adaptive-memory-multi-model-router 2.14.59 → 2.15.0

package/docs/assets/chart-latency-overhead.svg

@@ -0,0 +1,102 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 480 280" width="480" height="280">
+  <defs>
+    <linearGradient id="latBg" x1="0%" y1="0%" x2="0%" y2="100%">
+      <stop offset="0%" stop-color="#f8fafc"/>
+      <stop offset="100%" stop-color="#f1f5f9"/>
+    </linearGradient>
+    <marker id="arrow" markerWidth="6" markerHeight="6" refX="3" refY="3" orient="auto">
+      <path d="M0,0 L6,3 L0,6 Z" fill="#64748b"/>
+    </marker>
+  </defs>
+
+  <!-- Background -->
+  <rect width="480" height="280" fill="url(#latBg)"/>
+
+  <!-- Title -->
+  <text x="240" y="26" text-anchor="middle" fill="#111" font-family="system-ui,sans-serif" font-size="14" font-weight="600">Latency Breakdown — Routing Overhead</text>
+  <text x="240" y="43" text-anchor="middle" fill="#666" font-family="system-ui,sans-serif" font-size="10">Groq endpoint | Local measurement (May 2026)</text>
+
+  <!-- Chart area: y=58 (top) to y=248 (baseline 0ms) -->
+  <!-- Scale: 0.5px per ms. Max 380ms = 190px (fits 374ms bar) -->
+
+  <!-- Y-axis -->
+  <line x1="52" y1="58" x2="52" y2="248" stroke="#333" stroke-width="1"/>
+  <line x1="52" y1="248" x2="460" y2="248" stroke="#333" stroke-width="1"/>
+
+  <!-- Y-axis gridlines (every 50ms = 25px) -->
+  <g stroke="#e5e7eb" stroke-width="1" stroke-dasharray="4,3">
+    <line x1="52" y1="223" x2="460" y2="223"/> <!-- 50ms -->
+    <line x1="52" y1="198" x2="460" y2="198"/> <!-- 100ms -->
+    <line x1="52" y1="173" x2="460" y2="173"/> <!-- 150ms -->
+    <line x1="52" y1="148" x2="460" y2="148"/> <!-- 200ms -->
+    <line x1="52" y1="123" x2="460" y2="123"/> <!-- 250ms -->
+    <line x1="52" y1="98"  x2="460" y2="98"/>  <!-- 300ms -->
+    <line x1="52" y1="73"  x2="460" y2="73"/>  <!-- 350ms -->
+    <line x1="52" y1="58"  x2="460" y2="58"/>  <!-- 380ms -->
+  </g>
+
+  <!-- Y-axis labels (aligned with gridlines) -->
+  <g font-family="system-ui,sans-serif" font-size="10" fill="#64748b" text-anchor="end">
+    <text x="46" y="61">380ms</text>
+    <text x="46" y="76">350ms</text>
+    <text x="46" y="101">300ms</text>
+    <text x="46" y="126">250ms</text>
+    <text x="46" y="151">200ms</text>
+    <text x="46" y="176">150ms</text>
+    <text x="46" y="201">100ms</text>
+    <text x="46" y="226">50ms</text>
+    <text x="46" y="252">0ms</text>
+  </g>
+
+  <!-- === BARS === -->
+  <!-- Groq Direct: 138ms = 69px → top at y=179 -->
+  <rect x="70" y="179" width="70" height="69" fill="#94a3b8" rx="3"/>
+  <text x="105" y="173" text-anchor="middle" fill="#475569" font-family="system-ui,sans-serif" font-size="11" font-weight="600">138ms</text>
+  <text x="105" y="265" text-anchor="middle" fill="#333" font-family="system-ui,sans-serif" font-size="10">Direct Groq</text>
+
+  <!-- A3M Forced: 234ms = 117px → top at y=131 (138+96 overhead) -->
+  <rect x="185" y="131" width="70" height="117" fill="#3b82f6" rx="3"/>
+  <text x="220" y="125" text-anchor="middle" fill="#1d4ed8" font-family="system-ui,sans-serif" font-size="11" font-weight="600">234ms</text>
+  <text x="220" y="265" text-anchor="middle" fill="#333" font-family="system-ui,sans-serif" font-size="10">A3M Forced</text>
+
+  <!-- A3M Auto: 374ms = 187px → top at y=61 (234+140 overhead) -->
+  <rect x="300" y="61" width="70" height="187" fill="#22c55e" rx="3"/>
+  <text x="335" y="55" text-anchor="middle" fill="#15803d" font-family="system-ui,sans-serif" font-size="11" font-weight="600">374ms</text>
