA8BA,SAAgB,QACd,SACA,WACe;CACf,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MAAM,kCAAkC,QAAQ,OAAO,MAAM,UAAU,OAAO,EAAE;CAE5F,MAAM,IAAI,QAAQ;CAClB,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,cAAc,IAAI;CACxB,MAAM,YAAY,gBAAgB,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,IAAI,MAAM,IAAI;CACvE,OAAO;EAAE;EAAG;EAAa;EAAG;EAAG;EAAW,QAAQ,qBAAqB,GAAG,CAAC;CAAE;AAC/E;;;;;;;;;;;;;;;;AAmCA,SAAgB,qBACd,SACA,WACA,aAAa,KACS;CACtB,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MACR,+CAA+C,QAAQ,OAAO,MAAM,UAAU,OAAO,EACvF;CAEF,MAAM,IAAI,QAAQ;CAClB,IAAI,MAAM,GAAG,OAAO;EAAE,GAAG;EAAG,GAAG;EAAG,GAAG;EAAG,gBAAgB;EAAG,OAAO;EAAG,OAAO;EAAG;CAAW;CAC1F,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,MAAM,IAAI,KAAK;CACrB,MAAM,YAAY,IAAI,KAAK,IAAI,MAAM,IAAI,MAAM,IAAI;CAEnD,MAAM,OADI,UAAU,KAAK,IAAI,cAAc,CAC9B,IAAI,KAAK,KAAK,KAAK,IAAI,GAAG,QAAQ,CAAC;CAChD,OAAO;EACL;EACA;EACA;EACA,gBAAgB;EAChB,OAAO,KAAK,IAAI,IAAI,KAAK,IAAI;EAC7B,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI;EAC5B;CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsDA,SAAgB,0BACd,SACA,WACA,aAAa,KACc;CAC3B,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MACR,oDAAoD,QAAQ,OAAO,MAAM,UAAU,OAAO,EAC5F;CAEF,IAAI,cAAc,KAAK,cAAc,GACnC,MAAM,IAAI,MAAM,+DAA+D,YAAY;CAE7F,MAAM,IAAI,QAAQ;CAClB,IAAI,MAAM,GACR,OAAO;EACL,GAAG;EACH,GAAG;EACH,GAAG;EACH,aAAa;EACb,gBAAgB;EAChB,OAAO;EACP,OAAO;EACP;EACA,QAAQ;CACV;CAEF,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,IAAI,IAAI;CACd,MAAM,kBAAkB,IAAI,KAAK;CACjC,MAAM,SAAS,qBAAqB,GAAG,CAAC;CACxC,IAAI,MAAM,GACR,OAAO;EAAE;EAAG;EAAG;EAAG,aAAa;EAAG,gBAAgB;EAAG,OAAO;EAAG,OAAO;EAAG;EAAY;CAAO;CAE9F,MAAM,QAAQ,IAAI;CAClB,MAAM,QAAQ,MAAM,IAAI,IAAI,aAAa,QAAQ,GAAG,GAAG,IAAI,IAAI,CAAC;CAChE,MAAM,SAAS,MAAM,IAAI,IAAI,aAAa,IAAI,QAAQ,GAAG,IAAI,GAAG,IAAI,CAAC;CACrE,MAAM,QAAQ,IAAI;CAClB,OAAO;EACL;EACA;EACA;EACA,aAAa;EACb;EACA,OAAO,KAAK,IAAI,KAAK,IAAI,QAAQ,KAAK,KAAK;EAC3C,OAAO,KAAK,IAAI,IAAI,IAAI,SAAS,KAAK,KAAK;EAC3C;EACA;CACF;AACF;;;;;AAMA,SAAS,aAAa,GAAW,GAAW,GAAmB;CAC7D,IAAI,KAAK,GAAG,OAAO;CACnB,IAAI,KAAK,GAAG,OAAO;CACnB,IAAI,KAAK;CACT,IAAI,KAAK;CACT,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,KAAK;EAC3B,MAAM,OAAO,KAAK,MAAM;EACxB,IAAI,0BAA0B,KAAK,GAAG,CAAC,IAAI,GAAG,KAAK;OAC9C,KAAK;CACZ;CACA,QAAQ,KAAK,MAAM;AACrB;;;;;;;;;;AAgCA,SAAS,oBAAoB,GAAW,GAAW,GAAW,OAAuB;CACnF,MAAM,IAAI,IAAI,IAAI;CAClB,MAAM,YAAY,IAAI;CACtB,MAAM,SAAS,EAAE,IAAI,IAAI,SAAS,IAAI,IAAI,IAAI,IAAI;CAClD,MAAM,WAAW,CAAC,IAAI,SAAS,IAAI;CACnC,MAAM,eAAe,SAAS,SAAS,IAAI,YAAY;CACvD,MAAM,OAAO,eAAe,IAAI,KAAK,KAAK,YAAY,IAAI;CAC1D,MAAM,KAAK,CAAC,SAAS,SAAS,IAAI;CAElC,OAAO,KAAK,IAAI,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,KAAK,CAAC,GAAG,KAAK,IAAI,IAAI,IAAI,SAAS,CAAC,CAAC;AAChF;;;AAIA,SAAS,WAAW,GAAW,GAAW,GAAW,OAAuB;CAC1E,MAAM,YAAY,IAAI,IAAI,IAAI;CAE9B,MAAM,WAAW,KAAK,IADZ,oBAAoB,GAAG,GAAG,GAAG,KACb,IAAI,SAAS,IAAI;CAC3C,IAAI,EAAE,WAAW,IAAI;EACnB,IAAI,cAAc,GAAG,OAAO;EAC5B,OAAO,YAAY,IAAI,OAAO,oBAAoB,OAAO;CAC3D;CACA,OAAO,YAAY,KAAK,KAAK,QAAQ;AACvC;;;;;;;;;;;;;;;;;;;;;;;;;;AA2BA,SAAgB,0BACd,SACA,WACA,aAAa,KACc;CAC3B,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MACR,oDAAoD,QAAQ,OAAO,MAAM,UAAU,OAAO,EAC5F;CAEF,IAAI,cAAc,KAAK,cAAc,GACnC,MAAM,IAAI,MAAM,+DAA+D,YAAY;CAE7F,MAAM,IAAI,QAAQ;CAClB,IAAI,MAAM,GACR,OAAO;EAAE,GAAG;EAAG,GAAG;EAAG,GAAG;EAAG,aAAa;EAAG,gBAAgB;EAAG,OAAO;EAAI,OAAO;EAAG;CAAW;CAEhG,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,kBAAkB,IAAI,KAAK;CACjC,MAAM,IAAI,UAAU,KAAK,IAAI,cAAc,CAAC;CAK5C,IAAI,KAAK;CACT,IAAI,KAAK;CACT,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,KAAK;EAC5B,MAAM,OAAO,KAAK,MAAM;EACxB,IAAI,WAAW,GAAG,GAAG,GAAG,GAAG,IAAI,GAAG,KAAK;OAClC,KAAK;CACZ;CACA,MAAM,SAAS,KAAK,MAAM;CAG1B,IAAI,MAAM;CACV,IAAI,MAAM;CACV,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,KAAK;EAC5B,MAAM,OAAO,MAAM,OAAO;EAC1B,IAAI,WAAW,GAAG,GAAG,GAAG,GAAG,IAAI,CAAC,GAAG,MAAM;OACpC,MAAM;CACb;CACA,MAAM,SAAS,MAAM,OAAO;CAE5B,OAAO;EACL;EACA;EACA;EACA,aAAa,IAAI;EACjB;EACA,OAAO,KAAK,IAAI,IAAI,KAAK;EACzB,OAAO,KAAK,IAAI,GAAG,KAAK;EACxB;CACF;AACF;;;;;;;;;;;;;;;;;;;;AAqBA,SAAgB,kBACd,QACA,OACe;CACf,IAAI,QAAuB;CAC3B,KAAK,MAAM,OAAO,CAAC,QAAQ,KAAK,GAC9B,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;EACnC,MAAM,IAAI,IAAI;EACd,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;EAChC,IAAI,MAAM,GAAG;EACb,IAAI,IAAI,GAAG,OAAO;EAClB,IAAI,UAAU,MAAM,QAAQ;OACvB,IAAI,MAAM,OAAO,OAAO;CAC/B;CAEF,OAAO;AACT;;;;;;;;;;;AAYA,SAAgB,QAAQ,GAAW,GAAW,GAAmB;CAC/D,IAAI,CAAC,OAAO,UAAU,CAAC,KAAK,CAAC,OAAO,UAAU,CAAC,KAAK,CAAC,OAAO,UAAU,CAAC,GACrE,MAAM,IAAI,MAAM,4CAA4C,EAAE,MAAM,EAAE,MAAM,EAAE,EAAE;CAElF,IAAI,IAAI,KAAK,IAAI,KAAK,IAAI,KAAK,IAAI,GACjC,MAAM,IAAI,MAAM,mDAAmD,EAAE,MAAM,EAAE,MAAM,EAAE,EAAE;CAEzF,IAAI,IAAI,IAAI,GAAG,OAAO;CACtB,IAAI,OAAO;CACX,KAAK,IAAI,IAAI,IAAI,IAAI,GAAG,KAAK,GAAG,KAAK,QAAQ,IAAI,IAAI;CACrD,OAAO,IAAI;AACb;;;;;;;;;;ACzgBA,SAAgB,UAAU,GAAmB;CAC3C,IAAI,MAAM,GAAG,OAAO;CAEpB,MAAM,KAAK;CACX,MAAM,KAAK;CACX,MAAM,KAAK;CACX,MAAM,KAAK;CACX,MAAM,KAAK;CACX,MAAM,IAAI;CAEV,MAAM,SAAS,KAAK,IAAI,CAAC,IAAI,KAAK;CAClC,MAAM,IAAI,KAAK,IAAI,IAAI;CACvB,MAAM,iBAAiB,KAAK,IAAI,MAAM,IAAI,