diff --git a/docs/dashboard-request-knn-cap-evict.md b/docs/dashboard-request-knn-cap-evict.md
new file mode 100644
index 0000000..50faab1
--- /dev/null
+++ b/docs/dashboard-request-knn-cap-evict.md
@@ -0,0 +1,93 @@
+# Dashboard request — cap + evict for the KNN scatter, plus snapshot-replace semantics
+
+**Audience:** dashboard session (owns `training/dashboard/`).
+**Producer side:** `training/producers/knn.py stream` (currently
+running on the Pi).
+**Status:** request, not implementation. The producer side has reduced
+its cadence as a band-aid, but the underlying fix lives in
+`training/dashboard/static/dashboard.js` scene-9 (KNN scatter) handler.
+
+## Problem
+
+The current scene-9 handler is:
+
+```js
+on('embedding', m => {
+  if (typeof m.x !== 'number' || typeof m.y !== 'number') return;
+  const pt = { x: m.x, y: m.y, z: ..., phase: m.phase,
+               predicted: m.predicted, cluster: ... };
+  points.push(pt);
+  addStat(pt);
+  rebuildLegend();
+});
+```
+
+Every `embedding` event pushes onto the `points` array with no cap.
+The producer republishes its (deterministic, stable) point set on a
+cycle so reconnecting browsers eventually see the scatter populate;
+each cycle therefore pushes the same N points onto `points` again,
+and over time the in-memory point count grows without bound. After
+~10 minutes the browser starts slowing down.
+
+The producer side has band-aided this by reducing cycle cadence
+(200 points every 30 s, was 600 every 5 s). That's 18× slower
+accumulation, but still a leak.
+
+## Two complementary fixes the dashboard could land
+
+### A. Cap the points array (cheapest)
+
+Add a FIFO eviction:
+
+```js
+const MAX_POINTS = 4000;   // tune to taste
+on('embedding', m => {
+  // ...validate + build pt as before...
+  points.push(pt);
+  if (points.length > MAX_POINTS) {
+    points.shift();   // or splice(0, points.length - MAX_POINTS)
+  }
+  addStat(pt);
+  rebuildLegend();
+});
+```
+
+This bounds memory regardless of how often the producer publishes.
+Existing visual quality stays the same once the cap is reached
+(the most-recent N points are kept).
+
+### B. Snapshot-replace via a new event type
+
+For a cleaner architecture: the producer sends one `embedding_batch`
+event per cycle containing the full set of points; the handler
+*replaces* the contents of `points` rather than appending. Eliminates
+duplicate-publish leakage entirely and naturally supports the
+"refresh shows something" use case via state replay (see the
+companion request `dashboard-request-embedding-persistence.md`).
+
+Producer would emit:
+
+```json
+{ "type": "embedding_batch",
+  "points": [ {x, y, z, phase, predicted, cluster}, ... ],
+  "replace": true }
+```
+
+Handler:
+
+```js
+on('embedding_batch', m => {
+  if (m.replace) { points.length = 0; resetStats(); }
+  for (const pt of m.points) { points.push(pt); addStat(pt); }
+  rebuildLegend();
+});
+```
+
+If you want this, the producer side will switch to it; just confirm
+the event name and payload shape.
+
+## Suggested order
+
+A first (1-line change, fixes the leak today). B second when you have
+time — pairs naturally with the snapshot/sticky-cache request for
+refresh-time hydration.