diff --git a/pyproject.toml b/pyproject.toml
index 02d4f00..e911c0d 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -18,7 +18,7 @@ dependencies = [
 ]
 
 [project.optional-dependencies]
-dev = ["pytest>=8.0", "ruff>=0.5", "mypy>=1.10"]
+dev = ["pytest>=8.0", "ruff>=0.5,<0.16", "mypy>=1.10"]
 
 [build-system]
 requires = ["hatchling"]
diff --git a/src/trinity/adapters/drop.py b/src/trinity/adapters/drop.py
index 1314803..07fcbd4 100644
--- a/src/trinity/adapters/drop.py
+++ b/src/trinity/adapters/drop.py
@@ -130,9 +130,12 @@ def _maybe_unbox(segment: str) -> str:
     return boxed if boxed is not None else ""
 
 
-#: Surrounding punctuation stripped from a token, EXCLUDING the signs ``+``/``-`` — a
-#: leading sign is part of a number's value, not wrapping noise.
-_STRIP_EDGE = "".join(c for c in string.punctuation if c not in "+-")
+#: Surrounding punctuation stripped from a token, EXCLUDING the signs ``+``/``-`` and
+#: the decimal point ``.`` — a leading sign or decimal point is part of a number's
+#: value, not wrapping noise. A genuinely-trailing ``.`` (sentence period) is handled
+#: by the dedicated rstrip retry in :func:`_normalize_token`, which can tell it apart
+#: from a value-bearing leading point; a blanket edge-strip cannot.
+_STRIP_EDGE = "".join(c for c in string.punctuation if c not in "+-.")
 
 
 def _normalize_token(raw: str) -> str:
@@ -147,20 +150,27 @@ def _normalize_token(raw: str) -> str:
     dropped the ``-`` and left commas to break ``float()``) did not deliver.
 
     A token that is ALREADY a number is recognised before any punctuation is
-    stripped: the edge-strip set includes ``.``, so a leading-decimal token like
-    ``".5"`` would otherwise lose its point and normalize to ``"5.0"`` — equal to
-    a gold ``"5"`` (false positive) and unequal to the value-identical gold
-    ``"0.5"`` (false negative). The official DROP ``_remove_punc`` tests
-    ``_is_number`` first and leaves numbers untouched for exactly this reason."""
+    stripped: a leading-decimal token like ``".5"`` must not lose its point and
+    normalize to ``"5.0"`` — equal to a gold ``"5"`` (false positive) and unequal
+    to the value-identical gold ``"0.5"`` (false negative). The official DROP
+    ``_remove_punc`` tests ``_is_number`` first and leaves numbers untouched for
+    exactly this reason (issue #423). The float-first path alone only covers the
+    *bare* token: with ``.`` in the edge-strip set, wrapped forms like ``"$.5"``
+    and ``".5."`` still lost the leading point on the second-chance path. So the
+    edge strip excludes ``.`` entirely, and a genuinely-trailing period (``".5."``,
+    ``"16.."``) is retried with an explicit ``rstrip(".")`` — right-side dots are
+    sentence punctuation, left-side dots are value."""
     try:
         return str(float(raw.replace(",", "")))
     except ValueError:
         pass
     core = raw.strip(_STRIP_EDGE)
-    try:
-        return str(float(core.replace(",", "")))
-    except ValueError:
-        return _PUNCT.sub("", raw)
+    for cand in (core, core.rstrip(".")):
+        try:
+            return str(float(cand.replace(",", "")))
+        except ValueError:
+            continue
+    return _PUNCT.sub("", raw)
 
 
 def _split_internal_hyphens(token: str) -> list[str]:
diff --git a/src/trinity/orchestration/reward.py b/src/trinity/orchestration/reward.py
index 7192ef8..321d954 100644
--- a/src/trinity/orchestration/reward.py
+++ b/src/trinity/orchestration/reward.py
@@ -502,6 +502,15 @@ def extract_last_number(text: str) -> str | None:
     # comma so the thousands-separator branch below reads it as one number instead
     # of splitting it into "1" and "000".
     text = text.replace("{,}", ",")
+    # Thin-space thousands (``1\\,000`` / ``1\\:000`` / ``1\\;000``): collapse
+    # before the literal regex so the last match is the full number, not the
+    # trailing ``000`` group alone (issue #478). Boxed answers already work
+    # because ``normalize_math_answer`` strips ``\\,`` globally.
+    while True:
+        nxt = re.sub(r"(\d)\\[,;:](\d)", r"\1\2", text)
+        if nxt == text:
+            break
+        text = nxt
     candidates: list[tuple[int, str]] = []
     for _start, end, term in _iter_latex_frac_sqrt_spans(text):
         candidates.append((end, term))
diff --git a/tests/test_reward_thinspace_thousands.py b/tests/test_reward_thinspace_thousands.py
new file mode 100644
index 0000000..f06c16f
--- /dev/null
+++ b/tests/test_reward_thinspace_thousands.py
@@ -0,0 +1,13 @@
+"""LaTeX thin-space thousands must extract as one number (#478)."""
+from __future__ import annotations
+from trinity.orchestration import reward as R
+
+def test_extract_thinspace_thousands():
+    assert R.extract_last_number(r"1\,000") == "1000"
+    assert R.extract_last_number(r"1\,234\,567") == "1234567"
+    assert R.extract_last_number(r"1\:000") == "1000"
+    assert R.extract_last_number(r"1\;000") == "1000"
+
+def test_score_thinspace_thousands():
+    assert R.score_text("math500", r"1\,000", "1000") == 1.0
+    assert R.score_text("math500", r"The answer is 1\,000.", "1000") == 1.0