MAAM,IAAI,MAAM,IAAI,MAAM,IAAI,KAAK,IAAI,CAAC,SAAS,MAAM;CAEjG,OAAO,IAAI,IAAI,aAAa,IAAI,IAAI,aAAa;AACnD;;;;ACyBA,MAAa,gCAAgC;;AAE7C,MAAa,8BAA8B;;AAE3C,MAAa,uBAAuB;;AAEpC,MAAa,uBAAuB;;;;;;;;;;;;;AA2BpC,SAAgB,aACd,GACA,GACA,OAAwB,CAAC,GACN;CACnB,mBAAmB,gBAAgB,KAAK,CAAC;CACzC,mBAAmB,gBAAgB,KAAK,CAAC;CAEzC,MAAM,KAAK,EAAE;CACb,MAAM,KAAK,EAAE;CACb,IAAI,OAAO,KAAK,OAAO,GAAG,OAAO;EAAE,GAAG;EAAG,IAAI;EAAG,GAAG;EAAG,QAAQ;EAAS,QAAQ;CAAE;CAEjF,MAAM,QAAQ,KAAK;CACnB,MAAM,WAAW,CACf,GAAG,EAAE,KAAK,OAAO;EAAE;EAAG,OAAO;CAAK,EAAE,GACpC,GAAG,EAAE,KAAK,OAAO;EAAE;EAAG,OAAO;CAAM,EAAE,CACvC,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,IAAI,EAAE,CAAC;CAE1B,MAAM,EAAE,UAAU,YAAY,oBAAoB,SAAS,KAAK,UAAU,MAAM,CAAC,CAAC;CAClF,IAAI,WAAW;CACf,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KACzB,IAAI,SAAS,EAAE,CAAE,OAAO,YAAY,SAAS;CAG/C,MAAM,KAAK,WAAY,MAAM,KAAK,KAAM;CACxC,MAAM,IAAI,KAAK,IAAI,IAAI,KAAK,KAAK,EAAE;CAGnC,MAAM,UAAU,SAAS,KAAK,SAAS,KAAK,MAAM,OAAO,CAAC,CAAC;CAC3D,MAAM,mBAAmB,KAAK,IAAI,IAAI,KAAK,KAAK,EAAE;CAClD,MAAM,YAAY,KAAK,IAAI,IAAI,EAAE;CACjC,MAAM,SAAS,QAAQ;CACvB,MAAM,YAAY,mBAAmB,SAAS,SAAS;CAEvD,MAAM,cAAc,oBAAoB,SAAS,SAAS;CAC1D,MAAM,SAAS,qBACb,gBACA,KAAK,UAAU,QACf,MAAM,GAAG,OAAO,MAChB,UAAU,UAAA,QACR,UAAU,QAAA,MACZ,aACA,GAAG,8BAA8B,eAAe,OAAO,EAAE,cACpD,4BAA4B,eAAe,OAAO,EAAE,aAC3D;CAEA,IAAI,WAAW,SAAS;EACtB,MAAM,EAAE,GAAG,WAAW,gBAAgB,SAAS,WAAW,QAAQ,gBAAgB;EAClF,OAAO;GAAE;GAAG;GAAI;GAAG;GAAQ;EAAO;CACpC;CAEA,IAAI,WAAW,cACb,OAAO;EACL;EACA;EACA,GAAG,oBAAoB,mBAAmB,GAAG,eAAe,IAAI,IAAI,OAAO,OAAO,CAAC;EACnF;EACA,QAAQ;CACV;CAGF,MAAM,eAAe,oBAAoB,gBAAgB,KAAK,YAAY;CAC1E,MAAM,MAAM,KAAK,SAAS,KAAA,IAAY,QAAQ,uBAAuB,GAAG,CAAC,CAAC,IAAI,QAAQ,KAAK,IAAI;CAC/F,IAAI,mBAAmB;CACvB,MAAM,OAAO,CAAC,GAAG,OAAO;CACxB,KAAK,IAAI,YAAY,GAAG,YAAY,cAAc,aAAa;EAC7D,IAAI,iBAAiB;EACrB,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;GAClC,MAAM,OAAO,IAAI,KAAK,MAAM,IAAI,KAAK,QAAQ,EAAE;GAC/C,MAAM,UAAU,KAAK;GACrB,KAAK,QAAQ,KAAK;GAClB,KAAK,KAAK;GACV,kBAAkB;EACpB;EACA,IACE,KAAK,IAAI,iBAAiB,aAAa,YAAY,KAAK,YAAY,MAAM,KAC1E,kBAEA;CAEJ;CACA,MAAM,SAAS,KAAK,IAAI,KAAK,eAAe,IAAI,WAAW;CAC3D,OAAO;EACL;EACA;EACA,GAAG,KAAK,KAAK,IAAI,qBAAqB,eAAe,IAAI,MAAM;EAC/D;EACA;CACF;AACF;;;;;;;;;;;;;;AA6BA,SAAgB,mBACd,QACA,OACA,OAAwB,CAAC,GACC;CAC1B,IAAI,OAAO,WAAW,MAAM,QAC1B,MAAM,IAAI,gBACR,6CAA6C,OAAO,OAAO,MAAM,MAAM,OAAO,EAChF;CAEF,mBAAmB,sBAAsB,UAAU,MAAM;CACzD,mBAAmB,sBAAsB,SAAS,KAAK;CAEvD,MAAM,QAAQ,OAAO,KAAK,GAAG,MAAM,MAAM,KAAM,CAAC,CAAC,CAAC,QAAQ,MAAM,MAAM,CAAC;CACvE,MAAM,IAAI,MAAM;CAChB,IAAI,MAAM,GAAG,OAAO;EAAE,GAAG;EAAG,GAAG;EAAG,QAAQ;EAAS,QAAQ;EAAG,UAAU;CAAE;CAE1E,MAAM,QAAQ,MAAM,KAAK,GAAG,OAAO;EAAE,KAAK,KAAK,IAAI,CAAC;EAAG;CAAE,EAAE,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,MAAM,EAAE,GAAG;CACzF,MAAM,EAAE,UAAU,YAAY,oBAAoB,MAAM,KAAK,UAAU,MAAM,GAAG,CAAC;CACjF,MAAM,QAAkB,IAAI,MAAM,CAAC;CACnC,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,MAAM,MAAM,EAAE,CAAE,KAAK,SAAS;CAE1D,IAAI,QAAQ;CACZ,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,IAAI,MAAM,KAAM,GAAG,SAAS,MAAM;CAK9D,MAAM,UAAU,SAAS,KAAK,SAAS,KAAK,MAAM,OAAO,CAAC,CAAC;CAC3D,MAAM,mBAAmB,KAAK,IAAI,IAAI,QAAS,KAAK,IAAI,KAAM,CAAC;CAE/D,MAAM,cAAc,KAAK,IAAI,GAAG,MAAM,IAAI,EAAE;CAC5C,MAAM,SAAS,qBACb,sBACA,KAAK,UAAU,QACf,KAAK,EAAE,wBACP,KAAA,IACA,aACA,yBACF;CAEA,IAAI,WAAW,SAAS;EACtB,MAAM,EAAE,GAAG,WAAW,iBAAiB,SAAS,gBAAgB;EAChE,OAAO;GAAE,GAAG;GAAO;GAAG;GAAQ;GAAQ,UAAU;EAAE;CACpD;CAEA,IAAI,WAAW,cAAc;EAC3B,MAAM,WAAY,KAAK,IAAI,MAAM,IAAI,IAAI,KAAM,KAAK,UAAU;EAC9D,OAAO;GACL,GAAG;GACH,GAAG,oBAAoB,mBAAmB,GAAG,KAAK,KAAK,QAAQ,CAAC;GAChE;GACA,QAAQ;GACR,UAAU;EACZ;CACF;CAEA,MAAM,eAAe,oBAAoB,sBAAsB,KAAK,YAAY;CAChF,MAAM,MAAM,QAAQ,KAAK,MAAM,QAAQ,KAAK;CAC5C,MAAM,gBAAiB,KAAK,IAAI,KAAM;CACtC,IAAI,mBAAmB;CACvB,KAAK,IAAI,YAAY,GAAG,YAAY,cAAc,aAAa;EAC7D,IAAI,eAAe;EACnB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,IAAI,IAAI,IAAI,IAAK,gBAAgB,QAAQ;EACrE,IAAI,KAAK,IAAI,eAAe,aAAa,KAAK,kBAAkB;CAClE;CACA,OAAO;EACL,GAAG;EACH,IAAI,IAAI,qBAAqB,eAAe;EAC5C;EACA,QAAQ,KAAK,IAAI,KAAK,eAAe,IAAI,WAAW;EACpD,UAAU;CACZ;AACF;;;;;;AAOA,SAAS,oBAAoB,QAAoE;CAC/F,MAAM,WAAW,IAAI,MAAc,OAAO,MAAM;CAChD,IAAI,UAAU;CACd,IAAI,IAAI;CACR,OAAO,IAAI,OAAO,QAAQ;EACxB,IAAI,IAAI;EACR,OAAO,IAAI,OAAO,UAAU,OAAO,OAAO,OAAO,IAAI;EACrD,MAAM,WAAW,IAAI,IAAI,KAAK;EAC9B,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,SAAS,KAAK;EAC1C,MAAM,YAAY,IAAI;EACtB,IAAI,YAAY,GAAG,WAAW,aAAa,IAAI;EAC/C,IAAI;CACN;CACA,OAAO;EAAE;EAAU;CAAQ;AAC7B;AAEA,SAAS,qBACP,IACA,SACA,QACA,eACA,aACA,WACgB;CAChB,IAAI,YAAY,QAAQ,OAAO,gBAAgB,UAAU;CACzD,IAAI,YAAY,SAAS;EACvB,IAAI,eAAe,OAAO;EAC1B,MAAM,IAAI,gBACR,GAAG,GAAG,sCAAsC,OAAO,+BAC9C,UAAU,yFAEjB;CACF;CACA,IAAI,eACF,MAAM,IAAI,gBACR,GAAG,GAAG,sCAAsC,OAAO,+CACpC,kBAAkB,WAAW,EAAE,0HAEzC,UAAU,EACjB;CAEF,OAAO;AACT;AAEA,SAAS,oBAAoB,IAAY,cAA0C;CACjF,IAAI,iBAAiB,KAAA,GAAW,OAAO;CACvC,IAAI,CAAC,OAAO,UAAU,YAAY,KAAK,eAAe,GACpD,MAAM,IAAI,gBAAgB,GAAG,GAAG,iDAAiD,cAAc;CAEjG,OAAO;AACT;;AAGA,SAAS,oBAAoB,WAAmB,OAAuB;CACrE,IAAI,EAAE,QAAQ,IAAI,OAAO;CACzB,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,UAAU,KAAK,IAAI,GAAG,YAAY,EAAG,IAAI,KAAK,EAAE;AAC9E;;;;AAKA,SAAS,eAAe,IAAY,IAAY,OAAe,SAAyB;CACtF,IAAI,QAAQ,GAAG,OAAO;CACtB,MAAM,WAAa,KAAK,KAAM,MAAO,QAAQ,IAAI,WAAW,SAAS,QAAQ;CAC7E,OAAO,KAAK,KAAK,KAAK,IAAI,GAAG,QAAQ,CAAC;AACxC;AAEA,SAAS,UAAU,GAAW,GAAmB;CAC/C,OAAO,QAAQ,IAAI,CAAC,IAAI,QAAQ,IAAI,CAAC,IAAI,QAAQ,IAAI,IAAI,CAAC;AAC5D;AAEA,SAAS,kBAAkB,OAAuB;CAChD,OAAO,SAAS,QAAQ,UAAU,IAAI,MAAM,QAAQ,CAAC,IAAI,MAAM,cAAc,CAAC;AAChF;;;;;;;AAQA,SAAS,mBACP,cACA,WACkC;CAClC,MAAM,SAAS,aAAa,QAAQ,KAAK,SAAS,MAAM,MAAM,CAAC;CAC/D,IAAI,OAAO;CACX,KAAK,IAAI,SAAS,GAAG,SAAS,aAAa,QAAQ,UACjD,QAAQ,KAAK,IAAI,WAAW,SAAS,CAAC,KAAK,SAAS,aAAa,UAAW;CAE9E,OAAO;EACL,SAAS,YAAY,MAAM,SAAS;EACpC;CACF;AACF;;;;;;;;;AAUA,SAAS,oBAAoB,cAAiC,WAA2B;CACvF,MAAM,QAAQ,aAAa;CAC3B,MAAM,SAAS,QAAQ;CACvB,MAAM,aAAa,aAAa,MAAM,GAAG,SAAS,CAAC,CAAC,QAAQ,KAAK,SAAS,MAAM,MAAM,CAAC;CACvF,MAAM,aAAa,aAAa,MAAM,QAAQ,SAAS,CAAC,CAAC,QAAQ,KAAK,SAAS,MAAM,MAAM,CAAC;CAC5F,IAAI,eAAe,YAAY,OAAO;CAEtC,MAAM,SAAS,aAAa,YAAY,KAAK,YAAY;CACzD,MAAM,mBAAmB,KAAK,IAAI,aAAa,MAAM;CACrD,MAAM,mBAAmB,KAAK,IAAI,aAAa,MAAM;CACrD,MAAM,eAAe,UAAU,OAAO,SAAS;CAC/C,MAAM,cAAc,KAAK,IACvB,qBAAqB,cAAc,WAAW,SAAS,IAAI,YAC7D;CACA,MAAM,cAAc,KAAK,IACvB,qBAAqB,cAAc,WAAW,SAAS,IAAI,YAC7D;CAEA,IAAI,mBAAmB,kBAAkB,OAAO;CAChD,IAAI,mBAAmB,kBAAkB,OAAO;CAChD,OAAO,KAAK,IAAI,GAAG,cAAc,WAAW;AAC9C;AAEA,SAAS,qBACP,aACA,WACA,MACQ;CACR,MAAM,gBAAgB,SAAS,YAAY,YAAY,IAAI,YAAY,SAAS;CAChF,MAAM,WAAW,YAAY;CAC7B,IAAI,QAAQ;CACZ,IAAI,YAAY,gBAAgB;CAChC,OAAO,QAAQ,KAAK,YAAY,QAAQ,OAAO,UAAU;CACzD,OAAO,YAAY,YAAY,UAAU,YAAY,eAAe,UAAU;CAI9E,OAAO,UAFS,YAAY,OAEF,aADZ,SAAS,YAAY,QAAQ,YAAY,SAAS,UACrB;AAC7C;;;;;;;;;;;;;AAcA,SAAS,gBACP,cACA,IACA,IACA,kBAC+B;CAC/B,MAAM,SAAS,aAAa,QAAQ,KAAK,SAAS,MAAM,MAAM,CAAC;CAC/D,MAAM,QAAQ,SAAS;CAEvB,MAAM,OAAuB,MAAM,KAAK,EAAE,QAAQ,KAAK,EAAE,SAAS,IAAI,aAAa,KAAK,CAAC;CACzF,KAAK,EAAE,CAAE,KAAK;CACd,IAAI,SAAS;CACb,KAAK,MAAM,QAAQ,cAAc;EAC/B,KAAK,IAAI,IAAI,KAAK,IAAI,IAAI,SAAS,CAAC,GAAG,KAAK,GAAG,KAAK;GAClD,MAAM,OAAO,KAAK,IAAI;GACtB,MAAM,OAAO,KAAK;GAClB,KAAK,IAAI,MAAM,SAAS,MAAM,OAAO,GAAG,OAAO;IAC7C,MAAM,QAAQ,KAAK;IACnB,IAAI,UAAU,GAAG,KAAK,MAAM,SAAU;GACxC;EACF;EACA;CACF;CAIA,MAAM,QAAQ,MAAM,KAAK,KAAK,KAAK;CACnC,MAAM,SAAS,KAAK;CACpB,IAAI,YAAY;CAChB,IAAI,cAAc;CAClB,IAAI,WAAW;CACf,IAAI,eAAe;CACnB,KAAK,IAAI,MAAM,GAAG,MAAM,OAAO,OAAO;EACpC,MAAM,QAAQ,OAAO;EACrB,IAAI,UAAU,GAAG;EACjB,aAAa;EACb,MAAM,YAAY,KAAK,IAAI,MAAM,KAAK;EACtC,IAAI,aAAa,kBAAkB,YAAY;EAC/C,IAAI,YAAY,cAAc;GAC5B,eAAe;GACf,cAAc;EAChB,OAAO,IAAI,cAAc,cACvB,eAAe;CAEnB;CACA,OAAO;EAAE,GAAG,WAAW;EAAW,QAAQ,cAAc;CAAU;AACpE;;;;;;;;AASA,SAAS,iBACP,cACA,kBAC+B;CAC/B,MAAM,SAAS,aAAa,QAAQ,KAAK,SAAS,MAAM,MAAM,CAAC;CAC/D,MAAM,QAAQ,SAAS;CACvB,IAAI,OAAO,IAAI,aAAa,KAAK;CACjC,KAAK,KAAK;CACV,KAAK,MAAM,QAAQ,cAAc;EAC/B,MAAM,OAAO,IAAI,aAAa,KAAK;EACnC,KAAK,IAAI,MAAM,GAAG,MAAM,OAAO,OAAO;GACpC,MAAM,cAAc,KAAK;GACzB,IAAI,gBAAgB,GAAG;GACvB,KAAK,QAAS,cAAc;GAC5B,KAAK,MAAM,SAAU,cAAc;EACrC;EACA,OAAO;CACT;CAEA,MAAM,SAAS,SAAS;CACxB,IAAI,OAAO;CACX,IAAI,UAAU;CACd,IAAI,eAAe;CACnB,KAAK,IAAI,MAAM,GAAG,MAAM,OAAO,OAAO;EACpC,MAAM,cAAc,KAAK;EACzB,IAAI,gBAAgB,GAAG;EACvB,MAAM,YAAY,KAAK,IAAI,MAAM,MAAM;EACvC,IAAI,aAAa,kBAAkB,QAAQ;EAC3C,IAAI,YAAY,cAAc;GAC5B,eAAe;GACf,UAAU;EACZ,OAAO,IAAI,cAAc,cACvB,WAAW;CAEf;CACA,OAAO;EAAE,GAAG,KAAK,IAAI,GAAG,IAAI;EAAG,QAAQ,KAAK,IAAI,GAAG,OAAO;CAAE;AAC9D;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACtaA,SAAgB,SAAS,MAA+B,MAAuC;CAC7F,MAAM,EAAE,aAAa,iBAAiB;CACtC,IAAI,gBAAgB,cAClB,MAAM,IAAI,gBACR,oDAAoD,YAAY,wCAClE;CAIF,MAAM,wBAAQ,IAAI,IAAyC;CAC3D,MAAM,2BAAW,IAAI,IAAY;CACjC,KAAK,MAAM,OAAO,MAAM;EACtB,SAAS,IAAI,IAAI,GAAG;EACpB,IAAI,IAAI,QAAQ,eAAe,IAAI,QAAQ,cAAc;EACzD,MAAM,QAAQ,MAAM,IAAI,IAAI,GAAG,qBAAK,IAAI,IAA4B;EACpE,MAAM,QAAQ,MAAM,IAAI,IAAI,OAAO,KAAK,CAAC;EACzC,MAAM,KAAK,GAAG;EACd,MAAM,IAAI,IAAI,SAAS,KAAK;EAC5B,MAAM,IAAI,IAAI,KAAK,KAAK;CAC1B;CAEA,KAAK,MAAM,OAAO,CAAC,aAAa,YAAY,GAC1C,IAAI,CAAC,MAAM,IAAI,GAAG,GAEhB,MAAM,IAAI,gBAAgB,8BAA8B,IAAI,mBAD/C,CAAC,GAAG,QAAQ,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,IAAI,KAAK,SACoC,EAAE;CAI1F,MAAM,gBAAgB,MAAM,IAAI,WAAW;CAC3C,MAAM,iBAAiB,MAAM,IAAI,YAAY;CAE7C,MAAM,UAAU,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,cAAc,KAAK,GAAG,GAAG,eAAe,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK;CACvF,MAAM,QAAuB,CAAC;CAC9B,MAAM,mBAAmC,CAAC;CAC1C,MAAM,oBAAoC,CAAC;CAC3C,KAAK,MAAM,WAAW,SAAS;EAC7B,MAAM,IAAI,cAAc,IAAI,OAAO,KAAK,CAAC;EACzC,MAAM,IAAI,eAAe,IAAI,OAAO,KAAK,CAAC;EAE1C,IAAI,EAAE,UAAU,KAAK,EAAE,UAAU,GAAG;GAClC,IAAI,EAAE,WAAW,KAAK,EAAE,WAAW,GAAG;IACpC,MAAM,WAAW,EAAE;IACnB,MAAM,YAAY,EAAE;IACpB,IAAI,SAAS,WAAW,KAAA,KAAa,UAAU,WAAW,KAAA,GAEtD;SAAA,SAAS,WAAW,KAAA,KACpB,UAAU,WAAW,KAAA,KACrB,SAAS,WAAW,UAAU,QAC9B;MACA,iBAAiB,KAAK,QAAQ;MAC9B,kBAAkB,KAAK,SAAS;MAChC;KACF;;IAEF,MAAM,KAAK;KAAE;KAAS,UAAU;KAAG;KAAU;IAAU,CAAC;GAC1D,OAAO;IACL,iBAAiB,KAAK,GAAG,CAAC;IAC1B,kBAAkB,KAAK,GAAG,CAAC;GAC7B;GACA;EACF;EAEA,MAAM,SAAS,cAAc,GAAG,SAAS,WAAW;EACpD,MAAM,SAAS,cAAc,GAAG,SAAS,YAAY;EACrD,MAAM,UAAU,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,OAAO,KAAK,GAAG,GAAG,OAAO,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK;EACxE,IAAI,WAAW;EACf,KAAK,MAAM,UAAU,SAAS;GAC5B,MAAM,WAAW,OAAO,IAAI,MAAM;GAClC,MAAM,YAAY,OAAO,IAAI,MAAM;GACnC,IAAI,aAAa,KAAA,KAAa,cAAc,KAAA,GAC1C