diff --git a/infra/docs/04-arc-and-caching.md b/infra/docs/04-arc-and-caching.md
index 504ebff46..6c20776e0 100644
--- a/infra/docs/04-arc-and-caching.md
+++ b/infra/docs/04-arc-and-caching.md
@@ -201,12 +201,15 @@ template:
             mountPath: /opt/bazel-repo-cache
             subPath: bazel-repo-cache
         resources:
+          # Burstable on purpose: requests bin-pack 8 runners onto the
+          # 32-vCPU / 125 GiB host; limits are each Kata VM's hotplug
+          # ceiling. Keep sum(memory limits) under host RAM.
           requests:
-            cpu: "8"
-            memory: "24Gi"
+            cpu: "3"
+            memory: "10Gi"
           limits:
             cpu: "8"
-            memory: "24Gi"
+            memory: "14Gi"
     volumes:
       - name: runner-cache
         persistentVolumeClaim:
@@ -259,14 +262,17 @@ Field by field:
   mounts to the `arc-runners/runner-cache` PVC. `ReadWriteOnce` is enough on this
   single-node k3s host; use a RWX-capable storage class before spreading runners
   across nodes.
-- **`resources`** - requests `8` CPU / `24Gi`, limits `8` CPU / `24Gi`. Kata reads
-  these and sizes the guest accordingly: the VM now boots at the same
-  guaranteed floor (`default_vcpus: 2`, `default_memory: 4096`) and only
-  hotplugs beyond that toward the limits, with `default_maxvcpus: 0` allowing up
-  to all host CPUs. Effectively the **requests are the boot-time VM size** and
-  the **limits are the hotplug ceiling**. See [02-kata-runtime.md](02-kata-runtime.md)
-  for the runtime knobs and [`infra/tune-kata-runtime.sh`](../tune-kata-runtime.sh)
-  for the SSH-driven patch helper.
+- **`resources`** - requests `3` CPU / `10Gi`, limits `8` CPU / `14Gi`
+  (burstable; see the `maxRunners` bullet above). Kata sizes the guest from
+  these: every VM boots at the fixed floor from the runtime config
+  (`default_vcpus: 2`, `default_memory: 4096` — deliberately at or below the
+  pod request so boot stays cheap) and hotplugs beyond it toward the pod
+  **limits**, with `default_maxvcpus: 0` allowing up to all host CPUs.
+  Effectively the **boot shape is a fixed floor**, the **requests are the
+  scheduler's bin-packing unit**, and the **limits are the hotplug ceiling**.
+  See [02-kata-runtime.md](02-kata-runtime.md) for the runtime knobs and
+  [`infra/tune-kata-runtime.sh`](../tune-kata-runtime.sh) for the SSH-driven
+  patch helper.
 
 ---
 
diff --git a/infra/tune-kata-runtime.sh b/infra/tune-kata-runtime.sh
index 2766d7c24..9eda821e2 100755
--- a/infra/tune-kata-runtime.sh
+++ b/infra/tune-kata-runtime.sh
@@ -1,6 +1,11 @@
 #!/usr/bin/env bash
-# Patch the live Kata QEMU config to match the runner pod's guaranteed boot
-# shape, raise guest and host open-file limits, and enlarge the virtiofsd worker pool.
+# Patch the live Kata QEMU config: set the guest BOOT floor (default_vcpus /
+# default_memory), raise guest and host open-file limits, and enlarge the
+# virtiofsd worker pool. Runner pods are burstable (see reload-runner.sh):
+# the boot floor stays deliberately at or below the pod's CPU/memory REQUEST
+# so VMs boot small and fast, and Kata hotplugs each guest toward the pod
+# LIMIT under load. Keep BOOT_VCPUS/BOOT_MEMORY_MIB <= the request when
+# retuning either side.
 # The script runs over SSH, which keeps the desired values version-controlled.
 # It then smoke-tests a new kata-qemu pod.
 #