diff --git a/src/components/latency-chart.test.tsx b/src/components/latency-chart.test.tsx
index 3ca51ba..136aec1 100644
--- a/src/components/latency-chart.test.tsx
+++ b/src/components/latency-chart.test.tsx
@@ -367,6 +367,32 @@ describe('latency chart on sample data', () => {
     }
   })
 
+  it('agrees with the p95 tile, as a live platform would', () => {
+    // On a platform the tile and the chart's newest point are one number. The demo
+    // used to print 88 ms on the tile beside a curve ending in the 40s — sample
+    // data, but a false lesson about what the two panels mean.
+    const restoreEventSource = installFakeEventSource()
+    vi.useFakeTimers()
+    try {
+      const { result } = renderHook(() => useConsoleData({ sampleDelayMs: 500 }))
+      act(() => {
+        vi.advanceTimersByTime(600) // no platform answered: the demo takes over
+      })
+      expect(result.current.conn.status).toBe('sample')
+      const { latency, summary } = result.current.data!
+      const newest = latency[latency.length - 1]
+      const windowBack = latency.findLast((p) => p.at <= newest.at - summary!.windowMs)!
+      expect(summary!.p95LatencyMs).toBe(newest.p95)
+      expect(summary!.p95DeltaMs).toBe(newest.p95 - windowBack.p95)
+
+      const { getByText } = render(<KpiRow metrics={result.current.data!.metrics} summary={summary} />)
+      expect(getByText(String(newest.p95))).toBeTruthy()
+    } finally {
+      vi.useRealTimers()
+      restoreEventSource()
+    }
+  })
+
   it('drops the label once real samples replace it', () => {
     const { queryByText, rerender } = render(<LatencyChart points={mockLatencySeries(T0)} range="15m" sample />)
     rerender(<LatencyChart points={run(T0, 2)} range="15m" now={T0 + 1_000} />)
diff --git a/src/hooks/use-console-data.ts b/src/hooks/use-console-data.ts
index 812abf9..0397080 100644
--- a/src/hooks/use-console-data.ts
+++ b/src/hooks/use-console-data.ts
@@ -166,12 +166,13 @@ function int(v: unknown): number | null {
 }
 
 function mockData(now: number): LiveData {
+  const latency = mockLatencySeries(now)
   return {
     traces: mockTraces(now),
     metrics: mockMetrics(now),
     modules: mockModules(),
-    summary: mockSummary(),
-    latency: mockLatencySeries(now),
+    summary: mockSummary(latency),
+    latency,
   }
 }
 
diff --git a/src/lib/mock.ts b/src/lib/mock.ts
index c3c4943..6a18f40 100644
--- a/src/lib/mock.ts
+++ b/src/lib/mock.ts
@@ -123,19 +123,28 @@ export function mockModules(): ModuleStatus[] {
 
 // Windowed KPI header aggregate, mirroring the platform `summary` SSE event.
 // Trends are populated (not null) so the sample demo shows the fully-wired header.
-export function mockSummary(): TraceSummary {
+//
+// The p95 and its trend are read off the sample latency series rather than made up
+// beside it. On a real platform the p95 tile and the chart's newest point are the
+// same number, and a demo that showed them disagreeing would teach the wrong thing
+// about the screen — sample data is still a claim about how the console behaves.
+export function mockSummary(latency: readonly LatencyPoint[]): TraceSummary {
+  const windowMs = 60000
+  const newest = latency[latency.length - 1]
+  // The point one window back is what the platform compares against.
+  const prev = latency.findLast((p) => p.at <= newest.at - windowMs) ?? latency[0]
   return {
-    windowMs: 60000,
+    windowMs,
     count: 512,
     errorCount: 7,
     errorRate: 0.0137,
     p50LatencyMs: 12,
-    p95LatencyMs: 88,
+    p95LatencyMs: newest.p95,
     p99LatencyMs: 210,
     maxLatencyMs: 640,
     requestsDeltaPct: 0.12,
     errorRateDeltaPp: 0.31,
-    p95DeltaMs: -6,
+    p95DeltaMs: newest.p95 - prev.p95,
     activeModules: 6,
     modulesByMode: { worker: 3, container: 1, 'in-process': 2 },
   }