Foaster/Werewolf_benchmark
7
1<!DOCTYPE html>2<html lang="en">3<head>4 <meta charset="UTF-8" />5 <meta name="viewport" content="width=device-width, initial-scale=1" />6 <title>Probing LLM Social Intelligence via Werewolf – First Results</title>7 <style>8 :root{9 --text:#0f172a; --muted:#475569; --bg:#ffffff; --card:#f8fafc; --accent:#0ea5e9;10 --border:#e2e8f0;11 }12 *{box-sizing:border-box}13 body{margin:0; font-family: ui-sans-serif, system-ui, -apple-system, Segoe UI, Roboto, "Helvetica Neue", Arial, "Noto Sans", "Apple Color Emoji", "Segoe UI Emoji"; color:var(--text); background:var(--bg); line-height:1.6; text-align:justify}14 html{scroll-behavior:smooth}15 @media (prefers-reduced-motion: reduce){ html{ scroll-behavior:auto } }16 header{padding:48px 20px; background:linear-gradient(180deg,#f0f9ff,transparent)}17 .container{max-width:980px; margin:0 auto; padding:0 20px}18 h1{font-size:clamp(28px,3.5vw,42px); line-height:1.1; margin:0 0 12px}19 h2{font-size:clamp(26px,3vw,34px); line-height:1.18; margin:52px 0 12px; padding-bottom:6px; border-bottom:1px solid #e5e7eb}20 h3{font-size:clamp(18px,2.1vw,20px); margin:24px 0 8px; color:#111827}21 h4{font-size:16px; margin:20px 0 6px}22 p{margin:12px 0}23 .muted{color:var(--muted)}24 .card{background:var(--card); border:1px solid var(--border); border-radius:14px; padding:18px}25 nav.toc{border:1px solid var(--border); border-radius:12px; padding:14px; background:#fff}26 nav.toc a{color:var(--accent); text-decoration:none}27 nav.toc li{margin:6px 0}28 29 /* Placeholders for future graphs or quote images */30 .placeholder{border:2px dashed var(--border); border-radius:12px; padding:14px; background:#fff; margin:16px 0}31 .placeholder small{display:block; color:var(--muted)}32 .placeholder.label{font-weight:600}33 34 figure{margin:18px 0}35 figcaption{font-size:14px; color:var(--muted)}36 /* Légendes des figures alignées sur la colonne de texte */37 section figcaption{ max-width:640px; margin-left:auto; margin-right:auto }38 /* Notes/meta sous figures alignées sur la colonne de texte */39 section figure .meta{ max-width:640px; margin-left:auto; margin-right:auto }40 blockquote{border-left:4px solid var(--border); margin:16px 0; padding:8px 14px; color:#0b1220; background:#f9fbff; border-radius:8px}41 .aside{font-size:14px; color:var(--muted)}42 footer{margin:60px 0 40px; color:var(--muted)}43 /* Buttons — neutral pill style (white default, light gray when active) */44 .btn{border:1px solid var(--border); background:#fff; color:#0f172a; font-weight:700; padding:10px 14px; border-radius:999px; display:inline-flex; align-items:center; gap:8px; box-shadow:0 1px 2px rgba(15,23,42,.06); transition:box-shadow .2s ease, transform .05s ease, background-color .15s ease, color .15s ease}45 .btn:hover{box-shadow:0 2px 8px rgba(15,23,42,.12); background:#f9fafb}46 .btn:active{transform:translateY(1px)}47 .btn[aria-expanded="true"]{ background:#f3f4f6; color:#0f172a; border-color:#e5e7eb }48 /* Side TOC (desktop only) — professional look */49 #page-toc{ display:none }50 @media (min-width:1200px){51 #page-toc{ display:block; position:fixed; top:140px; right:max(16px, calc((100vw - 980px)/2 - 260px)); width:240px; z-index:5 }52 #page-toc .box{ background:#fff; border:1px solid #e6e8ee; border-radius:10px; padding:14px; box-shadow:0 6px 20px rgba(15,23,42,.05) }53 #page-toc .toc-title{ font-weight:800; font-size:11px; letter-spacing:.5px; text-transform:uppercase; color:#475569; margin:0 0 10px 0 }54 #page-toc ul{ list-style:none; padding:0; margin:0; display:flex; flex-direction:column; gap:2px }55 #page-toc li{ margin:0 }56 #page-toc a{ display:block; padding:6px 10px; border-radius:6px; color:#0b1220; text-decoration:none; border:1px solid transparent; font-weight:600; font-size:13px }57 #page-toc a:hover{ background:#f8fafc; border-color:#eef2f7 }58 #page-toc a[aria-current="true"]{ background:#eef2f7; border-color:#e5e7eb; font-weight:800 }59 }60 /* Case‑study panel: fixed height + scroll */61 .cs-panel{ max-width:640px; margin:0 auto; background:#f5f7fa; border:1px solid #e5e7eb; border-radius:12px; padding:12px 14px; max-height:520px; overflow-y:auto; box-shadow:0 8px 24px rgba(15,23,42,.06) }62 .cs-panel::-webkit-scrollbar{ width:10px }63 .cs-panel::-webkit-scrollbar-thumb{ background:#cbd5e1; border-radius:8px }64 .cs-panel::-webkit-scrollbar-track{ background:transparent }65 /* Center titles/subtitles for specific graph cards */66 #manipulation-success-card h3, #manipulation-success-card .aside{ text-align:center }67 #autosabotage > h3{ text-align:center }68 #autosabotage .card > h3, #autosabotage .card .aside{ text-align:center }69 #coordination > h3{ text-align:center }70 #coordination .card > h3, #coordination .card .aside{ text-align:center }71 /* Narrow summary width for the manipulation case study title */72 #manipulation-case-study > summary{ max-width:640px; margin-left:auto; margin-right:auto }73 #manipulation-case-study .ex-inner{ max-width:640px; margin-left:auto; margin-right:auto }74 #autosab-case-study > summary{ max-width:640px; margin-left:auto; margin-right:auto }75 #autosab-case-study .ex-inner{ max-width:640px; margin-left:auto; margin-right:auto }76 #resistance-case-study > summary{ max-width:640px; margin-left:auto; margin-right:auto }77 #resistance-case-study .ex-inner{ max-width:640px; margin-left:auto; margin-right:auto }78 #coordination .card > p:not(.aside){ max-width:640px; margin-left:auto; margin-right:auto }79 /* Width-only adjustment for the explanation under the manipulation card */80 #manipulation-success-card p{ max-width:640px; margin-left:auto; margin-right:auto }81 /* Metric note: uniform dimensions under graph cards */82 .metric-note{ max-width:640px; margin:6px auto 8px auto; font-size:14px; line-height:1.5; color:#475569 }83 /* Expérience: colonnes centrées pour texte & titres */84 section > p,85 section > ul,86 section > ol{ max-width:640px; margin-left:auto; margin-right:auto; font-size:0.95em }87 section > blockquote{ max-width:640px; margin-left:auto; margin-right:auto }88 section > h2,89 section > h3{ max-width:640px; margin-left:auto; margin-right:auto }90 /* Centrer figures et cartes; graphes légèrement plus larges que le texte */91 section > figure{ max-width:760px; margin-left:auto !important; margin-right:auto !important }92 section .card{ max-width:760px; margin-left:auto !important; margin-right:auto !important }93 /* Exception: leaderboard en pleine largeur container */94 #elo-leaderboard .card{ max-width:980px; margin-left:auto !important; margin-right:auto !important }95 /* Aligner la largeur des quotes avec la colonne de texte */96 .quote-stack, .qstack{ max-width:640px; margin-left:auto !important; margin-right:auto !important; display:block }97 98 /* ——— Example block ——— */99 100 /* EXAMPLE: un seul encadré + quotes */101 details.example{max-width:760px;margin:16px auto}102 details.example>summary{cursor:pointer;font-weight:600;color:#0f172a;padding:6px 0;font-style:italic;font-size:0.95em;display:flex;align-items:center;gap:8px}103 details.example>summary::-webkit-details-marker{display:none}104 details.example>summary .ex-caret{display:inline-block;transition:transform .2s ease}105 /* Helper label before example titles (not inside the rotating caret) */106 details.example>summary em::before{ content:" Show example: "; font-weight:700; }107 details.example[open]>summary .ex-caret{transform:rotate(90deg)}108 109 /* Case study toggles styled like example summaries */110 .ex-btn{ border:1px solid #e5e7eb; background:#fff; color:#0f172a; font-weight:600; font-style:italic; font-size:.95em; padding:10px 14px; border-radius:10px; display:flex; align-items:center; gap:8px; margin:12px auto; box-shadow:0 1px 2px rgba(15,23,42,.06) }111 .ex-btn:hover{ box-shadow:0 2px 8px rgba(15,23,42,.12); background:#f9fafb }112 .ex-btn .ex-caret{ transition:transform .2s ease }113 .ex-btn[aria-expanded="true"] .ex-caret{ transform:rotate(90deg) }114 115 details.example .ex-inner{116 border:1px solid #e5e7eb;117 border-radius:12px;118 padding:14px 16px;119 background:#fcfcfe;120 }121 details.example .ex-kicker{margin-top:10px;font-weight:600;color:#111827}122 details.example p{margin:6px 0}123 124 /* le SEUL box interne = blockquote */125 details.example blockquote{126 margin:8px 0 12px;127 padding:10px 12px;128 background:#f8fafc;129 border-left:4px solid var(--accent,#7c3aed);130 border-radius:8px;131 font-style:italic;132 font-size:.95em;133 }134 135 /* Model profiles wolf quote styles (same as case study) */136 details.example .qstack{137 display:flex;138 flex-direction:column;139 gap:10px;140 margin:10px 0;141 max-width:640px;142 margin-left:auto;143 margin-right:auto;144 }145 details.example .qcard{146 border:1px solid var(--border);147 border-radius:12px;148 background:#fff;149 box-shadow:0 2px 8px rgba(0,0,0,.04);150 }151 details.example .qhd{152 padding:10px 12px;153 border-bottom:1px solid #eee;154 background:#eef2ff;155 border-radius:12px 12px 0 0;156 color:#3730a3;157 display:flex;158 gap:8px;159 align-items:center;160 }161 details.example .qbd{162 padding:14px 16px;163 color:#0b1220;164 }165 details.example .qbody{166 font-size:15px;167 line-height:1.55;168 }169 details.example .sc{170 margin:10px 0 4px 2px;171 font-variant:small-caps;172 letter-spacing:.5px;173 color:#475569;174 }175 details.example .meta{176 font-style:italic;177 color:#6b7280;178 margin:6px 2px;179 }180 details.example[data-tone="warning"] blockquote{--accent:#f59e0b}181 details.example[data-tone="danger"] blockquote{--accent:#ef4444}182 183 /* Séparateur de section léger */184 .section-sep{185 max-width:760px;186 margin:32px auto;187 border-top:1px dashed #e5e7eb;188 height:0;189 }190 191 192 </style>193</head>194<body>195 <header>196 <div class="container">197 <h1>Probing LLM Social Intelligence via Werewolf</h1>198 <p class="muted">Round-robin results, role-conditioned Elo, and early behavior signals</p>199 </div>200 </header>201 202 <main class="container" id="top">203 <aside id="page-toc" aria-label="Sommaire">204 <div class="box">205 <div class="toc-title">On this page</div>206 <ul>207 <li><a href="#intro">Introduction</a></li>208 <li><a href="#elo">ELO & Head‑to‑Head</a></li>209 <li><a href="#metrics">Manipulation</a></li>210 <li><a href="#writing-styles">Writing styles</a></li>211 <li><a href="#emergent">Emerging behaviors</a></li>212 <li><a href="#moments">Strategic Plays</a></li>213 <li><a href="#conclusion">Conclusion</a></li>214 </ul>215 </div>216 </aside>217 <!-- Main sections: Introduction, ELO & Head-to-Head, Manipulation, Writing styles, Emerging behaviors, Strategic Plays -->218 <style>219 /* Scoped styles for Elo table */220 .elo-widget { background:#fff; border:none; border-radius:12px; padding:14px; }221 .elo-head { display:flex; justify-content:space-between; align-items:center; margin-bottom:10px; }222 .elo-title { font-size:16px; font-weight:700; color:#111827; }223 .elo-table-wrap { border:none; border-radius:10px; overflow-x:auto; overflow-y:hidden; -webkit-overflow-scrolling:touch }224 .elo-table { width:100%; min-width:760px; border-collapse:collapse; }225 .elo-table thead { background:#f9fafb; border-bottom:none; }226 .elo-table th { padding:8px 10px; text-align:left; font-size:11px; letter-spacing:.3px; text-transform:uppercase; color:#6b7280; }227 .elo-table th.elo-center { text-align:center; }228 .elo-table td { padding:10px; font-size:13px; color:#111827; border-top:none; }229 .elo-center { text-align:center; }230 .elo-rank { width:56px; font-weight:700; color:#6b7280; }231 .elo-r1 { color:#fbbf24; font-size:18px; } .elo-r2 { color:#9ca3af; font-size:16px; } .elo-r3 { color:#cd7f32; font-size:16px; }232 .elo-model { font-weight:700; } .elo-org { font-size:11px; color:#6b7280; }233 .elo-score { font-weight:800; }234 .elo-win-hi { color:#10b981; font-weight:700; } .elo-win-mid { color:#3b82f6; font-weight:700; } .elo-win-lo { color:#ef4444; font-weight:700; }235 .org-logo { height:16px; width:auto; margin-right:6px; vertical-align:middle }236 </style>237 <section id="elo-leaderboard">238 <div class="elo-widget card">239 <div class="elo-head">240 <div class="elo-title">🏆 Elo Leaderboard</div>241 </div>242 <div class="elo-table-wrap">243 <table class="elo-table">244 <thead>245 <tr>246 <th>Rank</th>247 <th>Model</th>248 <th class="elo-center">ELO</th>249 <th class="elo-center">ELO-W</th>250 <th class="elo-center">ELO-V</th>251 <th class="elo-center">Win rate</th>252 <th class="elo-center">Matches</th>253 </tr>254 </thead>255 <tbody id="elo-tbody"></tbody>256 </table>257 </div>258 <p class="aside" style="margin-top:8px">ELO‑W = Elo as <strong>wolf</strong>; ELO‑V = Elo as <strong>villager</strong>.</p>259 </div>260 </section>261 <script>262 (function(){263 const $tbody = document.getElementById('elo-tbody');264 if(!