agentix-cli 0.26.0 → 0.28.0

package/dist/{chunk-5U35WAZF.js → chunk-77OV7346.js}

@@ -1,11 +1,11 @@
-var m="claude-haiku-4-20250514",f=`You are a memory extraction system for an AI agent. Given a conversation exchange, extract facts worth remembering for future conversations across different chat sessions.
+var d="claude-haiku-4-20250514",f=`You are a memory extraction system for an AI agent. Given a conversation exchange, extract facts worth remembering for future conversations across different chat sessions.
 
 Extract ONLY facts that would help the agent in FUTURE, DIFFERENT conversations:
 - Credentials, tokens, API keys shared by the user
 - User preferences and instructions ("never do X", "always use Y")
 - Commitments the agent made ("I will deploy by Friday")
-- Relationships ("Anis is the admin", "Nadia handles content")
-- Task state ("GitLab tokens need deploying to clawd-server")
+- Relationships ("Alex is the admin", "Marketing handles content")
+- Task state ("GitLab tokens need deploying to peer-server")
 - Important facts about infrastructure, config, or processes
 
 SKIP:
@@ -17,7 +17,7 @@ SKIP:
 For each fact, output a JSON array:
 [{"category":"fact|secret|preference|commitment|task-state","content":"concise fact, 1-2 sentences","keywords":["keyword1","keyword2"]}]
 
-If nothing worth remembering, output: []`;async function h(t,s,a,n,r){if(s.length<20&&a.length<50)return;let{createProvider:i}=await import("./providers-L27BJ5A6.js"),c=(await i("claude").generate([{role:"system",content:f},{role:"user",content:`${n.sender} (via ${n.channel}): ${s}
+If nothing worth remembering, output: []`;async function h(t,s,a,n,r){if(s.length<20&&a.length<50)return;let{createProvider:i}=await import("./providers-KO2DODIO.js"),c=(await i("claude").generate([{role:"system",content:f},{role:"user",content:`${n.sender} (via ${n.channel}): ${s}
 
-Agent: ${a}`}],{model:m,maxTokens:512})).content.trim().match(/\[[\s\S]*\]/);if(!c)return;let o=JSON.parse(c[0]);if(!Array.isArray(o)||o.length===0)return;let d=new Date().toISOString().slice(0,10);for(let e of o)r.hasSimilar(t,e.content)||r.addMemory(t,{agentId:t,category:e.category,content:e.content,keywords:e.keywords||[],source:{...n,date:d},expiresAt:e.category==="task-state"?new Date(Date.now()+7*24*60*60*1e3).toISOString():void 0});Math.random()<.02&&r.prune(t)}export{h as a};
-//# sourceMappingURL=chunk-5U35WAZF.js.map
+Agent: ${a}`}],{model:d,maxTokens:512})).content.trim().match(/\[[\s\S]*\]/);if(!c)return;let o=JSON.parse(c[0]);if(!Array.isArray(o)||o.length===0)return;let m=new Date().toISOString().slice(0,10);for(let e of o)r.hasSimilar(t,e.content)||r.addMemory(t,{agentId:t,category:e.category,content:e.content,keywords:e.keywords||[],source:{...n,date:m},expiresAt:e.category==="task-state"?new Date(Date.now()+7*24*60*60*1e3).toISOString():void 0});Math.random()<.02&&r.prune(t)}export{h as a};
+//# sourceMappingURL=chunk-77OV7346.js.map

package/dist/chunk-77OV7346.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/agents/memory-extract.ts"],"sourcesContent":["import type { MemoryStore } from \"./memory-store\"\n\n// --- Haiku-powered memory extraction ---\n// Fire-and-forget after each agent response.\n// Extracts memorable facts and writes them to the persistent store.\n\nconst EXTRACTION_MODEL = \"claude-haiku-4-20250514\"\n\nconst EXTRACTION_PROMPT = `You are a memory extraction system for an AI agent. Given a conversation exchange, extract facts worth remembering for future conversations across different chat sessions.\n\nExtract ONLY facts that would help the agent in FUTURE, DIFFERENT conversations:\n- Credentials, tokens, API keys shared by the user\n- User preferences and instructions (\"never do X\", \"always use Y\")\n- Commitments the agent made (\"I will deploy by Friday\")\n- Relationships (\"Alex is the admin\", \"Marketing handles content\")\n- Task state (\"GitLab tokens need deploying to peer-server\")\n- Important facts about infrastructure, config, or processes\n\nSKIP:\n- Routine greetings, acknowledgments, status updates\n- Information derivable from code or config files\n- Transient conversation flow (\"let me check\", \"here's what I found\")\n- Facts already obvious from the agent's system prompt\n\nFor each fact, output a JSON array:\n[{\"category\":\"fact|secret|preference|commitment|task-state\",\"content\":\"concise fact, 1-2 sentences\",\"keywords\":[\"keyword1\",\"keyword2\"]}]\n\nIf nothing worth remembering, output: []`\n\nexport async function extractMemories(\n  agentId: string,\n  userMessage: string,\n  agentResponse: string,\n  source: { channel: string; chatId: string; sender: string },\n  store: MemoryStore,\n): Promise<void> {\n  // Skip very short exchanges (unlikely to contain memorable facts)\n  if (userMessage.length < 20 && agentResponse.length < 50) return\n\n  const { createProvider } = await import(\"@/agent/providers\")\n  const provider = createProvider(\"claude\")\n\n  const result = await provider.generate(\n    [\n      { role: \"system\", content: EXTRACTION_PROMPT },\n      {\n        role: \"user\",\n        content: `${source.sender} (via ${source.channel}): ${userMessage}\\n\\nAgent: ${agentResponse}`,\n      },\n    ],\n    { model: EXTRACTION_MODEL, maxTokens: 512 },\n  )\n\n  // Parse JSON array from response\n  const content = result.content.trim()\n  const jsonMatch = content.match(/\\[[\\s\\S]*\\]/)\n  if (!jsonMatch) return\n\n  const facts = JSON.parse(jsonMatch[0]) as Array<{\n    category: string\n    content: string\n    keywords: string[]\n  }>\n\n  if (!Array.isArray(facts) || facts.length === 0) return\n\n  const date = new Date().toISOString().slice(0, 10)\n\n  for (const fact of facts) {\n    // Skip duplicates\n    if (store.hasSimilar(agentId, fact.content)) continue\n\n    store.addMemory(agentId, {\n      agentId,\n      category: fact.category as any,\n      content: fact.content,\n      keywords: fact.keywords || [],\n      source: { ...source, date },\n      expiresAt: fact.category === \"task-state\"\n        ? new Date(Date.now() + 7 * 24 * 60 * 60 * 1000).toISOString()\n        : undefined,\n    })\n  }\n\n  // Prune occasionally (every ~50 calls, probabilistic)\n  if (Math.random() < 0.02) {\n    store.prune(agentId)\n  }\n}\n"],"mappings":"AAMA,IAAMA,EAAmB,0BAEnBC,EAAoB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,0CAqB1B,eAAsBC,EACpBC,EACAC,EACAC,EACAC,EACAC,EACe,CAEf,GAAIH,EAAY,OAAS,IAAMC,EAAc,OAAS,GAAI,OAE1D,GAAM,CAAE,eAAAG,CAAe,EAAI,KAAM,QAAO,yBAAmB,EAgBrDC,GAbS,MAFED,EAAe,QAAQ,EAEV,SAC5B,CACE,CAAE,KAAM,SAAU,QAASP,CAAkB,EAC7C,CACE,KAAM,OACN,QAAS,GAAGK,EAAO,eAAeA,EAAO,aAAaF;AAAA;AAAA,SAAyBC,GACjF,CACF,EACA,CAAE,MAAOL,EAAkB,UAAW,GAAI,CAC5C,GAGuB,QAAQ,KAAK,EACV,MAAM,aAAa,EAC7C,GAAI,CAACS,EAAW,OAEhB,IAAMC,EAAQ,KAAK,MAAMD,EAAU,CAAC,CAAC,EAMrC,GAAI,CAAC,MAAM,QAAQC,CAAK,GAAKA,EAAM,SAAW,EAAG,OAEjD,IAAMC,EAAO,IAAI,KAAK,EAAE,YAAY,EAAE,MAAM,EAAG,EAAE,EAEjD,QAAWC,KAAQF,EAEbH,EAAM,WAAWJ,EAASS,EAAK,OAAO,GAE1CL,EAAM,UAAUJ,EAAS,CACvB,QAAAA,EACA,SAAUS,EAAK,SACf,QAASA,EAAK,QACd,SAAUA,EAAK,UAAY,CAAC,EAC5B,OAAQ,CAAE,GAAGN,EAAQ,KAAAK,CAAK,EAC1B,UAAWC,EAAK,WAAa,aACzB,IAAI,KAAK,KAAK,IAAI,EAAI,EAAI,GAAK,GAAK,GAAK,GAAI,EAAE,YAAY,EAC3D,MACN,CAAC,EAIC,KAAK,OAAO,EAAI,KAClBL,EAAM,MAAMJ,CAAO,CAEvB","names":["EXTRACTION_MODEL","EXTRACTION_PROMPT","extractMemories","agentId","userMessage","agentResponse","source","store","createProvider","jsonMatch","facts","date","fact"]}

