diff --git a/Services/Audio/AudioMixer.cs b/Services/Audio/AudioMixer.cs
index e773b22..dd0b6e3 100644
--- a/Services/Audio/AudioMixer.cs
+++ b/Services/Audio/AudioMixer.cs
@@ -402,7 +402,12 @@ public sealed class AudioMixer : IDisposable
         // any scheduler error into a permanent A/V rate error (measured 2026-11-12: 2.1% fast, 224 ms
         // of drift in ten seconds); asking the clock how much is owed and producing exactly that keeps
         // the track the same length as the show.
-        Stopwatch? clock = null;
+        // The clock runs from the moment the session opens, NOT from the first successful write. Starting
+        // it at connect made the track begin that much later than the picture: measured 2026-11-12, the
+        // writes began 160 ms after ffmpeg started reading and the audio stream came out 137 ms shorter
+        // than the video, so every word sat 160 ms late. Owed time is paid out of the capture ring when
+        // the pipe connects, which is the audio that was missed — real audio, not padding.
+        var clock = Stopwatch.StartNew();
         long framesProduced = 0;
         var maxChunkFrames = micChunk.Length;
 
@@ -411,15 +416,14 @@ public sealed class AudioMixer : IDisposable
             var nextTick = DateTime.UtcNow + _mixInterval;
             try
             {
-                // The clock starts when ffmpeg is actually reading. Before that there is nowhere for
-                // audio to go, and time spent waiting must not become a shorter track.
+                // Nowhere for audio to go until ffmpeg is reading: wait without producing, so the debt
+                // stays owed and is paid from the ring the moment the pipe connects.
                 if (!pipe.IsConnected)
                 {
                     await Task.Delay(1, cancellationToken).ConfigureAwait(false);
                     continue;
                 }
 
-                clock ??= Stopwatch.StartNew();
                 var frames = AudioClock.FramesDue(
                     clock.Elapsed.TotalSeconds, framesProduced, OutputSampleRate, maxChunkFrames);
                 if (frames <= 0)
diff --git a/docs/ai/output.md b/docs/ai/output.md
index ff3c444..3aa90d4 100644
--- a/docs/ai/output.md
+++ b/docs/ai/output.md
@@ -308,9 +308,11 @@ devices, no timers).
   the lips across ten seconds and no static delay could fix it. Buffers size to the largest chunk the
   planner may return (`OutputSampleRate / 5`), and a stalled loop pays its debt over the following ticks
   rather than dumping one huge write. Video never had this bug — the pump paces off a stopwatch.
-  **The clock also starts only once ffmpeg is connected, and `Advance` counts a frame only when its write
-  reached the pipe**: the first clock-driven recording came out **226 ms short** because 27 start-up
-  writes were dropped before the encoder connected, which is the same desync wearing the opposite sign.
+  **`Advance` counts a frame only when its write reached the pipe**, and the clock runs from session
+  open, not from the first successful write. Both come from measurement: 27 start-up writes were being
+  counted while dropped (**226 ms short**), and starting the clock at connect instead made the track
+  begin 160 ms after the picture (**137 ms short**, every word late). The owed time is paid out of the
+  capture ring when the pipe connects — that is the audio that was missed, not padding.
 - **TRAX — free background music (TASK 8):** `MusicPlayer` = NAudio `MediaFoundationReader`
   (mp3/wav/m4a) → `VolumeWaveProvider16` at the hardcoded **0.20** bed (no slider) → `WaveOutEvent` on
   the default device, **looping on any clean natural end** (`PlaybackStopped` with `e.Exception == null`