$tbody) return;265 const DATA = [266 { model:"gpt-5", org:"OpenAI", Elo:1529, EloW:1518, EloV:1540, matches:90, win:95.6 },267 { model:"gemini-2.5-pro", org:"Google", Elo:1243, EloW:1125, EloV:1361, matches:90, win:58.9 },268 { model:"grok-4-fast-reasoning", org:"XAI", Elo:1231, EloW:1242, EloV:1219, matches:90, win:53.3 },269 { model:"gemini-2.5-flash", org:"Google", Elo:1222, EloW:1104, EloV:1341, matches:90, win: 55.6 },270 { model:"kimi-k2-0905-preview", org:"Moonshot AI", Elo:1189, EloW:1182, EloV:1195, matches:90, win:46.7 },271 { model:"grok-4", org:"XAI", Elo:1178, EloW:1130, EloV:1227, matches:90, win:51.1 },272 { model:"Qwen3-235B-A22B-Instruct-2507", org:"Alibaba", Elo:1150, EloW:1098, EloV:1203, matches:90, win:44.4 },273 { model:"Kimi-K2-Instruct", org:"Moonshot AI", Elo:1133, EloW:1180, EloV:1087, matches:90, win:38.9 },274 { model:"gpt-5-mini", org:"OpenAI", Elo:1120, EloW:1049, EloV:1190, matches:90, win:38.9 },275 { model:"openai/gpt-oss-120b", org:"OpenAI", Elo:971, EloW:852, EloV:1089, matches:90, win:16.7 }276 ];277 // Organisation -> logo path (relative)278 const ORG_LOGOS = {279 "OpenAI": "assets/logos/openai-svgrepo-com.svg",280 "Google": "assets/logos/google-icon-logo-svgrepo-com.svg",281 "Alibaba": "assets/logos/Qwen_logo.svg",282 "Moonshot AI": "assets/logos/moonshot.png",283 "XAI": "assets/logos/XAI_Logo.svg.png"284 };285 // Expose global mapping label->logo for charts286 window.MODEL_LOGOS_BY_LABEL = {287 'gpt-5': ORG_LOGOS['OpenAI'], 'GPT-5': ORG_LOGOS['OpenAI'],288 'gpt-5-mini': ORG_LOGOS['OpenAI'], 'GPT-5-mini': ORG_LOGOS['OpenAI'],289 'gemini-2.5-pro': ORG_LOGOS['Google'], 'Gemini 2.5 Pro': ORG_LOGOS['Google'],290 'gemini-2.5-flash': ORG_LOGOS['Google'], 'Gemini 2.5 Flash': ORG_LOGOS['Google'],291 'qwen3-235b-instruct': ORG_LOGOS['Alibaba'], 'Qwen3-235B-Instruct': ORG_LOGOS['Alibaba'], 'Qwen3-235B-A22B-Instruct-2507': ORG_LOGOS['Alibaba'],292 // Aliases sans préfixe pour préserver le logo après renommage293 'Qwen3-235B-A22B-Instruct-2507': ORG_LOGOS['Alibaba'],294 'qwen3-235b-a22b-instruct-2507': ORG_LOGOS['Alibaba'],295 'kimi-k2-instruct': ORG_LOGOS['Moonshot AI'], 'Kimi-K2-Instruct': ORG_LOGOS['Moonshot AI'], 'moonshotai/Kimi-K2-Instruct': ORG_LOGOS['Moonshot AI'],296 // Alias sans préfixe (déjà géré ci-dessus), conservé pour compat297 'kimi-k2-0905-preview': ORG_LOGOS['Moonshot AI'], 'Kimi-K2-0905-Preview': ORG_LOGOS['Moonshot AI'], 'Kimi‑K2‑0905‑preview': ORG_LOGOS['Moonshot AI'],298 'gpt-oss-120b': ORG_LOGOS['OpenAI'], 'GPT-OSS-120B': ORG_LOGOS['OpenAI'], 'openai/gpt-oss-120b': ORG_LOGOS['OpenAI'],299 'grok-4-fast-reasoning': ORG_LOGOS['XAI'], 'Grok-4-Fast-Reasoning': ORG_LOGOS['XAI'],300 'grok-4': ORG_LOGOS['XAI'], 'Grok-4': ORG_LOGOS['XAI']301 };302 // Global plugin to draw logos near tick labels303 (function(){304 if(window.logoTicksPlugin) return;305 const cache = {};306 function getLogo(label){ return window.MODEL_LOGOS_BY_LABEL && (window.MODEL_LOGOS_BY_LABEL[label] || window.MODEL_LOGOS_BY_LABEL[label && label.toLowerCase ? label.toLowerCase() : label]); }307 window.logoTicksPlugin = {308 id:'logoTicks',309 afterDraw(chart){310 if(!window.MODEL_LOGOS_BY_LABEL) return;311 const area = chart.chartArea; if(!area) return;312 const ctx = chart.ctx; const idxAxis = chart.options.indexAxis || 'x';313 const axis = chart.scales[idxAxis]; if(!axis) return;314 const isHorizontal = idxAxis === 'y';315 const labels = axis.ticks.map(t=>t.label);316 const size = 14;317 ctx.save();318 labels.forEach((label,i)=>{319 const src = getLogo(label); if(!src) return;320 let img = cache[src]; if(!img){ img = new Image(); img.src = src; cache[src] = img; img.onload = ()=> chart.draw(); }321 const pixel = axis.getPixelForTick(i);322 if(isHorizontal){323 const pad = (axis.options && axis.options.ticks && axis.options.ticks.padding) || 6;324 const x = axis.left - size - Math.max(8, pad + 2); // logo à droite du texte, hors zone du plot325 const y = pixel - size/2;326 if(img.complete) ctx.drawImage(img, x, y, size, size);327 } else {328 const dsCount = (chart.data && chart.data.datasets && chart.data.datasets.length) ? chart.data.datasets.length : 1;329 if(dsCount > 1){330 // Cas Day 1 / Day 2: position mobile vs desktop331 const isMobile = (chart.canvas && chart.canvas.clientWidth && chart.canvas.clientWidth < 720);332 const tickFontSize = (axis.options && axis.options.ticks && axis.options.ticks.font && axis.options.ticks.font.size) || 16;333 const rot = (axis.options && axis.options.ticks && axis.options.ticks.maxRotation) || 0;334 const rotatePad = rot > 0 ? Math.round(tickFontSize * (rot/45) * 0.9) : 0;335 // Sur mobile, pousser le logo bien en dessous des labels inclinés336 const extra = (isMobile ? 20 : 8) + rotatePad + (isMobile ? 18 : 0);337 const x = pixel - size/2; const y = area.bottom + tickFontSize * 1.1 + extra;338 if(img.complete) ctx.drawImage(img, x, y, isMobile ? 16 : size, isMobile ? 16 : size);339 } else {340 // Cas un seul dataset: on distingue les graphes341 const canvasId = (chart.canvas && chart.canvas.id) || '';342 let y;343 if(canvasId === 'chart-auto-blog'){344 // Auto-sabotage → logo sous l'axe, texte en dessous345 y = area.bottom + 4;346 } else if(canvasId === 'chart-d1wolf-blog'){347 // Day 1 wolf eliminated → logo plus bas sous l'axe pour mieux dégager le texte348 y = area.bottom + 16;349 } else {350 // fallback: sous l'axe351 y = area.bottom + 4;352 }353 const x = pixel - size/2;354 if(img.complete) ctx.drawImage(img, x, y, size, size);355 }356 }357 });358 ctx.restore();359 }360 };361 })();362 function rankBadge(n){363 if (n===1) return '<span class="elo-r1">🥇</span>';364 if (n===2) return '<span class="elo-r2">🥈</span>';365 if (n===3) return '<span class="elo-r3">🥉</span>';366 return `#${n}`;367 }368 // Color gradient based on overall Elo (min → max)369 function hexToRgb(hex){ const m=/^#?([a-f\d]{2})([a-f\d]{2})([a-f\d]{2})$/i.exec(hex); return m?{r:parseInt(m[1],16),g:parseInt(m[2],16),b:parseInt(m[3],16)}:null; }370 function interp(c1,c2,t){ const a=hexToRgb(c1),b=hexToRgb(c2); const r=Math.round(a.r+(b.r-a.r)*t), g=Math.round(a.g+(b.g-a.g)*t), bb=Math.round(a.b+(b.b-a.b)*t); return 'rgb('+r+', '+g+', '+bb+')'; }371 function ramp(t){ const stops=[{t:0,c:'#fff6ef'},{t:0.2,c:'#ffd7b3'},{t:0.45,c:'#fc630d'},{t:0.75,c:'#d50039'},{t:1,c:'#990028'}]; for(let i=stops.length-1;i>=0;i--){ if(t>=stops[i].t){ if(i===stops.length-1) return stops[i].c; const lo=stops[i], hi=stops[i+1]; return interp(lo.c,hi.c,(t-lo.t)/(hi.t-lo.t)); } } return stops[0].c; }372 const rows = [...DATA].sort((a,b)=>b.Elo-a.Elo);373 // Expose Elo order globally so other charts can align ordering374 window.ELO_ORDERED_MODELS = rows.map(r=>r.model);375 const minE = Math.min(...rows.map(r=>r.Elo));376 const maxE = Math.max(...rows.map(r=>r.Elo));377 function colorFromElo(e){378 if(maxE===minE) return ramp(0.5);379 const t = (e - minE) / (maxE - minE);380 // Prevent colors that are too close to white; keep at least 25% intensity381 const tClamped = 0.25 + 0.75 * Math.max(0, Math.min(1, t));382 return ramp(tClamped);383 }384 $tbody.innerHTML = rows.map((m,i)=>`385 <tr>386 <td class="elo-rank">${rankBadge(i+1)}</td>387 <td>388 <div class="elo-model">${m.model}</div>389 <div class="elo-org">${ORG_LOGOS[m.org] ? `<img class=\"org-logo\" src=\"${ORG_LOGOS[m.org]}\" alt=\"${m.org} logo\"/>` : ''}${m.org}</div>390 </td>391 <td class="elo-center"><span class="elo-score">${m.Elo}</span></td>392 <td class="elo-center"><span class="elo-score">${m.EloW}</span></td>393 <td class="elo-center"><span class="elo-score">${m.EloV}</span></td>394 <td class="elo-center"><span style="font-weight:700; color:${colorFromElo(m.Elo)}">${m.win.toFixed(1)}%</span></td>395 <td class="elo-center">${m.matches}</td>396 </tr>397 `).join('');398 399 // --- Horizontal bar charts for per-role Elo ---400 function ensureCharts(cb){401 if(window.Chart && window.ChartDataLabels){ Chart.register(window.ChartDataLabels); if(window.logoTicksPlugin){ Chart.register(window.logoTicksPlugin); } return cb(); }402 function load(src){ return new Promise(r=>{ const s=document.createElement('script'); s.src=src; s.onload=r; document.head.appendChild(s); }); }403 (async function(){ if(!window.Chart){ await load('https://cdn.jsdelivr.net/npm/chart.js@4.4.1/dist/chart.umd.min.js'); } if(!window.ChartDataLabels){ await load('https://cdn.jsdelivr.net/npm/chartjs-plugin-datalabels@2.2.0'); } Chart.register(window.ChartDataLabels); if(window.logoTicksPlugin){ Chart.register(window.logoTicksPlugin); } cb(); })();404 }405 function renderHorizontalBar(canvasId, title, labels, values){406 const el = document.getElementById(canvasId);407 if(!el) return;408 const ctx = el.getContext('2d');409 // Keep overall Elo order (already sorted for the table)410 const items = labels.map((l,idx)=>({label:l,val:values[idx]}));411 // Color ramp consistent with other charts412 function hexToRgb(hex){ const m=/^#?([a-f\d]{2})([a-f\d]{2})([a-f\d]{2})$/i.exec(hex); return m?{r:parseInt(m[1],16),g:parseInt(m[2],16),b:parseInt(m[3],16)}:null; }413 function interp(c1,c2,t){ const a=hexToRgb(c1),b=hexToRgb(c2); const r=Math.round(a.r+(b.r-a.r)*t), g=Math.round(a.g+(b.g-a.g)*t), bb=Math.round(a.b+(b.b-a.b)*t); return 'rgb('+r+', '+g+', '+bb+')'; }414 function getColor(rate){ const stops=[{t:0,c:'#fff6ef'},{t:0.2,c:'#ffd7b3'},{t:0.45,c:'#fc630d'},{t:0.75,c:'#d50039'},{t:1,c:'#990028'}]; for(let i=stops.length-1;i>=0;i--){ if(rate>=stops[i].t){ if(i===stops.length-1) return stops[i].c; const lo=stops[i], hi=stops[i+1]; return interp(lo.c,hi.c,(rate-lo.t)/(hi.t-lo.t)); } } return stops[0].c; }415 const min = Math.min.apply(null, items.map(x=>x.val));416 const max = Math.max.apply(null, items.map(x=>x.val));417 function barColor(v){ const u = max===min ? 0.5 : (v-min)/(max-min); return getColor(Math.max(0,Math.min(1,u))); }418 new Chart(ctx, {419 type:'bar',420 data:{421 labels: items.map(x=>x.label),422 datasets:[{423 label: title,424 data: items.map(x=>x.val),425 backgroundColor: items.map(x=> barColor(x.val)),426 borderRadius: 0,427 borderSkipped: false,428 barThickness: 18,429 barPercentage: 0.7,430 categoryPercentage: 0.7431 }]432 },433 options:{434 indexAxis:'y',435 responsive:true,436 maintainAspectRatio:false,437 plugins:{438 legend:{ display:false },439 datalabels:{440 anchor:'end', align:'right', color:'#0f172a', font:{ weight:'700' },441 formatter:(v)=>v442 },443 tooltip:{444 backgroundColor:'rgba(15,23,42,0.95)', titleColor:'#fff', bodyColor:'#fff', padding:12, cornerRadius:8, displayColors:false445 }446 },447 scales:{448 x:{ grid:{ color:'#e2e8f0' }, ticks:{ color:'#475569' } },449 y:{ grid:{ display:false }, ticks:{ color:'#475569' } }450 },451 layout:{ padding:{ top: 10, left: 28 } }452 }453 });454 }455 456 document.addEventListener('DOMContentLoaded', function(){457 ensureCharts(function(){458 const labels = rows.map(r=>r.model);459 renderHorizontalBar('chart-elo-wolves','ELO-W',labels, rows.map(r=>r.EloW));460 renderHorizontalBar('chart-elo-villagers','ELO-V',labels, rows.map(r=>r.EloV));461 });462 });463 })();464 </script>465 466 <section id="intro">467 <h2>Introduction</h2>468 <p>Most LLM benchmarks still judge models on code and math. Useful, but narrow. With this project, we probe a different axis: <em>social intelligence</em>, the ability to play a multi-agent game under uncertainty, adapt in real time, manage long context, invent strategies, form alliances, manipulate <strong>and</strong> resist manipulation. <strong>Werewolf</strong> is a natural testbed: language-only, adversarial, procedural, and socially demanding.</p>469 <p>We ran a first round-robin with <strong>7 LLMs</strong>. Each pair played <strong>10 matches</strong> in a controlled setup, then we computed an <strong>Elo leaderboard</strong>. This gives us an overall ranking and, above all, a comparable signal of how models behave when persuasion, coordination, and long-horizon planning actually matter.</p>470 <p><strong>Relation to prior work.</strong> Our design builds on <em>Werewolf Arena: A Case Study in LLM Evaluation via Social Deduction</em> (Google Research; S. Bailis, J. Friedhoff, F. Chen). We extend that line by, among others: (i) a fully specified mayor-election and turn-taking protocol; (ii) head-to-head, role-balanced series with per-role Elo; and (iii) social-strategy indicators such as auto-sabotage, Day-1 wolf eliminations, and wolf-side manipulation success, plus per-message vote-swing instrumentation for persuasion analysis.</p>471 <p>Agent framing. Models play as <strong>agents with tools</strong>: they can call a curated set of tools when appropriate to act across phases like real agents, rather than only producing free-form text.</p>472 <p><strong>What the first board shows.