package/dist/chunk-AJ5RMU7X.js

@@ -0,0 +1,2 @@
+import{a as c,b as d,f as u}from"./chunk-F4HZTZHW.js";u();var f="wiki-rerank",k=30,p=900;function b(t){let r=t.excerptChars??p;return{query:t.query,candidates:t.candidates.map(e=>({id:e.id,title:e.title,tags:e.tags?.length?e.tags.join(", "):null,excerpt:l(e.excerpt,r)}))}}function w(t){let r={};for(let e of t)r[e.id]=e.title;return{best:d(r,"Which candidate answers the query?"),answerable:c("Do any of the candidates answer the query?")}}function x(t,r){let e=t.best,s={};for(let n of r)s[n.id]=e.probabilities[n.id]??0;let o=new Map(r.map((n,a)=>[n.id,a]));return{ranked:[...r.map(n=>n.id)].sort((n,a)=>{let i=s[a]-s[n];return i!==0?i:(o.get(n)??0)-(o.get(a)??0)}),scores:s,answerable:t.answerable.noul,confidence:e.confidence}}function l(t,r){let e=String(t??"").replace(/\s+/g," ").trim();return e.length>r?e.slice(0,r):e}export{f as a,k as b,b as c,w as d,x as e};
+//# sourceMappingURL=chunk-AJ5RMU7X.js.map

package/dist/chunk-AJ5RMU7X.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/decisions/seats/wiki-rerank.ts"],"sourcesContent":["import { choice, noul } from \"../questions\"\nimport type { AnswersFor, ChoiceAnswer, NoulAnswer, StateValue } from \"../types\"\n\n// Re-ranking a BM25 shortlist.\n//\n// Measured on the operator's own wiki (11,656 articles, 12 queries):\n// BM25 alone gets R@1 33.3%, R@3 41.7%. Two thirds of queries do not put\n// the right article first. BM25 is a bag-of-words match, so it does well\n// on \"gitlab token expired\" and badly on \"the bot keeps repeating itself\n// what is wrong\" — the synonym cases score 33% / 33%.\n//\n// Shape follows TypeSafe's semantic_find cookbook rather than their\n// rerank one. Rerank scores each query-candidate pair separately, which\n// is N calls; semantic_find scores 218 line ids against one query with a\n// single Choice, and the per-option probabilities ARE the ranking. One\n// call for the whole shortlist, and the candidates are compared against\n// each other rather than each judged alone — which is what ranking means.\n//\n// The companion Noul is the part that keeps this honest. A Choice must\n// return something: asked which of 30 articles answers a question the\n// wiki has no answer to, it will still name one, with a confident-looking\n// probability. `answerable` is the question that can say no, and code\n// reads it first.\n\nexport const WIKI_RERANK_SEAT = \"wiki-rerank\"\n\n// Measured on the 12-query eval, sweeping candidate count against\n// excerpt length under a fixed state budget:\n//\n//   k=120 x 150 chars   R@1 25.0%   (worse than BM25's 33.3%)\n//   k=60  x 300 chars   R@1 41.7%\n//   k=30  x 600 chars   R@1 41.7%\n//   k=30  x 900 chars   R@1 50.0%   <- default\n//   k=20  x 1100 chars  R@1 50.0%\n//\n// Candidates and excerpt trade against one another, and excerpt wins.\n// More candidates raises the ceiling — what the shortlist could contain —\n// while starving the re-ranker of the text it needs to rank with. At 150\n// characters a candidate is barely more than its title, and the result is\n// worse than no re-ranking at all. Raise k only alongside the state\n// budget, never instead of the excerpt.\nexport const DEFAULT_SHORTLIST = 30\nexport const DEFAULT_EXCERPT_CHARS = 900\n\nexport interface RerankCandidate {\n  id: string\n  title: string\n  tags?: string[]\n  excerpt: string\n}\n\nexport interface RerankInput {\n  query: string\n  candidates: RerankCandidate[]\n  /** Per-candidate excerpt budget. See DEFAULT_EXCERPT_CHARS — this is\n   *  the lever that actually moves accuracy. */\n  excerptChars?: number\n}\n\nexport function rerankState(input: RerankInput): StateValue {\n  const budget = input.excerptChars ?? DEFAULT_EXCERPT_CHARS\n  return {\n    query: input.query,\n    candidates: input.candidates.map((c) => ({\n      id: c.id,\n      title: c.title,\n      tags: c.tags?.length ? c.tags.join(\", \") : null,\n      excerpt: clip(c.excerpt, budget),\n    })),\n  }\n}\n\n/** Built per call, because the option set IS the candidate set. */\nexport function rerankQuestions(candidates: RerankCandidate[]) {\n  const criteria: Record<string, string> = {}\n  for (const c of candidates) criteria[c.id] = c.title\n  return {\n    best: choice(criteria, \"Which candidate answers the query?\"),\n    answerable: noul(\"Do any of the candidates answer the query?\"),\n  }\n}\n\nexport type RerankAnswers = AnswersFor<ReturnType<typeof rerankQuestions>>\n\nexport interface RerankResult {\n  /** Candidate ids, best first, by the Choice's per-option probability. */\n  ranked: string[]\n  scores: Record<string, number>\n  /** P(the shortlist contains an answer at all). */\n  answerable: number\n  /** The Choice's own confidence — flat means it could not separate them. */\n  confidence: number\n}\n\n/**\n * Turn one answer into a ranking.\n *\n * Ties keep the incoming order, so a Choice that genuinely cannot\n * separate the candidates degrades to BM25's ranking rather than to an\n * arbitrary one.\n */\nexport function toRanking(\n  answers: RerankAnswers,\n  candidates: RerankCandidate[],\n): RerankResult {\n  const best = answers.best as ChoiceAnswer\n  const scores: Record<string, number> = {}\n  for (const c of candidates) scores[c.id] = best.probabilities[c.id] ?? 0\n\n  const order = new Map(candidates.map((c, i) => [c.id, i]))\n  const ranked = [...candidates.map((c) => c.id)].sort((a, b) => {\n    const diff = scores[b] - scores[a]\n    return diff !== 0 ? diff : (order.get(a) ?? 0) - (order.get(b) ?? 0)\n  })\n\n  return {\n    ranked,\n    scores,\n    answerable: (answers.answerable as NoulAnswer).noul,\n    confidence: best.confidence,\n  }\n}\n\nfunction clip(text: string, max: number): string {\n  const flat = String(text ?? \"\").replace(/\\s+/g, \" \").trim()\n  return flat.length > max ? flat.slice(0, max) : flat\n}\n"],"mappings":"sDAAAA,IAwBO,IAAMC,EAAmB,cAiBnBC,EAAoB,GACpBC,EAAwB,IAiB9B,SAASC,EAAYC,EAAgC,CAC1D,IAAMC,EAASD,EAAM,cAAgBF,EACrC,MAAO,CACL,MAAOE,EAAM,MACb,WAAYA,EAAM,WAAW,IAAKE,IAAO,CACvC,GAAIA,EAAE,GACN,MAAOA,EAAE,MACT,KAAMA,EAAE,MAAM,OAASA,EAAE,KAAK,KAAK,IAAI,EAAI,KAC3C,QAASC,EAAKD,EAAE,QAASD,CAAM,CACjC,EAAE,CACJ,CACF,CAGO,SAASG,EAAgBC,EAA+B,CAC7D,IAAMC,EAAmC,CAAC,EAC1C,QAAWJ,KAAKG,EAAYC,EAASJ,EAAE,EAAE,EAAIA,EAAE,MAC/C,MAAO,CACL,KAAMK,EAAOD,EAAU,oCAAoC,EAC3D,WAAYE,EAAK,4CAA4C,CAC/D,CACF,CAqBO,SAASC,EACdC,EACAL,EACc,CACd,IAAMM,EAAOD,EAAQ,KACfE,EAAiC,CAAC,EACxC,QAAWV,KAAKG,EAAYO,EAAOV,EAAE,EAAE,EAAIS,EAAK,cAAcT,EAAE,EAAE,GAAK,EAEvE,IAAMW,EAAQ,IAAI,IAAIR,EAAW,IAAI,CAACH,EAAGY,IAAM,CAACZ,EAAE,GAAIY,CAAC,CAAC,CAAC,EAMzD,MAAO,CACL,OANa,CAAC,GAAGT,EAAW,IAAKH,GAAMA,EAAE,EAAE,CAAC,EAAE,KAAK,CAACa,EAAGC,IAAM,CAC7D,IAAMC,EAAOL,EAAOI,CAAC,EAAIJ,EAAOG,CAAC,EACjC,OAAOE,IAAS,EAAIA,GAAQJ,EAAM,IAAIE,CAAC,GAAK,IAAMF,EAAM,IAAIG,CAAC,GAAK,EACpE,CAAC,EAIC,OAAAJ,EACA,WAAaF,EAAQ,WAA0B,KAC/C,WAAYC,EAAK,UACnB,CACF,CAEA,SAASR,EAAKe,EAAcC,EAAqB,CAC/C,IAAMC,EAAO,OAAOF,GAAQ,EAAE,EAAE,QAAQ,OAAQ,GAAG,EAAE,KAAK,EAC1D,OAAOE,EAAK,OAASD,EAAMC,EAAK,MAAM,EAAGD,CAAG,EAAIC,CAClD","names":["init_questions","WIKI_RERANK_SEAT","DEFAULT_SHORTLIST","DEFAULT_EXCERPT_CHARS","rerankState","input","budget","c","clip","rerankQuestions","candidates","criteria","choice","noul","toRanking","answers","best","scores","order","i","a","b","diff","text","max","flat"]}