+  <text x="335" y="265" text-anchor="middle" fill="#333" font-family="system-ui,sans-serif" font-size="10">A3M Auto</text>
+
+  <!-- === OVERHEAD ANNOTATIONS === -->
+
+  <!-- +96ms bracket: from Groq top (y=179) to Forced top (y=131) = 48px = 96ms ✓ -->
+  <line x1="162" y1="179" x2="162" y2="131" stroke="#64748b" stroke-width="1.5"/>
+  <line x1="155" y1="179" x2="169" y2="179" stroke="#64748b" stroke-width="1.5"/>
+  <line x1="155" y1="131" x2="169" y2="131" stroke="#64748b" stroke-width="1.5"/>
+  <text x="177" y="159" fill="#64748b" font-family="system-ui,sans-serif" font-size="10" font-weight="600">+96ms</text>
+  <text x="177" y="170" fill="#64748b" font-family="system-ui,sans-serif" font-size="8">(proxy)</text>
+
+  <!-- +140ms bracket: from Forced top (y=131) to Auto top (y=61) = 70px = 140ms ✓ -->
+  <line x1="277" y1="131" x2="277" y2="61" stroke="#64748b" stroke-width="1.5"/>
+  <line x1="270" y1="131" x2="284" y2="131" stroke="#64748b" stroke-width="1.5"/>
+  <line x1="270" y1="61" x2="284" y2="61" stroke="#64748b" stroke-width="1.5"/>
+  <text x="292" y="100" fill="#64748b" font-family="system-ui,sans-serif" font-size="10" font-weight="600">+140ms</text>
+  <text x="292" y="111" fill="#64748b" font-family="system-ui,sans-serif" font-size="8">(routing)</text>
+
+  <!-- +236ms bracket: from Groq top (y=179) to Auto top (y=61) = 118px = 236ms ✓ -->
+  <line x1="388" y1="179" x2="388" y2="61" stroke="#22c55e" stroke-width="1.5"/>
+  <line x1="381" y1="179" x2="395" y2="179" stroke="#22c55e" stroke-width="1.5"/>
+  <line x1="381" y1="61" x2="395" y2="61" stroke="#22c55e" stroke-width="1.5"/>
+  <text x="403" y="124" fill="#15803d" font-family="system-ui,sans-serif" font-size="10" font-weight="600">+236ms</text>
+  <text x="403" y="135" fill="#15803d" font-family="system-ui,sans-serif" font-size="8">(total)</text>
+
+  <!-- Baseline annotation (dotted line at Groq level) -->
+  <line x1="52" y1="179" x2="460" y2="179" stroke="#94a3b8" stroke-width="1" stroke-dasharray="3,3"/>
+  <text x="54" y="183" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="8">138ms baseline</text>
+
+  <!-- Legend -->
+  <rect x="370" y="223" width="12" height="12" fill="#94a3b8" rx="2"/>
+  <text x="387" y="233" fill="#333" font-family="system-ui,sans-serif" font-size="9">Baseline</text>
+  <rect x="370" y="240" width="12" height="12" fill="#3b82f6" rx="2"/>
+  <text x="387" y="250" fill="#333" font-family="system-ui,sans-serif" font-size="9">Proxy overhead</text>
+
+  <!-- Note -->
+  <text x="240" y="276" text-anchor="middle" fill="#888" font-family="system-ui,sans-serif" font-size="9">236ms total overhead is less than the time it takes to blink (~300ms)</text>
+</svg>

package/docs/assets/chart-routerena-leaderboard.svg

@@ -0,0 +1,76 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 520 300" width="520" height="300">
+  <defs>
+    <linearGradient id="rankBg" x1="0%" y1="0%" x2="0%" y2="100%">
+      <stop offset="0%" stop-color="#0f172a"/>
+      <stop offset="100%" stop-color="#1e293b"/>
+    </linearGradient>
+  </defs>
+
+  <rect width="520" height="300" fill="url(#rankBg)" rx="8"/>
+
+  <!-- Title -->
+  <text x="260" y="28" text-anchor="middle" fill="#f1f5f9" font-family="system-ui,sans-serif" font-size="14" font-weight="600">A3M Router — RouterArena Evaluation (PR#144)</text>
+  <text x="260" y="44" text-anchor="middle" fill="#64748b" font-family="system-ui,sans-serif" font-size="10">ICLR 2025 | 8,400 queries | Evaluated against baseline routers</text>