,MAAM,KAAK;IAAE;IAAS,UAAU;IAAY;IAAU;GAAU,CAAC;QAC5D,IAAI,aAAa,KAAA,GACtB,iBAAiB,KAAK,QAAQ;QACzB,IAAI,cAAc,KAAA,GACvB,kBAAkB,KAAK,SAAS;EAEpC;CACF;CAEA,OAAO;EAAE;EAAO;EAAkB;CAAkB;AACtD;;;AAIA,SAAS,cACP,OACA,SACA,KAC2B;CAC3B,MAAM,wBAAQ,IAAI,IAA0B;CAC5C,KAAK,MAAM,OAAO,OAAO;EACvB,IAAI,IAAI,WAAW,KAAA,GACjB,MAAM,IAAI,gBACR,sBAAsB,QAAQ,mDAAmD,IAAI,8KAGvF;EAEF,IAAI,MAAM,IAAI,IAAI,MAAM,GACtB,MAAM,IAAI,gBACR,+BAA+B,IAAI,OAAO,iBAAiB,QAAQ,YAAY,IAAI,iEAErF;EAEF,MAAM,IAAI,IAAI,QAAQ,GAAG;CAC3B;CACA,OAAO;AACT;;;;;;;;;;;;;;AAiEA,SAAgB,kBACd,MACA,MACsB;CACtB,MAAM,EAAE,OAAO,kBAAkB,sBAAsB,SAAS,MAAM,IAAI;CAE1E,IAAI,cAAwC;CAC5C,MAAM,eAAyB,CAAC;CAChC,MAAM,gBAA0B,CAAC;CACjC,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,KAAK,SAAS,SAAS,KAAA,KAAa,KAAK,UAAU,SAAS,KAAA,GAAW;EAC3E,aAAa,KAAK,KAAK,SAAS,OAAO,IAAI,CAAC;EAC5C,cAAc,KAAK,KAAK,UAAU,OAAO,IAAI,CAAC;CAChD;CACA,IAAI,aAAa,SAAS,GAAG;EAC3B,MAAM,KAAK,QAAQ,cAAc,aAAa;EAC9C,cAAc;GACZ,KAAK,GAAG;GACR,KAAK,GAAG;GACR,SAAS;GACT,gBAAgB,qBAAqB,cAAc,aAAa;EAClE;CACF;CAaA,MAAM,gBAVJ,KAAK,eACL,CACE,GAAG,IAAI,IACL,MAAM,SAAS,MAAM,CACnB,GAAG,OAAO,KAAK,EAAE,SAAS,WAAW,CAAC,CAAC,GACvC,GAAG,OAAO,KAAK,EAAE,UAAU,WAAW,CAAC,CAAC,CAC1C,CAAC,CACH,CACF,CAAC,CAAC,KAAK,EAAA,CAE6C,KAAK,SAAS;EAClE,MAAM,SAAmB,CAAC;EAC1B,MAAM,QAAkB,CAAC;EACzB,IAAI,WAAW;EACf,KAAK,MAAM,QAAQ,OAAO;GACxB,MAAM,IAAI,YAAY,KAAK,UAAU,IAAI;GACzC,MAAM,IAAI,YAAY,KAAK,WAAW,IAAI;GAC1C,IAAI,MAAM,KAAA,KAAa,MAAM,KAAA,GAAW;IACtC;IACA;GACF;GACA,OAAO,KAAK,CAAC;GACb,MAAM,KAAK,CAAC;EACd;EACA,MAAM,cAAc,OAAO,WAAW,IAAI,OAAO,gBAAgB,QAAQ,OAAO,KAAK,SAAS;EAC9F,OAAO;GACL;GACA,GAAG,OAAO;GACV;GACA,aAAa,aAAa,UAAU;GACpC,WAAW,aAAa,QAAQ;GAChC;GACA,UAAU,OAAO,WAAW,IAAI,OAAO,mBAAmB,QAAQ,KAAK;EACzE;CACF,CAAC;CAED,OAAO;EACL,QAAQ,MAAM;EACd,mBAAmB,iBAAiB;EACpC,oBAAoB,kBAAkB;EACtC;EACA;CACF;AACF;;;;;;;;AA2BA,SAAgB,eACd,cACA,eACsB;CACtB,MAAM,eAAe,iBAAiB,cAAc,UAAU;CAC9D,MAAM,gBAAgB,iBAAiB,eAAe,WAAW;CACjE,yBAAyB,cAAc,UAAU;CACjD,yBAAyB,eAAe,WAAW;CACnD,IAAI,aAAa,WAAW,KAAK,cAAc,WAAW,GACxD,OAAO;EACL,OAAO,CAAC;EACR,kBAAkB,aAAa,KAAK,QAAQ,IAAI,GAAG;EACnD,mBAAmB,cAAc,KAAK,QAAQ,IAAI,GAAG;CACvD;CAGF,MAAM,SAAS,SAAS,CAAC,GAAG,cAAc,GAAG,aAAa,GAAG;EAC3D,aAAa;EACb,cAAc;CAChB,CAAC;CACD,OAAO;EACL,OAAO,OAAO,MAAM,KAAK,SAAS;GAChC,MAAM,WAAW,KAAK;GACtB,MAAM,YAAY,KAAK;GACvB,OAAO;IACL,SAAS,KAAK;IACd,QAAQ,SAAS;IACjB,UAAU,SAAS;IACnB,WAAW,UAAU;GACvB;EACF,CAAC;EACD,kBAAkB,OAAO,iBAAiB,KAAK,QAAS,IAAwB,GAAG;EACnF,mBAAmB,OAAO,kBAAkB,KAAK,QAAS,IAAwB,GAAG;CACvF;AACF;AAEA,SAAS,iBAAiB,MAA4B,KAAgC;CACpF,OAAO,KAAK,KAAK,QAAQ;EACvB,MAAM,aAAa,IAAI,WAAW,KAAK;EACvC,IAAI,CAAC,YACH,MAAM,IAAI,gBACR,wBAAwB,IAAI,MAAM,+EACpC;EAEF,OAAO;GACL,SAAS,KAAK,UAAU,CAAC,IAAI,cAAc,UAAU,CAAC;GACtD,QAAQ,OAAO,IAAI,IAAI;GACvB;GACA;EACF;CACF,CAAC;AACH;AAEA,SAAS,yBAAyB,MAAkC,KAAmB;CACrF,MAAM,4BAAY,IAAI,IAA+B;CACrD,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,QAAQ,UAAU,IAAI,IAAI,OAAO,KAAK,CAAC;EAC7C,MAAM,KAAK,GAAG;EACd,UAAU,IAAI,IAAI,SAAS,KAAK;CAClC;CACA,KAAK,MAAM,CAAC,SAAS,UAAU,WAC7B,IAAI,MAAM,SAAS,GAAG,cAAc,OAAO,SAAS,GAAG;AAE3D;AAEA,SAAS,YAAY,KAAmB,MAAkC;CACxE,MAAM,IAAI,IAAI,UAAU;CACxB,IAAI,MAAM,KAAA,GAAW,OAAO,KAAA;CAC5B,IAAI,CAAC,OAAO,SAAS,CAAC,GACpB,MAAM,IAAI,gBACR,mDAAmD,KAAK,gBAAgB,IAAI,QAAQ,UAAU,IAAI,IAAI,MAAM,GAC9G;CAEF,OAAO;AACT"}

package/dist/promotion-policy-DWOm70gx.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"promotion-policy-DWOm70gx.js","names":[],"sources":["../src/campaign/gates/promotion-policy.ts"],"sourcesContent":["/**\n * Promotion policy over the evidence VECTOR — the substrate's answer to \"never\n * collapse the multi-objective promotion decision into one scalar.\" A\n * `defaultProductionGate` is one opinionated composition; this module factors\n * the decision into two reusable pieces so MANY policies can compete over the\n * SAME evidence (the quant-desk pattern: one evidence bus, plural strategies):\n *\n *   buildEvidenceVector(ctx, objectives, opts) -> EvidenceVector   // the bus\n *   PromotionPolicy = (ev: EvidenceVector) => GateResult           // a strategy\n *   paretoPolicy(ev)                                               // the default strategy\n *   paretoSignificanceGate(options): Gate                          // bus + policy as a Gate\n *\n * The Pareto policy is SYMMETRIC multi-objective: every objective is BOTH a\n * potential gain source AND a safety floor (unlike `defaultProductionGate`,\n * where only `composite` can win and `criticalDimensions` are pure floors). A\n * candidate ships iff it weakly DOMINATES the baseline at the confidence level —\n * no objective credibly worse (CI floor breach) AND at least one objective\n * credibly better (CI gain). Insufficient evidence on ANY axis -> need_more_work\n * (NOT folded into hold: \"gather more reps\" and \"reject\" are different actions).\n *\n * Cost/latency are NOT CI axes here — `GateContext` carries only an aggregate\n * per-side cost, no per-cell observation vector to bootstrap. Treat them as hard\n * constraints (compose with a budget gate via `composeGate`), not faked CIs.