diff --git a/README.md b/README.md
index 2a68581..3e522db 100644
--- a/README.md
+++ b/README.md
@@ -106,7 +106,7 @@ slugify(
 - `backend`: select a backend from the table above. Ignored for transliteration in Unicode mode, but still validated.
 
 `smart_truncate(string, max_length=0, word_boundary=False, separator=' ', save_order=False)`
-is also public and retains legacy behavior: `str.strip(separator)` strips a character set, empty separators raise `ValueError`, zero is unlimited, and negative limits retain slicing semantics. Modern slugify uses a separate private token-budget helper; it does not change this public function.
+is also public and retains legacy behavior: `str.strip(separator)` strips a character set, an empty separator raises `ValueError` when word-boundary truncation reaches the split step (the hard-cut path does not raise), zero is unlimited, and negative limits retain slicing semantics. Modern slugify uses a separate private token-budget helper; it does not change this public function.
 
 ```python
 from slugify import slugify
diff --git a/slugify/_legacy.py b/slugify/_legacy.py
index c0dfd67..158da86 100644
--- a/slugify/_legacy.py
+++ b/slugify/_legacy.py
@@ -75,7 +75,8 @@ def smart_truncate(
     """Historical public truncation behavior, including character-set stripping.
 
     Zero means unlimited; negative limits retain Python slicing semantics.
-    An empty separator raises ValueError, as in the legacy implementation.
+    An empty separator raises ValueError when word-boundary truncation splits
+    the string; the hard-cut path does not raise for an empty separator.
     """
     string = string.strip(separator)
     if not max_length:
diff --git a/tests/test_release.py b/tests/test_release.py
index 0daccd6..4a89b14 100644
--- a/tests/test_release.py
+++ b/tests/test_release.py
@@ -19,6 +19,12 @@
 
 
 class ReleaseRegressionTests(unittest.TestCase):
+    def test_smart_truncate_empty_separator_only_raises_on_word_boundary_path(self):
+        with self.assertRaisesRegex(ValueError, 'empty separator'):
+            smart_truncate('a b c', 3, word_boundary=True, separator='')
+        self.assertEqual(
+            smart_truncate('a b c', 3, word_boundary=False, separator=''), 'a b')
+
     def test_legacy_goldens_and_differential(self):
         from tools import legacy_reference
         from tools.check_algorithms import check_differential, check_goldens