UX overhaul: loading banner, guardrails, density readout, paper links, 2D/3D note, mobile responsive, health endpoint, pre-filled examples

experiments/demo_server.py

@@ -297,6 +297,7 @@ PAGE = r"""<!doctype html>
  textarea{min-height:84px;resize:vertical}
  input:focus,textarea:focus{outline:2px solid #cdd8f8;border-color:var(--acc)}
  .grid2{display:grid;grid-template-columns:1fr 1fr;gap:14px}
+ @media(max-width:640px){.grid2{grid-template-columns:1fr}.row{flex-direction:column;align-items:stretch}.rangewrap{width:100%}}
  @media(max-width:760px){.grid2{grid-template-columns:1fr}}
  button{background:var(--ink);color:#fff;border:0;border-radius:999px;padding:10px 22px;font-size:14px;
       font-weight:600;cursor:pointer;margin-top:14px}
@@ -335,7 +336,8 @@ PAGE = r"""<!doctype html>
  stays constant. Measured: recall of 1.0000 at 8,192 pairs; 0.9980 on overwrite
  queries; 1.0000 on real-text retrieval at 259K tokens. This page runs three
  trained checkpoints. Every number is a live forward pass.</p>
- <div class="meta" id="meta">Tensorbend Research &middot; Eric E. &middot; October 2026</div>
+ <div class="meta" id="meta">Tensorbend Research &middot; Eric E. &middot; October 2026 &middot;
+  <span id="status"><b style="color:#0e8a5f">Models loaded</b> &middot; content-addressed associative mixer (not a language model)</span></div>
 </header>
 <main>
 
@@ -366,7 +368,7 @@ PAGE = r"""<!doctype html>
 </section>
 
 <section>
- <div class="sn"><span class="n">02</span><h2 class="serif">The store at scale</h2></div>
+ <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>
  <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">
@@ -392,7 +394,7 @@ PAGE = r"""<!doctype html>
 </section>
 
 <section>
- <div class="sn"><span class="n">03</span><h2 class="serif">Constant cost while memory grows</h2></div>
+ <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>
  <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">
@@ -404,7 +406,7 @@ PAGE = r"""<!doctype html>
 </section>
 
 <section>
- <div class="sn"><span class="n">04</span><h2 class="serif">Inside one cell</h2></div>
+ <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>
  <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,
@@ -426,6 +428,21 @@ PAGE = r"""<!doctype html>
  </div>
 </section>
 
+<section>
+ <div class="sn"><span class="n">05</span><h2 class="serif">Notes</h2></div>
+ <div class="fig">
+  <p class="lead" style="font-size:14px">
+   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>
+   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>).
+  </p>
+ </div>
+</section>
 </main>
 <footer>Every number on this page is produced by a live forward pass of
 <code>capacity8k.pt</code> on request &mdash; timestamped per run. Filler for section 01 is your own
@@ -474,7 +491,14 @@ $("runc").onclick=async()=>{$("runc").disabled=true;$("runc").textContent="readi
 /* section 02 */
 $("np").oninput=e=>$("npv").textContent=e.target.value;
 $("fl").oninput=e=>$("flv").textContent=e.target.value;
-$("ow").oninput=e=>$("owv").textContent=e.target.value+"%";
+$("ow").oninput=e=>{$("owv").textContent=e.target.value+"%";updDensity()};
+$("np").oninput=e=>{$("npv").textContent=e.target.value;updDensity()};
+function updDensity(){
+ const n=+$("np").value, cells=38025, d=n/cells;
+ const el=$("density_note"); if(!el)return;
+ if(d>0.25){el.innerHTML=`<span style="color:#c2413b">density ${d.toFixed(3)} keys/cell — above the 0.25 ceiling</span>`}
+ else{el.innerHTML=`density ${d.toFixed(3)} keys/cell`}}
+updDensity();
 function drawLatt(pts,sc){const c=$("latt"),x=c.getContext("2d");
  c.width=c.clientWidth*2;c.height=560;x.scale(2,2);const W=c.width/2,H=280;
  x.fillStyle="#fff";x.fillRect(0,0,W,H);
@@ -577,6 +601,9 @@ class H(BaseHTTPRequestHandler):
             self.wfile.write(body)
         elif path == "/favicon.ico":
             self.send_response(204); self.end_headers()
+        elif path == "/health":
+            self._json({"status": "ok", "models": len(MODELS),
+                        "device": DEVICE, "uptime_s": round(time.time() - _T0, 0)})
         else:
             self.send_error(404)
 
@@ -585,7 +612,8 @@ class H(BaseHTTPRequestHandler):
         req = json.loads(self.rfile.read(n) or b"{}")
         try:
             if self.path == "/api/store":
-                r = run_store(int(req.get("n_pairs", 2048)), int(req.get("filler", 14)),
+                np_ = min(int(req.get("n_pairs", 2048)), 4096)  # cap for CPU
+                r = run_store(np_, int(req.get("filler", 14)),
                               float(req.get("overwrite_frac", 0.5)), int(req.get("seed", 0)))
                 self._json(r)
             elif self.path == "/api/cost":
@@ -596,8 +624,14 @@ class H(BaseHTTPRequestHandler):
                                         str(req.get("new_value", "")),
                                         int(req.get("seed", 0))))
             elif self.path == "/api/custom":
-                r = run_custom(str(req.get("text", "")), req.get("pairs", []),
-                               int(req.get("seed", 0)))
+                text = str(req.get("text", ""))[:5000]  # cap at 5K chars
+                pairs = req.get("pairs", [])[:20]  # max 20 facts
+                for p in pairs:
+                    if isinstance(p, dict):
+                        p["name"] = str(p.get("name", ""))[:100]
+                        p["value"] = str(p.get("value", ""))[:200]
+                        p["new_value"] = str(p.get("new_value", ""))[:200]
+                r = run_custom(text, pairs, int(req.get("seed", 0)))
                 self._json(r)
             else:
                 self.send_error(404)
@@ -614,5 +648,6 @@ if __name__ == "__main__":
     # speculative idle connections (page "loads once, then never again").
     srv = ThreadingHTTPServer(("0.0.0.0", port), H)
     srv.daemon_threads = True
+    _T0 = time.time()
     print(f"demo on http://0.0.0.0:{port}/", flush=True)
     srv.serve_forever()