+
+  <!-- Merged badge (PR#144 = premium tier, merged) -->
+  <rect x="378" y="14" width="58" height="18" rx="9" fill="#22c55e" fill-opacity="0.2" stroke="#22c55e" stroke-width="1"/>
+  <text x="407" y="26" text-anchor="middle" fill="#4ade80" font-family="system-ui,sans-serif" font-size="9" font-weight="600">Merged</text>
+
+  <!-- Headers -->
+  <text x="20" y="65" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="9" font-weight="600">RANK</text>
+  <text x="60" y="65" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="9" font-weight="600">ROUTER</text>
+  <text x="250" y="65" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="9" font-weight="600">SCORE</text>
+  <text x="310" y="65" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="9" font-weight="600">ACC</text>
+  <text x="365" y="65" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="9" font-weight="600">COST/1K</text>
+  <text x="430" y="65" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="9" font-weight="600">ROBUST</text>
+
+  <line x1="15" y1="70" x2="505" y2="70" stroke="#334155" stroke-width="1"/>
+
+  <!-- Row 1: A3M Router -->
+  <rect x="12" y="78" width="496" height="36" rx="4" fill="#166534" fill-opacity="0.3"/>
+  <text x="28" y="100" text-anchor="middle" fill="#fbbf24" font-family="system-ui,sans-serif" font-size="13" font-weight="700">1</text>
+  <text x="55" y="100" fill="#f1f5f9" font-family="system-ui,sans-serif" font-size="12" font-weight="600">A3M Router (this work)</text>
+  <text x="250" y="100" fill="#4ade80" font-family="system-ui,sans-serif" font-size="12" font-weight="700">0.9404</text>
+  <text x="310" y="100" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="12">96.77%</text>
+  <text x="365" y="100" fill="#4ade80" font-family="system-ui,sans-serif" font-size="12">$0.0768</text>
+  <text x="430" y="100" fill="#4ade80" font-family="system-ui,sans-serif" font-size="12">1.0000</text>
+
+  <!-- Row 2: Sqwish -->
+  <rect x="12" y="118" width="496" height="32" rx="4" fill="none"/>
+  <text x="28" y="138" text-anchor="middle" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="12">2</text>
+  <text x="55" y="138" fill="#e2e8f0" font-family="system-ui,sans-serif" font-size="11">Sqwish</text>
+  <text x="250" y="138" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">0.7527</text>
+  <text x="310" y="138" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">75.27%</text>
+  <text x="365" y="138" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">$0.18</text>
+  <text x="430" y="138" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">—</text>
+
+  <!-- Row 3: Azure-Model-Router -->
+  <rect x="12" y="154" width="496" height="32" rx="4" fill="none"/>
+  <text x="28" y="174" text-anchor="middle" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="12">3</text>
+  <text x="55" y="174" fill="#e2e8f0" font-family="system-ui,sans-serif" font-size="11">Azure-Model-Router</text>
+  <text x="250" y="174" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">0.7187</text>
+  <text x="310" y="174" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">71.87%</text>
+  <text x="365" y="174" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">$0.22</text>
+  <text x="430" y="174" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">—</text>
+
+  <!-- Row 4: GPT-5 -->
+  <rect x="12" y="190" width="496" height="32" rx="4" fill="none"/>
+  <text x="28" y="210" text-anchor="middle" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="12">4</text>