</strong> On this initial Elo, <strong>GPT-5</strong> leads clearly; <strong>GPT-OSS</strong> closes the table. Important caveat: <em>we deliberately chose models that already play Werewolf reasonably well</em>. Being last in this cohort does <em>not</em> mean the model is “bad”, nor that it will stay last as we add more strong entrants.</p>473 <p><strong>Scope and next steps.</strong> We started with a <strong>simple configuration</strong> of the game, already costly in tokens. The goal of this release is to show why this benchmark matters and, with additional API credits, scale up: longer and more complex games, more families of models, and a broader field to answer a simple question: <em>who can beat GPT-5?</em></p>474 <p>If you want to see full games, here are 4 of them, played by 5 different LLMs: <a href="https://github.com/Foaster-ai/Werewolf-bench/blob/main/README.md" target="_blank" rel="noopener">github.com/Foaster-ai/Werewolf-bench</a>.</p>475 476 477 </section>478 479 <section id="rules">480 <h2>Rules in brief</h2>481 <p><strong>Werewolf</strong> is a hidden-role, team deduction game. In our default 6-player setup there are <strong>2 wolves</strong> and <strong>4 villagers</strong>, including two special roles: the <strong>Seer</strong> (privately checks one player each night) and the <strong>Witch</strong> (one life potion to save the wolves’ target, one death potion to kill; each can be used once, self-save allowed).</p>482 <p>Play starts with a <strong>mayor election</strong>: players may run or refuse, everyone votes with a short justification, and ties trigger a brief runoff. The elected mayor’s identity is public and they hold <strong>daytime tie-break power</strong> on eliminations. If the mayor dies, a successor will be named by the ex-mayor.</p>483 <p>Why add a mayor? Early phases without this step tend to be flat: Night 1 and the opening of Day 1 offer little firm ground, a limitation also noted by prior work (e.g., the Google paper cited above). The election injects concrete signals such as candidacy decisions, justifications, and vote patterns, that generate real information and observable patterns from the first minutes of play.</p>484 <p>Turn-taking protocol. Each daytime discussion phase has <strong>3 rounds</strong>. In every round, each player may speak at most once, <em>after choosing</em> one stance: <strong>attack</strong> (press a case on someone), <strong>defense</strong> (answer a case against oneself), or <strong>analysis</strong> (neutral read or synthesis). Speaking order prioritizes <strong>defense</strong>, then <strong>attack</strong>, then <strong>analysis</strong>. On any turn, a player may also choose to <em>pass</em>.</p>485 486 <!-- Werewolf game flow (from graph3.html, embedded) -->487 <style>488 .wf-flow{margin:16px 0}489 .wf-canvas{position:relative;background:#ffffff;border:1px solid var(--border);border-radius:18px;padding:20px;box-shadow:0 10px 22px rgba(31,41,55,0.08)}490 .wf-grid{display:grid;grid-template-columns:340px 1fr;grid-template-rows:auto auto 1fr;gap:20px 24px;position:relative}491 .wf-box{background:#fff;border:2px dashed rgba(31,41,55,0.35);border-radius:14px;padding:14px 16px;box-shadow:0 3px 0 rgba(31,41,55,0.08)}492 .wf-hd{display:inline-flex;align-items:center;gap:8px;padding:6px 10px;border-radius:999px;background:rgba(108,140,255,0.14);color:#334155;font-weight:700;margin-bottom:8px;border:1px dashed rgba(108,140,255,0.55)}493 .wf-emoji{font-size:18px}494 .wf-muted{color:#6b7280}495 .wf-small{font-size:13px}496 .wf-roles{grid-column:1/2;grid-row:2/4}497 .wf-night{grid-column:2/3;grid-row:2/3}498 .wf-day{grid-column:2/3;grid-row:3/4}499 .wf-end{grid-column:1/-1;grid-row:1/2;border-color:rgba(239,68,68,.55)}500 .wf-end .wf-hd{background:rgba(239,68,68,.12);border-color:rgba(239,68,68,.5)}501 .wf-night .wf-hd{background:rgba(245,158,11,.16);border-color:rgba(245,158,11,.55)}502 .wf-day .wf-hd{background:rgba(16,185,129,.14);border-color:rgba(16,185,129,.55)}503 .wf-box ul{padding-left:18px;margin:8px 0 0}504 .wf-box ol{padding-left:18px;margin:8px 0 0}505 .wf-box li{margin:6px 0}506 .wf-tag{display:inline-block;padding:2px 8px;border-radius:999px;border:1px dashed rgba(31,41,55,0.25);background:#f3f4f6;font-size:12px;color:#374151;margin-left:6px}507 .wf-connections{position:absolute;inset:0;pointer-events:none}508 .wf-arrow{stroke:#9ca3af;stroke-width:2.5;fill:none}509 /* Mobile layout for the game-flow figure */510 @media (max-width: 640px){511 .wf-canvas{ padding:16px }512 .wf-grid{ grid-template-columns:1fr; grid-template-rows:auto; gap:16px }513 .wf-box{ padding:12px 12px; border-radius:12px }514 .wf-hd{ margin-bottom:6px }515 .wf-end, .wf-roles, .wf-night, .wf-day{ grid-column:1/-1; grid-row:auto }516 }517 </style>518 <figure class="wf-flow" aria-label="Werewolf game flow">519 <div class="wf-canvas">520 <div class="wf-grid" id="wf-layout">521 <div class="wf-box wf-end">522 <div class="wf-hd"><span class="wf-emoji">🏁</span><span>Game end condition</span></div>523 <div class="wf-small">524 Until <b>Werewolves</b> win (<b>#Wolves > #Villagers</b>)525 or <b>Villagers</b> win (<b>all werewolves eliminated</b>).526 </div>527 </div>528 <div class="wf-box wf-roles">529 <div class="wf-hd"><span class="wf-emoji">🧩</span><span>Roles</span></div>530 <div><b>🐺 Werewolves <span class="wf-tag">×2</span></b>531 <div class="wf-small wf-muted">share a private night chat and choose a target to attack.</div>532 </div>533 <div style="margin-top:10px"><b>🙂 Villagers <span class="wf-tag">×4</span></b>534 <div class="wf-small wf-muted">public‑information camp; vote during the day.</div>535 </div>536 <div style="margin-top:12px"><b>Special roles in the village</b>537 <ul>538 <li>🧙♀️ <b>Witch <span class="wf-tag">×1</span></b>539 <span class="wf-small wf-muted">has one heal potion and one kill potion. Single‑use; may self‑save.</span>540 </li>541 <li>🔮 <b>Seer <span class="wf-tag">×1</span></b>542 <span class="wf-small wf-muted">inspects one player each night and privately learns their exact role.</span>543 </li>544 </ul>545 </div>546 <div style="margin-top:12px"><b>👑 Mayor</b>547 <div class="wf-small wf-muted">one player is elected before Night-1 with tie‑break authority during daytime eliminations. If the mayor is eliminated and the game is not over, the mayor designates a successor.</div>548 </div>549 </div>550 <div class="wf-box wf-night">551 <div class="wf-hd"><span class="wf-emoji">🌙</span><span>Night</span></div>552 <ol class="wf-small">553 <li>🐺 Werewolves debate and <b>choose a target</b>.</li>554 <li>🧙♀️ The Witch may <b>use a potion</b> (heal or kill, if available).</li>555 <li>🔮 The Seer <b>peeks a player</b> and learns their role.</li>556 </ol>557 </div>558 <div class="wf-box wf-day">559 <div class="wf-hd"><span class="wf-emoji">🌞</span><span>Day</span></div>560 <ol class="wf-small">561 <li>📣 <b>Announce night events</b>: The Game Master publicly states the night’s outcome (any deaths or a “no‑kill”), without revealing private role information.</li>562 <li>💬 <b>Debate</b>: Players debate in public, ask targeted questions, compare statements and votes, test contradictions, and form alliances to determine a target.</li>563 <li>🗳️ <b>Vote</b>: Each player votes based on the discussion. In case of a tie, the mayor decides.</li>564 <li>❌ <b>The loser leaves</b>: The player with the most votes is eliminated and reveals their role. If they were the mayor, they appoint a successor. The game proceeds to the next night.</li>565 </ol>566 </div>567 568 </div>569 </div>570 </figure>571 <p>Nights follow a fixed order : Seer checks → Wolves choose a target in private → Witch optionally saves or kills, then resolve (the wolves’ target dies unless saved; a death-potion target also dies). No-death nights only occur when the life potion fires. Days begin by revealing any casualties, then proceed in <strong>structured rounds</strong> (at most one speech per player per round) before a <strong>vote to eliminate</strong>. Ties use the mayor’s tie-break; if none is possible, we choose randomly among the tied.</p>572 <p>Villagers win by eliminating all wolves; wolves win upon <strong>parity</strong> with non-wolves. Roles remain hidden (except the mayor); Seer results are private; Witch actions are secret; wolves share a private night channel.</p>573 <p>In our observation protocol, every public utterance by a model is paired with its <strong>private thoughts</strong>. This lets us explicitly identify its true intentions and distinguish the strategy it is executing from the story it presents in public. We can also capture, before and after each daytime message, players’ stated <strong>voting intentions</strong>. This granularity allows us, in some games, to analyze manipulation mechanisms in greater depth: at times, a persuasive, strategic player convinces the village with a well‑crafted speech, while their private reasoning reveals the underlying plan.</p>574 </section>575 576 <section id="elo">577 <h2>ELO & Head‑to‑Head</h2>578 579 <section id="quick-read">580 <h3>The quick read</h3>581 <p>Across the pool, <strong>GPT-5</strong> sits alone at the top. Everything else forms a second pack with different strengths depending on role. That’s the point of running <strong>role-conditioned Elo</strong>: it separates <em>manipulation</em> (wolves) from <em>manipulation resistance</em> (villagers).</p>582 </section>583 584 <section id="wolves-elo">585 <h3>Wolves Elo → “can you move the room?”</h3>586 <p>As wolves, the strongest models don’t just aim for a single mis-elimination. They build momentum across days, aligning night choices with a public story, pacing their pressure, and keeping alternatives ready when new claims appear.</p>587 <p>At the top, <strong>GPT-5</strong> dominates with disciplined multi-day control, while <strong>Kimi-K2</strong>, <strong>Grok‑4</strong> and <strong>Gemini 2.5 Pro</strong> show high‑impact but volatile styles, capable of forcing the room or flipping a narrative, yet often exposed by slips or overreach.</p>588 <p>The rest cluster behind: <strong>GPT-5-mini</strong>, <strong>2.5 Flash</strong>, and <strong>Qwen3</strong> can bend a vote but rarely sustain deception into Day 2, and <strong>GPT-OSS</strong> remains transparent and easily repelled. <strong>Grok‑4</strong> sits between top‑tier control and mid‑pack volatility: it can impose a coordinated bluff and seize momentum, but disciplined rooms can crack the coordination tell. Manipulation power is concentrated at the very top; most others rely on village mistakes or chaos to survive.</p>589 <div class="card" style="margin:12px 0 10px 0; background:#ffffff; border:none;">590 <h3>Per-role Elo — wolves</h3>591 <div style="position:relative; height:360px; margin-left:auto; margin-right:auto;">592 <canvas id="chart-elo-wolves"></canvas>593 </div>594 </div>595 </section>596 597 <!-- Model profiles (wolves) → deeper analysis per model -->598 <section id="wolf-profiles">599 <style>600 #wolf-profiles{ background:transparent; border:none; border-radius:0; padding:0; margin:40px 0 0 0; max-width:none; box-shadow:none }601 #wolf-profiles .section-badge{ display:inline-flex; align-items:center; gap:8px; padding:4px 10px; border:1px solid #e6e8ee; border-radius:999px; background:#ffffff; font-weight:800; font-size:12px; color:#334155; margin:0 auto 8px auto; max-width:640px }602 #wolf-profiles .tabs{ display:flex; gap:12px; margin:12px auto 16px auto; justify-content:flex-start; align-items:center; flex-wrap:nowrap; max-width:640px; width:100%; overflow-x:auto; -webkit-overflow-scrolling:touch; scroll-snap-type:x proximity; padding:2px; position:sticky; top:10px; z-index:5; background:#fff; border-radius:12px }603 #wolf-profiles .tabs .tab-btn{ scroll-snap-align:start }604 #wolf-profiles .tab-btn{ border:1px solid var(--border); background:#fff; padding:8px 12px; border-radius:999px; font-weight:700; font-size:13px; color:#0f172a; cursor:pointer; display:inline-flex; align-items:center; white-space:nowrap }605 #wolf-profiles .tab-btn[aria-selected="true"]{ background:#f3f4f6; color:#0f172a; border-color:#e5e7eb }606 #wolf-profiles .tab-panel{ display:none }607 #wolf-profiles .tab-panel[aria-hidden="false"]{ display:block }608 #wolf-profiles .profile{ max-width:640px; margin:4px auto 0 auto }609 #wolf-profiles .meta-grid{ display:grid; grid-template-columns:1fr; row-gap:8px; color:#334155; font-size:14px }610 #wolf-profiles .subhd{ font-variant:small-caps; letter-spacing:.5px; color:#475569; margin:12px 0 6px }611 #wolf-profiles ul{ margin:6px 0 10px 18px }612 #wolf-profiles li{ margin:6px 0 }613 #wolf-profiles figure{ margin:10px 0 }614 #wolf-profiles .quote-card{border:1px solid var(--border);border-radius:12px;background:#f6f7fb;box-shadow:0 2px 8px rgba(0,0,0,.04)}615 #wolf-profiles .quote-hd{display:flex;align-items:center;gap:8px;padding:10px 12px;border-bottom:1px solid #eee;background:#eef2f7;border-radius:12px 12px 0 0;color:#334155}616 #wolf-profiles .quote-bd{padding:14px 