diff --git a/tests/coverage-floor.json b/tests/coverage-floor.json
index 1587c5e..5097239 100644
--- a/tests/coverage-floor.json
+++ b/tests/coverage-floor.json
@@ -1,8 +1,8 @@
 {
   "anywhere": {
-    "Bws.Core": 71.9,
-    "Bws.Gui": 86.2,
-    "Bws.Cli": 9.5
+    "Bws.Core": 89.6,
+    "Bws.Gui": 91.5,
+    "Bws.Cli": 80.0
   },
   "everything": {
     "Bws.Core": 91.3,
diff --git a/tests/coverage-gate.ps1 b/tests/coverage-gate.ps1
index 01b72ca..7fcda25 100644
--- a/tests/coverage-gate.ps1
+++ b/tests/coverage-gate.ps1
@@ -41,6 +41,16 @@
 # 2103 covered, identically. The property tests draw random seeds, so this was worth checking -
 # a gate that wobbles teaches everybody to ignore it.
 #
+# THE anywhere FLOORS SIT HALF A POINT UNDER WHAT CI MEASURES, set by hand on 2026-11-02 rather
+# than with -Mark, which writes the measurement itself. Until that day they were the August
+# figures, 5.8 to 71.1 points under seven CI runs (Bws.Cli at 80.5% over a floor of 9.5), so a
+# test project could lose most of its coverage without a word. And the paragraph above stopped
+# being the whole truth: Bws.Cli measured 80.5% and 80.6% on the same tree in two CI runs, while
+# Bws.Core (90.1%) and Bws.Gui (92%) held still in all seven. A floor equal to a figure that moves
+# by a tenth fails a build that changed nothing, and half a point is five times that movement. To
+# raise them again, read the step "coverage against its floor" in CI and take half a point off the
+# lowest figure there.
+#
 # WHAT THIS DOES NOT TELL YOU, and it is the reason backlog 62 exists: coverage says what was
 # EXECUTED, never what was CHECKED. A test with no assertions covers everything it touches. The
 # answer to "was it checked" is the mutation registry, and this does not replace it.