package/dist/chunk-AJONZWLT.js

@@ -0,0 +1,2 @@
+import{b as o}from"./chunk-FIOB3MEX.js";function g(e){if(!e)return null;let s=l(e);if(s!==void 0)return s;let c=e.match(/```(?:json)?\s*([\s\S]*?)```/);if(c){let n=l(c[1]);if(n!==void 0)return n}let f=[["{","}"],["[","]"]].map(([n,i])=>({open:n,close:i,at:e.indexOf(n)})).filter(n=>n.at>=0).sort((n,i)=>n.at-i.at);for(let{open:n,close:i}of f){let t=a(e,n,i);if(t!==null){let r=l(t);if(r!==void 0)return r}}return null}function l(e){try{return JSON.parse(e.trim())}catch{return}}function a(e,s,c){let f=e.indexOf(s);if(f<0)return null;let n=0,i=!1,t=!1;for(let r=f;r<e.length;r++){let u=e[r];if(t){t=!1;continue}if(u==="\\"){i&&(t=!0);continue}if(u==='"'){i=!i;continue}if(!i){if(u===s)n++;else if(u===c&&(n--,n===0))return e.slice(f,r+1)}}return null}var d=o(()=>{"use strict"});export{g as a,d as b};
+//# sourceMappingURL=chunk-AJONZWLT.js.map