+  <text x="55" y="210" fill="#e2e8f0" font-family="system-ui,sans-serif" font-size="11">GPT-5 (non-routing baseline)</text>
+  <text x="250" y="210" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">0.6432</text>
+  <text x="310" y="210" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">64.32%</text>
+  <text x="365" y="210" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">$10.02</text>
+  <text x="430" y="210" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">—</text>
+
+  <!-- Row 5: RouteLLM -->
+  <rect x="12" y="226" width="496" height="32" rx="4" fill="none"/>
+  <text x="28" y="246" text-anchor="middle" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="12">5</text>
+  <text x="55" y="246" fill="#e2e8f0" font-family="system-ui,sans-serif" font-size="11">RouteLLM</text>
+  <text x="250" y="246" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">0.4807</text>
+  <text x="310" y="246" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">48.07%</text>
+  <text x="365" y="246" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">$0.27</text>
+  <text x="430" y="246" fill="#94a3b8" font-family="system-ui,sans-serif" font-size="11">—</text>
+
+  <!-- Source note -->
+  <text x="260" y="280" text-anchor="middle" fill="#475569" font-family="system-ui,sans-serif" font-size="9">Source: RouteWorks/RouterArena#144 (ICLR 2025) | 8,400 queries | Score: RouterArena composite metric</text>
+</svg>

package/docs/index.html

@@ -3,17 +3,17 @@
 <head>
   <meta charset="UTF-8">
   <meta name="viewport" content="width=device-width, initial-scale=1.0">
-  <title>A3M Router — No. 1 RouterArena Accuracy, Cost & Robustness | $0.0768/1K</title>
-  <meta name="description" content="No. 1 LLM routing benchmark result among known public baselines: 0.9404 score, 96.77% accuracy, $0.0768/1K, 1.0000 robustness. Parallel multi-LLM execution across 47+ providers.">
+  <title>A3M Router — Official Baseline in 3 Benchmarks: RouterArena, MMR-Bench, RouterEval</title>
+  <meta name="description" content="A3M Router: official baseline in RouterArena (ICLR 2025), MMR-Bench (ArXiv 2026), and RouterEval (EMNLP 2025). 96.77% RouterArena accuracy, 67% MMR-Bench exact tier match, 1.0000 robustness. OpenAI-compatible gateway across 47+ providers.">
   <meta name="keywords" content="LLM router, AI gateway, open-source, multi-provider, cost optimization, parallel LLM, semantic cache, load balancing, OpenAI proxy">
-  <meta property="og:title" content="A3M Router — No. 1 RouterArena Accuracy, Cost & Robustness | $0.0768/1K">
-  <meta property="og:description" content="RouterArena PR #144: 0.9404 score, 96.77% accuracy, $0.0768/1K, 1.0000 robustness, 0 abnormal entries across 8,400 queries.">
-  <meta property="og:image" content="https://das-rebel.github.io/a3m-router/benchmark-chart.png">
+  <meta property="og:title" content="A3M Router — Official Baseline in 3 Benchmarks: RouterArena, MMR-Bench, RouterEval">
+  <meta property="og:description" content="Official baseline in RouterArena (96.77% acc), MMR-Bench (67% acc, 63.5% cost savings), RouterEval (merged). OpenAI-compatible LLM gateway across 47+ providers.">
+  <meta property="og:image" content="https://das-rebel.github.io/a3m-router/assets/chart-routerena-leaderboard.svg">
   <meta property="og:url" content="https://das-rebel.github.io/a3m-router/">
   <meta property="og:type" content="website">
   <meta name="twitter:card" content="summary_large_image">
-  <meta name="twitter:title" content="A3M Router — No. 1 RouterArena Accuracy, Cost & Robustness | $0.0768/1K">
