diff --git a/SKILL.md b/SKILL.md
index 852756a..d800440 100644
--- a/SKILL.md
+++ b/SKILL.md
@@ -190,6 +190,13 @@ is the bridge: it writes the repo's input format, shells out to
    The one thing that *does* skip work: deleting a chapter from `extracted/`
    before the run. It is then never sent, and `rebuild` keeps the original
    English for it — the right treatment for an index.
+**Numbered exercise items come back untranslated.** Measured on Stroustrup
+2026-08-21: 182 paragraphs of 10151 (1.8%) shaped `[2] Expanding on what you
+have learned…` were echoed back in English. The document-wide Cyrillic check
+cannot see this — 1.8% drowns in it — so `rebuild` counts them separately as
+"осталось на языке оригинала". If the count is high, add an explicit line to the
+prompt that numbered items are prose and must be translated too.
+
 4. **Read the reported counts.** "без перевода" above zero means a chapter came
    back with a different paragraph count and kept its original text; "разметка
    потеряна" counts paragraphs where the model mangled the inline-tag markers
diff --git a/scripts/test_translate.py b/scripts/test_translate.py
index 7e3d6d5..4698e35 100644
--- a/scripts/test_translate.py
+++ b/scripts/test_translate.py
@@ -132,6 +132,32 @@ def test_junk_translation_falls_back_to_original(tmp: Path):
     assert "Второй абзац" in result, "нормальный перевод должен остаться"
 
 
+def test_untouched_paragraphs_are_counted(tmp: Path):
+    """Абзац, вернувшийся по-английски, доля кириллицы по документу не ловит:
+    у Страуструпа так осталось 182 упражнения из 10151 блока."""
+    src = tmp / "book.html"
+    en = ("Expanding on what you have learned, write a program that lists the "
+          "instructions for a computer to find the upstairs bedroom. ") * 2
+    ru_src = ("Этот абзац достаточно длинный, чтобы попасть под проверку языка "
+              "и быть переведённым как положено. ") * 2
+    src.write_text("\n".join(["<h1>Chapter</h1>", "<p>%s</p>" % en, "<p>%s</p>" % en]),
+                   encoding="utf-8")
+    chapters = t.split_chapters(t.parse_blocks(src))
+    trans = tmp / "translations"
+    trans.mkdir()
+    (trans / "chapter_000_translated.json").write_text(json.dumps(
+        {"number": 0, "paragraphs": ["Глава", en, ru_src]}, ensure_ascii=False),
+        encoding="utf-8")
+
+    out = tmp / "out.html"
+    import io, contextlib
+    buf = io.StringIO()
+    with contextlib.redirect_stdout(buf):
+        t.rebuild(src, chapters, tmp, out)
+    report = buf.getvalue()
+    assert "осталось на языке оригинала: 1" in report, report
+
+
 def test_workdir_belongs_to_one_book(tmp: Path):
     """Имена chapter_NNN.json у всех книг одинаковы: чужой workdir склеит
     перевод одной книги с текстом другой."""
@@ -183,6 +209,7 @@ if __name__ == "__main__":
     test_language_detection()
     for case in (test_split_and_rebuild, test_code_listings_are_left_alone,
                  test_junk_translation_falls_back_to_original,
+                 test_untouched_paragraphs_are_counted,
                  test_workdir_belongs_to_one_book, test_stale_chapters_removed,
                  test_verify_output_catches_untranslated):
         with tempfile.TemporaryDirectory() as d:
diff --git a/scripts/translate.py b/scripts/translate.py
index 4f56878..f4e5187 100644
--- a/scripts/translate.py
+++ b/scripts/translate.py
@@ -156,7 +156,7 @@ MIN_LEN_SHARE = 0.25
 
 def rebuild(src, chapters, workdir, out):
     lines = src.read_text(encoding="utf-8").splitlines()
-    broken = missing = translated = junk = 0
+    broken = missing = translated = junk = untouched = 0
     for n, ch in enumerate(chapters):
         f = workdir / "translations" / ("chapter_%03d_translated.json" % n)
         if not f.exists():
@@ -174,13 +174,20 @@ def rebuild(src, chapters, workdir, out):
                 junk += 1
                 lines[idx] = "<%s%s>%s</%s>" % (tag, attrs, original, tag)
                 continue  # оставляем оригинал: английский абзац лучше огрызка
+            # Абзац, вернувшийся на языке оригинала. Доля кириллицы по всему
+            # документу такое не ловит: полтора процента в ней тонут, а на
+            # странице это заметный кусок английского текста.
+            if len(plain_src) > 150 and detect_language(re.sub("<[^>]+>", "", text))[0] \
+                    == detect_language(plain_src)[0] != "unknown":
+                untouched += 1
             body, ok = from_marks(text)
             broken += not ok
             translated += 1
             lines[idx] = "<%s%s>%s</%s>" % (tag, attrs, body, tag)
     out.write_text("\n".join(lines) + "\n", encoding="utf-8")
     print("переведено блоков: %d, без перевода: %d, разметка потеряна в %d, "
-          "огрызков заменено оригиналом: %d" % (translated, missing, broken, junk))
+          "огрызков заменено оригиналом: %d, осталось на языке оригинала: %d"
+          % (translated, missing, broken, junk, untouched))
     return missing == 0 and junk * 200 <= translated