package/dist/chunk-AJONZWLT.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/utils/extract-json.ts"],"sourcesContent":["// Pull a JSON value out of text a model wrote.\n//\n// Consolidated from four byte-identical copies (src/graph/classifier.ts,\n// src/agents/context-planner.ts, src/commands/graph.ts, and\n// src/daemon/board-dashboard.ts). Callers that return prose keep their own\n// bespoke variants — this is for call sites whose whole reply is supposed\n// to be a structured value.\n//\n// One behavioural difference from the copies it replaces: the balanced-brace\n// scan is string-aware. The old version counted every `{` and `}` including\n// ones inside string literals, so a perfectly valid `{\"glob\": \"src/**/{a,b}\"}`\n// truncated at the wrong byte and returned null.\n\n/** Parse `text` as JSON, tolerating code fences and surrounding prose.\n *  Returns `null` when nothing parseable is found — never throws. */\nexport function extractJson<T = any>(text: string): T | null {\n  if (!text) return null\n\n  const direct = tryParse<T>(text)\n  if (direct !== undefined) return direct\n\n  const fenced = text.match(/```(?:json)?\\s*([\\s\\S]*?)```/)\n  if (fenced) {\n    const parsed = tryParse<T>(fenced[1])\n    if (parsed !== undefined) return parsed\n  }\n\n  // Last-ditch: the first balanced object or array in the text. Whichever\n  // bracket appears first wins, so `here: [{\"a\":1}]` yields the array and\n  // not the object nested inside it.\n  const pairs = ([[\"{\", \"}\"], [\"[\", \"]\"]] as const)\n    .map(([open, close]) => ({ open, close, at: text.indexOf(open) }))\n    .filter((p) => p.at >= 0)\n    .sort((a, b) => a.at - b.at)\n\n  for (const { open, close } of pairs) {\n    const slice = balancedSlice(text, open, close)\n    if (slice !== null) {\n      const parsed = tryParse<T>(slice)\n      if (parsed !== undefined) return parsed\n    }\n  }\n\n  return null\n}\n\nfunction tryParse<T>(text: string): T | undefined {\n  try {\n    return JSON.parse(text.trim()) as T\n  } catch {\n    return undefined\n  }\n}\n\nfunction balancedSlice(text: string, open: string, close: string): string | null {\n  const start = text.indexOf(open)\n  if (start < 0) return null\n\n  let depth = 0\n  let inString = false\n  let escaped = false\n\n  for (let i = start; i < text.length; i++) {\n    const ch = text[i]\n\n    if (escaped) {\n      escaped = false\n      continue\n    }\n    if (ch === \"\\\\\") {\n      if (inString) escaped = true\n      continue\n    }\n    if (ch === '\"') {\n      inString = !inString\n      continue\n    }\n    if (inString) continue\n\n    if (ch === open) depth++\n    else if (ch === close) {\n      depth--\n      if (depth === 0) return text.slice(start, i + 1)\n    }\n  }\n\n  return null\n}\n"],"mappings":"wCAeO,SAASA,EAAqBC,EAAwB,CAC3D,GAAI,CAACA,EAAM,OAAO,KAElB,IAAMC,EAASC,EAAYF,CAAI,EAC/B,GAAIC,IAAW,OAAW,OAAOA,EAEjC,IAAME,EAASH,EAAK,MAAM,8BAA8B,EACxD,GAAIG,EAAQ,CACV,IAAMC,EAASF,EAAYC,EAAO,CAAC,CAAC,EACpC,GAAIC,IAAW,OAAW,OAAOA,EAMnC,IAAMC,EAAS,CAAC,CAAC,IAAK,GAAG,EAAG,CAAC,IAAK,GAAG,CAAC,EACnC,IAAI,CAAC,CAACC,EAAMC,CAAK,KAAO,CAAE,KAAAD,EAAM,MAAAC,EAAO,GAAIP,EAAK,QAAQM,CAAI,CAAE,EAAE,EAChE,OAAQE,GAAMA,EAAE,IAAM,CAAC,EACvB,KAAK,CAACC,EAAGC,IAAMD,EAAE,GAAKC,EAAE,EAAE,EAE7B,OAAW,CAAE,KAAAJ,EAAM,MAAAC,CAAM,IAAKF,EAAO,CACnC,IAAMM,EAAQC,EAAcZ,EAAMM,EAAMC,CAAK,EAC7C,GAAII,IAAU,KAAM,CAClB,IAAMP,EAASF,EAAYS,CAAK,EAChC,GAAIP,IAAW,OAAW,OAAOA,GAIrC,OAAO,IACT,CAEA,SAASF,EAAYF,EAA6B,CAChD,GAAI,CACF,OAAO,KAAK,MAAMA,EAAK,KAAK,CAAC,CAC/B,MAAE,CACA,MACF,CACF,CAEA,SAASY,EAAcZ,EAAcM,EAAcC,EAA8B,CAC/E,IAAMM,EAAQb,EAAK,QAAQM,CAAI,EAC/B,GAAIO,EAAQ,EAAG,OAAO,KAEtB,IAAIC,EAAQ,EACRC,EAAW,GACXC,EAAU,GAEd,QAASC,EAAIJ,EAAOI,EAAIjB,EAAK,OAAQiB,IAAK,CACxC,IAAMC,EAAKlB,EAAKiB,CAAC,EAEjB,GAAID,EAAS,CACXA,EAAU,GACV,SAEF,GAAIE,IAAO,KAAM,CACXH,IAAUC,EAAU,IACxB,SAEF,GAAIE,IAAO,IAAK,CACdH,EAAW,CAACA,EACZ,SAEF,GAAI,CAAAA,GAEJ,GAAIG,IAAOZ,EAAMQ,YACRI,IAAOX,IACdO,IACIA,IAAU,GAAG,OAAOd,EAAK,MAAMa,EAAOI,EAAI,CAAC,GAInD,OAAO,IACT,CAvFA,IAAAE,EAAAC,EAAA","names":["extractJson","text","direct","tryParse","fenced","parsed","pairs","open","close","p","a","b","slice","balancedSlice","start","depth","inString","escaped","i","ch","init_extract_json","__esmMin"]}

package/dist/chunk-BH3GTCZ6.js

@@ -0,0 +1,2 @@
+import{b as h}from"./chunk-FIOB3MEX.js";function p(i){let o=i.filter(l=>l.incumbent!==void 0),a=o.length;if(a===0)return{n:0,agree:0,rate:0,kappa:0};let t=0,n=new Map,r=new Map;for(let l of o)l.predicted===l.incumbent&&t++,n.set(l.predicted,(n.get(l.predicted)??0)+1),r.set(l.incumbent,(r.get(l.incumbent)??0)+1);let e=t/a,u=0;for(let[l,f]of n)u+=f/a*((r.get(l)??0)/a);let c=u<1?(e-u)/(1-u):1;return{n:a,agree:t,rate:e,kappa:c}}function s(i,o=10,a="width"){let t=i.filter(e=>e.truth!==void 0);if(t.length===0)return[];let n=[...t].sort((e,u)=>e.confidence-u.confidence);return(a==="mass"?v(n,o):y(n,o)).filter(e=>e.length>0).map(e=>{let u=e.map(c=>c.confidence);return{lo:Math.min(...u),hi:Math.max(...u),n:e.length,meanConfidence:m(u),accuracy:m(e.map(c=>c.predicted===c.truth?1:0))}})}function d(i,o=10,a="width"){let t=s(i,o,a),n=t.reduce((r,e)=>r+e.n,0);return n===0?0:t.reduce((r,e)=>r+e.n/n*Math.abs(e.accuracy-e.meanConfidence),0)}function g(i,o=10,a="width"){let t=s(i,o,a);return t.length===0?0:Math.max(...t.map(n=>Math.abs(n.accuracy-n.meanConfidence)))}function R(i){let o=i.filter(a=>a.truth!==void 0);return o.length===0?0:m(o.map(a=>{let t=0;for(let[n,r]of Object.entries(a.probabilities)){let e=n===a.truth?1:0;t+=(r-e)**2}return t}))}function b(i,o=1e-12){let a=i.filter(t=>t.truth!==void 0);return a.length===0?0:m(a.map(t=>{let n=t.probabilities[t.truth]??0;return-Math.log(Math.max(n,o))}))}function M(i,o=20){let a=i.filter(r=>r.truth!==void 0),t=a.length;if(t===0)return[];let n=[];for(let r=0;r<=o;r++){let e=r/o,u=a.filter(c=>c.confidence>=e);n.push({threshold:e,coverage:u.length/t,accuracy:u.length>0?m(u.map(c=>c.predicted===c.truth?1:0)):0,n:u.length})}return n}function x(i,o){let a=Math.max(o,1e-6),t={},n=0;for(let[r,e]of Object.entries(i)){let u=Math.pow(Math.max(e,0),1/a);t[r]=u,n+=u}if(n<=0){let r=Object.keys(i);for(let e of r)t[e]=1/r.length;return t}for(let r of Object.keys(t))t[r]=t[r]/n;return t}function w(i){let o=i.filter(e=>e.truth!==void 0),a=b(o);if(o.length===0)return{temperature:1,nll:0,nllBefore:0};let t=e=>m(o.map(u=>{let c=x(u.probabilities,e);return-Math.log(Math.max(c[u.truth]??0,1e-12))})),n=1,r=t(1);for(let e=.1;e<=10.0001;e+=.1){let u=t(e);u<r&&(r=u,n=e)}for(let e=Math.max(.01,n-.1);e<=n+.1;e+=.01){let u=t(e);u<r&&(r=u,n=e)}return{temperature:Number(n.toFixed(3)),nll:r,nllBefore:a}}function G(i,o,a=1e3,t=.05,n=Math.random){if(i.length===0)return{lo:0,hi:0};let r=[];for(let e=0;e<a;e++){let u=[];for(let c=0;c<i.length;c++)u.push(i[Math.floor(n()*i.length)]);r.push(o(u))}return r.sort((e,u)=>e-u),{lo:r[Math.floor(t/2*r.length)],hi:r[Math.min(r.length-1,Math.floor((1-t/2)*r.length))]}}function C(i,o={}){let a=o.minN??100,t=o.bins??10,n=i.filter(e=>e.truth!==void 0),r={insufficient:n.length<a,n:i.length,nLabeled:n.length,minN:a,agreement:p(i)};return r.insufficient?r:{...r,accuracy:m(n.map(e=>e.predicted===e.truth?1:0)),ece:d(n,t,"width"),eceEqualMass:d(n,t,"mass"),eceInterval:G(n,e=>d(e,t,"width"),o.bootstrap??1e3,.05,o.rng),mce:g(n,t,"width"),brier:R(n),logLoss:b(n),bins:s(n,t,"width"),binsEqualMass:s(n,t,"mass"),coverage:M(n),temperature:w(n)}}function y(i,o){let a=Array.from({length:o},()=>[]);for(let t of i){let n=Math.min(o-1,Math.max(0,Math.floor(t.confidence*o)));a[n].push(t)}return a}function v(i,o){let a=Array.from({length:o},()=>[]);return i.forEach((t,n)=>{a[Math.min(o-1,Math.floor(n*o/i.length))].push(t)}),a}function m(i){return i.length===0?0:i.reduce((o,a)=>o+a,0)/i.length}var B=h(()=>{});export{p as a,s as b,d as c,g as d,R as e,b as f,M as g,x as h,w as i,G as j,C as k,B as l};
+//# sourceMappingURL=chunk-BH3GTCZ6.js.map

