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..4caf94c 100644
--- a/src/trinity/adapters/drop.py
+++ b/src/trinity/adapters/drop.py
@@ -156,13 +156,34 @@ def _normalize_token(raw: str) -> str:
         return str(float(raw.replace(",", "")))
     except ValueError:
         pass
-    core = raw.strip(_STRIP_EDGE)
+    core = _strip_edges_keeping_leading_decimal(raw)
     try:
         return str(float(core.replace(",", "")))
     except ValueError:
         return _PUNCT.sub("", raw)
 
 
+def _strip_edges_keeping_leading_decimal(raw: str) -> str:
+    """Strip edge punctuation like raw.strip(_STRIP_EDGE), except that a .
+    starting a leading-decimal number (. directly followed by a digit) is kept.
+
+    A plain strip treats the . of ".5." or "$.5" as wrapping noise and
+    yields "5", so the token normalizes to "5.0" — equal to a gold "5"
+    (false positive) and unequal to the value-identical "0.5" (false negative).
+    The official DROP normalizer never corrupts a number this way, so the left-edge
+    scan stops as soon as stripping one more character would break a leading decimal.
+    """
+    end = len(raw)
+    while end > 0 and raw[end - 1] in _STRIP_EDGE:
+        end -= 1
+    start = 0
+    while start < end and raw[start] in _STRIP_EDGE:
+        if raw[start] == "." and start + 1 < end and raw[start + 1].isdigit():
+            break
+        start += 1
+    return raw[start:end]
+
+
 def _split_internal_hyphens(token: str) -> list[str]:
     """Split ``token`` on INTERNAL hyphens, matching DROP's ``re.split(" |-", ...)``.
 
diff --git a/src/trinity/standings.py b/src/trinity/standings.py
index f535a7d..9f6dfff 100644
--- a/src/trinity/standings.py
+++ b/src/trinity/standings.py
@@ -117,12 +117,16 @@ def compute_standings(leaderboard: Mapping[str, Any]) -> Standings:
         per_benchmark = h.get("per_benchmark")
         if miner is None or not isinstance(per_benchmark, dict):
             continue
-        best = per_miner.setdefault(str(miner), {})
+        miner_key = str(miner)
         for bench, score in per_benchmark.items():
             if not _is_num(score):
                 continue
             b = str(bench)
             seen_benches.add(b)
+            # Register the miner only once a real numeric score is confirmed. A merged win
+            # whose per_benchmark is empty or all-non-numeric contributes nothing and must
+            # not create a phantom standing (matching the non-dict per_benchmark case).
+            best = per_miner.setdefault(miner_key, {})
             if b not in best or float(score) > best[b]:
                 best[b] = float(score)
 
diff --git a/tests/test_standings.py b/tests/test_standings.py
index 7bca49f..22cf582 100644
--- a/tests/test_standings.py
+++ b/tests/test_standings.py
@@ -151,6 +151,32 @@ def test_non_dict_per_benchmark_is_skipped():
     assert [m.miner for m in s.miners] == ["r"]
 
 
+def test_win_with_no_numeric_score_creates_no_phantom_leader():
+    # Regression: a merged win whose per_benchmark is empty (or entirely non-numeric)
+    # must NOT register the miner. Pre-fix, setdefault ran before any numeric score was
+    # confirmed, so such a win produced a phantom MinerStanding(overall=0.0, n_competed=0,
+    # rank=1) that was falsely reported as the standings leader. This must match the
+    # non-dict per_benchmark case, which drops the record entirely.
+    lb = _lb([
+        {"miner": "ghost", "merged": True, "per_benchmark": {}, "pr": 1,
+         "score": 0.0, "generation": 1, "timestamp": "2026-07-13T00:00:00Z"},
+        {"miner": "nul", "merged": True, "per_benchmark": {"math500": None, "mmlu": "x"},
+         "pr": 2, "score": 0.0, "generation": 1, "timestamp": "2026-07-13T00:00:00Z"},
+    ])
+    s = compute_standings(lb)
+    assert s.miners == []                                   # neither win registers a miner
+    assert s.leader is None                                 # no phantom leader
+
+    # And a phantom win must not shadow a real one: only the scored miner ranks/leads.
+    lb2 = _lb([
+        {"miner": "ghost", "merged": True, "per_benchmark": {}, "pr": 1,
+         "score": 0.0, "generation": 1, "timestamp": "2026-07-13T00:00:00Z"},
+        _win("real", {"math500": 0.7, "mmlu": 0.8}, 2),
+    ])
+    s2 = compute_standings(lb2)
+    assert [m.miner for m in s2.miners] == ["real"] and s2.leader == "real"
+
+
 # --------------------------------------------------------------------------- #
 # load + render
 # --------------------------------------------------------------------------- #