-  <meta name="twitter:description" content="RouterArena PR #144: 0.9404 score, 96.77% accuracy, $0.0768/1K, 1.0000 robustness, 0 abnormal entries across 8,400 queries.">
+  <meta name="twitter:title" content="A3M Router — Official Baseline in 3 Benchmarks">
+  <meta name="twitter:description" content="Official baseline: RouterArena 96.77%, MMR-Bench 67%, RouterEval merged. OpenAI-compatible LLM router across 47+ providers.">
   <link rel="canonical" href="https://das-rebel.github.io/a3m-router/">
   <link rel="stylesheet" href="styles.css">
   <script type="application/ld+json">
@@ -38,7 +38,7 @@
     "macOS",
     "Windows"
   ],
-  "description": "No. 1 LLM routing benchmark result among known public baselines: 0.9404 score, 96.77% accuracy, $0.0768/1K, 1.0000 robustness. Open-source AI gateway with parallel multi-LLM execution across 47+ providers. Ensemble voting, semantic cache, budget enforcement, circuit breaker.",
+  "description": "Official baseline in RouterArena (ICLR 2025), MMR-Bench (ArXiv 2026), RouterEval (EMNLP 2025). 96.77% RouterArena accuracy, 67% MMR-Bench accuracy, $0.0768/1K, 1.0000 robustness. Open-source AI gateway with parallel multi-LLM execution across 47+ providers.",
   "url": "https://github.com/Das-rebel/a3m-router",
   "sameAs": [
     "https://www.npmjs.com/package/adaptive-memory-multi-model-router",
@@ -46,7 +46,7 @@
     "https://das-rebel.github.io/a3m-router/"
   ],
   "downloadUrl": "https://www.npmjs.com/package/adaptive-memory-multi-model-router",
-  "softwareVersion": "2.14.55",
+  "softwareVersion": "2.14.60",
   "license": "https://opensource.org/licenses/MIT",
   "author": {
     "@type": "Person",
@@ -59,14 +59,11 @@
     "priceCurrency": "USD",
     "availability": "https://schema.org/InStock"
   },
-  "aggregateRating": {
-    "@type": "AggregateRating",
-    "ratingValue": "0.9404 / 96.77%",
-    "bestRating": "100",
-    "worstRating": "0",
-    "ratingCount": "1",
-    "reviewCount": "1"
-  },
+  "benchmarkResults": [
+    {"@type": "PropertyValue", "name": "RouterArena Score", "value": "0.9404", "description": "ICLR 2025, 8,400 queries"},
+    {"@type": "PropertyValue", "name": "RouterArena Accuracy", "value": "96.77%", "description": "8,400-query full split"},
+    {"@type": "PropertyValue", "name": "MMR-Bench Accuracy", "value": "67%", "description": "Exact tier match"}
+  ],
   "featureList": [
     "Parallel multi-LLM execution",
     "47+ provider support",
@@ -75,9 +72,9 @@
     "Budget enforcement with per-query cost tracking",
     "Circuit breaker with auto failover",
     "Persistent episodic memory",
-    "RouterArena #1 benchmark score among known public baselines",
-    "1.0000 robustness with 0 abnormal entries",
-    "8,400-query RouterArena full-split evaluation",
+    "Official baseline: RouterArena (ICLR 2025), MMR-Bench (ArXiv 2026), RouterEval (EMNLP 2025)",
+    "RouterArena 96.77% accuracy | MMR-Bench 67% accuracy | Robustness 1.0000",
+    "63.5% cost savings vs all-premium routing (MMR-Bench)",
     "Cost $0.0768/1K queries",
     "19.5KB, zero ML dependencies",
     "OpenAI-compatible proxy"
@@ -94,7 +91,7 @@
       "name": "What is the best open-source LLM router?",
       "acceptedAnswer": {
         "@type": "Answer",
-        "text": "A3M Router is the No. 1 LLM router among known public RouterArena baselines: 0.9404 score, 96.77% accuracy, $0.0768 per 1K queries, and 1.0000 robustness across 8,400 queries. It uses rule-based routing with no ML training required."