package/dist/chunk-BH3GTCZ6.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/decisions/calibration.ts"],"sourcesContent":["import type { GradedRow } from \"./store\"\n\n// Does the confidence mean anything?\n//\n// Every metric here is a pure function of GradedRow[]. Three choices are\n// deliberate and worth defending:\n//\n//   1. `minN` (default 100). Below it the report refuses to print an ECE.\n//      An expected calibration error computed on thirty rows is noise, and\n//      a noisy number on a dashboard gets acted on exactly as if it were a\n//      real one.\n//\n//   2. `coverageCurve` is the headline, not ECE. The question a call site\n//      actually asks is \"what threshold do I set for autoRunThreshold\" —\n//      and that is answered by accuracy-at-threshold against the fraction\n//      of traffic retained, not by a single aggregate score.\n//\n//   3. Equal-mass bins are reported alongside equal-width. Equal-width ECE\n//      is dominated by whichever bin holds most of the mass, which for an\n//      overconfident model is the top bin — precisely the regime we are\n//      trying to detect.\n\nexport interface AgreementResult {\n  n: number\n  agree: number\n  rate: number\n  /** Cohen's kappa: agreement corrected for what chance alone would give.\n   *  Two systems that both answer \"no\" 95% of the time agree 90% of the\n   *  time while sharing no information, and kappa says so. */\n  kappa: number\n}\n\nexport function agreement(rows: GradedRow[]): AgreementResult {\n  const usable = rows.filter((r) => r.incumbent !== undefined)\n  const n = usable.length\n  if (n === 0) return { n: 0, agree: 0, rate: 0, kappa: 0 }\n\n  let agree = 0\n  const predicted = new Map<string, number>()\n  const incumbent = new Map<string, number>()\n  for (const row of usable) {\n    if (row.predicted === row.incumbent) agree++\n    predicted.set(row.predicted, (predicted.get(row.predicted) ?? 0) + 1)\n    incumbent.set(row.incumbent!, (incumbent.get(row.incumbent!) ?? 0) + 1)\n  }\n\n  const observed = agree / n\n  let expected = 0\n  for (const [label, count] of predicted) {\n    expected += (count / n) * ((incumbent.get(label) ?? 0) / n)\n  }\n  const kappa = expected < 1 ? (observed - expected) / (1 - expected) : 1\n\n  return { n, agree, rate: observed, kappa }\n}\n\nexport interface ReliabilityBin {\n  lo: number\n  hi: number\n  n: number\n  meanConfidence: number\n  accuracy: number\n}\n\nexport type BinScheme = \"width\" | \"mass\"\n\nexport function reliabilityBins(\n  rows: GradedRow[],\n  bins = 10,\n  scheme: BinScheme = \"width\",\n): ReliabilityBin[] {\n  const labeled = rows.filter((r) => r.truth !== undefined)\n  if (labeled.length === 0) return []\n\n  const sorted = [...labeled].sort((a, b) => a.confidence - b.confidence)\n  const groups: GradedRow[][] =\n    scheme === \"mass\" ? equalMassGroups(sorted, bins) : equalWidthGroups(sorted, bins)\n\n  return groups\n    .filter((g) => g.length > 0)\n    .map((group) => {\n      const confidences = group.map((r) => r.confidence)\n      return {\n        lo: Math.min(...confidences),\n        hi: Math.max(...confidences),\n        n: group.length,\n        meanConfidence: mean(confidences),\n        accuracy: mean(group.map((r) => (r.predicted === r.truth ? 1 : 0))),\n      }\n    })\n}\n\n/** Expected calibration error: the mass-weighted gap between how confident\n *  the model was and how often it was right. Zero is perfect. */\nexport function ece(rows: GradedRow[], bins = 10, scheme: BinScheme = \"width\"): number {\n  const grouped = reliabilityBins(rows, bins, scheme)\n  const total = grouped.reduce((sum, b) => sum + b.n, 0)\n  if (total === 0) return 0\n  return grouped.reduce(\n    (sum, b) => sum + (b.n / total) * Math.abs(b.accuracy - b.meanConfidence),\n    0,\n  )\n}\n\n/** The worst single bin. ECE can look fine while one region is badly wrong. */\nexport function mce(rows: GradedRow[], bins = 10, scheme: BinScheme = \"width\"): number {\n  const grouped = reliabilityBins(rows, bins, scheme)\n  if (grouped.length === 0) return 0\n  return Math.max(...grouped.map((b) => Math.abs(b.accuracy - b.meanConfidence)))\n}\n\n/** Multiclass Brier score. A proper scoring rule, so unlike ECE it cannot\n *  be gamed by a model that reports a constant confidence. Lower is better. */\nexport function brier(rows: GradedRow[]): number {\n  const labeled = rows.filter((r) => r.truth !== undefined)\n  if (labeled.length === 0) return 0\n  return mean(\n    labeled.map((row) => {\n      let sum = 0\n      for (const [label, p] of Object.entries(row.probabilities)) {\n        const actual = label === row.truth ? 