\n */\n\nimport {\n  decidePairedPromotion,\n  type PairedDecisionMethod,\n  type PairedDecisionStatistic,\n  type PairedMcNemarEvidence,\n} from '../../paired-promotion-decision'\nimport type { Direction } from '../../pareto'\nimport {\n  DECISION_PAIRED_DELTA_STATISTIC,\n  type PairedBootstrapResult,\n  pairedBootstrap,\n} from '../../statistics'\nimport type { Gate, GateContext, GateDecision, GateResult, JudgeScore, Scenario } from '../types'\nimport { detectScale, pairHoldout } from './statistical-heldout'\n\n/** Where an objective's per-cell scalar comes from. `composite` reads the\n *  judge's composite; `dimension` reads a named per-dimension score. */\nexport type ObjectiveSource = { kind: 'composite' } | { kind: 'dimension'; dimension: string }\n\nexport interface PromotionObjective {\n  /** Stable label used in reports + `contributingGates`. */\n  name: string\n  source: ObjectiveSource\n  /** 'maximize' (quality dims) or 'minimize' (error/risk/length dims). Orients\n   *  the paired delta so a positive bootstrap always means \"candidate better\". */\n  direction: Direction\n  /** The good-direction paired-delta CI lower bound must EXCEED this to count\n   *  as a significant gain on this axis. Interpreted in the judge's native\n   *  scale. Default 0 (⇒ \"confidently better\"). */\n  gainThreshold?: number\n  /** A floor breach (regression) is declared when the good-direction CI lower\n   *  bound is below −floorTolerance, or when the exact small-sample test proves\n   *  a drop past it. When omitted it auto-scales off observed magnitudes\n   *  (0.05 on [0,1], 5 on 0-100), matching `dimensionRegressions`. */\n  floorTolerance?: number\n}\n\n/** Per-axis verdict from the good-direction paired bootstrap. */\nexport type AxisVerdict = 'improved' | 'regressed' | 'flat' | 'few_runs'\n\nexport interface AxisEvidence {\n  name: string\n  source: ObjectiveSource\n  direction: Direction\n  /** Paired bootstrap on the GOOD-DIRECTION delta (oriented by `direction`):\n   *  a positive value means the candidate is better on this axis.\n   *\n   *  DIAGNOSTIC on a pass/fail axis: there the verdict is decided on Tango's\n   *  score interval instead, because a percentile bootstrap over a three-atom\n   *  delta lattice is not a valid interval at the nonzero margin `floorTolerance`\n   *  and `gainThreshold` create. `ci` carries the interval that decided. */\n  bootstrap: PairedBootstrapResult\n  /** Which paired statistic `bootstrap.low`/`.high` bracket. `'mean'` unless the\n   *  caller asked for the median — on a pass/fail axis the median and its whole\n   *  CI are pinned at 0 by tie domination and can see neither a gain nor a\n   *  regression. `bootstrap.median` still carries the median point estimate. */\n  bootstrapStatistic: 'median' | 'mean'\n  /** The interval the axis verdict was actually decided on, good-direction and\n   *  in the axis's native units. */\n  ci: { low: number; high: number }\n  /** Which estimator produced `ci`. */\n  decisionStatistic: PairedDecisionStatistic\n  /** McNemar's exact evidence on a pass/fail axis; null otherwise. */\n  mcnemar: PairedMcNemarEvidence | null\n  /** `ci` has zero width — no evidence in either direction, so the axis is\n   *  neither improved nor regressed however the point estimate sits. */\n  indeterminate: boolean\n  /** Paired observations contributing to this axis. */\n  n: number\n  minimumRequired: number\n  decisionMethod: PairedDecisionMethod\n  gainThreshold: number\n  floorTolerance: number\n  verdict: AxisVerdict\n}\n\nexport interface EvidenceVector {\n  /** One entry per objective — NOTHING averaged across axes. */\n  axes: AxisEvidence[]\n  /** Smallest paired n across axes that produced observations — the binding\n   *  evidence-sufficiency constraint. 0 when no axis produced observations. */\n  minN: number\n  /** Aggregate per-side cost from the gate context (a constraint input, not a\n   *  CI axis — see the module header). */\n  cost: { candidate: number; baseline: number }\n}\n\n/** A promotion strategy: a pure function from the evidence vector to a verdict.\n *  Many policies can run over the same `EvidenceVector` and disagree — that's\n *  the point (competing strategies, shared evidence). */\nexport type PromotionPolicy = (ev: EvidenceVector) => GateResult\n\nexport interface BuildEvidenceVectorOptions {\n  /** Minimum paired observations before an axis can claim significance; below\n   *  it the axis is `few_runs`. The exact small-sample test may require more\n   *  observations at the selected confidence. */\n  minProductiveRuns?: number\n  /** Confidence level for every axis bootstrap. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples. Default 2000. */\n  resamples?: number\n  /** Fixed bootstrap seed for a deterministic, reproducible verdict. Default 1337. */\n  seed?: number\n  /** Paired statistic every axis CI is computed on. Default `'mean'` — see\n   *  {@link DECISION_PAIRED_DELTA_STATISTIC} for why the median is not. */\n  statistic?: 'mean' | 'median'\n}\n\n/**\n * The Evidence Bus. For each objective, pair candidate vs baseline by full\n * cellId and bootstrap a CI on the good-direction paired delta. Reuses the\n * exact `pairHoldout` + `pairedBootstrap` machinery the held-out gate uses, so\n * a single source of truth governs pairing granularity + scale handling.