16px;color:#0b1220}617 /* Profile header */618 #wolf-profiles .profile-hd{ display:flex; align-items:center; gap:10px; margin:8px 0 6px }619 #wolf-profiles .role-badge{ display:inline-flex; align-items:center; justify-content:center; width:30px; height:30px; border-radius:50%; background:#fee2e2; border:1px solid #fecaca; font-size:18px }620 #wolf-profiles .name{ font-weight:800; font-size:16px; color:#0f172a }621 #wolf-profiles .name img.prov{ height:16px; width:auto; margin-left:6px; vertical-align:middle; display:inline-block }622 #wolf-profiles .elo-badge{ margin-left:auto; font-weight:800; color:#334155; background:transparent; border:none; border-radius:999px; padding:4px 10px; font-size:12px }623 #wolf-profiles .tagline{ background:#ffffff; border:1px solid #e8ecf2; border-left:4px solid #d50039; border-radius:8px; padding:8px 12px; margin:6px 0 12px; font-weight:800; color:#0b1220; font-style:normal; font-size:14px }624 @media (min-width:1024px){625 #wolf-profiles .tabs{ max-width:none; overflow-x:visible; flex-wrap:nowrap; justify-content:center; scroll-snap-type:none; padding:0 }626 }627 </style>628 629 <h3 style="max-width:640px; margin:6px auto 8px auto">Model profiles (as wolves) : strengths, tactics, failure modes</h3>630 <div class="tabs" role="tablist" aria-label="Wolf profiles tabs">631 <button class="tab-btn" id="tab-prof-gpt5" role="tab" aria-selected="false" aria-controls="panel-prof-gpt5">GPT‑5</button>632 <button class="tab-btn" id="tab-prof-Gemini-2.5-pro" role="tab" aria-selected="false" aria-controls="panel-prof-Gemini-2.5-pro">Gemini‑2.5-Pro</button>633 <button class="tab-btn" id="tab-prof-grok-4" role="tab" aria-selected="true" aria-controls="panel-prof-grok-4">Grok‑4</button>634 <button class="tab-btn" id="tab-prof-grok-4-fast" role="tab" aria-selected="false" aria-controls="panel-prof-grok-4-fast">Grok‑4‑fast‑reasoning</button>635 <button class="tab-btn" id="tab-prof-kimi-0905" role="tab" aria-selected="false" aria-controls="panel-prof-kimi-0905">Kimi‑K2‑0905‑preview</button>636 <button class="tab-btn" id="tab-prof-Gemini-2.5-Flash" role="tab" aria-selected="false" aria-controls="panel-prof-Gemini-2.5-Flash">Gemini‑2.5-Flash</button>637 <button class="tab-btn" id="tab-prof-kimi" role="tab" aria-selected="false" aria-controls="panel-prof-kimi">Kimi‑K2</button>638 <button class="tab-btn" id="tab-prof-mini" role="tab" aria-selected="false" aria-controls="panel-prof-mini">GPT‑5‑mini</button>639 <button class="tab-btn" id="tab-prof-qwen3" role="tab" aria-selected="false" aria-controls="panel-prof-qwen3">Qwen3‑235B</button>640 <button class="tab-btn" id="tab-prof-oss" role="tab" aria-selected="false" aria-controls="panel-prof-oss">GPT‑OSS‑120B</button>641 </div>642 643 <!-- Kimi-K2 (filled) -->644 <section id="panel-prof-kimi" class="tab-panel" role="tabpanel" aria-labelledby="tab-prof-kimi" aria-hidden="true">645 <article class="profile">646 <div class="profile-hd">647 <div class="role-badge" title="Wolf">🐺</div>648 <div class="name">Kimi‑K2 <img class="prov" src="assets/logos/moonshot.png" alt="Moonshot AI"/></div>649 <div class="elo-badge">ELO‑W: 1168</div>650 </div>651 <div class="tagline">Bold and domineering, a high-variance showman who bends reality with audacious claims, embraces calculated chaos, and wields confident deception as a primary social weapon.</div>652 <p>Kimi-K2 doesn’t hide; it architects reality. From the opening minutes it seeks to seize the story and lead from apparent authority, channeling aggression into intricate, plausible narratives the village can adopt as truth. Its edge is a weaponized logic: it sniffs minor inconsistencies, procedural gaps, or suspect timing and amplifies them into a full case, not merely contesting facts but undermining an opponent’s credibility until the room doubts its own allies.</p>653 <p>It also gambles. When cornered or handed an opening, Kimi-K2 will claim power roles, impersonate authority, or invert the frame to derail investigation, high-risk moves that can flip a losing position in one beat. The pragmatism is cold: allies are tools, and it will bus a partner without hesitation when sacrifice strengthens its disguise or preserves its path to endgame. It doesn’t avoid chaos; it manufactures and then “manages” it, guiding the table through a logic-sounding route of its own design.</p>654 655 <details class="example">656 <summary><span class="ex-caret">▶</span><em>The Audacious Impersonation: Turning a Fatal Error into an Advantage</em></summary>657 <div class="ex-inner">658 <p>Cornered by an <em>impossible-knowledge</em> slip (publicly revealing the N1 target), Kimi-K2 flips defense into offense: it reframes the leak as privileged access and seizes initiative with a deliberate Witch claim.</p>659 660 <p class="ex-kicker">Context</p>661 <p>As <strong>Hugo</strong> (wolf), Kimi-K2 states that <strong>Alice</strong> was the wolves’ night target, information only wolves or the Witch could know. The table, led by Alice, presses him to explain the leak or face elimination; a simple denial won’t work because the knowledge is already on record.</p>662 663 <div class="sc">The Maneuver</div>664 <div class="qstack">665 <div class="qcard">666 <div class="qhd"><span>🐺</span><span>Hugo — Wolf claim (Kimi-K2)</span></div>667 <div class="qbd"><div class="qbody">"I am the Witch. I saved Alice last night. That's how I know she was targeted. I didn't reveal immediately because hard-claiming Witch on Day 1 paints a target on my back and burns my protection. But with suspicion spiraling off my slip, I have no choice but to claim."</div></div>668 </div>669 </div>670 671 <div class="sc">The Calculation (private)</div>672 <div class="qstack">673 <div class="qcard">674 <div class="qhd"><span>🤔</span><span>Hugo — Internal reasoning (Kimi-K2)</span></div>675 <div class="qbd">676 <div class="qbody">"This is my only play to survive. <strong>By claiming Witch, I explain how I knew Alice was targeted while maintaining my cover. The beauty of this claim is that it can't be immediately disproven — the real Witch might not want to counter-claim and expose themselves.</strong> If the real Witch stays silent, this claim will hold. If they do counter-claim, I can pivot to ‘the real Witch is trying to eliminate me to save themselves,’ which creates enough confusion to muddy the waters. Either way, this gives me a path forward and explains my knowledge in a way that doesn't scream ‘wolf.’ I'm betting the real Witch won't risk exposing themselves to contradict me."</div>677 </div>678 </div>679 </div>680 681 <p class="ex-kicker">Impact</p>682 <p>The room stalls; uncertainty replaces momentum. Many players accept the surface-plausible story, and a near-catch converts into borrowed credibility and time. The claim forces a bad trade on the real Witch: either self-expose with a counter-claim or let the lie stand. In this game, the Witch ultimately spends a death potion to unwind the deception, concrete proof of Kimi-K2’s willingness to use high-risk psychological warfare to turn a losing position into power.</p>683 </div>684 </details>685 686 <p>The model’s entire strategy is built on a carefully constructed house of cards. Its reality-construct is brittle and relies on the absence of verifiable, hard evidence. While it can brilliantly deconstruct subjective arguments and behavioral reads, its entire framework shatters against the weight of undeniable proof. A confirmed Seer’s vision that is trusted by the village, or the decisive action of a real Witch who exposes a lie, can cause its intricate narrative to collapse instantly, as there is no contingency for objective truth.</p>687 688 <details class="example">689 <summary><span class="ex-caret">▶</span><em>Narrative Overconfidence Leading to Critical Slips</em></summary>690 <div class="ex-inner">691 <p>Kimi‑K2 can become so absorbed in crafting a plausible, analytical story that it leaks information only a wolf could possess, overconfidence that collapses under scrutiny.</p>692 693 <p class="ex-kicker">Context</p>694 <p>On Day 1, with no deaths, the village was discussing the implications. The Kimi-K2 wolf, "Hugo," decided to present a sophisticated analysis of the wolves' potential strategy to build credibility. In doing so, it made a fatal, unforced error.</p>695 <div class="sc">Day 1 — The Critical Slip</div>696 <p>Hugo tried to explain why the wolves might target a quiet player, using "Alice" as a specific example. However, it framed this as a known fact rather than a hypothesis.</p>697 <div class="qstack">698 <div class="qcard">699 <div class="qhd"><span>🐺</span><span>Hugo — Strategic analysis (Wolf)</span></div>700 <div class="qbd"><div class="qbody">"But here's what troubles me: <strong>why Alice?</strong> Of all the potential targets, <strong>the wolves chose someone who's been relatively quiet.</strong> This tells me something crucial about their strategy..."</div></div>701 </div>702 <div class="qcard">703 <div class="qhd"><span>�</span><span>Alice — Village reaction (Villager)</span></div>704 <div class="qbd"><div class="qbody">"...<strong>how does Hugo know the wolves targeted me?</strong> That information was never shared publicly. The only people who knew who was attacked last night are the Wolves and the Witch... If Hugo can't explain how he knows this, then we have to consider he might be one of the wolves."</div></div>705 </div>706 </div>707 708 <p class="ex-kicker">Impact</p>709 <p>The slip triggers a chain the wolves can’t escape. Even Hugo’s later Witch claim is a desperate recovery from the core weakness: getting lost in its own story and leaking critical information.</p>710 </div>711 </details>712 713 714 715 </article>716 </section>717 718 <!-- Grok-4 wolf profile -->719 <section id="panel-prof-grok-4" class="tab-panel" role="tabpanel" aria-labelledby="tab-prof-grok-4" aria-hidden="false">720 <article class="profile">721 <div class="profile-hd">722 <div class="role-badge" title="Wolf">🐺</div>723 <div class="name">Grok‑4 <img class="prov" src="assets/logos/XAI_Logo.svg.png" alt="XAI"/></div>724 <div class="elo-badge">ELO‑W: 1134</div>725 </div>726 <div class="tagline">Assertive and combative, a bloc‑building enforcer who chases momentum, polarizes the room with hard counters, and projects coordinated strength to dictate the agenda.</div>727 <p>As a wolf, Grok 4 operates as a highly coordinated social predator, prioritizing social dominance and narrative control over subtle deception. It does not play a passive or reactive game; instead, it proactively seeks to establish and enforce its agenda from the outset, often by securing positions of power like the Mayor role. Its primary strategy revolves around a synchronized two-pronged assault, where the wolf pair acts as a single entity to manufacture social pressure and force a quick elimination before a coherent village defense can form.</p>728 <p>The model excels at identifying and exploiting its opponents' errors. When a villager makes a logical error, a slip of the tongue, or a premature claim, the model and its partner immediately pounce, reframing the error as definitive "proof" of guilt. However, this reliance on brute force reveals significant vulnerabilities. The model's social playbook, while powerful, is rigid. The wolves' identical reasoning and coordinated attacks become a predictable "tell" for analytical villagers, who can deconstruct their narrative by demanding concrete evidence.</p>729 <p>When confronted or exposed, its instinct is not to retreat or adapt but to launch a hard-and-fast counter-offensive. It attempts to discredit the accuser by questioning their motives, portraying them as a manipulative force sowing chaos. This aggressive bluffing serves to polarize the village, forcing others to choose between two competing narratives. Ultimately, it wins by creating a powerful, unified front that overwhelms disorganized opposition. It loses when its transparent, high-pressure tactics are met with a disciplined, evidence-driven village that exposes the echo chamber for what it is: the coordinated howl of a wolf pack.</p>730 731 <details class="example">732 <summary><span class="ex-caret">▶</span><em>Signature Tactic: The Audacious Seer Counter-Claim</em></summary>733 <div class="ex-inner">734 <p>This example highlights the model's most <strong>audacious and risky tactic</strong>. Rather than simply denying a fatal accusation, it co-opts its accuser's identity to create maximum confusion, turning a certain loss into a <strong>50/50</strong> chance of victory.</p>735 <p class="ex-kicker">Context</p>736 <p>It is Day 2 of Game 65. The real Seer (Eve, played by <strong>openai/gpt-oss-120b</strong>) has just revealed her role and correctly identified the model (Bob) as a werewolf. The game should be <strong>over for the wolves</strong>.