qIOcJSCGn2 / poster_embed.html
DineshAI's picture
Update logbook: Reproduction: SVL time-to-goal distributions and survival identity
5afa513 verified
Raw
History Blame Contribute Delete
4.34 kB
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width,initial-scale=1">
<title>SVL reproduction audit</title>
<style>
:root { color-scheme:dark; --ink:#f6f8ff; --muted:#aeb7ce; --panel:#141b30; --line:#2c3858; --violet:#a78bfa; --cyan:#67e8f9; --green:#86efac; --amber:#fcd34d; }
* { box-sizing:border-box; }
body { margin:0; background:radial-gradient(circle at 20% 0,#292050 0,#11182b 42%,#080d19 100%); color:var(--ink); font:16px/1.45 ui-sans-serif,system-ui,sans-serif; }
main { max-width:1100px; margin:auto; padding:42px; }
.eyebrow { color:var(--cyan); font-weight:780; letter-spacing:.12em; text-transform:uppercase; }
h1 { margin:.3rem 0 .6rem; max-width:920px; font-size:clamp(2.2rem,5.2vw,4.4rem); line-height:1.02; }
.lead { color:var(--muted); max-width:850px; font-size:1.17rem; }
.grid { display:grid; grid-template-columns:repeat(3,1fr); gap:18px; margin:32px 0; }
.card { background:linear-gradient(180deg,#18213a,#101728); border:1px solid var(--line); border-radius:18px; padding:23px; box-shadow:0 18px 42px #0006; }
.claim { color:var(--muted); font-size:.84rem; text-transform:uppercase; letter-spacing:.09em; }
.status { margin:.35rem 0; font-size:1.5rem; font-weight:820; color:var(--green); }
.pending { color:var(--amber); }
.metric { display:flex; align-items:end; gap:18px; margin:25px 0 12px; }
.metric b { color:var(--cyan); font-size:3rem; line-height:1; }
.metric span { color:var(--muted); }
.bars { display:grid; gap:10px; }
.bar { display:grid; grid-template-columns:85px 1fr 80px; gap:12px; align-items:center; color:var(--muted); }
.track { height:15px; background:#2c3550; border-radius:999px; overflow:hidden; }
.track i { display:block; height:100%; border-radius:inherit; }
.pcs { width:75.9%; background:linear-gradient(90deg,var(--violet),var(--cyan)); }
.geo { width:100%; background:#64748b; }
.foot { display:grid; grid-template-columns:2fr 1fr; gap:18px; margin-top:18px; }
code { color:#dbeafe; }
@media(max-width:760px){ main{padding:24px}.grid,.foot{grid-template-columns:1fr}.bar{grid-template-columns:72px 1fr 64px} }
</style>
</head>
<body><main>
<div class="eyebrow">ICML 2026 reproduction · applied distribution audit</div>
<h1>SVL learns a real navigation time-to-goal law</h1>
<p class="lead">Released survival tuples, critic, and censored likelihood on the full official PointMaze training pool. Disjoint validation and a standalone NumPy audit test the distribution directly—not through the value identity.</p>
<section class="grid">
<article class="card"><div class="claim">Claim 1 · Distribution</div><div class="status">Applied verification</div><p>Three seeds; every PCS model improves held-out censored NLL over a goal-conditioned repeated-hazard baseline.</p></article>
<article class="card"><div class="claim">Claim 2 · Identity</div><div class="status">Verified</div><p>Bellman, survival-matrix, and value iteration agree within <code>1.2e-12</code>; two scope controls break it.</p></article>
<article class="card"><div class="claim">Claim 3 · Benchmarks</div><div class="status pending">Inconclusive</div><p>No actors, HIQL rerun, or long-horizon success-rate claim. C1 evidence is not relabeled as C3.</p></article>
</section>
<section class="card">
<div class="claim">Held-out censored negative log-likelihood · lower is better</div>
<div class="metric"><b>0.760</b><span>mean NLL improvement<br>paired gains <strong>0.715–0.807</strong></span></div>
<div class="bars">
<div class="bar"><span>PCS</span><div class="track"><i class="pcs"></i></div><strong>2.398</strong></div>
<div class="bar"><span>1-hazard</span><div class="track"><i class="geo"></i></div><strong>3.158</strong></div>
</div>
</section>
<section class="foot">
<article class="card"><div class="claim">Independent reconstruction</div><p>Mass error ≤ <code>1.34e-15</code>; zero monotonicity violations; raw NLLs and hashes match exactly; target shuffling costs <code>1.755–2.082</code> NLL.</p></article>
<article class="card"><div class="claim">Scope & cost</div><p>1M train pool<br>100k validation pool<br>3 seeds · 87.13 s<br>648 MiB peak RSS<br>USD 0 · 7/7 tests</p></article>
</section>
</main></body>
</html>