1 : 0\n        sum += (p - actual) ** 2\n      }\n      return sum\n    }),\n  )\n}\n\nexport function logLoss(rows: GradedRow[], epsilon = 1e-12): number {\n  const labeled = rows.filter((r) => r.truth !== undefined)\n  if (labeled.length === 0) return 0\n  return mean(\n    labeled.map((row) => {\n      const p = row.probabilities[row.truth!] ?? 0\n      return -Math.log(Math.max(p, epsilon))\n    }),\n  )\n}\n\nexport interface CoveragePoint {\n  threshold: number\n  /** Fraction of rows at or above the threshold. */\n  coverage: number\n  /** Accuracy among those rows. */\n  accuracy: number\n  n: number\n}\n\n/** The curve a call site actually reads: \"if I only act above confidence X,\n *  how often am I right, and how much traffic do I still handle?\" */\nexport function coverageCurve(rows: GradedRow[], steps = 20): CoveragePoint[] {\n  const labeled = rows.filter((r) => r.truth !== undefined)\n  const total = labeled.length\n  if (total === 0) return []\n\n  const points: CoveragePoint[] = []\n  for (let i = 0; i <= steps; i++) {\n    const threshold = i / steps\n    const kept = labeled.filter((r) => r.confidence >= threshold)\n    points.push({\n      threshold,\n      coverage: kept.length / total,\n      accuracy: kept.length > 0 ? mean(kept.map((r) => (r.predicted === r.truth ? 1 : 0))) : 0,\n      n: kept.length,\n    })\n  }\n  return points\n}\n\n/** Sharpen or soften a distribution: p^(1/T), renormalized. T > 1 softens\n *  (the fix for overconfidence), T < 1 sharpens. */\nexport function applyTemperature(\n  probs: Record<string, number>,\n  temperature: number,\n): Record<string, number> {\n  const t = Math.max(temperature, 1e-6)\n  const scaled: Record<string, number> = {}\n  let sum = 0\n  for (const [label, p] of Object.entries(probs)) {\n    const v = Math.pow(Math.max(p, 0), 1 / t)\n    scaled[label] = v\n    sum += v\n  }\n  if (sum <= 0) {\n    const labels = Object.keys(probs)\n    for (const label of labels) scaled[label] = 1 / labels.length\n    return scaled\n  }\n  for (const label of Object.keys(scaled)) scaled[label] = scaled[label] / sum\n  return scaled\n}\n\nexport interface TemperatureFit {\n  temperature: number\n  nll: number\n  nllBefore: number\n}\n\n/** Fit the single parameter that turns a well-formed distribution into a\n *  calibrated one ON THIS TRAFFIC. This, not the backend, is where\n *  calibration comes from — which is why an uncalibrated backend plus a few\n *  hundred labeled rows is a defensible position and an uncalibrated\n *  backend alone is not.\n *\n *  Coarse grid then local refinement: one parameter, a smooth convex-ish\n *  objective, and a few hundred rows. Anything fancier is false precision. */\nexport function fitTemperature(rows: GradedRow[]): TemperatureFit {\n  const labeled = rows.filter((r) => r.truth !== undefined)\n  const nllBefore = logLoss(labeled)\n  if (labeled.length === 0) return { temperature: 1, nll: 0, nllBefore: 0 }\n\n  const nllAt = (t: number) =>\n    mean(\n      labeled.map((row) => {\n        const scaled = applyTemperature(row.probabilities, t)\n        return -Math.log(Math.max(scaled[row.truth!] ?? 0, 1e-12))\n      }),\n    )\n\n  let best = 1\n  let bestNll = nllAt(1)\n  for (let t = 0.1; t <= 10.0001; t += 0.1) {\n    const value = nllAt(t)\n    if (value < bestNll) {\n      bestNll = value\n      best = t\n    }\n  }\n  for (let t = Math.max(0.01, best - 0.1); t <= best + 0.1; t += 0.01) {\n    const value = nllAt(t)\n    if (value < bestNll) {\n      bestNll = value\n      best = t\n    }\n  }\n\n  return { temperature: Number(best.toFixed(3)), nll: bestNll, nllBefore }\n}\n\nexport interface Interval {\n  lo: number\n  hi: number\n}\n\n/** Percentile bootstrap. Without an interval, a difference between two\n *  backends' ECEs is not a result. */\nexport function bootstrapInterval(\n  rows: GradedRow[],\n  statistic: (sample: GradedRow[]) => number,\n  resamples = 1000,\n  alpha = 0.05,\n  rng: () => number = Math.random,\n): Interval {\n  if (rows.length === 0) return { lo: 0, hi: 0 }\n  const values: number[] = []\n  for (let i = 0; i < resamples; i++) {\n    const sample: GradedRow[] = []\n    for (let j = 0; j < rows.length; j++) {\n      sample.push(rows[Math.floor(rng() * rows.length)])\n    }\n    values.push(statistic(sample))\n  }\n  values.sort((a, b) => a - b)\n  return {\n    lo: values[Math.floor((alpha / 2) * values.length)],\n    hi: values[Math.min(values.length - 1, Math.floor((1 - alpha / 2) * values.length))],\n  }\n}\n\nexport interface CalibrationReport {\n  insufficient: boolean\n  n: number\n  nLabeled: number\n  minN: number\n  agreement: AgreementResult\n  accuracy?: number\n  ece?: number\n  eceEqualMass?: number\n  eceInterval?: Interval\n  mce?: number\n  brier?: number\n  logLoss?: number\n  bins?: ReliabilityBin[]\n  binsEqualMass?: ReliabilityBin[]\n  coverage?: CoveragePoint[]\n  temperature?: TemperatureFit\n}\n\nexport function calibrationReport(\n  rows: GradedRow[],\n  opts: { minN?: number; bins?: number; bootstrap?: number; rng?: () => number } = {},\n): CalibrationReport {\n  const minN = opts.minN ?? 