diff --git a/platform/cluster/flux/apps/stateless/assistant-api/deployment-enschede-ws.yaml b/platform/cluster/flux/apps/stateless/assistant-api/deployment-enschede-ws.yaml
index e41ed3f2..1a7275c3 100644
--- a/platform/cluster/flux/apps/stateless/assistant-api/deployment-enschede-ws.yaml
+++ b/platform/cluster/flux/apps/stateless/assistant-api/deployment-enschede-ws.yaml
@@ -69,15 +69,13 @@ spec:
       serviceAccountName: assistant-api
       priorityClassName: platform-critical-app
       nodeSelector:
-        personal-stack/site: enschede
-        # The assistant-api image is built amd64-only. The Enschede site
-        # spans both amd64 nodes and the arm64 Pis (pi-1/2/3); without
-        # this constraint the scheduler can place the surge pod on a Pi,
-        # where the image fails with "no match for platform in manifest"
-        # → ImagePullBackOff while the old pod is torn down. Keel's
-        # force-poll rolls a new pod every 2 min, so that coin-flip took
-        # the terminal WS down repeatedly. Pin to amd64.
-        kubernetes.io/arch: amd64
+        # Pin to the always-on t1000 server. The other amd64 Enschede
+        # nodes (rx7900xtx, gtx-960m) are desktops that are powered off
+        # when not in use. When the pod lands on a desktop and that
+        # desktop goes offline, the Service endpoint dies and the WS stays
+        # broken for the full k8s node eviction window (~5 min) before the
+        # pod is rescheduled. Pinning to the t1000 prevents this entirely.
+        personal-stack/node: enschede-t1000-1
       initContainers:
         - name: download-pyroscope-agent
           image: curlimages/curl:8.10.1