fix(index): refresh stale claims (v0.6.11 -> v0.6.16, 14 PPL-verified -> 20 with comparator caveat, Mamba 1.0119 -> 1.0059 canonical) + remove failed-HN banner + remove tier-3 trainer-measurement leak token

index.html

@@ -76,15 +76,6 @@
     color: var(--accent2);
     margin: .25rem .25rem 0 0;
   }
-  .hn-banner {
-    background: linear-gradient(135deg, #ff6600 0%, #ff8533 100%);
-    color: #fff;
-    padding: .6rem 1rem;
-    border-radius: 6px;
-    margin: 1rem 0;
-    font-weight: 600;
-  }
-  .hn-banner a { color: #fff; text-decoration: underline; }
 </style>
 </head>
 <body>
@@ -93,15 +84,13 @@
 <h1>Sipsa Labs, Inc.</h1>
 <blockquote>An <b>experimental and deep tech-and-software company</b> (Delaware C-corp, incorporated May 2026). We invent and ship across the full breadth of tech and software &mdash; deep research, runtime systems, novel substrates, infrastructure, hardware-adjacent stacks, and software products that don&#39;t fit anywhere else yet. <b>UltraCompress</b> &mdash; lossless 5-bit transformer compression &mdash; is our first flagship publicly-shipped product. More products in flight.</blockquote>
 
-<div class="hn-banner">Live on Hacker News today: <a href="https://news.ycombinator.com/item?id=48099107">news.ycombinator.com/item?id=48099107</a></div>
-
-<span class="badge">v0.6.11 on PyPI</span>
-<span class="badge">22 architectures shipped; 14 PPL-verified end-to-end</span>
+<span class="badge">v0.6.16 on PyPI</span>
+<span class="badge">22 architectures shipped; 20 PPL-verified (19 transformer + 1 SSM, with comparator-note caveat)</span>
 <span class="badge">BUSL-1.1 + Additional Use Grant</span>
 
-<h2>2026-05-11 &mdash; 22 architectures shipped; 14 PPL-verified end-to-end at 5 bpw, three new sub-1.005&times; records this week</h2>
+<h2>Production-grade lossless 5-bit transformer compression</h2>
 
-<p>Production-grade lossless 5-bit transformer compression across 22 architectures, dense + Mixture-of-Experts + state-space (Mamba), 0.6B to 405B parameters. Mathematically lossless customer-side reconstruction: SHA-256 over reconstructed tensor bytes matches the trainer&#39;s measurement, verified by <code>uc verify</code>.</p>
+<p>22 architectures shipped, 20 PPL-verified end-to-end (19 transformer + 1 SSM with comparator-note caveat). Dense + Mixture-of-Experts + state-space, 0.6B to 405B parameters. Mathematically lossless customer-side reconstruction: SHA-256 over reconstructed tensor bytes matches the reference manifest, verified by <code>uc verify</code>.</p>
 
 <h2>Tightest verified PPL ratios at 5 bpw</h2>
 <p><i>(perplexity ratio = compressed PPL / bf16 baseline PPL; FineWeb-edu held-out tail; seq_len = 1024; n = 30-50; seed = 42)</i></p>
@@ -114,12 +103,12 @@
   <tr><td>Qwen3-14B</td><td>14B</td><td>dense</td><td><b>1.00403&times;</b></td><td>14B-class record</td></tr>
   <tr><td>Yi-1.5-9B</td><td>8.8B</td><td>dense</td><td><b>1.00414&times;</b></td><td>&gt;8B record</td></tr>
   <tr><td>Qwen3-8B</td><td>8B</td><td>dense</td><td><b>1.00440&times;</b></td><td>8B-class record</td></tr>
-  <tr><td>Mistral-7B-v0.3</td><td>7B</td><td>dense</td><td><b>1.00548&times;</b></td><td>NEW this week &mdash; 9.16&times; tighter than prior</td></tr>
+  <tr><td>Mistral-7B-v0.3</td><td>7B</td><td>dense</td><td><b>1.00548&times;</b></td><td>a strong dense 7B-class result</td></tr>
   <tr class="headline"><td><b>Hermes-3-Llama-3.1-405B</b></td><td>405B</td><td>dense</td><td><b>1.0066&times;</b></td><td>largest dense 5-bit lossless on the Hub</td></tr>
   <tr><td>Qwen3-0.6B</td><td>0.6B</td><td>dense</td><td><b>1.0069&times;</b></td><td></td></tr>
   <tr><td>OLMo-2-0425-1B</td><td>1B</td><td>dense</td><td><b>1.0073&times;</b></td><td></td></tr>
   <tr><td>SmolLM2-1.7B-Instruct</td><td>1.7B</td><td>dense</td><td><b>1.0075&times;</b></td><td></td></tr>
-  <tr><td>Mamba-2.8B</td><td>2.8B</td><td>SSM</td><td><b>1.0119&times;</b></td><td>first published 5-bit lossless on a state-space model</td></tr>
+  <tr><td>Mamba-2.8B</td><td>2.8B</td><td>SSM</td><td><b>1.0059&times;</b></td><td>state-space model (comparator-note caveat: GSQ-only, no V18-C overlay)</td></tr>
   <tr><td>Llama-3.1-8B</td><td>8B</td><td>dense</td><td><b>1.0125&times;</b></td><td>standard eval</td></tr>
 </table>
 
@@ -128,7 +117,7 @@
 <h2>Quick start (3 commands, no GPU required for verify)</h2>
 <pre>pip install ultracompress
 hf download SipsaLabs/qwen3-8b-uc-v3-bpw5 --local-dir ./qwen3-8b
-uc bench ./qwen3-8b</pre>
+uc verify ./qwen3-8b</pre>
 
 <h2>Inference API</h2>
 <p><code>api.sipsalabs.com</code> is live, OpenAI-compatible:</p>