100\n  const bins = opts.bins ?? 10\n  const labeled = rows.filter((r) => r.truth !== undefined)\n  const base: CalibrationReport = {\n    insufficient: labeled.length < minN,\n    n: rows.length,\n    nLabeled: labeled.length,\n    minN,\n    agreement: agreement(rows),\n  }\n\n  // Agreement needs no ground truth, so it is always reported. Everything\n  // that does need ground truth is withheld until there is enough of it.\n  if (base.insufficient) return base\n\n  return {\n    ...base,\n    accuracy: mean(labeled.map((r) => (r.predicted === r.truth ? 1 : 0))),\n    ece: ece(labeled, bins, \"width\"),\n    eceEqualMass: ece(labeled, bins, \"mass\"),\n    eceInterval: bootstrapInterval(\n      labeled,\n      (sample) => ece(sample, bins, \"width\"),\n      opts.bootstrap ?? 1000,\n      0.05,\n      opts.rng,\n    ),\n    mce: mce(labeled, bins, \"width\"),\n    brier: brier(labeled),\n    logLoss: logLoss(labeled),\n    bins: reliabilityBins(labeled, bins, \"width\"),\n    binsEqualMass: reliabilityBins(labeled, bins, \"mass\"),\n    coverage: coverageCurve(labeled),\n    temperature: fitTemperature(labeled),\n  }\n}\n\nfunction equalWidthGroups(sorted: GradedRow[], bins: number): GradedRow[][] {\n  const groups: GradedRow[][] = Array.from({ length: bins }, () => [])\n  for (const row of sorted) {\n    const index = Math.min(bins - 1, Math.max(0, Math.floor(row.confidence * bins)))\n    groups[index].push(row)\n  }\n  return groups\n}\n\nfunction equalMassGroups(sorted: GradedRow[], bins: number): GradedRow[][] {\n  const groups: GradedRow[][] = Array.from({ length: bins }, () => [])\n  sorted.forEach((row, i) => {\n    groups[Math.min(bins - 1, Math.floor((i * bins) / sorted.length))].push(row)\n  })\n  return groups\n}\n\nfunction mean(values: number[]): number {\n  if (values.length === 0) return 0\n  return values.reduce((a, b) => a + b, 0) / values.length\n}\n"],"mappings":"wCAgCO,SAASA,EAAUC,EAAoC,CAC5D,IAAMC,EAASD,EAAK,OAAQE,GAAMA,EAAE,YAAc,MAAS,EACrDC,EAAIF,EAAO,OACjB,GAAIE,IAAM,EAAG,MAAO,CAAE,EAAG,EAAG,MAAO,EAAG,KAAM,EAAG,MAAO,CAAE,EAExD,IAAIC,EAAQ,EACNC,EAAY,IAAI,IAChBC,EAAY,IAAI,IACtB,QAAWC,KAAON,EACZM,EAAI,YAAcA,EAAI,WAAWH,IACrCC,EAAU,IAAIE,EAAI,WAAYF,EAAU,IAAIE,EAAI,SAAS,GAAK,GAAK,CAAC,EACpED,EAAU,IAAIC,EAAI,WAAaD,EAAU,IAAIC,EAAI,SAAU,GAAK,GAAK,CAAC,EAGxE,IAAMC,EAAWJ,EAAQD,EACrBM,EAAW,EACf,OAAW,CAACC,EAAOC,CAAK,IAAKN,EAC3BI,GAAaE,EAAQR,IAAOG,EAAU,IAAII,CAAK,GAAK,GAAKP,GAE3D,IAAMS,EAAQH,EAAW,GAAKD,EAAWC,IAAa,EAAIA,GAAY,EAEtE,MAAO,CAAE,EAAAN,EAAG,MAAAC,EAAO,KAAMI,EAAU,MAAAI,CAAM,CAC3C,CAYO,SAASC,EACdb,EACAc,EAAO,GACPC,EAAoB,QACF,CAClB,IAAMC,EAAUhB,EAAK,OAAQE,GAAMA,EAAE,QAAU,MAAS,EACxD,GAAIc,EAAQ,SAAW,EAAG,MAAO,CAAC,EAElC,IAAMC,EAAS,CAAC,GAAGD,CAAO,EAAE,KAAK,CAACE,EAAGC,IAAMD,EAAE,WAAaC,EAAE,UAAU,EAItE,OAFEJ,IAAW,OAASK,EAAgBH,EAAQH,CAAI,EAAIO,EAAiBJ,EAAQH,CAAI,GAGhF,OAAQQ,GAAMA,EAAE,OAAS,CAAC,EAC1B,IAAKC,GAAU,CACd,IAAMC,EAAcD,EAAM,IAAKrB,GAAMA,EAAE,UAAU,EACjD,MAAO,CACL,GAAI,KAAK,IAAI,GAAGsB,CAAW,EAC3B,GAAI,KAAK,IAAI,GAAGA,CAAW,EAC3B,EAAGD,EAAM,OACT,eAAgBE,EAAKD,CAAW,EAChC,SAAUC,EAAKF,EAAM,IAAKrB,GAAOA,EAAE,YAAcA,EAAE,MAAQ,EAAI,CAAE,CAAC,CACpE,CACF,CAAC,CACL,CAIO,SAASwB,EAAI1B,EAAmBc,EAAO,GAAIC,EAAoB,QAAiB,CACrF,IAAMY,EAAUd,EAAgBb,EAAMc,EAAMC,CAAM,EAC5Ca,EAAQD,EAAQ,OAAO,CAACE,EAAKV,IAAMU,EAAMV,EAAE,EAAG,CAAC,EACrD,OAAIS,IAAU,EAAU,EACjBD,EAAQ,OACb,CAACE,EAAKV,IAAMU,EAAOV,EAAE,EAAIS,EAAS,KAAK,IAAIT,EAAE,SAAWA,EAAE,cAAc,EACxE,CACF,CACF,CAGO,SAASW,EAAI9B,EAAmBc,EAAO,GAAIC,EAAoB,QAAiB,CACrF,IAAMY,EAAUd,EAAgBb,EAAMc,EAAMC,CAAM,EAClD,OAAIY,EAAQ,SAAW,EAAU,EAC1B,KAAK,IAAI,GAAGA,EAAQ,IAAKR,GAAM,KAAK,IAAIA,EAAE,SAAWA,EAAE,cAAc,CAAC,CAAC,CAChF,CAIO,SAASY,EAAM/B,EAA2B,CAC/C,IAAMgB,EAAUhB,EAAK,OAAQE,GAAMA,EAAE,QAAU,MAAS,EACxD,OAAIc,EAAQ,SAAW,EAAU,EAC1BS,EACLT,EAAQ,IAAKT,GAAQ,CACnB,IAAIsB,EAAM,EACV,OAAW,CAACnB,EAAOsB,CAAC,IAAK,OAAO,QAAQzB,EAAI,aAAa,EAAG,CAC1D,IAAM0B,EAASvB,IAAUH,EAAI,MAAQ,EAAI,EACzCsB,IAAQG,EAAIC,IAAW,EAEzB,OAAOJ,CACT,CAAC,CACH,CACF,CAEO,SAASK,EAAQlC,EAAmBmC,EAAU,MAAe,CAClE,IAAMnB,EAAUhB,EAAK,OAAQE,GAAMA,EAAE,QAAU,MAAS,EACxD,OAAIc,EAAQ,SAAW,EAAU,EAC1BS,EACLT,EAAQ,IAAKT,GAAQ,CACnB,IAAMyB,EAAIzB,EAAI,cAAcA,EAAI,KAAM,GAAK,EAC3C,MAAO,CAAC,KAAK,IAAI,KAAK,IAAIyB,EAAGG,CAAO,CAAC,CACvC,CAAC,CACH,CACF,CAaO,SAASC,EAAcpC,EAAmBqC,EAAQ,GAAqB,CAC5E,IAAMrB,EAAUhB,EAAK,OAAQ,GAAM,EAAE,QAAU,MAAS,EAClD4B,EAAQZ,EAAQ,OACtB,GAAIY,IAAU,EAAG,MAAO,CAAC,EAEzB,IAAMU,EAA0B,CAAC,EACjC,QAASC,EAAI,EAAGA,GAAKF,EAAOE,IAAK,CAC/B,IAAMC,EAAYD,EAAIF,EAChBI,EAAOzB,EAAQ,OAAQd,GAAMA,EAAE,YAAcsC,CAAS,EAC5DF,EAAO,KAAK,CACV,UAAAE,EACA,SAAUC,EAAK,OAASb,EACxB,SAAUa,EAAK,OAAS,EAAIhB,EAAKgB,EAAK,IAAKvC,GAAOA,EAAE,YAAcA,EAAE,MAAQ,EAAI,CAAE,CAAC,EAAI,EACvF,EAAGuC,EAAK,MACV,CAAC,EAEH,OAAOH,CACT,CAIO,SAASI,EACdC,EACAC,EACwB,CACxB,IAAMC,EAAI,KAAK,IAAID,EAAa,IAAI,EAC9BE,EAAiC,CAAC,EACpCjB,EAAM,EACV,OAAW,CAACnB,EAAOsB,CAAC,IAAK,OAAO,QAAQW,CAAK,EAAG,CAC9C,IAAMI,EAAI,KAAK,IAAI,KAAK,IAAIf,EAAG,CAAC,EAAG,EAAIa,CAAC,EACxCC,EAAOpC,CAAK,EAAIqC,EAChBlB,GAAOkB,EAET,GAAIlB,GAAO,EAAG,CACZ,IAAMmB,EAAS,OAAO,KAAKL,CAAK,EAChC,QAAWjC,KAASsC,EAAQF,EAAOpC,CAAK,EAAI,EAAIsC,EAAO,OACvD,OAAOF,EAET,QAAWpC,KAAS,OAAO,KAAKoC,CAAM,EAAGA,EAAOpC,CAAK,EAAIoC,EAAOpC,CAAK,EAAImB,EACzE,OAAOiB,CACT,CAgBO,SAASG,EAAejD,EAAmC,CAChE,IAAMgB,EAAUhB,EAAK,OAAQE,GAAMA,EAAE,QAAU,MAAS,EAClDgD,EAAYhB,EAAQlB,CAAO,EACjC,GAAIA,EAAQ,SAAW,EAAG,MAAO,CAAE,YAAa,EAAG,IAAK,EAAG,UAAW,CAAE,EAExE,IAAMmC,EAASN,GACbpB,EACET,EAAQ,IAAKT,GAAQ,CACnB,IAAMuC,EAASJ,EAAiBnC,EAAI,cAAesC,CAAC,EACpD,MAAO,CAAC,KAAK,IAAI,KAAK,IAAIC,EAAOvC,EAAI,KAAM,GAAK,EAAG,KAAK,CAAC,CAC3D,CAAC,CACH,EAEE6C,EAAO,EACPC,EAAUF,EAAM,CAAC,EACrB,QAASN,EAAI,GAAKA,GAAK,QAASA,GAAK,GAAK,CACxC,IAAMS,EAAQH,EAAMN,CAAC,EACjBS,EAAQD,IACVA,EAAUC,EACVF,EAAOP,GAGX,QAASA,EAAI,KAAK,IAAI,IAAMO,EAAO,EAAG,EAAGP,GAAKO,EAAO,GAAKP,GAAK,IAAM,CACnE,IAAMS,EAAQH,EAAMN,CAAC,EACjBS,EAAQD,IACVA,EAAUC,EACVF,EAAOP,GAIX,MAAO,CAAE,YAAa,OAAOO,EAAK,QAAQ,CAAC,CAAC,EAAG,IAAKC,EAAS,UAAAH,CAAU,CACzE,CASO,SAASK,EACdvD,EACAwD,EACAC,EAAY,IACZC,EAAQ,IACRC,EAAoB,KAAK,OACf,CACV,GAAI3D,EAAK,SAAW,EAAG,MAAO,CAAE,GAAI,EAAG,GAAI,CAAE,EAC7C,IAAM4D,EAAmB,CAAC,EAC1B,QAASrB,EAAI,EAAGA,EAAIkB,EAAWlB,IAAK,CAClC,IAAMsB,EAAsB,CAAC,EAC7B,QAASC,EAAI,EAAGA,EAAI9D,EAAK,OAAQ8D,IAC/BD,EAAO,KAAK7D,EAAK,KAAK,MAAM2D,EAAI,EAAI3D,EAAK,MAAM,CAAC,CAAC,EAEnD4D,EAAO,KAAKJ,EAAUK,CAAM,CAAC,EAE/B,OAAAD,EAAO,KAAK,CAAC1C,EAAGC,IAAMD,EAAIC,CAAC,EACpB,CACL,GAAIyC,EAAO,KAAK,MAAOF,EAAQ,EAAKE,EAAO,MAAM,CAAC,EAClD,GAAIA,EAAO,KAAK,IAAIA,EAAO,OAAS,EAAG,KAAK,OAAO,EAAIF,EAAQ,GAAKE,EAAO,MAAM,CAAC,CAAC,CACrF,CACF,CAqBO,SAASG,EACd/D,EACAgE,EAAiF,CAAC,EAC/D,CACnB,IAAMC,EAAOD,EAAK,MAAQ,IACpBlD,EAAOkD,EAAK,MAAQ,GACpBhD,EAAUhB,EAAK,OAAQE,GAAMA,EAAE,QAAU,MAAS,EAClDgE,EAA0B,CAC9B,aAAclD,EAAQ,OAASiD,EAC/B,EAAGjE,EAAK,OACR,SAAUgB,EAAQ,OAClB,KAAAiD,EACA,UAAWlE,EAAUC,CAAI,CAC3B,EAIA,OAAIkE,EAAK,aAAqBA,EAEvB,CACL,GAAGA,EACH,SAAUzC,EAAKT,EAAQ,IAAKd,GAAOA,EAAE,YAAcA,EAAE,MAAQ,EAAI,CAAE,CAAC,EACpE,IAAKwB,EAAIV,EAASF,EAAM,OAAO,EAC/B,aAAcY,EAAIV,EAASF,EAAM,MAAM,EACvC,YAAayC,EACXvC,EACC6C,GAAWnC,EAAImC,EAAQ/C,EAAM,OAAO,EACrCkD,EAAK,WAAa,IAClB,IACAA,EAAK,GACP,EACA,IAAKlC,EAAId,EAASF,EAAM,OAAO,EAC/B,MAAOiB,EAAMf,CAAO,EACpB,QAASkB,EAAQlB,CAAO,EACxB,KAAMH,EAAgBG,EAASF,EAAM,OAAO,EAC5C,cAAeD,EAAgBG,EAASF,EAAM,MAAM,EACpD,SAAUsB,EAAcpB,CAAO,EAC/B,YAAaiC,EAAejC,CAAO,CACrC,CACF,CAEA,SAASK,EAAiBJ,EAAqBH,EAA6B,CAC1E,IAAMqD,EAAwB,MAAM,KAAK,CAAE,OAAQrD,CAAK,EAAG,IAAM,CAAC,CAAC,EACnE,QAAWP,KAAOU,EAAQ,CACxB,IAAMmD,EAAQ,KAAK,IAAItD,EAAO,EAAG,KAAK,IAAI,EAAG,KAAK,MAAMP,EAAI,WAAaO,CAAI,CAAC,CAAC,EAC/EqD,EAAOC,CAAK,EAAE,KAAK7D,CAAG,EAExB,OAAO4D,CACT,CAEA,SAAS/C,EAAgBH,EAAqBH,EAA6B,CACzE,IAAMqD,EAAwB,MAAM,KAAK,CAAE,OAAQrD,CAAK,EAAG,IAAM,CAAC,CAAC,EACnE,OAAAG,EAAO,QAAQ,CAACV,EAAKgC,IAAM,CACzB4B,EAAO,KAAK,IAAIrD,EAAO,EAAG,KAAK,MAAOyB,EAAIzB,EAAQG,EAAO,MAAM,CAAC,CAAC,EAAE,KAAKV,CAAG,CAC7E,CAAC,EACM4D,CACT,CAEA,SAAS1C,EAAKmC,EAA0B,CACtC,OAAIA,EAAO,SAAW,EAAU,EACzBA,EAAO,OAAO,CAAC1C,EAAGC,IAAMD,EAAIC,EAAG,CAAC,EAAIyC,EAAO,MACpD,CA7VA,IAAAS,EAAAC,EAAA","names":["agreement","rows","usable","r","n","agree","predicted","incumbent","row","observed","expected","label","count","kappa","reliabilityBins","bins","scheme","labeled","sorted","a","b","equalMassGroups","equalWidthGroups","g","group","confidences","mean","ece","grouped","total","sum","mce","brier","p","actual","logLoss","epsilon","coverageCurve","steps","points","i","threshold","kept","applyTemperature","probs","temperature","t","scaled","v","labels","fitTemperature","nllBefore","nllAt","best","bestNll","value","bootstrapInterval","statistic","resamples","alpha","rng","values","sample","j","calibrationReport","opts","minN","base","groups","index","init_calibration","__esmMin"]}