diff --git a/news/score.py b/news/score.py
index 259f139..098f240 100644
--- a/news/score.py
+++ b/news/score.py
@@ -14,6 +14,7 @@ so a plain max hands one block 78% of all items. See profile/CALIBRATION.md.
 from __future__ import annotations
 
 import math
+import re
 import sqlite3
 from datetime import datetime
 from dataclasses import dataclass, field
@@ -109,6 +110,20 @@ class ProfileVectors:
         return {}
 
 
+def _gate_terms(block: dict) -> list[str]:
+    return [t.lower() for t in block.get("requires_any", [])]
+
+
+def _mentions(text: str, terms: list[str]) -> bool:
+    """Word-boundary match, never substring.
+
+    The old system's blocklist matched substrings and so vetoed every Iran-war
+    story because "nfl" appears inside "conflict". Same trap here: "rome" is
+    inside "chrome", "mali" inside "normalise".
+    """
+    return any(re.search(rf"\b{re.escape(t)}\b", text) for t in terms)
+
+
 def _z(values: np.ndarray) -> np.ndarray:
     """Z-score, safe when a block is flat across the population."""
     std = float(values.std())
@@ -197,7 +212,47 @@ def score_clusters(
             weights[g] = max(weights.get(g, 0.0), profile_vectors.weight(kind, block_id))
         return {g: norm(v) * weights[g] for g, v in by_group.items()}
 
+    # Place gate. Embeddings model meaning well and named entities poorly:
+    # "Italian politics" and "Australian politics" sit almost on top of each
+    # other, so 66% of the winners of a place-based block never named the place
+    # -- Glastonbury and Anthony Albanese were scoring as italy-mexico.
+    #
+    # Place is identity, not meaning, so it is gated lexically. A block with
+    # `requires_any` cannot win for an item whose text names none of its terms.
+    gated: dict[str, set[int]] = {}
+    texts: list[str] = []
+    ids = [r["id"] for r in rows]
+    member_titles: dict[str, list[str]] = {}
+    if ids:
+        marks = ",".join("?" * len(ids))
+        for m in con.execute(
+            f"""SELECT cm.cluster_id AS cid, i.title FROM cluster_members cm
+                JOIN items i ON i.id = cm.item_id
+                WHERE cm.cluster_id IN ({marks})""", ids):
+            member_titles.setdefault(m["cid"], []).append(m["title"])
+    for r in rows:
+        texts.append(" ".join([r["lead"] or ""] + member_titles.get(r["id"], [])).lower())
+    for block_id, _, block in profile_vectors.blocks["interest"]:
+        terms = _gate_terms(block)
+        if terms:
+            gated[block_id] = {i for i, t in enumerate(texts) if not _mentions(t, terms)}
+
     interest_z = normalise("interest", interest_raw)
+
+    # The gate is applied AFTER normalisation, on purpose. Setting the raw score
+    # to a sentinel first does not work: when most items are gated, that
+    # sentinel becomes the block's own MEAN, so a gated item z-scores to roughly
+    # zero and can still win. Normalisation erases a pre-normalisation gate.
+    for block_id, blocked in gated.items():
+        group = profile_vectors.group("interest", block_id)
+        # Only gate a group whose members ALL require a mention; otherwise an
+        # ungated sibling in the same group would be silenced along with it.
+        siblings = [b for b, _, _ in profile_vectors.blocks["interest"]
+                    if profile_vectors.group("interest", b) == group]
+        if not all(_gate_terms(profile_vectors.block("interest", b)) for b in siblings):
+            continue
+        for i in blocked:
+            interest_z[group][i] = -np.inf
     importance_z = normalise("importance", importance_raw)
     # Veto stays on z-scores: it is a comparison against the interest signal,
     # and both sides must be on the same scale.
diff --git a/profile/profile.toml b/profile/profile.toml
index 2d01b02..d713c93 100644
--- a/profile/profile.toml
+++ b/profile/profile.toml
@@ -204,6 +204,13 @@ probes = [
   "Federal agencies are restructured by executive order, prompting legal challenges.",
 ]
 
+# `requires_any` gates a block on a literal, word-boundary mention. Embeddings
+# model meaning well and named entities poorly -- "Italian politics" and
+# "Australian politics" are nearly the same vector -- so 66% of the winners of a
+# place block named no place at all: Glastonbury and Anthony Albanese were
+# scoring as italy-mexico, and a singing-apraxia story as ohio-state-politics.
+# Place is identity, not meaning, so it is matched lexically.
+#
 # --- place -------------------------------------------------------------------
 [[interest]]
 id = "urbanism"
@@ -217,6 +224,7 @@ probes = [
 
 [[interest]]
 id = "cincinnati"
+requires_any = ["cincinnati", "hamilton county", "northern kentucky", "covington", "newport", "over-the-rhine", "ohio"]
 group = "local"
 weight = 0.9
 local = true
@@ -230,6 +238,7 @@ probes = [
 
