diff --git a/AGENTS.md b/AGENTS.md
index aa84aa9..87c3c82 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -215,15 +215,15 @@ band transcription create <call-id> <rec-id> --wait             # blocks until t
 All `--wait` commands support `--timeout <seconds>`. Exit code 5 on timeout.
 
 **`--timeout` bounds when polling stops, not wall-clock duration.** The
-underlying poll loop (`internal/cmdutil/poll.go`) checks the deadline
-*between* poll attempts, not during one, and always sleeps the full
-`--interval` rather than whatever time remains before the deadline. So a
-call can run past `--timeout` by up to one poll interval plus one in-flight
-request — and if that late attempt happens to succeed, the command exits
-**0**, not 5, seconds after the timeout you asked for. Concretely: a 5s
-poll interval with `--timeout 1` can still succeed at t≈5s. Do not treat
-`--timeout` as a precise deadline; treat it as a lower bound on how long the
-CLI will keep trying.
+underlying poll loop (`internal/cmdutil/poll.go`) starts no new poll after
+the deadline, and the sleep before the deadline is capped so the final poll
+lands on the deadline rather than a full interval past it. A poll that
+starts on time and succeeds returns exit **0** with the result even if it
+finishes slightly after the deadline — the operation genuinely completed.
+So total overshoot is bounded by one in-flight request (the API client's
+HTTP timeout in the worst case), not by the poll interval: a 5s poll
+interval with `--timeout 1` resolves at t≈1s plus one request, as success
+or as exit 5.
 
 ## Output
 
diff --git a/internal/cmdutil/poll.go b/internal/cmdutil/poll.go
index 78ba7ef..5d062be 100644
--- a/internal/cmdutil/poll.go
+++ b/internal/cmdutil/poll.go
@@ -28,6 +28,14 @@ type PollConfig struct {
 // Poll runs cfg.Check repeatedly at cfg.Interval until it returns done=true or
 // cfg.Timeout is exceeded. On success it returns the result from Check.
 // On timeout it returns ErrPollTimeout.
+//
+// cfg.Timeout bounds when polling stops, not total wall-clock duration: no
+// check starts after the deadline, and the sleep before the deadline is
+// capped so the final check lands on the deadline rather than a full
+// interval past it. A check that starts on time and completes successfully
+// returns its result even if it finishes after the deadline — the operation
+// genuinely completed, and discarding that would be worse than being
+// slightly late. Total overshoot is bounded by one in-flight request.
 func Poll(cfg PollConfig) (interface{}, error) {
 	ctx := cfg.Context
 	if ctx == nil {
@@ -43,9 +51,14 @@ func Poll(cfg PollConfig) (interface{}, error) {
 		if done {
 			return result, nil
 		}
-		if time.Now().After(deadline) {
+		remaining := time.Until(deadline)
+		if remaining <= 0 {
 			return nil, fmt.Errorf("timed out after %s: %w", cfg.Timeout, ErrPollTimeout)
 		}
+		wait := cfg.Interval
+		if remaining < wait {
+			wait = remaining
+		}
 		// A fresh timer per iteration. Go 1.23+ made timer channels
 		// unbuffered and Stop() cancels any in-flight send, so there is
 		// nothing left to drain after Stop() returns — draining an
@@ -53,7 +66,7 @@ func Poll(cfg PollConfig) (interface{}, error) {
 		// Stop() on the cancellation path is a courtesy; an abandoned
 		// timer is garbage collected once unreferenced, so there is no
 		// leak either way.
-		timer := time.NewTimer(cfg.Interval)
+		timer := time.NewTimer(wait)
 		select {
 		case <-ctx.Done():
 			timer.Stop()
diff --git a/internal/cmdutil/poll_test.go b/internal/cmdutil/poll_test.go
index 1a79f28..e8251a5 100644
--- a/internal/cmdutil/poll_test.go
+++ b/internal/cmdutil/poll_test.go
@@ -77,6 +77,51 @@ func TestPollRespectsContextCancellation(t *testing.T) {
 	}
 }
 
+// The final sleep must be capped at the time remaining before the deadline,
+// not a full interval — a 250ms interval with a 20ms timeout should report
+// the timeout at ~20ms, not ~250ms.
+func TestPollTimeoutDoesNotOvershootByFullInterval(t *testing.T) {
+	start := time.Now()
+	_, err := Poll(PollConfig{
+		Interval: 250 * time.Millisecond,
+		Timeout:  20 * time.Millisecond,
+		Check:    func() (bool, interface{}, error) { return false, nil, nil },
+	})
+	elapsed := time.Since(start)
+	if !errors.Is(err, ErrPollTimeout) {
+		t.Fatalf("err = %v, want ErrPollTimeout", err)
+	}
+	if elapsed >= 150*time.Millisecond {
+		t.Errorf("elapsed = %v, want ~20ms (final sleep must be capped at the deadline, not the full interval)", elapsed)
+	}
+}
+
+// A check that starts at the deadline and succeeds still returns the result —
+// the operation genuinely completed, and it does so at ~deadline rather than
+// a full interval later.
+func TestPollFinalCheckAtDeadlineCanSucceed(t *testing.T) {
+	start := time.Now()
+	calls := 0
+	got, err := Poll(PollConfig{
+		Interval: 250 * time.Millisecond,
+		Timeout:  20 * time.Millisecond,
+		Check: func() (bool, interface{}, error) {
+			calls++
+			return calls >= 2, "done", nil
+		},
+	})
+	elapsed := time.Since(start)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if got != "done" {
+		t.Fatalf("got %v, want done", got)
+	}
+	if elapsed >= 150*time.Millisecond {
+		t.Errorf("elapsed = %v, want ~20ms (the pre-deadline sleep must be capped so the final check runs at the deadline)", elapsed)
+	}
+}
+
 // A nil Context must behave exactly as before — all existing callers omit it.
 func TestPollNilContextStillTimesOut(t *testing.T) {
 	_, err := Poll(PollConfig{