speed: pre-warm models, reduce defaults, remove GitHub links (X-Frame-Options)

experiments/demo_server.py

@@ -368,13 +368,13 @@ PAGE = r"""<!doctype html>
 </section>
 
 <section>
- <div class="sn"><span class="n">02</span><h2 class="serif">The store at scale <a href="https://github.com/eelbaz/mLattice/blob/main/PAPER.md#42-results" style="font-size:14px;color:#2547d0;text-decoration:none">&nearr;</a></h2></div>
+ <div class="sn"><span class="n">02</span><h2 class="serif">The store at scale &nearr;</h2></div>
  <p class="lead">The same mechanism at scale: store up to 8,192 pairs in one 133,120-token
  causal pass, overwrite half of them, then query. Measured recall: 1.0000.</p>
  <div class="fig">
   <div class="row">
    <div class="rangewrap"><label style="margin:0">pairs</label>
-    <input id="np" type="range" min="64" max="8192" step="64" value="2048"><span class="rv" id="npv">2048</span></div>
+    <input id="np" type="range" min="64" max="2048" step="64" value="128"><span class="rv" id="npv">2048</span></div>
    <div class="rangewrap"><label style="margin:0">filler/pair</label>
     <input id="fl" type="range" min="0" max="14" value="14"><span class="rv" id="flv">14</span></div>
    <div class="rangewrap"><label style="margin:0">changed mid-stream</label>
@@ -394,7 +394,7 @@ PAGE = r"""<!doctype html>
 </section>
 
 <section>
- <div class="sn"><span class="n">03</span><h2 class="serif">Constant cost while memory grows <a href="https://github.com/eelbaz/mLattice/blob/main/PAPER.md#43-cost-measurements-head-to-head-with-attention" style="font-size:14px;color:#2547d0;text-decoration:none">&nearr;</a></h2></div>
+ <div class="sn"><span class="n">03</span><h2 class="serif">Constant cost while memory grows &nearr;</h2></div>
  <p class="lead">Each token touches exactly nine memory cells, so per-token work cannot grow with
  context. Measure it: the curve below runs the model at increasing stream lengths, live.</p>
  <div class="fig">
@@ -406,7 +406,7 @@ PAGE = r"""<!doctype html>
 </section>
 
 <section>
- <div class="sn"><span class="n">04</span><h2 class="serif">Inside one cell <a href="https://github.com/eelbaz/mLattice/blob/main/PAPER.md#31-architecture" style="font-size:14px;color:#2547d0;text-decoration:none">&nearr;</a></h2></div>
+ <div class="sn"><span class="n">04</span><h2 class="serif">Inside one cell &nearr;</h2></div>
  <p class="lead">Where does a value physically live, and where does the <em>new</em> value go on
  overwrite? Pick one fact from section 01 (or any name/value). The model&rsquo;s learned address
  for that key, the 3&times;3 cells it touches, and the before/after answers &mdash; same cells,
@@ -435,11 +435,11 @@ PAGE = r"""<!doctype html>
    This is a <b>content-addressed associative mixer</b>, not a language model.
    It stores key→value associations and retrieves them at constant per-token cost.
    Perfect scores are on the synthetic MQAR distribution (see
-   <a href="https://github.com/eelbaz/mLattice/blob/main/PAPER.md" style="color:#2547d0">PAPER.md</a>).<br><br>
+   PAPER.md).<br><br>
    Current demo is <b>2D</b> (9-cell neighborhood per token).
    <b>3D</b> (27 cells) is supported by the kernel math without architectural change —
    it reduces key density by 1000&times; and raises the estimated capacity from ~8K to ~495K keys
-   (<a href="https://github.com/eelbaz/mLattice/blob/main/PAPER.md#47-address-space-scaling-3d-analysis" style="color:#2547d0">details</a>).
+   (details).
   </p>
  </div>
 </section>
@@ -544,7 +544,7 @@ function drawCC(data){const c=$("cc"),x=c.getContext("2d");c.width=c.clientWidth
  x.fillText("stream length (tokens, log scale)",W/2-80,H-2);
  x.fillText("&mu;s/token (blue, flat = constant) · cells (green) · attention O(T²) (dashed red)",P,16)}
 $("cost").onclick=async()=>{$("cost").disabled=true;$("cost").textContent="measuring...";
- try{const sizes=[64,256,1024,2048,4096,8192].map(n=>({n_pairs:n,filler:14}));
+ try{const sizes=[64,256,512,1024].map(n=>({n_pairs:n,filler:0}));
   const d=await api("/api/cost",{sizes});drawCC(d)}catch(e){alert(e)}
  $("cost").disabled=false;$("cost").textContent="measure the cost curve"};
 /* section 04 */