diff --git a/docs/tower-installation.md b/docs/tower-installation.md
index edfff4d..447eb19 100644
--- a/docs/tower-installation.md
+++ b/docs/tower-installation.md
@@ -734,11 +734,12 @@ itself when a camera has `ExposureTimeAbs` and no `ExposureTime`: frame trigger
 off (the legacy cameras booted with `TriggerMode = On`, `Line1`), `BayerRG8` on
 the colour camera and `Mono8` on a mono one, 60 MB/s each, the frame rate at
 the camera's maximum (rietschen's RGB camera came set to 0.955 fps), and
-auto-exposure on the frame mean (`ExposureAutoAlg = Mean`) at 50 % +- 5 %, down
-to 20 us by day. That last one matters: the keep rule waits for the frame mean
-to reach 50 % +- 5 %. At the cameras' old target of 25 every waypoint ran into
-`frame_timeout_ms`, and so it did with the RGB camera's `FitRange` and 10 %
-tolerance, which fit the histogram wherever its mean lands.
+auto-exposure on the frame mean (`ExposureAutoAlg = Mean`) at 35 % +- 5 %, down
+to 20 us by day. The keep rule waits for the frame mean to reach the camera's
+own target and tolerance, read once the camera is open. With the RGB camera's
+`FitRange`, which fits the histogram wherever its mean lands, no waypoint
+settled. 35 rather than the Alvium's 50: at 50 the sky over the horizon burnt
+out at noon, where smoke would rise; the old program ran these cameras at 23-25.
 
 Give `--camera` the id: with several cameras the first one the SDK lists is
 not the colour one, and the order changes from run to run. `install.sh
diff --git a/scripts/export_camera_profile.py b/scripts/export_camera_profile.py
index 2b972b4..f572fb5 100755
--- a/scripts/export_camera_profile.py
+++ b/scripts/export_camera_profile.py
@@ -151,15 +151,15 @@ _GIGE_COMMON = [
     ("ExposureTimeAbs",            1000.0),
     ("AcquisitionFrameRateAbs",    1000.0),      # clamped to the camera's maximum
     ("ChunkModeActive",            False),
-    # The keep rule waits for the frame mean to reach the camera's target. The
-    # host starts on its default band of 50 % - rietschen's cameras were set to
-    # 25, and every waypoint then ran into frame_timeout_ms (7.5 s, 2026-10-20).
-    # The camera has to aim at the mean, and stop adjusting only inside the same
-    # +-5 %: the RGB camera came with FitRange (which fits the histogram, wherever
-    # the mean lands) and a 10 % tolerance, and no waypoint settled (2026-10-21).
+    # The keep rule waits for the frame mean to reach the camera's target, +- its
+    # tolerance, which it reads from the camera. The camera has to aim at the
+    # mean: the RGB camera came with FitRange (which fits the histogram, wherever
+    # the mean lands) and no waypoint settled (2026-10-21). 35 % rather than the
+    # Alvium's 50: at 50 the sky over the horizon burnt out at noon, where smoke
+    # would rise; the old program ran these cameras at 23-25.
     ("ExposureAutoAlg",            "Mean"),
     ("ExposureAutoAdjustTol",      5),
-    ("ExposureAutoTarget",         50),
+    ("ExposureAutoTarget",         35),
     ("ExposureAuto",               "Continuous"),
     ("GainAuto",                   "Continuous"),
 ]
diff --git a/src/core/Application.cpp b/src/core/Application.cpp
index 7d94ba2..11cbb1b 100644
--- a/src/core/Application.cpp
+++ b/src/core/Application.cpp
@@ -3627,6 +3627,12 @@ struct Application::Impl {
             return 3;  // distinct from the config (1) and lock (2) exits
         }
 
+        // Again, now the camera is open: the keep rule reads the camera's own
+        // auto-exposure target, tolerance and limits, and before open() it got
+        // none of them and judged every frame against its 50 +- 5 % default -
+        // until the first profile switch. Rietschen's cameras aim at 35 %.
+        configureSettle();
+
         pipeline->start();
         channel->publishStatus(static_cast<int>(Mode::Manual));   // until [Startup] mode says otherwise