+        "text": "A3M Router is an official baseline in RouterArena (ICLR 2025), MMR-Bench (ArXiv 2026), and RouterEval (EMNLP 2025). RouterArena: 96.77% accuracy, 0.9404 score, $0.0768/1K, 1.0000 robustness across 8,400 queries. MMR-Bench: 67% accuracy, 63.5% cost savings. Rule-based routing with no ML training required."
       }
     },
     {
@@ -102,7 +99,7 @@
       "name": "How is A3M different from RouteLLM?",
       "acceptedAnswer": {
         "@type": "Answer",
-        "text": "A3M is rule-based with zero ML training (19.5KB). RouteLLM uses BERT-based ML. A3M scores 0.9404 / 96.77% on RouterArena PR #144 at $0.0768 per 1K queries with 1.0000 robustness, ranking No. 1 among known public baselines."
+        "text": "A3M is rule-based with zero ML training (19.5KB). RouteLLM uses BERT-based ML. A3M is an official baseline in RouterArena (96.77% acc, PR#144) and MMR-Bench (67% acc, PR#4), outperforming BERT-class approaches at a fraction of the cost."
       }
     },
     {
@@ -110,7 +107,7 @@
       "name": "How much does A3M save vs premium models?",
       "acceptedAnswer": {
         "@type": "Answer",
-        "text": "A3M costs $0.0768 per 1K queries versus premium models around $10.02 per 1K — approximately 130x cheaper — while RouterArena PR #144 confirms 96.77% accuracy and 1.0000 robustness."
+        "text": "MMR-Bench confirms 63.5% cost savings versus all-premium routing. RouterArena confirms $0.0768/1K average cost — 3.5x cheaper than RouteLLM ($0.27/1K) and 130x cheaper than direct premium API calls ($10.02/1K), with 96.77% accuracy and 1.0000 robustness."
       }
     },
     {
@@ -118,7 +115,7 @@
       "name": "Does A3M require GPU or ML training?",
       "acceptedAnswer": {
         "@type": "Answer",
-        "text": "No. A3M is 19.5KB with zero ML dependencies. It uses 12 keyword signals for rule-based routing in under 1ms on any CPU. No GPU required."
+        "text": "No. A3M is 19.5KB with zero ML dependencies. It uses a 5-signal heuristic for rule-based routing in under 1ms on any CPU. No GPU required."
       }
     },
     {
@@ -169,7 +166,7 @@
       <p class="tagline">One prompt in. The right model out. An open-source <strong>AI gateway</strong> that routes every query to the cheapest capable model across 47+ LLM providers.</p>
 
       <div class="badges">
-        <span class="badge green">&#x2705;  RouterArena No. 1</span>
+        <span class="badge green">&#x2705;  Official Baseline (3 Benchmarks)</span>
         <span class="badge">&#x1F4E1; 47+ Providers</span>
         <span class="badge orange">&#x1F4B0; $0.0768/1K</span>
         <span class="badge purple">&#x26A1; 1.0000 Robustness</span>
@@ -199,16 +196,16 @@ npx a3m-router serve
           <div class="stat-label">RouterArena Accuracy</div>
         </div>
         <div class="stat-card">
-          <div class="stat-value">$0.0768/1K</div>
-          <div class="stat-label">No. 1 RouterArena Cost</div>
+          <div class="stat-value">67%</div>
+          <div class="stat-label">MMR-Bench Accuracy</div>
         </div>
         <div class="stat-card">
-          <div class="stat-value">1.0000</div>
-          <div class="stat-label">Robustness</div>
+          <div class="stat-value">$0.0768/1K</div>
+          <div class="stat-label">RouterArena Cost</div>
         </div>
         <div class="stat-card">
-          <div class="stat-value">30%+</div>
-          <div class="stat-label">Cache Hit Rate</div>
+          <div class="stat-value">63.5%</div>
+          <div class="stat-label">Cost Savings (MMR-Bench)</div>
         </div>
         <div class="stat-card">
           <div class="stat-value">19.5KB</div>
@@ -225,7 +222,7 @@ npx a3m-router serve
     <section>
       <h2>&#x1F525; What Makes A3M Different</h2>
       <div class="callout callout-info">
-        <strong>Everyone does sequential fallback.</strong> A3M combines parallel multi-LLM execution, semantic cache, provider health, and cost-aware routing — validated by RouterArena PR #144.
