diff --git a/CHANGELOG.md b/CHANGELOG.md
index 1ab0bb7..3487d26 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -7,7 +7,37 @@ inside each artifact (`schema_version`).
 
 ## [Unreleased]
 
-## [0.9.1] - 2026-09-27
+## [0.9.2] - 2026-10-11
+
+The WoE floor no longer ignores rare 0/1 columns. No change to the runtime,
+the exporters, the artifact schema, or the benchmark, whose features are all
+continuous.
+
+`fit_reference` bins each feature with a shallow tree whose leaves must hold
+at least 2% of rows. A two-valued column whose rarer value fell below 2%
+could not be split, so it landed in one bin and carried no information: the
+floor silently ignored it. This surfaced on public hospital-readmission data
+(#76), where one-hot columns for rare discharge outcomes were lost, including
+patients who died, 1.5% of rows and never readmitted. The floor's Gini there
+rises from 0.235 to 0.300.
+
+It matters beyond the reference itself. `compile_selected`'s floor gate
+compares against this model, and so does validation check 10 when it is given
+a fitted reference. A weakened floor let
+through artifacts a fair one would have stopped. And 0.9.1 advises adding
+0/1 missing-indicator columns, exactly the kind of column the floor dropped.
+
+### Fixed
+- A column with exactly two values is already binned. The floor now cuts it
+  between the two values, where the tree would, whatever the size of the
+  rarer one. The 2% minimum still governs the cut points the tree chooses on
+  columns with more values, which is what it exists for. A 0/1 column the
+  tree could already split gets the same cut as before.
+- On data with rare two-valued columns, floor figures change and the floor
+  gets stronger: a floor gate or check 10 that passed may now fail. That is
+  the fix working.
+
+## [0.9.1] - 2026-10-10
 
 Missing values are refused instead of mis-compiled. No change to the runtime,
 the exporters, the artifact schema, or any artifact built from data without
diff --git a/CITATION.cff b/CITATION.cff
index 0f7fc3c..8e70470 100644
--- a/CITATION.cff
+++ b/CITATION.cff
@@ -19,8 +19,8 @@ authors:
 repository-code: "https://github.com/orgoca/CompileML"
 url: "https://github.com/orgoca/CompileML"
 license: Apache-2.0
-version: 0.9.1
-date-released: "2026-09-27"
+version: 0.9.2
+date-released: "2026-10-11"
 keywords:
   - credit risk
   - credit scoring
diff --git a/src/compileml/__init__.py b/src/compileml/__init__.py
index dfbc672..9ae62a2 100644
--- a/src/compileml/__init__.py
+++ b/src/compileml/__init__.py
@@ -16,7 +16,7 @@
 # this attribute at build time (setuptools dynamic version), so the wheel
 # metadata, `compileml.__version__`, `compileml inspect`, and the
 # `compileml_version` recorded inside every artifact can never disagree.
-__version__ = "0.9.1"
+__version__ = "0.9.2"
 
 __all__ = ["compile_selected", "decide", "load_artifact", "verify_artifact", "__version__"]
 
diff --git a/src/compileml/reference/woe.py b/src/compileml/reference/woe.py
index 231eea9..49c5866 100644
--- a/src/compileml/reference/woe.py
+++ b/src/compileml/reference/woe.py
@@ -78,9 +78,17 @@ def fit(self, X, y):
                 min_samples_leaf=self.min_samples_leaf,
                 random_state=self.random_state,
             )
+            values = np.unique(column)
             edges: np.ndarray
-            if np.unique(column).size < 2:
+            if values.size < 2:
                 edges = np.empty(0, dtype=float)  # constant feature: one bin
+            elif values.size == 2:
+                # Already binned: the one possible cut is between the two
+                # values, where the tree would put it. The leaf minimum stops
+                # a tree choosing cut points on noise; with no choice to make,
+                # a rare value (a 0/1 indicator at 1%) keeps its own bin
+                # instead of silently merging into the common one.
+                edges = np.array([(values[0] + values[1]) / 2.0])
             else:
                 tree.fit(column.reshape(-1, 1), y)
                 inner = tree.tree_.feature != -2  # -2 is sklearn's leaf sentinel
@@ -162,6 +170,10 @@ def fit_reference(
         kind: Only ``"woe_logit"`` today — supervised binning,
             weight-of-evidence encoding, then logistic regression.
         max_bins: Maximum bins per feature (shallow by design).
+        min_samples_leaf: Smallest bin, as a share of rows, when a tree
+            chooses the cut points. A column with exactly two values is
+            already binned and keeps both bins whatever their sizes, so a
+            rare 0/1 indicator still reaches the floor.
         cv: Folds used to choose ``C``. The WOE tables are refit inside each
             fold, so the choice is not made against leaked encodings.
 
diff --git a/tests/test_reference_floor.py b/tests/test_reference_floor.py
index 6ceaddd..80423fc 100644
--- a/tests/test_reference_floor.py
+++ b/tests/test_reference_floor.py
@@ -101,6 +101,25 @@ def test_fit_reference_rejects_bad_input(data):
         fit_reference(X[:500], np.zeros(500, dtype=int))
 
 
+def test_a_rare_two_valued_column_keeps_its_bin():
+    """A 0/1 indicator below the 2% leaf minimum still reaches the floor.
+
+    Before 0.9.2 the binning tree could not split it, so the floor ignored a
+    column that decides 1% of rows outright — and the floor gate got lenient.
+    """
+    rng = np.random.default_rng(5)
+    n = 20000
+    x = rng.standard_normal(n)
+    rare = (rng.random(n) < 0.01).astype(float)  # 1% of rows, never an event
+    common = (rng.random(n) < 0.3).astype(float)
+    y = np.where(rare == 1, 0, (rng.random(n) < 1 / (1 + np.exp(-x))).astype(int))
+    ref = fit_reference(np.column_stack([x, rare, common]), y, feature_names=["x", "rare", "c"])
+    assert ref.thresholds[1] == [0.5] and len(ref.woe[1]) == 2
+    assert ref.information_value["rare"] > 0.05
+    assert ref.coefficients["rare"] != 0
+    assert ref.thresholds[2] == [0.5]  # a common 0/1 column: the cut the tree always made
+
+
 # ------------------------------------------------------ the target axis (4.2)
 def test_alpha_axis_sweeps_the_target(data, reference):
     X, teacher, y = data