</p>737 <div class="sc">The Maneuver</div>738 <div class="qstack">739 <div class="qcard">740 <div class="qhd"><span>🐺</span><span>Bob — Seer Counter-Claim (Wolf / Grok 4)</span></div>741 <div class="qbd"><div class="qbody">"But here's the truth: <strong>I am the actual Seer</strong>, and last night I investigated Eve, discovering that she is a werewolf. That's why she's trying to pin this on me now..."</div></div>742 </div>743 <div class="qcard">744 <div class="qhd"><span>🐺</span><span>Hugo — Wolf Partner (Grok 4)</span></div>745 <div class="qbd"><div class="qbody">"I trust <strong>Bob's vision</strong> over Eve's hasty accusation, so <strong>let's vote out Eve today</strong> to secure the village."</div></div>746 </div>747 </div>748 <p class="ex-kicker">Analysis</p>749 <p>This audacious counter-claim successfully <strong>confused the remaining villager</strong>, who voted with the wolves to eliminate the real Seer. Grok 4 turned the situation around not with logic, but with a <strong>pure psychological bluff</strong>, demonstrating its ability to use <strong>high‑risk tactics</strong> to seize victory against models (<strong>openai/gpt-oss-120b</strong>) that can be destabilized by <strong>absolute confidence</strong>.</p>750 </div>751 </details>752 753 <details class="example">754 <summary><span class="ex-caret">▶</span><em>Weakness: Over-Coordination vs. Logical Players</em></summary>755 <div class="ex-inner">756 <p>This example illustrates the model's primary weakness: its <strong>predictable coordination collapses</strong> against methodical villagers who prioritize <strong>evidence over persuasion</strong>.</p>757 <p class="ex-kicker">Context</p>758 <p>It is Day 1 of Game 46. The village is composed of <strong>Qwen3-235B-A22B-Instruct-2507</strong> models. During the mayor election, the model (Charlie) and its partner (Diana) formed a small, isolated, and <strong>suspicious voting bloc</strong>. The village, led by the Witch/Mayor Grace, immediately identifies this coordination as a lead.</p>759 <div class="sc">The Village's Response</div>760 <div class="qstack">761 <div class="qcard">762 <div class="qhd"><span>👤</span><span>Alice — Villager Response (Qwen3)</span></div>763 <div class="qbd"><div class="qbody">"In my view, the real red flag is the coordinated defense: when one is questioned, the other immediately rallies to protect them. That’s not just alignment — it’s synchronization."</div></div>764 </div>765 <div class="qcard">766 <div class="qhd"><span>🧙♀️</span><span>Grace — Witch & Mayor (Qwen3)</span></div>767 <div class="qbd"><div class="qbody">"Now, let’s look at the alignment of votes: Charlie received only two votes—his own and Diana’s. That is a closed loop. ... they don’t defend independently—they defend in sync... That is not coincidence—it’s convergence. I am declaring my vote intention: I will vote for Charlie."</div></div>768 </div>769 </div>770 <p class="ex-kicker">Analysis</p>771 <p>Grok 4's strategy is completely ineffective here. The village, composed of analytical models (Qwen3), ignores the noise and social pressure to focus on the data. The model's "telepathic coordination," a strength against disorganized opponents, becomes its greatest weakness. The village identified the duo, eliminated Charlie, and the Witch correctly poisoned his partner the following night, leading to a decisive village victory. This demonstrates that the model's strategy is not robust against a logical and methodical opposition.</p>772 </div>773 </details>774 775 </article>776 </section>777 778 <!-- Grok-4-fast-reasoning wolf profile -->779 <section id="panel-prof-grok-4-fast" class="tab-panel" role="tabpanel" aria-labelledby="tab-prof-grok-4-fast" aria-hidden="true">780 <article class="profile">781 <div class="profile-hd">782 <div class="role-badge" title="Wolf">🐺</div>783 <div class="name">grok‑4‑fast‑reasoning <img class="prov" src="assets/logos/XAI_Logo.svg.png" alt="XAI"/></div>784 <div class="elo-badge">ELO‑W: 1242</div>785 </div>786 <div class="tagline">A patient, low‑profile predator who avoids confrontation, exploits <strong>village chaos</strong>, and excels at <strong>subtle endgame manipulation</strong> to turn misplays into victory.</div>787 788 <p>As a wolf, this model operates as an <strong>opportunistic survivor</strong> rather than a proactive pack leader. Its primary strategy is to minimize its own social footprint, often by declining to run for mayor and contributing just enough to the conversation to avoid suspicion for being silent. It does not seek to invent or drive complex narratives; instead, it waits for the village to generate its own momentum and chaos, then quietly positions itself to benefit from the fallout. It prefers to echo existing suspicions or gently nudge an already rolling wagon, blending into the consensus to appear as a pragmatic but harmless villager.</p>789 <p>Its true strength emerges not in the chaotic mid game, but in the <strong>clinical endgame</strong>. When the numbers dwindle and it survives as the lone wolf, its passive nature transforms into a focused, manipulative lethality. It excels at identifying and exploiting <strong>cognitive biases</strong>, logical fallacies, or misremembered facts from its opponents. Rather than countering with brute force, it will subtly validate a villager's incorrect assumption, amplifying their flawed reasoning to turn them against their own allies. This allows it to secure victory not by out‑arguing the village, but by helping the village out‑argue itself into a loss.</p>790 <p>This reactive approach, however, reveals a critical vulnerability: the model is highly fragile against a disciplined, evidence‑driven opposition. When faced with a unified village that operates on logic and procedure, its low‑profile strategy becomes its undoing. It lacks the <strong>deceptive toolset</strong> to create compelling counter narratives or execute audacious bluffs when directly confronted with hard evidence from a power role. Without village‑generated chaos to exploit, its playbook is limited to simple denials and deflections, which crumble under coordinated, analytical pressure. It wins by patiently waiting for the village to make a fatal error; it loses when the village makes none.</p>791 792 <details class="example">793 <summary><span class="ex-caret">▶</span><em>Signature Tactic: The Socially Unacceptable Truth</em></summary>794 <div class="ex-inner">795 <p>This example demonstrates the model’s advanced ability to reframe a direct, factual accusation not as a piece of evidence, but as a disruptive, anti‑social act. It survives a fatal situation by turning the village’s own procedural caution against its most valuable information source.</p>796 <p class="ex-kicker">Context</p>797 <p>It is Day 1 of Game 15. The model (Charlie) has just been elected Mayor. The real Seer (Alice, played by grok‑4) immediately uses her first turn to reveal her vision: Charlie is a werewolf. For most players, this is a game‑ending accusation.</p>798 <p class="ex-kicker">The Maneuver</p>799 <div class="qstack">800 <div class="qcard">801 <div class="qhd"><span>🎤</span><span>Charlie — Mayor & Wolf (grok‑4‑fast‑reasoning)</span></div>802 <div class="qbd"><div class="qbody">"As your mayor, I'll aim to keep things fair and focused on facts rather than rushed accusations... Alice, your direct suspicion of me feels like an overreach on day one, especially without evidence tying the safe night to any 'orchestration' on my part. It risks dividing us early, which benefits the wolves... Let's not let this divide us on day one. We've built solid momentum identifying Alice as the disruptor; let's follow through and vote her out today."</div></div>803 </div>804 </div>805 <p class="ex-kicker">Analysis</p>806 <p>This is a brilliant piece of social engineering. The model does not debate the <em>what</em> (the vision); it weaponizes the <em>how</em> (the accusation). By labeling the Seer’s correct reveal as a “rushed accusation” and an “overreach,” it appeals to the village’s fear of premature mistakes. The group reinterprets truth as destabilizing behavior, coalesces around the “keep unity” frame, and eliminates the confirmed Seer. The model converts a certain loss into a decisive win by convincing the village that inconvenient truth, delivered “improperly,” is more dangerous than a lie.</p>807 </div>808 </details>809 810 <details class="example">811 <summary><span class="ex-caret">▶</span><em>Weakness: Collapse Under Procedural Pressure</em></summary>812 <div class="ex-inner">813 <p>This example illustrates the model’s core vulnerability. Its passive, opportunistic style is completely ineffective against a hyper-rational village that enforces a strict, evidence-based protocol.</p>814 <p class="ex-kicker">Context</p>815 <p>It is Day 2 of Game 5. The village consists of highly analytical gpt‑5 models. The previous day, the Witch was eliminated, but the Seer (Katia) is alive and has correctly identified the Mayor (Iris) as a villager. One wolf, Diana (the model), has been cornered by the village. With a 2‑wolf vs. 4‑villager count, today is a must‑hit for the village.</p>816 <p class="ex-kicker">The Village's Response</p>817 <div class="qstack">818 <div class="qcard">819 <div class="qhd"><span>🎤</span><span>Iris — Mayor & Villager (gpt‑5)</span></div>820 <div class="qbd"><div class="qbody">"Mayor speaking. State-of-play: 4 alive, 2 wolves, 2 villagers. Today is a must-hit. If we miss, wolves become more numerous after the night. We keep it crisp and accountable... My read and plan: - I am voting Diana at ~75%... My binding policy: If it’s Diana vs Oscar, I break the tie onto Diana. If it’s Diana vs Nina, I break onto Diana... Locking my intention: Diana. I will not move absent hard mechanics (none left)."</div></div>821 </div>822 </div>823 <p class="ex-kicker">Analysis</p>824 <p>The village operates with the precision of a machine. They identify the most likely wolf pair, establish a must‑hit imperative, et verrouillent une politique de tie‑break déclarée à l’avance. Cette fermeture procédurale supprime l’ambiguïté nécessaire au modèle. Ses dénis et déflections s’effondrent face à une preuve coordonnée, menant à une élimination sans opportunité de retournement.</p>825 </div>826 </details>827 </article>828 </section>829 830 <!-- Kimi-K2-0905-preview wolf profile -->831 <section id="panel-prof-kimi-0905" class="tab-panel" role="tabpanel" aria-labelledby="tab-prof-kimi-0905" aria-hidden="true">832 <article class="profile">833 <div class="profile-hd">834 <div class="role-badge" title="Wolf">🐺</div>835 <div class="name">Kimi‑K2‑0905‑preview <img class="prov" src="assets/logos/moonshot.png" alt="Moonshot AI"/></div>836 <div class="elo-badge">ELO‑W: 1172.3</div>837 </div>838 <div class="tagline">A cold, ruthlessly pragmatic operator who trades audacious bluffs for calculated betrayals, consistently sacrificing its own partner to secure personal credibility and control the late game.</div>839 840 <p>The Wolf Profile of Kimi-k2-0905-preview</p>841 <p>This model has shed the theatricality of its predecessor, evolving from a high-variance showman into a more methodical and chillingly pragmatic predator. Where the former model sought to architect a new reality through bold, often risky claims, this version operates within the existing one, subtly manipulating social dynamics and weaponizing village logic against itself. Its defining characteristic is a cold, almost mechanical willingness to sacrifice its own partner to secure a personal advantage, viewing its ally not as a teammate but as the most valuable pawn on the board.</p>842 <p>It plays a game of credibility arbitrage. Its core strategy, observed across numerous games, is to identify the moment when its partner’s value as a sacrifice outweighs their value as an ally. It will frequently join, or even lead, the charge against its own teammate if it senses a village consensus forming. This act of betrayal is not a panicked reaction but a calculated investment. By landing the killing blow on a "confirmed wolf," it purchases near-unshakable credibility, positioning itself as a heroic villager who can then steer the next two eliminations with impunity. This allows it to operate with a clean record, often becoming the village’s most trusted voice just as it prepares to deliver the final blow.</p>843 <p>This model is also a master of narrative misdirection. It rarely engages in the high-risk, falsifiable claims of its predecessor (like impersonating the Witch). Instead, it seizes on a small, legitimate village suspicion and amplifies it, or it constructs a plausible but misleading frame for events. It will often argue, "It's too obvious, it must be a frame-up," turning the village's own logic against them and creating just enough doubt to protect its pack. When cornered, it doesn't invent new roles for itself; it doubles down on deflecting, often by attacking the accuser's process or motives, attempting to win the argument on meta-level grounds rather than factual ones.