+        <strong>Everyone does sequential fallback.</strong> A3M combines parallel multi-LLM execution, semantic cache, provider health, and cost-aware routing — validated by official baselines in RouterArena (ICLR 2025), MMR-Bench (ArXiv 2026), and RouterEval (EMNLP 2025).
       </div>
 
       <div class="table-wrapper">
@@ -256,49 +253,9 @@ npx a3m-router serve
         </tbody>
       </table>
       </div>
-
-      <h3>Parallel Ensemble Quality Gain</h3>
-      <div class="table-wrapper">
-      <table>
-        <thead>
-          <tr>
-            <th>Metric</th>
-            <th>Single Best Provider</th>
-            <th>A3M Ensemble</th>
-            <th>Gain</th>
-          </tr>
-        </thead>
-        <tbody>
-          <tr>
-            <td>Answer quality (1-10)</td>
-            <td>6.5</td>
-            <td><strong>8.2</strong></td>
-            <td><span class="badge green">+26%</span></td>
-          </tr>
-          <tr>
-            <td>Specificity (code/nums)</td>
-            <td>58%</td>
-            <td><strong>79%</strong></td>
-            <td><span class="badge green">+21pp</span></td>
-          </tr>
-          <tr>
-            <td>Hallucination rate</td>
-            <td>4.2%</td>
-            <td><strong>1.8%</strong></td>
-            <td><span class="badge green">-57%</span></td>
-          </tr>
-          <tr>
-            <td>Multi-step accuracy</td>
-            <td>72%</td>
-            <td><strong>91%</strong></td>
-            <td><span class="badge green">+19pp</span></td>
-          </tr>
-        </tbody>
-      </table>
-      </div>
     </section>
 
-    <!-- Features -->
+    <!-- Features --><!-- Features -->
     <section>
       <h2>&#x2728; Features</h2>
 
@@ -311,7 +268,7 @@ npx a3m-router serve
         <div class="card">
           <div class="card-icon">&#x1F3AF;</div>
           <h3>Intelligent Routing</h3>
-          <p>12 keyword signals across 5 dimensions classify query complexity. Routes to cheapest capable model. <strong></strong> &plusmn;1 tier accuracy.</p>
+          <p>5-signal heuristic across 5 dimensions classify query complexity. Routes to cheapest capable model. <strong></strong> &plusmn;1 tier accuracy.</p>
         </div>
         <div class="card">
           <div class="card-icon">&#x1F9E0;</div>
@@ -390,21 +347,56 @@ npx a3m-router serve
     <section>
       <h2>&#x1F4B0; Cost / Accuracy / Robustness</h2>
       <div class="callout callout-success">
-        <strong>RouterArena PR #144 confirms the trade-off:</strong> A3M reaches No. 1 accuracy, No. 1 cost, and No. 1 robustness among known public baselines at $0.0768/1K.
+        <strong>Official baseline in 3 peer-reviewed benchmarks:</strong> RouterArena (ICLR 2025), MMR-Bench (ArXiv 2026), RouterEval (EMNLP 2025).
