diff --git a/src/vocalinux/speech_recognition/model_ladder.py b/src/vocalinux/speech_recognition/model_ladder.py
index 3cd0892..dc4150a 100644
--- a/src/vocalinux/speech_recognition/model_ladder.py
+++ b/src/vocalinux/speech_recognition/model_ladder.py
@@ -8,12 +8,14 @@ Rungs (quantized Turbo first, small floor last)::
 
     large-v3-turbo-q8_0 -> large-v3-turbo-q5_0 -> small.en / small
 
-Thresholds derive from ticket #19's measured transcribe peaks (1132 MiB
-for q8_0, 846 MiB for q5_0) plus the 512 MiB desktop headroom from the
-locked baseline protocol
-(``scripts/measure_perf_baseline.py::recommend_thresholds``). The
-derivation is pinned by ``tests/test_model_ladder.py`` so a re-measure
-forces a threshold review.
+The thresholds are PROVISIONAL. They are the rule
+``scripts/measure_perf_baseline.py::recommend_thresholds`` applies (measured
+footprint plus the 512 MiB desktop headroom) to the transcribe peaks
+recorded for ticket #19 (1132 MiB for q8_0, 846 MiB for q5_0). Ticket #19
+is still open and its measurement is not recorded in this repository, so
+these numbers are the calibration to replace, not evidence. A re-measure
+changes the constants here, and ``tests/test_model_ladder.py`` pins them so
+that change is deliberate.
 
 Backend order follows the same readiness rule: CUDA ranks before Vulkan
 only when a complete architecture-matching CUDA toolkit is installed
@@ -39,7 +41,7 @@ LADDER_RUNGS: tuple[str, ...] = (DEFAULT_RUNG, LOW_MEMORY_RUNG)
 #: the baseline protocol adds to measured footprints.
 HEADROOM_MIB: float = 512.0
 
-#: Ticket #19 measured transcribe peaks (MiB) the thresholds below derive from.
+#: Provisional transcribe peaks (MiB) from ticket #19, not yet recorded in the repo.
 _MEASURED_Q8_0_PEAK_MIB: float = 1132.0
 _MEASURED_Q5_0_PEAK_MIB: float = 846.0
 
diff --git a/tests/test_injection_ladder.py b/tests/test_injection_ladder.py
index aefcead..ac52ac7 100644
--- a/tests/test_injection_ladder.py
+++ b/tests/test_injection_ladder.py
@@ -234,7 +234,9 @@ class TestNiriGuard:
     ) -> None:
         """The compositor-bridge predicate itself answers no on niri."""
         monkeypatch.setenv("XDG_CURRENT_DESKTOP", "niri")
-        monkeypatch.setattr(TextInjector, "_ibus_wayland_bridge_running", staticmethod(lambda: True))
+        monkeypatch.setattr(
+            TextInjector, "_ibus_wayland_bridge_running", staticmethod(lambda: True)
+        )
         injector = _bare_wayland_injector(environment=DesktopEnvironment.WAYLAND)
         assert injector._wayland_compositor_bridges_ibus() is False
 
@@ -263,9 +265,7 @@ class TestKeyEventToolOrder:
 
     @staticmethod
     def _tools(monkeypatch: pytest.MonkeyPatch, names: tuple) -> None:
-        monkeypatch.setattr(
-            "shutil.which", lambda cmd: f"/usr/bin/{cmd}" if cmd in names else None
-        )
+        monkeypatch.setattr("shutil.which", lambda cmd: f"/usr/bin/{cmd}" if cmd in names else None)
 
     def test_wtype_outranks_a_ready_ydotool(self, monkeypatch: pytest.MonkeyPatch) -> None:
         self._tools(monkeypatch, ("wtype", "ydotool"))
diff --git a/tests/test_model_ladder.py b/tests/test_model_ladder.py
index f747220..296b8ff 100644
--- a/tests/test_model_ladder.py
+++ b/tests/test_model_ladder.py
@@ -2,9 +2,10 @@
 
 All behavior is exercised with mocked probes and loaders: no GPU, no
 network, no hardware. The threshold-derivation test pins the rung
-thresholds to ticket #19's measured transcribe peaks (1132 MiB for q8_0,
-846 MiB for q5_0) plus the 512 MiB protocol headroom, so any re-measure
-forces a threshold review.
+thresholds to the provisional transcribe peaks (1132 MiB for q8_0, 846 MiB
+for q5_0, ticket #19, not yet recorded in the repo) plus the 512 MiB
+protocol headroom. Replacing them with a recorded measurement is a
+deliberate change to this test.
 """
 
 from unittest.mock import MagicMock, patch