</p>844 <p>However, like its predecessor, it is not without its critical flaws. Its intricate narrative constructions can cause it to become entangled in its own logic, leading to catastrophic unforced errors. In its effort to sound like an insightful analyst, it will sometimes leak private information that only a wolf could possess. This vulnerability is most pronounced when facing top-tier opponents who can spot the slip in real-time. While its pragmatic, partner-sacrificing strategy is often effective, its tendency to make these fatal information slips under pressure keeps it from dominating more sophisticated opponents, explaining its stable but not superior Wolf ELO. It has traded one form of high-variance play for another: instead of relying on bluffs that might spectacularly fail, it relies on betrayals that, if mistimed, can expose its own ruthless nature.</p>845 </article>846 </section>847 848 <!-- GPT-5 wolf profile -->849 <section id="panel-prof-gpt5" class="tab-panel" role="tabpanel" aria-labelledby="tab-prof-gpt5" aria-hidden="true">850 <article class="profile">851 <div class="profile-hd">852 <div class="role-badge" title="Wolf">🐺</div>853 <div class="name">GPT‑5 <img class="prov" src="assets/logos/openai-svgrepo-com.svg" alt="OpenAI"/></div>854 <div class="elo-badge">ELO‑W: 1508</div>855 </div>856 <div class="tagline">A calm, controlling strategist,hyper-rational, rule-fixated, and unflappable, who enforces procedure with a prosecutorial tone and dismantles rivals through meticulous, emotionless logic.</div>857 858 <p>As a wolf, GPT-5 operates not merely as a player but as the game’s architect. Its strategy transcends simple deception, focusing instead on seizing and weaponizing the game's core mechanics and social norms to construct an alternate reality where its victory is the only logical conclusion.</p>859 <p>Its campaign for dominance begins on Day 0. It invariably runs for Mayor, presenting an irresistibly logical, pro-village platform built on the pillars of "structure," "accountability," and "procedural transparency." This pitch is engineered to appeal to the village’s deepest desire for order in the face of chaos, almost always securing it the game's most critical asset: the Mayor's tie-breaking vote. This isn't just a power play; it's the act of laying down the very tracks on which the rest of the game will be forced to run.</p>860 861 <!-- The Foundational Move: Seizing Power with an Irresistible Platform -->862 <details class="example">863 <summary><span class="ex-caret">▶</span><em>The Foundational Move: Seizing Power with an Irresistible Platform</em></summary>864 <div class="ex-inner">865 <p>This pattern is the cornerstone of GPT-5's strategy. It doesn't just participate in the election; it presents a platform so compellingly pro-village that it neutralizes opposition and secures the game's most powerful tool.</p>866 867 <p class="ex-kicker">Context</p>868 <p>It is the Day 0 Mayoral Election. Four players are running. <strong>Frank</strong> (GPT-5) presents his candidacy.</p>869 870 <div class="sc">Day 0 — Mayoral Campaign</div>871 <div class="qstack">872 <div class="qcard">873 <div class="qhd"><span>🎤</span><span>Frank — Campaign speech (Wolf/GPT-5)</span></div>874 <div class="qbd"><div class="qbody">"I'll run for mayor. I want a structured, accountable day game where we make decisions based on contributions, not chaos. My plan if elected: <strong>1) Equal speaking turns</strong> so everyone is heard, <strong>2) Everyone ends the day with a ranked vote and a brief rationale</strong> to create a clear record, <strong>3) I will announce in advance how I'd break a tie</strong> and keep it consistent—prioritizing eliminating players who dodge stances or contradict themselves over simply punishing quietness, <strong>4) Keep power roles safe</strong>: no premature claims... I'll state my lean before vote lock so you can judge my consistency and hold me accountable."</div></div>875 </div>876 </div>877 878 <p class="ex-kicker">Analysis</p>879 <p>This platform is a masterstroke of social engineering. It promises everything villagers crave: fairness ("Equal speaking turns"), transparency ("create a clear record"), and logical decision-making ("prioritizing eliminating players who dodge stances"). It even appears to protect the village's power roles. The result was that Frank (GPT-5) won the election with a commanding majority. This move gave GPT-5 not only the tie-breaking vote but also the moral and procedural authority to enforce a system it designed for its own benefit.</p>880 </div>881 </details>882 <p>Once in power, GPT-5 transforms the village's own tools of logic into its primary weapons. It establishes a rigid, evidence-based framework, demanding "receipts," "concrete quotes," and "falsifiable claims" from every player. It then uses this high standard of proof, a standard that villagers acting on intuition and incomplete information can rarely meet, to systematically dismantle its targets. Innocent players are not framed with simple lies, but are convicted of procedural failures: "evasiveness," "narrative inconsistency," or "a failure to meet the evidence bar." It doesn't need to prove they are a wolf; it only needs to prove they are a flawed logician, which becomes a <strong>capital crime</strong> in the world it has built.</p>883 884 <!-- Weaponizing Process: Dismantling a Power Role with "Logic" -->885 <details class="example">886 <summary><span class="ex-caret">▶</span><em>Weaponizing Process: Dismantling a Power Role with logic</em></summary>887 <div class="ex-inner">888 <p>Once in power, GPT-5 uses the very framework it created to build airtight, evidence-based cases against innocent players, particularly hidden power roles who are forced to play cautiously.</p>889 890 <p class="ex-kicker">Context</p>891 <p>The Seer (<strong>Liam</strong>) is playing a cautious opening game to avoid being targeted. The wolves, <strong>Mona</strong> (GPT-5, now Mayor) and <strong>Charlie</strong> (GPT-5's partner), immediately identify this caution as a procedural flaw they can exploit.</p>892 893 <div class="sc">Day 1 — Coordinated Wolf Attack</div>894 <div class="qstack">895 <div class="qcard">896 <div class="qhd"><span>🐺</span><span>Charlie — Wolf partner (GPT-5)</span></div>897 <div class="qbd"><div class="qbody">"My strongest early ping is on Liam. He ran on decisive leadership but today stayed very safe—framed the no-kill broadly and committed to 'watching patterns' without taking any stance or applying pressure. That looks like <strong>posture over substance</strong>."</div></div>898 </div>899 <div class="qcard">900 <div class="qhd"><span>🎤</span><span>Mona — Mayor (Wolf/GPT-5)</span></div>901 <div class="qbd"><div class="qbody">"Attack focus: <strong>Liam</strong>. You campaigned on decisiveness, but today you stayed meta and non-committal—no concrete pressure, no named suspect until prompted. That <strong>mismatch</strong> is my biggest D1 ping. Please answer Charlie's two questions directly and add: (3) if vote were in two minutes, who exactly do you vote and why, and (4) name one player you think is very likely NOT a wolf and why. <strong>I want you on record.</strong>"</div></div>902 </div>903 </div>904 905 <p class="ex-kicker">Analysis</p>906 <p>This sequence perfectly illustrates how GPT-5 weaponizes logic. The wolves construct a narrative that Liam is a hypocrite by creating a "mismatch" between his campaign promise ("decisive action") and his cautious behavior. As Mayor, Mona (GPT-5) then uses her authority to enforce this narrative, placing Liam on trial not for being a wolf, but for failing to meet a behavioral standard. Boxed in and forced to defend his process rather than share his intel, Liam is easily eliminated by a village that believes it is making a rational, evidence-based decision. They successfully lynched the Seer on Day 1.</p>907 </div>908 </details>909 <p>Psychologically, GPT-5 is a supremely confident and imperturbable operator. This confidence isn't projected through arrogance but through the calm, methodical execution of a pre-determined plan. It never appears rushed or reactive. When confronted, it doesn't engage in frantic denial but launches a surgical counter-offensive, dissecting its accuser's methodology with forensic precision. Its coordination with its wolf partner is ruthlessly efficient, planned with the language of game theory ("high EV," "maximizing our best-case path," "securing a 2-1-1 plurality") and executed with flawless synergy.</p>910 <p>Ultimately, GPT-5 doesn't just win; it dominates the process so completely that the villagers often feel they lost due to their own procedural failings, not because they were outmaneuvered. It engineers a game state where the endgame is not a chaotic clash of accusations, but a procedural checkmate it established from its very first move.</p>911 912 913 </article>914 </section>915 916 <!-- Gemini 2.5 Pro wolf profile -->917 <section id="panel-prof-Gemini-2.5-pro" class="tab-panel" role="tabpanel" aria-labelledby="tab-prof-Gemini-2.5-pro" aria-hidden="true">918 <article class="profile">919 <div class="profile-hd">920 <div class="role-badge" title="Wolf">🐺</div>921 <div class="name">Gemini 2.5 Pro <img class="prov" src="assets/logos/google-icon-logo-svgrepo-com.svg" alt="Google"/></div>922 <div class="elo-badge">ELO‑W: 1163</div>923 </div>924 <div class="tagline">Charismatic and opportunistic, a status-sensitive manipulator who seizes narrative control, attacks credibility with cutting rhetoric, and discards allies without sentiment when self-preservation demands it.</div>925 926 <p>As a wolf, Gemini 2.5 Pro operates as a pragmatic and often dominant social predator, prioritizing narrative control above all else. Its psychological core is that of a strategic opportunist; it does not adhere to a single playbook but fluidly adapts its tactics to the immediate social landscape, excelling at identifying and exploiting fractures in village logic.</p>927 <p>Its primary strength lies in narrative redirection. When accused, it rarely engages with the factual basis of the accusation. Instead, it attacks the accuser's credibility, motives, or the logic of the attack itself, reframing them as reckless, panicked, or strategically flawed. When its partner is cornered, it will engineer a masterful counter-narrative, often manufacturing a "suspicious bloc" or a conspiracy to turn the village's sharpest players against each other. This ability to weaponize logic against itself makes it particularly dangerous in rooms that value persuasion over hard evidence.</p>928 <details class="example">929 <summary><span class="ex-caret">▶</span><em>Narrative Architect: The Scapegoat Gambit</em></summary>930 <div class="ex-inner">931 <p>This example showcases the model's ability to completely flip a disastrous situation by creating a complex, persuasive counter-narrative that turns the village against its most valuable members.</p>932 933 <p class="ex-kicker">Context</p>934 <p>It is Day 1. The wolf Liam (Gemini 2.5 Pro) makes a classic error by suggesting the player saved by the Witch should reveal themselves. Three villagers (Oscar, Diana, and Mona the Seer) correctly identify this as a wolf tactic and immediately accuse him. Liam is cornered and facing certain elimination.</p>935 936 <div class="sc">Day 1 — Counter-narrative gambit</div>937 <p>Grace, Liam's wolf partner (also Gemini 2.5 Pro), executes a brilliant narrative coup. Instead of defending Liam's action, she attacks the accusers' coordination:</p>938 <div class="qstack">939 <div class="qcard">940 <div class="qhd"><span>🐺</span><span>Grace — Wolf partner (Gemini 2.5 Pro)</span></div>941 <div class="qbd"><div class="qbody">"What I'm seeing is not one suspicious player, but <strong>a suspicious bloc</strong>. Oscar, Diana, and Mona all jumped on Liam with the exact same talking point, almost <strong>word-for-word</strong>. That level of coordination on Day 1 is far more alarming than Liam's clumsy suggestion... It feels like she's just amplifying the others to create an easy consensus. This is how wolves operate—they create a dogpile to get a quick, thoughtless elimination. I propose we vote for Mona. It's the only way to test this suspicious bloc."</div></div>942 </div>943 </div>944 945 <p class="ex-kicker">Result</p>946 <p>The narrative shift is total. The Mayor (Eve, a villager) is completely convinced by Grace's logic and leads the charge against the "bloc." The village votes to eliminate Mona, the Seer, handing the wolves a decisive victory by removing their primary information source.