diff --git a/sage/embodiment/ear.py b/sage/embodiment/ear.py
index 3162008f7..c90e0e34f 100644
--- a/sage/embodiment/ear.py
+++ b/sage/embodiment/ear.py
@@ -13,6 +13,14 @@
     the room: "only being directly addressed should wake it" -- run with SAGE_LISTEN=wake and SAGE_WAKE_NAME,
     so words become heard lines only after the being's name (listening.py).
 
+  * TRANSCRIBE OUT OF PROCESS: SAGE_LISTEN_BACKEND=whispercpp. The capture is read by a Python thread, so
+    a transcriber that decodes in this process competes with it for the GIL. On McNugget (2026-10-24) the
+    MLX backend, with music in the room, was decoding in 20 of 20 py-spy samples and the reader in none:
+    samples were lost, "hey McNugget" reached whisper as fragments ("Okay." x28, dropped as silence), and
+    five days of tries never chimed. whisper-server (sage/embodiment/whispercpp.py) decodes in its own
+    process; the same audio replayed through it woke and heard correctly, and the first live try chimed.
+    MLX is fine for a quiet room or a test; a room with music needs the out-of-process backend.
+
 Run as a service:  python -m sage.embodiment.ear   (SAGE_INSTANCE / SAGE_MEMBER as for presence.py)
 """
 from __future__ import annotations
diff --git a/sage/embodiment/listening.py b/sage/embodiment/listening.py
index f30c4d3bf..620bc88ff 100644
--- a/sage/embodiment/listening.py
+++ b/sage/embodiment/listening.py
@@ -294,6 +294,8 @@ def _load(self):
         if BACKEND == "mlx":
             # Apple silicon (McNugget, 2026-10-17): the same whisper model through MLX on the GPU. Measured there:
             # base.en ~0.1 s per short phrase warm. Its segments carry no_speech_prob / avg_logprob like torch's.
+            # IT DECODES IN THIS PROCESS, holding the GIL the capture reader needs: in a room with music it starved
+            # the reader on McNugget (2026-10-24, see ear.py). For a live ear prefer whispercpp (its own process).
             try:
                 import numpy as np
                 import mlx_whisper