diff --git a/src/pyrecest/utils/association_features.py b/src/pyrecest/utils/association_features.py
index 447592c5dc..ebc30076ce 100644
--- a/src/pyrecest/utils/association_features.py
+++ b/src/pyrecest/utils/association_features.py
@@ -403,7 +403,10 @@ def _finite_feature_plane(values: Any, feature_name: str) -> Any:
 
 
 def _flatten_prediction_features(features: Any) -> tuple[Any, tuple[int, ...]]:
-    features = asarray(features, dtype=float64)
+    features = _as_real_numeric_backend_array(
+        features,
+        message="features must contain real numeric values",
+    )
     if features.ndim == 0:
         raise ValueError("features must be at least one-dimensional")
     if features.ndim == 1:
diff --git a/tests/utils/test_calibrated_association_feature_validation.py b/tests/utils/test_calibrated_association_feature_validation.py
new file mode 100644
index 0000000000..ab4e8d89ef
--- /dev/null
+++ b/tests/utils/test_calibrated_association_feature_validation.py
@@ -0,0 +1,31 @@
+"""Regression tests for calibrated association feature validation."""
+
+import pytest
+
+from pyrecest.backend import array
+from pyrecest.utils import CalibratedPairwiseAssociationModel
+
+
+class _RecordingPredictProbaModel:
+    classes_ = array([0, 1])
+
+    def __init__(self):
+        self.called = False
+
+    def predict_proba(self, features):
+        self.called = True
+        return array([[0.25, 0.75]])
+
+
+def test_predict_proba_rejects_complex_direct_features_before_model_call():
+    model = _RecordingPredictProbaModel()
+    calibrated_model = CalibratedPairwiseAssociationModel(
+        model,
+        feature_names=("distance", "similarity"),
+    )
+    complex_features = array([[1.0 + 2.0j, 0.5]])
+
+    with pytest.raises(ValueError, match="real numeric"):
+        calibrated_model.predict_match_probability(complex_features)
+
+    assert not model.called