\n */\nexport function buildEvidenceVector<TArtifact, TScenario extends Scenario>(\n  ctx: GateContext<TArtifact, TScenario>,\n  objectives: PromotionObjective[],\n  opts: BuildEvidenceVectorOptions = {},\n): EvidenceVector {\n  if (objectives.length === 0) {\n    throw new Error('buildEvidenceVector: at least 1 objective required')\n  }\n  const confidence = opts.confidence ?? 0.95\n  const resamples = opts.resamples ?? 2000\n  const seed = opts.seed ?? 1337\n  const baseline = ctx.baselineJudgeScores ?? ctx.judgeScores\n  const scenarioIds = new Set(ctx.scenarios.map((s) => s.id))\n\n  const axes: AxisEvidence[] = []\n  for (const obj of objectives) {\n    let select: (s: JudgeScore) => number | undefined\n    if (obj.source.kind === 'composite') {\n      select = (s) => s.composite\n    } else {\n      const dim = obj.source.dimension\n      select = (s) => s.dimensions[dim]\n    }\n    const paired = pairHoldout(ctx.judgeScores, baseline, scenarioIds, select)\n    // Orient to the good direction: maximize ⇒ bootstrap (candidate − baseline);\n    // minimize ⇒ bootstrap (baseline − candidate) by swapping args, so a\n    // positive bootstrap always reads as \"candidate better on this axis\".\n    const before = obj.direction === 'maximize' ? paired.before : paired.after\n    const after = obj.direction === 'maximize' ? paired.after : paired.before\n    const n = paired.before.length\n    const floorTolerance =\n      obj.floorTolerance ?? 0.05 * detectScale([...paired.before, ...paired.after])\n    const gainThreshold = obj.gainThreshold ?? 0\n    // Axes are decided on the MEAN paired delta — which for a pass/fail axis is\n    // exactly the change in success rate. The median is structurally blind on\n    // the shapes eval data lands in: with most pairs tied its bootstrap CI\n    // collapses to [0,0] and the axis reads 'flat', hiding real gains AND real\n    // regressions — pass/fail axes on ANY encoding ({0,1} and the 0-100 one\n    // `detectScale` exists for), and low-cardinality axes even below half ties.\n    // Orthogonal to `pairedDeltaTest`'s own small-sample switch: that picks the\n    // TEST (bootstrap CI at n ≥ 20, exact sign test below), this picks the\n    // ESTIMATOR the test is applied to. Both are needed — an exact sign test on\n    // a tie-pinned median is still blind.\n    const bootstrapStatistic = opts.statistic ?? DECISION_PAIRED_DELTA_STATISTIC\n    // Both burdens of proof route through the ONE shared rule\n    // (`decidePairedPromotion`), so a pass/fail axis is judged on Tango's score\n    // interval — the only paired-binary construction valid at the nonzero\n    // margins `gainThreshold` / `floorTolerance` create — and a zero-width\n    // interval cannot buy a verdict in either direction.\n    const improvement = decidePairedPromotion(before, after, {\n      confidence,\n      resamples,\n      statistic: bootstrapStatistic,\n      seed,\n      threshold: gainThreshold,\n      minPairs: opts.minProductiveRuns,\n    })\n    const regression = decidePairedPromotion(after, before, {\n      confidence,\n      resamples,\n      statistic: bootstrapStatistic,\n      seed,\n      threshold: floorTolerance,\n      minPairs: opts.minProductiveRuns,\n    })\n    const bootstrap =\n      improvement.bootstrap ??\n      pairedBootstrap(before, after, {\n        confidence,\n        resamples,\n        statistic: bootstrapStatistic,\n        seed,\n      })\n    // A floor breach fires on EITHER burden of proof, because they cover\n    // different failures and the floor is the anti-Goodhart guard:\n    //   - `improvement.low < -floorTolerance` — the CREDIBLE WORST CASE exceeds\n    //     the tolerance. This is the contract `AxisEvidence.floorTolerance` and\n    //     `paretoPolicy`'s own reason string state, and it is the conservative\n    //     posture a safety axis needs: block unless the data can rule the\n    //     breach out, rather than waiting for the breach to be proven. Read off\n    //     the DECIDING interval, so a pass/fail axis is not screened by a\n    //     bootstrap that is pinned wherever ties dominate.\n    //   - `regression.promote` — a PROVEN drop past the tolerance. Adds the\n    //     small-sample path, where the decision is an exact sign test because\n    //     the bootstrap interval is descriptive only.