 [[interest]]
 id = "kentucky-louisville"
+requires_any = ["kentucky", "louisville", "frankfort", "lexington"]
 group = "local"
 weight = 0.6
 local = true
@@ -241,6 +250,7 @@ probes = [
 # coverage alongside Cincinnati city government.
 [[interest]]
 id = "ohio-state-politics"
+requires_any = ["ohio", "columbus", "cleveland", "toledo", "dayton", "akron", "statehouse"]
 group = "local"
 weight = 0.75
 local = true
@@ -257,6 +267,7 @@ probes = [
 # the consumer-lifestyle veto must not swallow it.
 [[interest]]
 id = "italy-mexico"
+requires_any = ["italy", "italian", "rome", "milan", "naples", "sicily", "meloni", "mexico", "mexican", "mexico city", "sheinbaum", "guadalajara", "monterrey"]
 weight = 0.75
 probes = [
   "Italian politics, government coalitions and social policy in Italy.",
diff --git a/tests/test_place_gate.py b/tests/test_place_gate.py
new file mode 100644
index 0000000..1a8d485
--- /dev/null
+++ b/tests/test_place_gate.py
@@ -0,0 +1,105 @@
+"""Place gating.
+
+Embeddings model meaning well and named entities poorly: "Italian politics" and
+"Australian politics" sit almost on top of each other. Measured on live data,
+66% of the winners of a place-based block named no place at all -- Glastonbury
+and Anthony Albanese were scoring as `italy-mexico`.
+
+Place is identity, not meaning, so it is matched lexically.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+import pytest
+
+from news import store
+from news.score import ProfileVectors, _mentions, score_clusters
+
+
+def test_word_boundaries_not_substrings():
+    """The lesson from "nfl" matching co-NFL-ict, which vetoed every Iran story."""
+    assert _mentions("meloni visits rome today", ["rome"])
+    assert not _mentions("open the chrome browser", ["rome"])
+    assert not _mentions("efforts to normalise ties", ["mali"])
+    assert _mentions("fighting spreads in mali", ["mali"])
+
+
+class _FakeVectors:
+    """Two blocks in different groups: one gated on a place, one ungated."""
+
+    def __init__(self):
+        near = np.array([[1.0, 0.0]], dtype=np.float32)
+        self.blocks = {
+            "interest": [
+                ("italy-mexico", near, {"id": "italy-mexico", "weight": 1.0,
+                                        "requires_any": ["italy", "mexico"]}),
+                ("urbanism", np.array([[0.0, 1.0]], dtype=np.float32),
+                 {"id": "urbanism", "weight": 1.0}),
+            ],
+            "veto": [], "demote": [], "importance": [
+                ("x", np.array([[0.0, 0.0]], dtype=np.float32), {"id": "x", "weight": 1.0})],
+        }
+
+    def raw_matrix(self, kind, vecs):
+        return {bid: (vecs @ m.T).max(axis=1) for bid, m, _ in self.blocks[kind]}
+
+    def weight(self, kind, bid):
+        return 1.0
+
+    def group(self, kind, bid):
+        return bid
+
+    def block(self, kind, bid):
+        return next(b for i, _, b in self.blocks[kind] if i == bid)
+
+
+@pytest.fixture
+def con(tmp_path):
+    c = store.connect(tmp_path / "g.db")
+    store.upsert_source(c, id="s1", name="A", url="http://a", kind="rss", bucket="a",
+                        weight=1.0, prominence=0, enabled=1, poll_interval_secs=900)
+    return c
+
+
+def _add(con, cid, title):
+    when = "2026-01-01T00:00:00+00:00"
+    con.execute("""INSERT INTO items (id,source_id,title,summary,url,canonical_url,
+                   published_at,first_seen_at,content_hash) VALUES (?,?,?,'',?,?,?,?,?)""",
+                (cid, "s1", title, f"http://x/{cid}", f"http://x/{cid}", when, when, cid))
+    con.execute("""INSERT INTO clusters (id,centroid,size,first_seen_at,last_seen_at,lead_item_id)
+                   VALUES (?,?,1,?,?,?)""",
+                (cid, store.to_blob(np.array([1.0, 0.0], dtype=np.float32)), when, when, cid))
+    con.execute("INSERT INTO cluster_members VALUES (?,?,?)", (cid, cid, when))
+
+
+CFG = {"veto_absolute": 0.9, "veto_margin_over_interest": 9.0,
+       "corroboration_pickup_weight": 0.3, "corroboration_depth_weight": 0.2}
+CLUST = {"syndication_threshold": 0.97}
+
+
+def test_a_gated_block_cannot_win_without_a_mention(con):
+    """Both items point at the italy-mexico probe; only one names a place."""
+    _add(con, "yes", "Meloni's government becomes Italy's longest-lasting")
+    _add(con, "no", "Anthony Albanese held an undeclared honorary membership")
+    scores = {s.cluster_id: s for s in score_clusters(con, _FakeVectors(), CFG, CLUST)}
+    assert scores["yes"].interest.block == "italy-mexico"
+    assert scores["no"].interest.block != "italy-mexico"
+
+
+def test_gate_survives_normalisation(con):
+    """Regression for a real bug.
+
+    The gate was first applied by writing a sentinel into the RAW score before
+    normalisation. When most items are gated that sentinel becomes the block's
+    own mean, so a gated item z-scored to roughly zero and could still win --
+    the normalisation erased the gate. With 30 gated items against 1 mention,
+    a pre-normalisation gate lets the gated ones through.
+    """
+    _add(con, "mention", "Sheinbaum speaks in Mexico")
+    for i in range(30):
+        _add(con, f"other{i}", f"Unrelated story number {i}")
+    scores = score_clusters(con, _FakeVectors(), CFG, CLUST)
+    winners = [s for s in scores if s.interest.block == "italy-mexico"]
+    assert [s.cluster_id for s in winners] == ["mention"], (
+        "only the item naming a place may win the gated block")