diff --git a/training/dashboard/static/dashboard.js b/training/dashboard/static/dashboard.js
index 0d5b233..e7555fe 100644
--- a/training/dashboard/static/dashboard.js
+++ b/training/dashboard/static/dashboard.js
@@ -224,20 +224,12 @@
   const demoBtn = document.getElementById('demo-btn');
   let demoTimer = null;
   let demoActive = false;
-  // Demo mode is intentionally narrow: it does NOT synthesize
-  // `episode` events (those would clobber the real ingest counter,
-  // host bars, and database explorer — all of which work fine in or
-  // out of demo mode), and it does NOT synthesize `phase` events
-  // (the baseline scene reads dataset-derived `phase_mix` instead).
-  // The only periodic side-effect is occasional model_metric jitter
-  // so the bars don't sit frozen during a long talk.
-  function demoTick() {
-    if (Math.random() < 0.05) {
-      const m = ['knn', 'rnn', 'gru', 'lstm', 'bert'][Math.floor(Math.random() * 5)];
-      const base = { knn: 0.736, rnn: 0.872, gru: 0.911, lstm: 0.928, bert: 0.954 }[m];
-      dispatch({ type: 'model_metric', model: m, accuracy: base + (Math.random() - 0.5) * 0.012 });
-    }
-  }
+  // Demo mode is intentionally narrow. Real data wins everywhere —
+  // demo only seeds widgets that haven't received any real producer
+  // output yet. demoTick is now a no-op; the per-widget demo_start
+  // handlers do the (one-shot) seeding, so we don't drown out real
+  // data with periodic synthetic noise.
+  function demoTick() {}
   function setDemo(active) {
     if (demoActive === active) return;
     demoActive = active;
@@ -1785,19 +1777,22 @@ def train_nn(*, model, X_train, y_train, X_val, y_val,
       r.fill.style.width = (visible * 100).toFixed(1) + '%';
       r.acc.textContent = accuracy.toFixed(3);
     }
+    // Real data wins. Demo only fills in when no real model_metric
+    // has arrived yet — once the producer publishes for any model,
+    // demo never overwrites it.
+    let hasRealMetric = false;
     on('demo_start', () => {
-      // KNN sits below the recurrent/transformer family; it memorizes
-      // the train host's feature space and generalizes worse than a
-      // model that learned temporal structure. Bar visible-scale starts
-      // at 0.5 so the real cross-host F1 (~0.43) reads as 0% — that's
-      // honest, just visually flat. Demo value here is the in-distribution
-      // ballpark for a healthier display.
+      if (hasRealMetric) return;
       [ ['knn', 0.736], ['rnn', 0.872], ['gru', 0.911], ['lstm', 0.928], ['bert', 0.954] ]
         .forEach(([m, a]) => render(m, a));
     });
-    on('demo_stop', () => { rows.clear(); emptyState(); });
+    on('demo_stop', () => {
+      // Don't wipe real data on demo toggle off.
+      if (!hasRealMetric) { rows.clear(); emptyState(); }
+    });
     on('model_metric', m => {
       if (!m.model || typeof m.accuracy !== 'number') return;
+      hasRealMetric = true;
       render(m.model, m.accuracy);
     });
     emptyState();
@@ -2208,8 +2203,12 @@ def train_nn(*, model, X_train, y_train, X_val, y_val,
     }
     setInterval(updateStats, 500);
 
-    function handleDetection(d) {
+    // Track whether real live_detection events have arrived. Demo
+    // only fills in when nothing real is flowing — never overwrites.
+    let hasRealDetection = false;
+    function handleDetection(d, fromReal) {
       if (!d.host_id || !d.predicted) return;
+      if (fromReal) hasRealDetection = true;
       eventTimes.push(Date.now());
       if (d.model) lastModel = d.model;
       if (d.actual) {
@@ -2220,11 +2219,12 @@ def train_nn(*, model, X_train, y_train, X_val, y_val,
       paintLatest(d);
     }
 
-    on('live_detection', handleDetection);
+    on('live_detection', m => handleDetection(m, true));
 
     // Synthetic demo: 5 hosts, walk through phases, ~92% accuracy.
     let demoTimer = null;
     function demoStart() {
+      if (hasRealDetection) return;
       if (demoTimer) clearInterval(demoTimer);
       const HOSTS = [
         { id: 'elliott-lab',       profile: 'cpu-saturate',     phaseIdx: 0 },
@@ -2258,6 +2258,8 @@ def train_nn(*, model, X_train, y_train, X_val, y_val,
     }
     function demoStop() {
       if (demoTimer) { clearInterval(demoTimer); demoTimer = null; }
+      // Don't wipe real data on demo toggle off.
+      if (hasRealDetection) return;
       lanes.forEach(l => l.row.remove());
       lanes.clear();
       eventTimes.length = 0;
@@ -2515,7 +2517,11 @@ def train_nn(*, model, X_train, y_train, X_val, y_val,
       rec.g.querySelector('text').textContent = model;
       repaintLabels();
     }
+    // Real data wins. Demo only fills in when no real model_perf has
+    // arrived yet.
+    let hasRealPerf = false;
     on('demo_start', () => {
+      if (hasRealPerf) return;
       [
         { model: 'knn',  latency_us:  90, accuracy: 0.84 },
         { model: 'rnn',  latency_us: 380, accuracy: 0.87 },
@@ -2524,9 +2530,10 @@ def train_nn(*, model, X_train, y_train, X_val, y_val,
         { model: 'bert', latency_us: 3200, accuracy: 0.95 },
       ].forEach(p => render(p.model, p.latency_us, p.accuracy));
     });
-    on('demo_stop', emptyState);
+    on('demo_stop', () => { if (!hasRealPerf) emptyState(); });
     on('model_perf', m => {
       if (!m.model || typeof m.latency_us !== 'number' || typeof m.accuracy !== 'number') return;
+      hasRealPerf = true;
       render(m.model, m.latency_us, m.accuracy);
     });
     emptyState();
diff --git a/training/dashboard/static/index.html b/training/dashboard/static/index.html
index 1b96fdc..f8f4885 100644
--- a/training/dashboard/static/index.html
+++ b/training/dashboard/static/index.html
@@ -1314,6 +1314,6 @@
     </article>
   </div>
 
-  <script src="/static/dashboard.js?v=7f398906"></script>
+  <script src="/static/dashboard.js?v=d4323f22"></script>
 </body>
 </html>