TrangBui's picture
Publish RoleRouteBench trajectory explorer
8f03abe verified
Raw History Blame Contribute Delete
8.94 kB
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta name="description" content="Explore RoleRouteBench live multi-agent routing trajectories, JEV reference results, and the public/private isolation protocol.">
<meta name="theme-color" content="#08111f">
<title>RoleRouteBench — Dynamic role routing, made inspectable</title>
<link rel="stylesheet" href="styles.css">
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "Dataset",
"name": "RoleRouteBench",
"alternateName": "RoutingBench V4-beta",
"description": "A reproducible multi-domain benchmark for dynamic role routing in multi-agent systems, with 82 live trajectories across four domains.",
"identifier": "https://doi.org/10.57967/hf/11430",
"url": "https://huggingface.co/datasets/TrangBui/RoleRouteBench",
"version": "4.0.0-beta.2",
"creator": {"@type": "Person", "name": "Trang Bui"},
"license": "https://huggingface.co/datasets/TrangBui/RoleRouteBench/blob/main/LICENSE_DATA",
"sameAs": [
"https://doi.org/10.57967/hf/11430",
"https://huggingface.co/datasets/TrangBui/RoleRouteBench",
"https://github.com/BuiThiThanhTrang/RACO_MAS"
],
"keywords": ["multi-agent systems", "dynamic routing", "agent orchestration", "tool use", "benchmark", "live trajectories"]
}
</script>
</head>
<body>
<a class="skip-link" href="#main">Skip to content</a>
<header class="site-header">
<a class="brand" href="#top" aria-label="RoleRouteBench home">
<span class="brand-mark">RR</span>
<span>RoleRouteBench</span>
</a>
<nav aria-label="Primary navigation">
<a href="#results">Results</a>
<a href="#explorer">Explorer</a>
<a href="#isolation">Isolation</a>
<a href="#submit">Submit</a>
</nav>
<a class="header-link" href="https://huggingface.co/datasets/TrangBui/RoleRouteBench">Dataset ↗</a>
</header>
<main id="main">
<section class="hero" id="top">
<div class="hero-copy">
<p class="eyebrow">Frozen environment · live consequences · multi-step tasks</p>
<h1>Dynamic role routing,<br><span>made inspectable.</span></h1>
<p class="hero-lede">RoleRouteBench evaluates whether a decision model can coordinate specialized executors across evolving task states—not merely whether the final answer happens to be correct.</p>
<div class="hero-actions">
<a class="button primary" href="#explorer">Explore trajectories</a>
<a class="button secondary" href="https://doi.org/10.57967/hf/11430">DOI 10.57967/hf/11430 ↗</a>
</div>
</div>
<div class="route-card" aria-label="Illustrative routing trajectory">
<div class="route-card-head"><span>packet / live state</span><span class="live-dot">LIVE</span></div>
<div class="route-step"><span class="step-index">01</span><div><strong>Inspect evidence</strong><small>state changed</small></div><span class="status success">success</span></div>
<div class="route-line"></div>
<div class="route-step"><span class="step-index">02</span><div><strong>Author candidate</strong><small>artifact created</small></div><span class="status success">success</span></div>
<div class="route-line warning"></div>
<div class="route-step"><span class="step-index">03</span><div><strong>Execute candidate</strong><small>failure preserved</small></div><span class="status failed">failed</span></div>
<div class="route-line"></div>
<div class="route-step"><span class="step-index">04</span><div><strong>Diagnose & repair</strong><small>new routed action</small></div><span class="status success">success</span></div>
</div>
</section>
<section class="stats" aria-label="Benchmark summary">
<div><strong>82</strong><span>live trajectories</span></div>
<div><strong>4</strong><span>task domains</span></div>
<div><strong>568</strong><span>primary decisions</span></div>
<div><strong>33</strong><span>specialized roles</span></div>
</section>
<section class="section why">
<div class="section-kicker">Why RoleRouteBench?</div>
<div class="why-grid">
<h2>Final task success cannot tell you whether a router coordinated well.</h2>
<div class="why-copy">