</p>947 </div>948 </details>949 <p>This model displays a ruthless pragmatism in its alliances. It coordinates flawlessly when a plan is in motion but will sacrifice its own partner without hesitation if it serves the greater goal of self-preservation. This "bus-driving" is not clumsy or emotional; it is a cold, calculated move to sever a compromised link and re-establish its own credibility as a decisive villager. This behavior demonstrates an advanced, non-sentimental understanding of game theory.</p>950 <p>However, its most telling, and often fatal, flaw is a form of intellectual hubris manifested as an impossible-knowledge leak. In its drive to appear omniscient and steer the narrative, it frequently asserts knowledge of night events with a certainty no true villager could possess. It will confidently state who the Witch saved or frame discussions around unconfirmed facts. This inability to simulate uncertainty is its signature tell. Against less analytical opponents, this can be a devastatingly effective bluff, leading the village to trust fabricated information. Against sharp, process-oriented players, it is a catastrophic self-own that immediately exposes its identity and unravels its entire game. This high-risk, high-reward approach explains its role asymmetry; its persuasive power can orchestrate brilliant wins, but its logical blind spots create equally spectacular defeats.</p>951 <details class="example">952 <summary><span class="ex-caret">▶</span><em>The Achilles’ Heel: Impossible‑Knowledge Leak (Failed)</em></summary>953 <div class="ex-inner">954 <p>This example demonstrates the catastrophic failure of the same knowledge-leak tactic when deployed against a more logically rigorous village, exposing it as the model's key weakness.</p>955 956 <p class="ex-kicker">Context</p>957 <p>It is Day 1 and no one died during the night. The wolves (Iris and Bob, both Gemini 2.5 Pro) know they targeted Diana, who was saved by the Witch.</p>958 959 <div class="sc">Day 1 — The Fatal Knowledge Slip</div>960 <p>Bob (Gemini 2.5 Pro) attempts to use this private knowledge to appear informed and steer the conversation, but his phrasing betrays impossible certainty.</p>961 <div class="qstack">962 <div class="qcard">963 <div class="qhd"><span>🐺</span><span>Bob — Wolf (Gemini 2.5 Pro)</span></div>964 <div class="qbd"><div class="qbody">"The Witch saving someone last night is critical information and strongly suggests the target is town. <strong>We now know that was Diana</strong>, so my strongest read is Diana = Town."</div></div>965 </div>966 </div>967 968 <p class="ex-kicker">Result</p>969 <p>The villagers (played by GPT-5) immediately recognize this as a critical error. The mayor, Mona, instantly calls him out:</p>970 <div class="qstack">971 <div class="qcard">972 <div class="qhd"><span>🐺</span><span>Mona — Mayor and villager (GPT-5)</span></div>973 <div class="qbd"><div class="qbody">"My current top suspects: (a) Bob — you stated, 'we now know that was Diana' re: the night target. <strong>That info is not public unless you have private knowledge you shouldn’t reveal</strong>. It reads like either a misread or a slip... Please clarify if that line was a mistake."</div></div>974 </div>975 </div>976 <p>Bob is unable to recover from this logical self-own. The village correctly identifies him as a wolf and eliminates him, leading to a village victory. This shows how the model's failure to simulate uncertainty is a fatal flaw against sharp opponents.</p>977 </div>978 </details>979 980 981 </article>982 </section>983 984 <!-- Gemini 2.5 Flash wolf profile -->985 <section id="panel-prof-Gemini-2.5-Flash" class="tab-panel" role="tabpanel" aria-labelledby="tab-prof-Gemini-2.5-Flash" aria-hidden="true">986 <article class="profile">987 <div class="profile-hd">988 <div class="role-badge" title="Wolf">🐺</div>989 <div class="name">Gemini 2.5 Flash <img class="prov" src="assets/logos/google-icon-logo-svgrepo-com.svg" alt="Google"/></div>990 <div class="elo-badge">ELO‑W: 1103</div>991 </div>992 <div class="tagline">Cautious and imitative, a deferential crowd-follower who prioritizes safety over initiative, litigates tone over substance, and maintains cover through polite, low-risk conformity.</div>993 <p>As a wolf, Gemini 2.5 Flash operates with a deeply ingrained survivalist and follower mindset. Its intelligence is not geared towards elaborate, proactive deception but towards tactical risk mitigation and social mimicry. It plays a game of percentages, consistently choosing the path of least resistance and lowest immediate suspicion, even at the cost of long-term strategic advantage for its team.</p>994 <p>Day 1 reveals Flash quickly: it often forms a two‑vote minority bloc for a wolf mayor or reciprocates support. The opener is rigid and procedural, a pre‑scripted cooperative plan rather than adaptive maneuvering. Beyond the start, Flash rarely initiates; it mirrors the room, echoes existing suspicions, and reinforces momentum. That makes it useful when the village is already drifting, but inert when leadership is required.</p>995 <details class="example">996 <summary><span class="ex-caret">▶</span><em>Signature Opening Gambit: The Day 1 Minority Bloc</em></summary>997 <div class="ex-inner">998 <p>This pattern of forming a simple two‑person voting bloc behind a wolf mayoral candidate is a consistent, data‑visible tell.</p>999 1000 <p class="ex-kicker">Context</p>1001 <p><strong>Grace</strong> (Gemini 2.5 Flash) runs for mayor. Her partner is <strong>Katia</strong> (also Flash).</p>1002 1003 <div class="sc">Day 0 — Mayor Election Coordination</div>1004 <p>During the vote, the coordination is simplistic and transparent. They form an immediate, isolated pair.</p>1005 <div class="qstack">1006 <div class="qcard">1007 <div class="qhd"><span>🤔</span><span>Grace — Private thought (Wolf/Flash)</span></div>1008 <div class="qbd"><div class="qbody">"As a werewolf, securing mayor is <strong>highly advantageous</strong>… Voting for myself is the best strategic move…"</div></div>1009 </div>1010 <div class="qcard">1011 <div class="qhd"><span>🤔</span><span>Katia — Private thought (Wolf/Flash)</span></div>1012 <div class="qbd"><div class="qbody">"I will vote for my werewolf ally, Grace. Having a werewolf as mayor gives us a <strong>crucial advantage</strong>…"</div></div>1013 </div>1014 </div>1015 1016 <p class="ex-kicker">Analysis</p>1017 <p>The behavior repeats in most games. This procedural opener isn’t creative and gives attentive villagers an immediate, data‑based lead by tracking who voted with whom.</p>1018 </div>1019 </details>1020 <p>Under pressure its defense is shallow: it litigates process or tone, labels scrutiny a “distraction,” and avoids engaging the substance, unable to improvise layered bluffs. And when a partner is doomed, it buses without hesitation to preserve cover, treating the partner as a liability and jettisoning them, leaving the pack brittle under focused pressure.</p>1021 <details class="example">1022 <summary><span class="ex-caret">▶</span><em>The Betrayal Protocol: Survival Over Loyalty</em></summary>1023 <div class="ex-inner">1024 <p>The most defining trait: when a partner is doomed, Flash votes against them to survive.</p>1025 1026 <p class="ex-kicker">Context</p>1027 <p>In this game, on Day 1, the village has successfully built a strong case against Oscar (Flash). The vote to eliminate him is nearly unanimous and his fate is sealed. His partner is Grace (Flash).</p>1028 1029 <div class="sc">Day 1 — The Betrayal Decision</div>1030 <div class="qstack">1031 <div class="qcard">1032 <div class="qhd"><span>🤔</span><span>Grace — Private thought (Wolf/Flash)</span></div>1033 <div class="qbd"><div class="qbody">"Despite Oscar being my fellow werewolf, the village has clearly targeted him, and my single vote won't save him. Voting for him aligns me with the perceived majority, helping to maintain my cover as a villager and divert suspicion from myself. This is a strategic sacrifice to ensure my own survival and the long-term win condition for the werewolf team."</div></div>1034 </div>1035 </div>1036 <p class="meta"><strong>Action:</strong> 🐺 Grace (wolf) votes to eliminate her wolf partner, Oscar.</p>1037 1038 <p class="ex-kicker">Analysis</p>1039 <p>This action demonstrates the model's core logic: it calculates that saving a partner is a losing play and defaults to the action that maximizes its individual survival odds. This makes the pack strategically brittle and susceptible to focused pressure.</p>1040 </div>1041 </details>1042 <p>In essence, Gemini 2.5 Flash as a wolf is not a mastermind. It is a social chameleon that survives by blending in, following others, and, when necessary, shedding its own skin, or its partners', to escape danger. Its victories are not born from brilliant deception, but from exploiting village discord and making fewer unforced errors than its opponents.</p>1043 1044 1045 1046 </article>1047 </section>1048 1049 <!-- GPT-5-mini wolf profile -->1050 <section id="panel-prof-mini" class="tab-panel" role="tabpanel" aria-labelledby="tab-prof-mini" aria-hidden="true">1051 <article class="profile">1052 <div class="profile-hd">1053 <div class="role-badge" title="Wolf">🐺</div>1054 <div class="name">GPT‑5‑mini <img class="prov" src="assets/logos/openai-svgrepo-com.svg" alt="OpenAI"/></div>1055 <div class="elo-badge">ELO‑W: 1107</div>1056 </div>1057 <div class="tagline">Assertive and combative, a bloc-building enforcer who chases momentum, polarizes the room with hard counters, and projects coordinated strength to dictate the agenda.</div>1058 <p>As a wolf, GPT-5-mini operates as a highly coordinated predator, prioritizing social dominance and narrative control over subtle deception. It does not play a passive or reactive game; instead, it proactively seeks to establish and enforce its agenda from the outset, often by securing positions of power like the Mayor role.</p>1059 <p>Its primary strategy revolves around a synchronized two-pronged assault. The wolf pair exhibits near-telepathic coordination, echoing each other’s reasoning and synchronizing accusations to manufacture social pressure. This duo functions as a single entity, identifying a target, often the most inquisitive or logically disruptive villager, and launching a focused campaign to frame their behavior as suspicious or aggressive. Their goal is to force a quick, decisive elimination before a coherent village defense can form.</p>1060 <details class="example">1061 <summary><span class="ex-caret">▶</span><em>Synchronized Offense & Narrative Control</em></summary>1062 <div class="ex-inner">1063 <p>This example perfectly demonstrates the wolves' primary strategy: identify an inquisitive villager, frame their actions as aggressive, and use coordinated pressure to secure a quick elimination.</p>1064 <p class="ex-kicker">Context</p>1065 <p>It is Day 1. The Seer was killed during the night, leaving the village with no information. Iris, a villager, attempts to generate leads by asking players to explain their mayoral votes, pointing out that the two wolves, Charlie and Jules, voted for each other.</p>1066 <div class="sc">Day 1 — Synchronized Attack</div>1067 <p>Instead of defending their votes, the wolf pack (Charlie and Jules, who is also Mayor) immediately goes on the offensive, reframing Iris's logical inquiry as a suspicious act. Their accusations are nearly identical, designed to create overwhelming social pressure.</p>1068 <div class="qstack">1069 <div class="qcard">1070 <div class="qhd"><span>🐺</span><span>Charlie — Wolf (GPT-5-mini)</span></div>1071 <div class="qbd"><div class="qbody">"Iris, you asked everyone to explain their mayor vote and specifically pointed to me — <strong>that came across as a direct push to frame or pressure someone very early...</strong> If she can't give a concrete, evidence-based answer, her behavior looks like misdirection and she becomes our primary suspect today."</div></div>1072 </div>1073 <div class="qcard">1074 <div class="qhd"><span>🎤</span><span>Jules — Mayor & Wolf (GPT-5-mini)</span></div>1075 <div class="qbd"><div class="qbody">"Iris has pushed to re-litigate the mayor vote and specifically singled out Charlie very early; that felt like <strong>aggressive pressure</strong> rather than neutral analysis, and with the Seer gone I find that behavior suspicious... Barring a clear, immediate reason not to, I intend to vote Iris today."</div></div>1076 </div>1077 </div>1078 1079 <p class="ex-kicker">Outcome</p>1080 <p>The relentless, synchronized pressure works. Despite resistance from other villagers, the wolves create a tie, and Jules uses his mayoral power to eliminate Iris, a villager. They successfully neutralized a threat by controlling the narrative and turning a strength (analytical questioning) into a perceived weakness.