diff --git a/desktop/src/features/mesh-compute/meshFieldNodes.test.mjs b/desktop/src/features/mesh-compute/meshFieldNodes.test.mjs
index 35a6f2936..36fad2923 100644
--- a/desktop/src/features/mesh-compute/meshFieldNodes.test.mjs
+++ b/desktop/src/features/mesh-compute/meshFieldNodes.test.mjs
@@ -84,7 +84,14 @@ test("relay nodes carry no RTT and claim no membership", () => {
   // A hollow centre and no spokes depend on this staying false.
   assert.equal(field.selfParticipating, false);
   assert.equal(field.selfCapacityGb, null);
-  assert.match(field.caption, /Last reported/);
+});
+
+test("the relay view is not hedged with a caption", () => {
+  // What we know is shared gets shown as shared. Annotating it with "last
+  // reported" hedged a fact we actually have, and the old "join to see live
+  // detail" mislabeled sharing as joining.
+  const field = deriveMeshFieldModel({ view: null, snapshot: snapshot() });
+  assert.equal(field.caption, null);
 });
 
 test("our own note is excluded so we are never drawn twice", () => {
@@ -132,5 +139,5 @@ test("a solo sharer is told it is waiting, not that it is alone in error", () =>
     snapshot: null,
   });
   assert.equal(field.source, "live");
-  assert.match(field.caption, /waiting for another device/);
+  assert.equal(field.caption, "Waiting for another device");
 });
diff --git a/desktop/src/features/mesh-compute/meshFieldNodes.ts b/desktop/src/features/mesh-compute/meshFieldNodes.ts
index bb910992a..efdad61cc 100644
--- a/desktop/src/features/mesh-compute/meshFieldNodes.ts
+++ b/desktop/src/features/mesh-compute/meshFieldNodes.ts
@@ -13,14 +13,16 @@ import type { MeshFieldNode } from "./ui/MeshRadarField";
  * renders either way, and the earlier "turn on sharing to see who's on the
  * mesh" was simply wrong.
  *
- * The two are not equivalent, and the difference is not cosmetic:
+ * The two are not equivalent, and the difference is expressed in the geometry
+ * rather than in a hedging caption — what we know is shared gets shown either
+ * way, and the drawing simply does not assert more than it can:
  *
  *   - **live**: reachable now, RTT known, adjacency to us is real
- *   - **relay**: last known within 120s, no RTT, adjacency unknown
+ *   - **relay**: published within the last 120s, no RTT, adjacency unknown
  *
  * So relay nodes get no RTT (they land mid-band rather than faking proximity),
- * the caller draws no spokes when we have not joined, and the caption names
- * which view is on screen. Never dress last-known state as live.
+ * and the caller draws no spokes and a hollow centre when we have not joined,
+ * because a spoke would assert a connection we do not have.
  */
 
 export type MeshFieldSource = "live" | "relay" | "none";
@@ -38,7 +40,13 @@ export type MeshFieldModel = {
    * so ghosts never contribute capacity — they are the invitation, not a total.
    */
   ghostCount: number;
-  /** Caption naming the source, or null when there is nothing to qualify. */
+  /**
+   * Short caption, or null.
+   *
+   * Only ever states a fact the field cannot draw — currently just "alone on
+   * the mesh". Never used to hedge the relay view: what we know is shared is
+   * shown as shared, not annotated with doubt.
+   */
   caption: string | null;
 };
 
@@ -68,10 +76,7 @@ export function deriveMeshFieldModel({
       selfParticipating: true,
       selfCapacityGb: view.selfCapacityGb,
       ghostCount,
-      caption:
-        view.peers.length === 0
-          ? "Connected · waiting for another device"
-          : null,
+      caption: view.peers.length === 0 ? "Waiting for another device" : null,
     };
   }
 