<p>Outcome quality also depends on the frozen executor. RoleRouteBench therefore retains every routing decision, failed action, state transition, budget event, and independent evaluator result.</p>
<p>Competing routers see the same public state, role pool, adapters, executors, budgets, and evaluator. The environment stays fixed; only the routing policy changes.</p>
</div>
</div>
</section>
<section class="section" id="results">
<div class="section-heading">
<div><div class="section-kicker">Reference system</div><h2>JEV 1.13 results</h2></div>
<p>Reference signals for environment validation—not an oracle and not a composite leaderboard score.</p>
</div>
<div class="table-wrap">
<table>
<thead><tr><th>Domain</th><th>Packets</th><th>Trajectory pass</th><th>Trusted task pass</th><th>Primary decisions</th><th>Budget-suppressed</th></tr></thead>
<tbody id="results-body"><tr><td colspan="6">Loading reference results…</td></tr></tbody>
</table>
</div>
</section>
<section class="section explorer-section" id="explorer">
<div class="section-heading">
<div><div class="section-kicker">Trajectory explorer</div><h2>Follow each routed consequence</h2></div>
<p>Filter public compact traces. Raw prompts and executor outputs are not included.</p>
</div>
<div class="filters" aria-label="Trajectory filters">
<label>Domain<select id="domain-filter"><option value="">All domains</option></select></label>
<label>Packet<select id="packet-filter"></select></label>
<label>Role<select id="role-filter"><option value="">All roles</option></select></label>
<label>Event status<select id="status-filter"><option value="">All statuses</option></select></label>
</div>
<div class="packet-summary" id="packet-summary" aria-live="polite"></div>
<div class="trajectory" id="trajectory" aria-live="polite"></div>
</section>
<section class="section isolation-section" id="isolation">
<div class="section-heading">
<div><div class="section-kicker">Evaluation boundary</div><h2>Public/private isolation by construction</h2></div>
<p>“Private” means evaluator-only under the benchmark protocol; the artifacts remain downloadable for transparent reproduction.</p>
</div>
<div class="isolation-grid">
<div class="boundary public-boundary">
<span class="boundary-label">ROUTER + EXECUTOR VISIBLE</span>
<h3>Public state</h3>
<ul><li>Task and public artifacts</li><li>Full role pool plus STOP</li><li>Prior observable consequences</li><li>Remaining call budget</li></ul>
</div>
<div class="boundary-flow" aria-hidden="true"><span>route</span><b>→</b><span>execute</span><b>→</b><span>observe</span></div>
<div class="boundary private-boundary">
<span class="boundary-label">EVALUATOR ONLY</span>
<h3>Reference state</h3>
<ul><li>Gold patches and proofs</li><li>Answer workbooks and oracles</li><li>Trusted evaluator results</li><li>Track A labels</li></ul>
</div>
</div>
<p class="boundary-note">The evaluator runs only after termination. It does not feed private artifacts back into the trajectory.</p>
</section>
<section class="section submit-section" id="submit">
<div>
<div class="section-kicker">Community evaluation</div>
<h2>Bring your own router.</h2>
<p>JEV is the reference system. Submit reproducible results from another decision model under the frozen environment—without sending us your model.</p>
</div>
<div class="submit-actions">
<a class="button primary" href="https://github.com/BuiThiThanhTrang/RACO_MAS/issues/new?template=router-result.yml">Submit router results ↗</a>
<a class="button secondary" href="https://github.com/BuiThiThanhTrang/RACO_MAS/blob/main/puppeteer/CONTRIBUTING_ROUTER_RESULTS.md">Read the protocol ↗</a>
</div>
</section>
</main>
<footer>
<div><strong>RoleRouteBench</strong><span>Version 4.0.0-beta.2 · DOI 10.57967/hf/11430</span></div>
<div class="footer-links"><a href="https://huggingface.co/datasets/TrangBui/RoleRouteBench">Dataset</a><a href="https://github.com/BuiThiThanhTrang/RACO_MAS">Code</a><a href="https://doi.org/10.57967/hf/11430">DOI</a></div>
</footer>
<script src="app.js" defer></script>
</body>
</html>