diff --git a/cmd/tester/probe.go b/cmd/tester/probe.go
index 0836b9a..e5f63b4 100644
--- a/cmd/tester/probe.go
+++ b/cmd/tester/probe.go
@@ -123,6 +123,17 @@ func newHTTPClient(opts probeOpts) *http.Client {
 // cancelled. Each completed probe posts a probeResultMsg into the
 // tea.Program via send.
 //
+// The probe duration is subtracted from the sleep so the loop's
+// *period* is the interval, not interval + probe time. If a 100ms
+// interval probe takes 30ms, the next sleep is 70ms (minus jitter),
+// not 100ms — otherwise a VIP with 30ms probes would actually fire
+// every 130ms and the --interval flag would quietly lie. On the
+// pathological case where a probe overruns the interval (slow VIP
+// under load) we clamp the sleep to zero and fire the next probe
+// immediately, but we never try to "catch up" by firing multiple
+// back-to-back probes — that would flood an already-struggling
+// backend right when it's slow.
+//
 // The loop honors the global `paused` flag by simply skipping the
 // probe call while paused — the ticker keeps ticking so a resume
 // picks up at the next natural tick boundary instead of fast-
@@ -141,6 +152,10 @@ func runProbeLoop(ctx context.Context, vip *vipInfo, opts probeOpts, send func(t
 		if !paused.Load() {
 			result := doProbe(ctx, vip, opts)
 			send(result)
+			sleepFor -= result.Duration
+			if sleepFor < 0 {
+				sleepFor = 0
+			}
 		}
 		select {
 		case <-ctx.Done():