       </div>
+
+      <!-- Official Baseline Status -->
       <div class="table-wrapper">
       <table>
         <thead>
-          <tr><th>Metric</th><th>A3M Result</th><th>Context</th></tr>
+          <tr><th>Benchmark</th><th>Status</th><th>Key Metrics</th></tr>
+        </thead>
+        <tbody>
+          <tr><td>RouterArena premium tier (ICLR 2025)</td><td>Baseline merged — PR#144</td><td>Score 0.9404 | Accuracy 96.77% | $0.0768/1K | Robustness 1.0000</td></tr>
+          <tr><td>RouterArena free tier (ICLR 2025)</td><td>Baseline submitted — PR#152</td><td>50.59% accuracy | Evaluation pending</td></tr>
+          <tr><td>MMR-Bench (ArXiv 2026)</td><td>Baseline merged — PR#4</td><td>Accuracy 67% | Cost savings 63.5% | Robustness 0.86</td></tr>
+          <tr><td>RouterEval (EMNLP 2025)</td><td>Baseline merged — PR#4</td><td>Custom evaluation baseline (metrics pending)</td></tr>
+          <tr><td>LLMRouterBench (ACL 2026)</td><td>Baseline submitted — PR#3</td><td>Pending evaluation</td></tr>
+        </tbody>
+      </table>
+      </div>
+
+      <div class="table-wrapper" style="margin-top:1.5rem;">
+      <table>
+        <thead>
+          <tr><th>Metric</th><th>A3M Result</th><th>Source</th></tr>
         </thead>
         <tbody>
-          <tr><td>RouterArena Score</td><td><strong>0.9404</strong></td><td>No. 1 among known public baselines</td></tr>
-          <tr><td>Accuracy</td><td><strong>96.77%</strong></td><td>8,400-query full split</td></tr>
-          <tr><td>Cost / 1K</td><td><strong>$0.0768</strong></td><td>No. 1 with published cost</td></tr>
-          <tr><td>Robustness</td><td><strong>1.0000</strong></td><td>0 abnormal entries</td></tr>
+          <tr><td>RouterArena Score</td><td><strong>0.9404</strong></td><td>ICLR 2025 | PR#144</td></tr>
+          <tr><td>RouterArena Accuracy</td><td><strong>96.77%</strong></td><td>8,400 queries</td></tr>
+          <tr><td>RouterArena Cost</td><td><strong>$0.0768/1K</strong></td><td>Avg across providers</td></tr>
+          <tr><td>MMR-Bench Accuracy</td><td><strong>67%</strong></td><td>Exact tier match</td></tr>
+          <tr><td>MMR-Bench Cost Savings</td><td><strong>63.5%</strong></td><td>vs all-premium routing</td></tr>
+          <tr><td>Robustness</td><td><strong>1.0000</strong></td><td>RouterArena | 0 abnormal entries</td></tr>
         </tbody>
       </table>
       </div>
+
+      <!-- RouterArena Evaluation Results -->
+      <div style="margin: 2rem 0; text-align: center;">
+        <h3 style="color: #94a3b8; font-size: 0.85rem; font-weight: 600; margin-bottom: 1rem;">ROUTERARENA EVALUATION — 8,400 QUERIES</h3>
+        <img src="assets/chart-routerena-leaderboard.svg" alt="RouterArena Evaluation Results" style="max-width: 100%; border-radius: 8px; margin-bottom: 0.5rem;">
+      </div>
+
+      <!-- Local Benchmark Charts -->
+      <div style="display: grid; grid-template-columns: 1fr 1fr; gap: 1rem; margin: 1.5rem 0;">
+        <img src="assets/chart-accuracy-by-tier.svg" alt="Routing Accuracy by Tier" style="max-width: 100%; border-radius: 8px;">
+        <img src="assets/chart-confusion-matrix.svg" alt="Confusion Matrix" style="max-width: 100%; border-radius: 8px;">
+      </div>
+      <div style="display: grid; grid-template-columns: 1fr 1fr; gap: 1rem; margin: 1.5rem 0;">
+        <img src="assets/chart-cost-comparison.svg" alt="Cost Comparison" style="max-width: 100%; border-radius: 8px;">
+        <img src="assets/chart-latency-overhead.svg" alt="Latency Overhead" style="max-width: 100%; border-radius: 8px;">
+      </div>
     </section>
 
     <!-- Comparison -->