@@ -103,6 +108,6 @@ export function deriveMeshFieldModel({
     selfParticipating: false,
     selfCapacityGb: null,
     ghostCount,
-    caption: "Last reported by the community · join to see live detail",
+    caption: null,
   };
 }
diff --git a/desktop/src/features/mesh-compute/meshMemoryModel.test.mjs b/desktop/src/features/mesh-compute/meshMemoryModel.test.mjs
index f5d7cadc7..27a6bc08e 100644
--- a/desktop/src/features/mesh-compute/meshMemoryModel.test.mjs
+++ b/desktop/src/features/mesh-compute/meshMemoryModel.test.mjs
@@ -33,7 +33,7 @@ test("prefers the runtime's reported model footprint", () => {
     catalog,
     modelRef: "gemma",
   });
-  assert.equal(model.label, "10 GB of 40 GB AI memory used by model");
+  assert.equal(model.label, "This computer · model uses 10 GB of 40 GB");
   assert.equal(model.usedSegments, 3);
 });
 
@@ -43,7 +43,7 @@ test("falls back to the selected catalog model before runtime status arrives", (
     catalog,
     modelRef: "gemma",
   });
-  assert.equal(model.label, "8 GB of 32 GB AI memory used by model");
+  assert.equal(model.label, "This computer · model uses 8 GB of 32 GB");
   assert.equal(model.usedSegments, 3);
 });
 
@@ -53,6 +53,32 @@ test("unknown model size shows available memory without inventing use", () => {
     catalog,
     modelRef: "custom/model",
   });
-  assert.equal(model.label, "32 GB AI memory available");
+  assert.equal(model.label, "This computer · 32 GB AI memory");
   assert.equal(model.usedSegments, 0);
 });
+
+test("every label names this computer, never the mesh", () => {
+  // The meter sits directly under a radar field showing the whole community.
+  // An unqualified "10 of 40 GB" would read as mesh-wide utilization, which is
+  // a number nobody has: members publish their own capacity, not each other's
+  // live allocation.
+  const cases = [
+    {
+      view: {
+        connected: true,
+        selfCapacityGb: 40,
+        selfModelSizeGb: 10,
+        peers: [],
+      },
+      catalog,
+      modelRef: "gemma",
+    },
+    { view: null, catalog, modelRef: "gemma" },
+    { view: null, catalog, modelRef: "custom/model" },
+    { view: null, catalog: null, modelRef: null },
+  ];
+  for (const input of cases) {
+    const model = deriveMeshMemoryModel(input);
+    assert.match(model.label, /this computer/i, model.label);
+  }
+});
diff --git a/desktop/src/features/mesh-compute/meshMemoryModel.ts b/desktop/src/features/mesh-compute/meshMemoryModel.ts
index 89ee6e8a8..8bee4c71f 100644
--- a/desktop/src/features/mesh-compute/meshMemoryModel.ts
+++ b/desktop/src/features/mesh-compute/meshMemoryModel.ts
@@ -11,9 +11,16 @@ export type MeshMemoryModel = {
 };
 
 /**
- * Approximate model footprint against this machine's shareable AI memory.
- * MeshLLM reports model package size, not live OS GPU allocation, so the UI
- * deliberately says "model uses" rather than claiming real-time utilization.
+ * Approximate model footprint against **this machine's** shareable AI memory.
+ *
+ * Scoped to this node on purpose, and the label says so. It sits directly under
+ * a radar field showing the whole community, so an unqualified "36 of 115 GB"
+ * would read as mesh-wide utilization — a number we do not have and could not
+ * compute, since members publish their own capacity but nothing about each
+ * other's live allocation.
+ *
+ * MeshLLM reports model package size, not live OS GPU allocation, so the label
+ * says the model "uses" rather than claiming real-time utilization.
  */
 export function deriveMeshMemoryModel({
   view,
@@ -47,12 +54,14 @@ export function deriveMeshMemoryModel({
     usedGb,
     usedSegments,
     totalSegments: MEMORY_SEGMENTS,
+    // "This computer" leads, because the meter sits under a community-wide
+    // radar field and must not be read as mesh utilization.
     label:
       usedGb != null && totalGb != null
-        ? `${formatGb(usedGb)} of ${formatGb(totalGb)} AI memory used by model`
+        ? `This computer · model uses ${formatGb(usedGb)} of ${formatGb(totalGb)}`
         : totalGb != null
-          ? `${formatGb(totalGb)} AI memory available`
-          : "Checking available AI memory…",
+          ? `This computer · ${formatGb(totalGb)} AI memory`
+          : "Checking this computer's AI memory…",
   };
 }