diff --git a/src/remove_ai_watermarks/_internal/text_restoration.py b/src/remove_ai_watermarks/_internal/text_restoration.py
index b956add..58431d2 100644
--- a/src/remove_ai_watermarks/_internal/text_restoration.py
+++ b/src/remove_ai_watermarks/_internal/text_restoration.py
@@ -185,6 +185,20 @@ def restore_verified_text(
     return Image.fromarray(restored)
 
 
+def _glyph_crop_box(box: tuple[int, int, int, int], width: int, height: int) -> tuple[int, int, int, int]:
+    """Widen the detector box so descenders stay inside the silhouette crop.
+
+    Paddle line boxes sit 2-5 px above the true ink bottom on the poster
+    fixtures. The crop is the box itself, so those pixels never reached the
+    donor composite. Expand only on Y: 8% up, 25% down, clamped to the frame.
+    """
+    x1, y1, x2, y2 = box
+    line_h = max(1, y2 - y1)
+    pad_top = max(1, round(line_h * 0.08))
+    pad_bot = max(2, round(line_h * 0.25))
+    return max(0, x1), max(0, y1 - pad_top), min(width, x2), min(height, y2 + pad_bot)
+
+
 def source_silhouette_mask(
     source_rgb: NDArray[Any],
     box: tuple[int, int, int, int],
@@ -192,7 +206,7 @@ def source_silhouette_mask(
 ) -> NDArray[Any]:
     """Recover a thresholded glyph shape without retaining source amplitudes."""
     height, width = source_rgb.shape[:2]
-    x1, y1, x2, y2 = _clip_box(box, width, height)
+    x1, y1, x2, y2 = _glyph_crop_box(box, width, height)
     gray = cv2.cvtColor(source_rgb[y1:y2, x1:x2], cv2.COLOR_RGB2GRAY)
     support = np.ones(gray.shape, dtype=np.uint8)
     if angle:
diff --git a/src/remove_ai_watermarks/text_draft.py b/src/remove_ai_watermarks/text_draft.py
index 96256a6..742c009 100644
--- a/src/remove_ai_watermarks/text_draft.py
+++ b/src/remove_ai_watermarks/text_draft.py
@@ -223,15 +223,22 @@ def draft_text_lines(
     min_score: float = 0.85,
     detector: Any | None = None,
     engines: dict[str, Any] | None = None,
+    stable: bool = True,
 ) -> TextDraft:
     """Propose verified-text manifest lines for ``image``; never verified ones.
 
     Args:
         image: path of the source image (draft boxes are in ITS pixel space).
-        min_score: recognition confidence floor for every jittered read.
+        min_score: recognition confidence floor. With ``stable=True`` every
+            jittered read must clear it; with ``stable=False`` the single
+            probe for the chosen language must.
         detector/engines: injectable Paddle objects (tests use fakes); when
             None they are built from the ``text-draft`` extra, which must be
             installed (``draft_available()`` reports it).
+        stable: when True (default), accept a line only if three crop paddings
+            normalize identically. When False, take one recognition trio and
+            accept on score alone. Automatic restoration uses box/script only,
+            so the jitter gate is optional there.
 
     Returns:
         ``TextDraft`` with crop-stable ``accepted`` proposals and unstable
@@ -262,8 +269,14 @@ def draft_text_lines(
             probes[language] = _recognize(engine, source_rgb, box, script, 0.1)
         language = choose_language(probes)
         script = "cjk" if language == "ch" else "alphabetic"
-        reads = [_recognize(engines[language], source_rgb, box, script, ratio) for ratio in JITTER_RATIOS]
-        text = stable_recognition(reads, min_score)
+        if stable:
+            reads = [_recognize(engines[language], source_rgb, box, script, ratio) for ratio in JITTER_RATIOS]
+            text = stable_recognition(reads, min_score)
+        else:
+            text, score = probes[language]
+            reads = [(text, score)]
+            if not text.strip() or score < min_score:
+                text = None
         if text is None:
             rejected.append(
                 RejectedLine(
diff --git a/tests/test_text_draft.py b/tests/test_text_draft.py
index 6a1abc0..4a8f3aa 100644
--- a/tests/test_text_draft.py
+++ b/tests/test_text_draft.py
@@ -139,3 +139,20 @@ class TestDraftTextLines:
         (line,) = draft.accepted
         assert line.language == "ch"
         assert line.script == "cjk"
+
+    def test_unstable_geometry_draft_accepts_a_single_high_score_read(self, tmp_path: Path):
+        path = self._poster(tmp_path)
+        jitter = _JitterEngine(["Hello", "Hallo", "Hullo"])
+        stable = _FakeEngine("Hello")
+        # Probes disagree across engines so language is en; jitter would fail
+        # the stable gate. Geometry mode uses the one en probe and accepts.
+        draft = draft_text_lines(
+            path,
+            min_score=0.75,
+            stable=False,
+            detector=_FakeDetector([(20, 20, 200, 60)]),
+            engines={"en": jitter, "ru": stable, "ch": stable},
+        )
+        (line,) = draft.accepted
+        assert line.text == "Hello"
+        assert line.min_score >= 0.75
diff --git a/tests/test_text_restoration.py b/tests/test_text_restoration.py
index 9ea4887..bdc0e2e 100644
--- a/tests/test_text_restoration.py
+++ b/tests/test_text_restoration.py
@@ -15,6 +15,7 @@ from remove_ai_watermarks._internal.text_restoration import (
     load_verified_text_manifest,
     restore_verified_text,
     source_pixel_sha256,
+    source_silhouette_mask,
 )
 
 
@@ -126,3 +127,14 @@ def test_restoration_uses_lama_and_qwen_vae_core(monkeypatch) -> None:
     assert calls
     assert np.all(restored[16, 20] == donor[16, 20])
     assert np.all(restored[0, 0] == candidate[0, 0])
+
+
+def test_silhouette_includes_descender_below_the_detector_box() -> None:
+    source = np.full((40, 50, 3), 240, dtype=np.uint8)
+    source[10:22, 18:24] = 20
+    source[22:27, 18:22] = 20  # tail of a y / Cyrillic u, under the box
+
+    mask = source_silhouette_mask(source, (10, 10, 40, 22))
+
+    assert mask[24, 20] == 255
+    assert mask[16, 20] == 255