\n    // The credible-worst-case arm stays on the BOOTSTRAP, deliberately. Reading\n    // it off the score interval instead would change what the floor MEANS on a\n    // pass/fail axis: with every pair concordant the score interval is\n    // ±z²/(n+z²) — ±0.39 at n=6, ±0.16 at n=20 — so a completely unchanged\n    // safety axis would breach a 0.05 floor at any realistic n, and the gate\n    // would refuse everything. That the bootstrap arm is instead fail-OPEN on a\n    // tied pass/fail axis is a real and separate weakness: the honest fix is a\n    // minimum-power requirement on the floor, not a wider interval, because the\n    // data genuinely cannot rule a 5pp drop out at n=20 and a gate that says so\n    // by blocking every candidate is not usable. `regression.promote` — the\n    // PROVEN-drop arm — does route through the shared rule, so a real pass/fail\n    // regression is now caught on an interval valid at the nonzero tolerance.\n    const floorBreached = bootstrap.low < -floorTolerance || regression.promote\n    // Floor check precedes the gain check: a credible regression must never be\n    // masked as \"improved\". With the defaults (gainThreshold 0, positive floor)\n    // the regions are disjoint and order is moot, but a consumer who sets a\n    // negative gainThreshold (\"accept small dips\") could otherwise have a real\n    // floor breach classified as a gain — anti-Goodhart wins the tie.\n    const verdict: AxisVerdict = !improvement.sufficient\n      ? 'few_runs'\n      : floorBreached\n        ? 'regressed'\n        : improvement.promote\n          ? 'improved'\n          : 'flat'\n    axes.push({\n      name: obj.name,\n      source: obj.source,\n      direction: obj.direction,\n      bootstrap,\n      bootstrapStatistic,\n      ci: { low: improvement.low, high: improvement.high },\n      decisionStatistic: improvement.statistic,\n      mcnemar: improvement.mcnemar,\n      indeterminate: improvement.indeterminate,\n      n,\n      minimumRequired: improvement.minimumPairs,\n      decisionMethod: improvement.method,\n      gainThreshold,\n      floorTolerance,\n      verdict,\n    })\n  }\n  const ns = axes.map((a) => a.n).filter((n) => n > 0)\n  const minN = ns.length > 0 ? Math.min(...ns) : 0\n  return { axes, minN, cost: { candidate: ctx.cost.candidate, baseline: ctx.cost.baseline } }\n}\n\n/**\n * The default strategy: symmetric multi-objective Pareto significance. Ship iff\n * the candidate weakly dominates the baseline at the confidence level — no axis\n * credibly worse AND ≥1 axis credibly better. Floor breach on any axis → hold\n * (anti-Goodhart, dominates everything). Insufficient evidence on any axis →\n * need_more_work. Statistically equivalent → hold (never ship noise).\n */\nexport const paretoPolicy: PromotionPolicy = (ev) => {\n  const contributingGates = ev.axes.map((ax) => ({\n    name: `objective:${ax.name}`,\n    status:\n      ax.verdict === 'regressed'\n        ? ('fail' as const)\n        : ax.verdict === 'few_runs'\n          ? ('not_evaluated' as const)\n          : ('pass' as const),\n    detail: {\n      direction: ax.direction,\n      source: ax.source,\n      verdict: ax.verdict,\n      n: ax.n,\n      deltaMedian: ax.bootstrap.median,\n      ciLow: ax.ci.low,\n      ciHigh: ax.ci.high,\n      decisionStatistic: ax.decisionStatistic,\n      decisionMethod: ax.decisionMethod,\n      mcnemar: ax.mcnemar,\n      indeterminate: ax.indeterminate,\n      bootstrapCiLow: ax.bootstrap.low,\n      bootstrapCiHigh: ax.bootstrap.high,\n      confidence: ax.bootstrap.confidence,\n      gainThreshold: ax.gainThreshold,\n      floorTolerance: ax.floorTolerance,\n    },\n  }))\n\n  const regressed = ev.axes.filter((a) => a.verdict === 'regressed')\n  const fewRuns = ev.axes.filter((a) => a.verdict === 'few_runs')\n  const improved = ev.axes.filter((a) => a.verdict === 'improved')\n\n  let decision: GateDecision\n  const reasons: string[] = []\n  if (regressed.length > 0) {\n    // Floor breach dominates: a credible regression on ANY axis blocks ship even\n    // if another axis improved. This makes the +gain/−safety false positive\n    // structurally impossible whenever the safety dim is an objective.\n    decision = 'hold'\n    for (const a of regressed) {\n      reasons.push(\n        `objective '${a.name}' regressed: good-direction CI.low ${a.ci.low.toFixed(3)} < -${a.floorTolerance} (n=${a.n})`,\n      )\n    }\n  } else if (fewRuns.length > 0) {\n    // No credible regression on the scored axes, but ≥1 axis lacks the evidence\n    // to claim a gain ⇒ gather more reps, do NOT reject.\n    decision = 'need_more_work'\n    for (const a of fewRuns) {\n      reasons.push(\n        `objective '${a.name}' has only n=${a.n} paired runs — insufficient evidence to claim significance`,\n      )\n    }\n  } else if (improved.length > 0) {\n    // Weakly dominates (no axis worse) AND strictly better on ≥1 axis ⇒ a Pareto\n    // improvement at the confidence level.\n    decision = 'ship'\n    reasons.push(\n      `Pareto improvement at the confidence level: ${improved\n        .map(\n          (a) =>\n            `'${a.name}' +${a.ci.low > 0 ? a.ci.low.toFixed(3) : a.bootstrap.mean.toFixed(3)} (CI.low ${a.ci.low.toFixed(3)})`,\n        )\n        .join(', ')}; no objective regressed`,\n    )\n  } else {\n    // Enough evidence, nothing credibly better or worse ⇒ statistically\n    // equivalent. Do NOT ship a no-op.