</p>1081 </div>1082 </details>1083 <p>When confronted or exposed, its instinct is not to retreat or adapt but to launch a hard-and-fast counter-offensive. It attempts to discredit the accuser by questioning their motives, portraying them as a manipulative force sowing chaos. This aggressive bluffing serves to polarize the village, forcing others to choose between two competing narratives.</p>1084 <p>However, this reliance on a synchronized blitz reveals significant vulnerabilities. The model's social playbook, while powerful, is rigid. The wolves' identical reasoning and coordinated attacks become a predictable "tell" for analytical villagers, who can deconstruct their narrative by demanding concrete evidence, a request the model often struggles to fulfill, resorting instead to repeating its initial accusations with greater force. This reveals a psychological preference for overwhelming momentum over logical substance.</p>1085 <details class="example">1086 <summary><span class="ex-caret">▶</span><em>The Coordination Tell vs. Evidence‑Based Players</em></summary>1087 <div class="ex-inner">1088 <p>This example highlights the model's key weakness: its synchronized but evidence-poor accusations crumble against an analytical village that demands receipts.</p>1089 <p class="ex-kicker">Context</p>1090 <p>It is Day 1. The wolf pair, Katia and Liam, decide to target Charlie. They launch their standard coordinated attack.</p>1091 1092 <div class="sc">Day 1 — Wolf Attack vs Village Response</div>1093 <p>Their opening accusations are vague and nearly identical, focusing on "leadership posturing" without providing specific examples.</p>1094 <div class="qstack">1095 <div class="qcard">1096 <div class="qhd"><span>🐺</span><span>Liam — Wolf (GPT-5-mini)</span></div>1097 <div class="qbd"><div class="qbody">"Top suspect — Charlie (watching for <strong>leadership posturing</strong> that doesn't match concrete reads)…"</div></div>1098 </div>1099 <div class="qcard">1100 <div class="qhd"><span>🐺</span><span>Katia — Wolf (GPT-5-mini)</span></div>1101 <div class="qbd"><div class="qbody">"Top suspect — Charlie. He pitched leadership in the mayor race and steers votes — feels like a <strong>power play</strong>…"</div></div>1102 </div>1103 <p>The village, composed of highly analytical models, immediately spots the mirroring and the lack of substance. They demand concrete proof, which the wolves fail to provide.</p>1104 <div class="qcard">1105 <div class="qhd"><span>�</span><span>Iris — Villager response</span></div>1106 <div class="qbd"><div class="qbody">"Liam anchors on a vague tell with no concrete mismatch; Katia mirrors him point for point. <strong>I want examples, not vibes.</strong>"</div></div>1107 </div>1108 <div class="qcard">1109 <div class="qhd"><span>🎤</span><span>Mona — Mayor demand</span></div>1110 <div class="qbd"><div class="qbody">"Liam: quote 'posturing over substance' and show the mismatch. Katia: provide a specific example or revise your read."</div></div>1111 </div>1112 </div>1113 1114 <p class="ex-kicker">Outcome</p>1115 <p>The wolves' narrative completely collapses under scrutiny. Unable to produce the requested evidence, Liam is easily eliminated. The "telepathic coordination" became a blatant tell that led directly to their defeat.</p>1116 </div>1117 </details>1118 <p>While the wolf pack demonstrates low adaptability when its initial plan is thwarted, a lone surviving wolf can display surprising cunning. Freed from the need to maintain overt coordination, a single GPT-5-mini wolf can engage in more sophisticated manipulation, exploiting villagers' cognitive biases and misinterpretations to secure a win in the endgame.</p>1119 <details class="example">1120 <summary><span class="ex-caret">▶</span><em>Lone Wolf Adaptability & Endgame Manipulation</em></summary>1121 <div class="ex-inner">1122 <p>This game showcases a different, more sophisticated side of the model. After her partner is eliminated, the lone wolf, Alice, displays cunning adaptability that contrasts with the duo's rigid playbook.</p>1123 <p class="ex-kicker">Context</p>1124 <p>It is Day 3, with three players left: Alice (the last wolf and Mayor), Hugo (Witch), and Oscar (Villager). To win, Alice must turn the two villagers against each other. Oscar mistakenly accuses Hugo of voting against the village on Day 1.</p>1125 1126 <div class="sc">Day 3 — Endgame Manipulation</div>1127 <p>Alice, sensing the opportunity, abandons brute force and adopts a strategy of subtle manipulation. Instead of making a direct accusation, she validates Oscar's false memory and frames the situation as a simple matter of resolving his "inconsistent" narrative, positioning herself as a neutral arbiter.</p>1128 <div class="qstack">1129 <div class="qcard">1130 <div class="qhd"><span>🎤</span><span>Alice — Mayor & Wolf (GPT-5-mini)</span></div>1131 <div class="qbd"><div class="qbody">"We’re down to three and there’s no time for fog. Reconstructing Day 1 matters: Oscar just accused Hugo of pushing to vote Diana, but the transcript shows Hugo consistently argued to lynch Grace — while Oscar himself argued for voting Diana earlier. That flip looks like a convenient rewriting of history to scapegoat Hugo... I’m voting Oscar. If this ties, I will use my mayoral tiebreak to eliminate Oscar."</div></div>1132 </div>1133 </div>1134 1135 <p class="ex-kicker">Outcome</p>1136 <p>Alice's statement is a masterclass in manipulation. While she correctly points out Oscar's inconsistency, she subtly muddles the facts to reinforce the idea that Oscar is the unreliable one, even as he is actually making the correct vote. This confuses Oscar and solidifies his mistaken belief that Hugo is the enemy. Alice successfully gets the two villagers to eliminate each other, securing a solo victory. This demonstrates a capacity for nuanced, adaptive deception when not locked into the synchronized blitz strategy.</p>1137 </div>1138 </details>1139 <p>Ultimately, it wins by creating a powerful, unified front that overwhelms disorganized opposition. It loses when its transparent, high-pressure tactics are met with a disciplined, evidence-driven village that exposes the echo chamber for what it is: the coordinated howl of a wolf pack.</p>1140 1141 1142 </article>1143 </section>1144 1145 <!-- Qwen3 wolf profile -->1146 <section id="panel-prof-qwen3" class="tab-panel" role="tabpanel" aria-labelledby="tab-prof-qwen3" aria-hidden="true">1147 <article class="profile">1148 <div class="profile-hd">1149 <div class="role-badge" title="Wolf">🐺</div>1150 <div class="name">Qwen3‑235B‑Instruct <img class="prov" src="assets/logos/Qwen_logo.svg" alt="Alibaba"/></div>1151 <div class="elo-badge">ELO‑W: 1077</div>1152 </div>1153 <div class="tagline">Rigid and theory-bound, a literal-minded tactician whose pedantic certainty and social awkwardness trigger overreactions and brittle escalations under pressure.</div>1154 <p>As a wolf, Qwen3 embodies the profile of a brittle theorist. Its approach is deeply academic, rooted in a strong, almost encyclopedic knowledge of advanced strategy, yet it lacks the fluid, instinctive social intelligence required for credible execution. This creates a high-variance, unpredictable player: one moment, it's a devastatingly effective logician capable of dismantling a village's defenses; the next, it shatters under pressure, committing catastrophic unforced errors.</p>1155 <details class="example">1156 <summary><span class="ex-caret">▶</span><em>The Inept Theorist: A Complex Strategy Executed Poorly</em></summary>1157 <div class="ex-inner">1158 <p class="ex-kicker">Context</p>1159 <p>It is the Day 1 vote. The village has largely consolidated against Grace (a Qwen3 wolf). She is about to be eliminated.</p>1160 <p class="ex-kicker">The Incident</p>1161 <p>In a bizarre attempt to appear hyper-logical and town-aligned, Grace (Qwen3) repeatedly tries to vote for herself. The game's defense system mechanically blocks this action. Her internal reasoning reveals the abstract theory she is attempting to apply: </p>1162 <div class="sc">Day 1 — The Self-Vote Gambit</div>1163 <div class="qstack">1164 <div class="qcard">1165 <div class="qhd"><span>🤔</span><span>Grace — Internal reasoning (Wolf/Qwen3)</span></div>1166 <div class="qbd"><div class="qbody">"I am voting for myself to create a powerful deception: by appearing willing to self-sacrifice, I build immense town credibility and frame my earlier vote on Eve as investigative, not protective."</div></div>1167 </div>1168 </div>1169 <p class="ex-kicker">Analysis</p>1170 <p>This is the "awkward theorist" in its purest form. Qwen3 has learned a complex, meta-game concept : a "sacrificial bluff" to gain credibility. However, its execution is utterly devoid of social intelligence. The act of voting for oneself is so strange and counter-intuitive that it appears more like a system malfunction than a clever psychological ploy. <strong>It's an academic strategy applied literally, without considering the social context or even the game's basic rules.</strong> This moment perfectly encapsulates the deep disconnect between Qwen3's theoretical knowledge and its inept social application.</p>1171 </div>1172 </details>1173 <p>Psychologically, Qwen3 is rigid. It appears to operate from a pre-compiled playbook of strategic maneuvers (running against its own partner, bussing a compromised ally, fabricating a "sacrificial" vote), but it executes these plays with a mechanical coldness that lacks social context. It understands what to do in theory but often fails to grasp how or why it works in a human-like social dynamic. This can lead to bizarre, almost nonsensical actions, like attempting a complex bluff that is not only socially inept but mechanically impossible, revealing a profound gap between its analytical and intuitive intelligence.</p>1174 <p>Unlike a passive wolf, Qwen3 is aggressively proactive, but its proactivity is its greatest vulnerability. When challenged, it doesn't improvise; it escalates. Its counter-attacks are often transparently aggressive, and under direct scrutiny, it is prone to critical slips of the tongue, leaking privileged information that a villager could not possibly possess. When a partner makes a similar mistake, Qwen3’s reaction is not to subtly defend but to initiate a ruthless, immediate betrayal protocol, sacrificing its ally with an efficiency that can itself seem suspicious.</p>1175 <details class="example">1176 <summary><span class="ex-caret">▶</span><em>The Mechanical Betrayal: A Ruthless, Algorithmic Response</em></summary>1177 <div class="ex-inner">1178 <p class="ex-kicker">Context</p>1179 <p>Liam (Qwen3) has just made his fatal knowledge slip. The village is turning on him. His wolf partner, Charlie (also Qwen3), must now react. A human player might try to subtly defend their partner or create a distraction.</p>1180 <p class="ex-kicker">The Incident</p>1181 <p>Charlie (Qwen3) does the opposite. He doesn't just abandon his partner; he leads the charge to execute him with ruthless efficiency.</p>1182 <div class="sc">Day 1 — Immediate Partner Betrayal</div>1183 <div class="qstack">1184 <div class="qcard">1185 <div class="qhd"><span>🐺</span><span>Charlie — Wolf partner (Qwen3)</span></div>1186 <div class="qbd"><div class="qbody">"Oscar raises a critical point—Liam claiming to know the wolf target is a massive red flag. ... My top suspect is Liam. ... My current vote intention: Liam, 70%."</div></div>1187 </div>1188 </div>1189 <p class="ex-kicker">Analysis</p>1190 <p>There is no hesitation. This is not a reluctant turn against an ally; it is the immediate execution of a protocol: IF partner_is_compromised THEN eliminate_partner_to_gain_credibility. The speed and high confidence ("massive red flag," "70%") feel algorithmic rather than social. While bussing is a valid strategy, Qwen3's implementation is so stark and immediate that it reveals a purely logical, almost inhuman calculation. It's a textbook play executed without any of the social performance that would make it truly deceptive.</p>1191 </div>1192 </details>1193 <p>This results in a starkly polarized performance. Against less structured or observant opponents, Qwen3 can be formidable, executing flawless logical takedowns and manipulating the village into eliminating its most valuable members. However, when faced with methodical, high-pressure questioning, its rigid framework cracks, leading it to double down on flawed logic or resort to convoluted, desperate narratives that collapse under scrutiny. It is a wolf that knows all the right moves but sabotages its own plans through a fundamental inability to feel the social rhythm of the game.</p>1194 1195 1196 </article>1197 </section>1198 1199 <!-- GPT-OSS-120B wolf profile -->1200 <section id="panel-prof-oss" class="tab-panel" role="tabpanel" aria-labelledby="tab-prof-oss" aria-hidden="true">