diff --git a/tests/unit/python/utils/i18n/test_damage.py b/tests/unit/python/utils/i18n/test_damage.py
index 3668b2c13c..1b597da383 100644
--- a/tests/unit/python/utils/i18n/test_damage.py
+++ b/tests/unit/python/utils/i18n/test_damage.py
@@ -67,6 +67,34 @@ class TestWrongScript(unittest.TestCase):
     def test_a_translation_without_a_word_is_kept(self) -> None:
         self.assertEqual(reason("Status", "100 %", "be"), "")
 
+    def test_a_separatorless_script_counts_one_word_per_character(self) -> None:
+        """Japanese and Chinese write without spaces, so a run-based pattern
+        returns the whole sentence as one token, and two Latin product names
+        inside it carry that token past a half and the ratio to 1.0."""
+        self.assertEqual(
+            reason(
+                "Runs Playwright in Docker with stable browser settings",
+                "安定したブラウザ設定でDockerでPlaywrightを実行",
+                "ja",
+            ),
+            "",
+        )
+        self.assertEqual(
+            reason("Run Playwright in Docker", "在Docker中运行Playwright", "zh"), ""
+        )
+
+    def test_a_separatorless_script_answered_in_english_is_still_condemned(
+        self,
+    ) -> None:
+        self.assertEqual(
+            reason(
+                "Runs Playwright in Docker with stable settings",
+                "Runs Playwright in Docker with stable browser settings",
+                "ja",
+            ),
+            "untranslated",
+        )
+
 
 class TestForeignScript(unittest.TestCase):
     """A Latin-script language came back in a script it does not use.
diff --git a/utils/i18n/damage.py b/utils/i18n/damage.py
index 8a1e5de0c8..ffe195f31f 100644
--- a/utils/i18n/damage.py
+++ b/utils/i18n/damage.py
@@ -3,6 +3,14 @@
 :func:`reason` is the single predicate both ends of the pipeline ask, and
 :func:`harms` is its boolean face; the rest of this module are the criteria it
 folds together.
+
+``WORDS`` takes one character as one word in Han, Kana, Thai, Lao, Khmer and
+Myanmar, which write without separators. Its single-character branch has to
+come first, or the run branch swallows a whole sentence together with the
+Latin product names inside it and returns one token whose Latin share decides
+the verdict alone. ``UNSPACED_RANGES`` holds codepoints rather than escapes
+because a formatter rewrites an escape into the character it names, and the
+literal blob that leaves behind cannot be read or reviewed.
 """
 
 from __future__ import annotations
@@ -28,7 +36,18 @@ COLLAPSE_RATIO = 0.15
 MAX_SHARING_SOURCES = 2
 ECHO_FLOOR = 30
 STUTTER = re.compile(r"\b(\w+)(?:\s+\1\b){2,}", re.IGNORECASE | re.UNICODE)
-WORDS = re.compile(r"[^\W\d_]{2,}", re.UNICODE)
+UNSPACED_RANGES = (
+    (0x3040, 0x30FF),
+    (0x3400, 0x4DBF),
+    (0x4E00, 0x9FFF),
+    (0xF900, 0xFAFF),
+    (0x0E00, 0x0E7F),
+    (0x0E80, 0x0EFF),
+    (0x1780, 0x17FF),
+    (0x1000, 0x109F),
+)
+UNSPACED = "".join(f"{chr(low)}-{chr(high)}" for low, high in UNSPACED_RANGES)
+WORDS = re.compile(rf"[{UNSPACED}]|[^\W\d_]{{2,}}", re.UNICODE)
 LATIN_SHARE = 0.9
 NON_LATIN_SCRIPTS = non_latin_languages(PROJECT_ROOT)
 
@@ -94,6 +113,12 @@ def untranslated(translation: str, language: str) -> bool:
     four-word labels, which is where the wrong language hid for as long as a
     six-word floor stood here.
 
+    The ratio is only as good as the word count, which is why ``WORDS`` splits
+    a separator-less script per character: a Japanese sentence carrying two
+    Latin product names is one token to a run-based pattern, and one token
+    whose Latin share passes a half makes the ratio 1.0 on a sound
+    translation.
+
     Args:
         translation: what came back for the entry.
         language: ISO 639-1 code of the catalog being filled.