\n    decision = 'hold'\n    reasons.push(\n      'no Pareto improvement: candidate statistically equivalent to baseline on every objective',\n    )\n  }\n\n  // `delta` surfaces the composite axis if present, else the first axis — a\n  // single convenience scalar; the vector lives in `contributingGates`.\n  const composite = ev.axes.find((a) => a.source.kind === 'composite') ?? ev.axes[0]\n  return { decision, reasons, contributingGates, delta: composite?.bootstrap.median }\n}\n\nexport interface ParetoSignificanceGateOptions extends BuildEvidenceVectorOptions {\n  /** The objective vector. Every axis is both a gain source and a safety floor. */\n  objectives: PromotionObjective[]\n  /** Strategy applied to the evidence vector. Default `paretoPolicy`. Override\n   *  to run a stricter/looser strategy over the SAME bus (competing policies). */\n  policy?: PromotionPolicy\n  /** Override the gate name in reports. */\n  name?: string\n}\n\n/**\n * Wrap the bus + a policy as a `Gate`. Plugs into the existing\n * `runImprovementLoop({ gate })` slot and composes via `composeGate`; default\n * loop behavior is unchanged because consumers opt in by passing this gate.\n */\nexport function paretoSignificanceGate<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  options: ParetoSignificanceGateOptions,\n): Gate<TArtifact, TScenario> {\n  if (options.objectives.length === 0) {\n    throw new Error('paretoSignificanceGate: at least 1 objective required')\n  }\n  const policy = options.policy ?? paretoPolicy\n  return {\n    name: options.name ?? 'paretoSignificanceGate',\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      const ev = buildEvidenceVector(ctx, options.objectives, options)\n      return policy(ev)\n    },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA2IA,SAAgB,oBACd,KACA,YACA,OAAmC,CAAC,GACpB;CAChB,IAAI,WAAW,WAAW,GACxB,MAAM,IAAI,MAAM,oDAAoD;CAEtE,MAAM,aAAa,KAAK,cAAc;CACtC,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,OAAO,KAAK,QAAQ;CAC1B,MAAM,WAAW,IAAI,uBAAuB,IAAI;CAChD,MAAM,cAAc,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,EAAE,EAAE,CAAC;CAE1D,MAAM,OAAuB,CAAC;CAC9B,KAAK,MAAM,OAAO,YAAY;EAC5B,IAAI;EACJ,IAAI,IAAI,OAAO,SAAS,aACtB,UAAU,MAAM,EAAE;OACb;GACL,MAAM,MAAM,IAAI,OAAO;GACvB,UAAU,MAAM,EAAE,WAAW;EAC/B;EACA,MAAM,SAAS,YAAY,IAAI,aAAa,UAAU,aAAa,MAAM;EAIzE,MAAM,SAAS,IAAI,cAAc,aAAa,OAAO,SAAS,OAAO;EACrE,MAAM,QAAQ,IAAI,cAAc,aAAa,OAAO,QAAQ,OAAO;EACnE,MAAM,IAAI,OAAO,OAAO;EACxB,MAAM,iBACJ,IAAI,kBAAkB,MAAO,YAAY,CAAC,GAAG,OAAO,QAAQ,GAAG,OAAO,KAAK,CAAC;EAC9E,MAAM,gBAAgB,IAAI,iBAAiB;EAW3C,MAAM,qBAAqB,KAAK,aAAA;EAMhC,MAAM,cAAc,sBAAsB,QAAQ,OAAO;GACvD;GACA;GACA,WAAW;GACX;GACA,WAAW;GACX,UAAU,KAAK;EACjB,CAAC;EACD,MAAM,aAAa,sBAAsB,OAAO,QAAQ;GACtD;GACA;GACA,WAAW;GACX;GACA,WAAW;GACX,UAAU,KAAK;EACjB,CAAC;EACD,MAAM,YACJ,YAAY,aACZ,gBAAgB,QAAQ,OAAO;GAC7B;GACA;GACA,WAAW;GACX;EACF,CAAC;EAyBH,MAAM,gBAAgB,UAAU,MAAM,CAAC,kBAAkB,WAAW;EAMpE,MAAM,UAAuB,CAAC,YAAY,aACtC,aACA,gBACE,cACA,YAAY,UACV,aACA;EACR,KAAK,KAAK;GACR,MAAM,IAAI;GACV,QAAQ,IAAI;GACZ,WAAW,IAAI;GACf;GACA;GACA,IAAI;IAAE,KAAK,YAAY;IAAK,MAAM,YAAY;GAAK;GACnD,mBAAmB,YAAY;GAC/B,SAAS,YAAY;GACrB,eAAe,YAAY;GAC3B;GACA,iBAAiB,YAAY;GAC7B,gBAAgB,YAAY;GAC5B;GACA;GACA;EACF,CAAC;CACH;CACA,MAAM,KAAK,KAAK,KAAK,MAAM,EAAE,CAAC,CAAC,CAAC,QAAQ,MAAM,IAAI,CAAC;CAEnD,OAAO;EAAE;EAAM,MADF,GAAG,SAAS,IAAI,KAAK,IAAI,GAAG,EAAE,IAAI;EAC1B,MAAM;GAAE,WAAW,IAAI,KAAK;GAAW,UAAU,IAAI,KAAK;EAAS;CAAE;AAC5F;;;;;;;;AASA,MAAa,gBAAiC,OAAO;CACnD,MAAM,oBAAoB,GAAG,KAAK,KAAK,QAAQ;EAC7C,MAAM,aAAa,GAAG;EACtB,QACE,GAAG,YAAY,cACV,SACD,GAAG,YAAY,aACZ,kBACA;EACT,QAAQ;GACN,WAAW,GAAG;GACd,QAAQ,GAAG;GACX,SAAS,GAAG;GACZ,GAAG,GAAG;GACN,aAAa,GAAG,UAAU;GAC1B,OAAO,GAAG,GAAG;GACb,QAAQ,GAAG,GAAG;GACd,mBAAmB,GAAG;GACtB,gBAAgB,GAAG;GACnB,SAAS,GAAG;GACZ,eAAe,GAAG;GAClB,gBAAgB,GAAG,UAAU;GAC7B,iBAAiB,GAAG,UAAU;GAC9B,YAAY,GAAG,UAAU;GACzB,eAAe,GAAG;GAClB,gBAAgB,GAAG;EACrB;CACF,EAAE;CAEF,MAAM,YAAY,GAAG,KAAK,QAAQ,MAAM,EAAE,YAAY,WAAW;CACjE,MAAM,UAAU,GAAG,KAAK,QAAQ,MAAM,EAAE,YAAY,UAAU;CAC9D,MAAM,WAAW,GAAG,KAAK,QAAQ,MAAM,EAAE,YAAY,UAAU;CAE/D,IAAI;CACJ,MAAM,UAAoB,CAAC;CAC3B,IAAI,UAAU,SAAS,GAAG;EAIxB,WAAW;EACX,KAAK,MAAM,KAAK,WACd,QAAQ,KACN,cAAc,EAAE,KAAK,qCAAqC,EAAE,GAAG,IAAI,QAAQ,CAAC,EAAE,MAAM,EAAE,eAAe,MAAM,EAAE,EAAE,EACjH;CAEJ,OAAO,IAAI,QAAQ,SAAS,GAAG;EAG7B,WAAW;EACX,KAAK,MAAM,KAAK,SACd,QAAQ,KACN,cAAc,EAAE,KAAK,eAAe,EAAE,EAAE,2DAC1C;CAEJ,OAAO,IAAI,SAAS,SAAS,GAAG;EAG9B,WAAW;EACX,QAAQ,KACN,+CAA+C,SAC5C,KACE,MACC,IAAI,EAAE,KAAK,KAAK,EAAE,GAAG,MAAM,IAAI,EAAE,GAAG,IAAI,QAAQ,CAAC,IAAI,EAAE,UAAU,KAAK,QAAQ,CAAC,EAAE,WAAW,EAAE,GAAG,IAAI,QAAQ,CAAC,EAAE,EACpH,CAAC,CACA,KAAK,IAAI,EAAE,yBAChB;CACF,OAAO;EAGL,WAAW;EACX,QAAQ,KACN,0FACF;CACF;CAIA,MAAM,YAAY,GAAG,KAAK,MAAM,MAAM,EAAE,OAAO,SAAS,WAAW,KAAK,GAAG,KAAK;CAChF,OAAO;EAAE;EAAU;EAAS;EAAmB,OAAO,WAAW,UAAU;CAAO;AACpF;;;;;;AAiBA,SAAgB,uBACd,SAC4B;CAC5B,IAAI,QAAQ,WAAW,WAAW,GAChC,MAAM,IAAI,MAAM,uDAAuD;CAEzE,MAAM,SAAS,QAAQ,UAAU;CACjC,OAAO;EACL,MAAM,QAAQ,QAAQ;EACtB,MAAM,OAAO,KAA6D;GACxE,MAAM,KAAK,oBAAoB,KAAK,QAAQ,YAAY,OAAO;GAC/D,OAAO,OAAO,EAAE;EAClB;CACF;AACF"}