diff --git a/domain/epc/property_overrides/glazing_mix_guard.py b/domain/epc/property_overrides/glazing_mix_guard.py
index 36d8fce58..ea6aa7cdc 100644
--- a/domain/epc/property_overrides/glazing_mix_guard.py
+++ b/domain/epc/property_overrides/glazing_mix_guard.py
@@ -34,7 +34,17 @@ def glazing_mix_guard(description: str) -> Optional[GlazingType]:
     if len(base_types) < 2:
         # One (or no) glazing type named — a uniform assertion, not a mix.
         return None
-    if max(int(percent) for percent, _ in matches) >= _UNIFORM_THRESHOLD_PERCENT:
-        # One type dominates — treat as that (near-)uniform type, not a mix.
+    dominant_percent, dominant_type = max(
+        ((int(percent), base_type) for percent, base_type in matches),
+        key=lambda pair: pair[0],
+    )
+    if dominant_percent >= _UNIFORM_THRESHOLD_PERCENT:
+        # One type dominates — a near-uniform assertion, not a mix. Single glazing
+        # is era-free, so a dominant-single split is fully determined: apply SINGLE
+        # deterministically (the LLM otherwise flattens it onto the minority double
+        # type). A dominant double/triple still carries era ambiguity (pre-2002 /
+        # 2002 / unknown age), so it stays the LLM's job.
+        if dominant_type == "single":
+            return GlazingType.SINGLE
         return None
     return GlazingType.MIXED
diff --git a/scripts/lisasrequest/reclassify_dominant_single_glazing.py b/scripts/lisasrequest/reclassify_dominant_single_glazing.py
new file mode 100644
index 000000000..295ae88a1
--- /dev/null
+++ b/scripts/lisasrequest/reclassify_dominant_single_glazing.py
@@ -0,0 +1,123 @@
+"""Backfill glazing overrides where a >= 90%-single dwelling was flattened onto its
+*minority* double type.
+
+Sibling to ``scripts/reclassify_mixed_glazing.py``. That script rescued genuine
+mixes (no type >= 90%) to ``MIXED``; it deliberately left near-uniform rows to the
+LLM classifier. But for a dominant-*single* split ("4% Double glazing 2002 or
+later, 96% Single Glazing") the LLM had latched onto the 4% minority and written a
+"Double glazing…" value. ``glazing_mix_guard`` now resolves a dominant-single
+split deterministically to ``SINGLE`` (single glazing is era-free), so this
+backfill fixes the rows written before that guard change.
+
+Uses the SAME guard as the live path, so the backfill and the classifier cannot
+drift. SCOPED TO ONE PORTFOLIO (``--portfolio``, default 796 = Hyde) and only
+touches ``property_overrides.override_value`` (TEXT — what the modelling reads);
+the global ``landlord_glazing_overrides`` classifier cache is left alone.
+
+DRY-RUN BY DEFAULT: prints what it would change and writes nothing. Pass
+``--apply`` to execute inside a transaction; it also writes an audit CSV of every
+row changed (property_id, uprn, old value, new value) so the change is reversible.
+Idempotent — only rows whose stored value differs from the guard's target member
+are touched.
+
+    python -m scripts.lisasrequest.reclassify_dominant_single_glazing            # dry run
+    python -m scripts.lisasrequest.reclassify_dominant_single_glazing --apply    # write
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import sys
+from collections import Counter
+from pathlib import Path
+
+from sqlalchemy import text
+
+_REPO_ROOT = Path(__file__).resolve().parents[2]
+sys.path.insert(0, str(_REPO_ROOT))
+
+from domain.epc.property_overrides.glazing_mix_guard import glazing_mix_guard  # noqa: E402
+from scripts.e2e_common import ENV_PATH, build_engine, load_env  # noqa: E402
+
+_SELECT = text(
+    """
+    SELECT po.property_id, pr.uprn,
+           po.original_spreadsheet_description AS description,
+           po.override_value AS value
+    FROM property_overrides po
+    JOIN property pr ON pr.id = po.property_id
+    WHERE po.portfolio_id = :portfolio
+      AND po.override_component = 'glazing'
+    """
+)
+_UPDATE = text(
+    """
+    UPDATE property_overrides
+       SET override_value = :new_value
+     WHERE portfolio_id = :portfolio
+       AND override_component = 'glazing'
+       AND property_id = :property_id
+       AND override_value <> :new_value
+    """
+)
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--portfolio", type=int, default=796)
+    parser.add_argument(
+        "--apply",
+        action="store_true",
+        help="execute the updates (default: dry-run, writes nothing)",
+    )
+    args = parser.parse_args()
+
+    load_env(ENV_PATH)
+    engine = build_engine()
+
+    audit: list[tuple[int, object, str, str, str]] = []  # pid, uprn, descr, old, new
+    tally: Counter[str] = Counter()
+    with engine.begin() as conn:
+        conn.execute(text("SET statement_timeout = 120000"))
+        for property_id, uprn, description, value in conn.execute(
+            _SELECT, {"portfolio": args.portfolio}
+        ):
+            member = glazing_mix_guard(description or "")
+            if member is None or value == member.value:
+                continue
+            tally[f"{value!r} -> {member.value!r}"] += 1
+            audit.append((property_id, uprn, description, value, member.value))
+            if args.apply:
+                conn.execute(
+                    _UPDATE,
+                    {
+                        "portfolio": args.portfolio,
+                        "property_id": property_id,
+                        "new_value": member.value,
+                    },
+                )
+        if not args.apply:
+            conn.rollback()
+
+    verb = "re-classified" if args.apply else "would re-classify"
+    print(f"portfolio {args.portfolio}: {verb} {len(audit)} glazing override row(s)")
+    for change, n in tally.most_common():
+        print(f"  {n:5d}  {change}")
+
+    if args.apply and audit:
+        out = _REPO_ROOT / "scripts" / "lisasrequest" / (
+            f"reclassify_dominant_single_glazing_{args.portfolio}_audit.csv"
+        )
+        with out.open("w", newline="") as fh:
+            w = csv.writer(fh)
+            w.writerow(["property_id", "uprn", "description", "old_value", "new_value"])
+            w.writerows(audit)
+        print(f"\naudit trail: {out}")
+    if not args.apply:
+        print("\nDRY-RUN — nothing written. Re-run with --apply to execute.")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/tests/domain/epc/test_glazing_mix_guard.py b/tests/domain/epc/test_glazing_mix_guard.py
index 04e351cd3..2e839a5df 100644
--- a/tests/domain/epc/test_glazing_mix_guard.py
+++ b/tests/domain/epc/test_glazing_mix_guard.py
@@ -18,11 +18,31 @@ def test_a_genuine_percentage_mix_classifies_as_mixed() -> None:
     assert result is GlazingType.MIXED
 
 
+@pytest.mark.parametrize(
+    "description",
+    [
+        "4% double glazing 2002 or later, 96% single glazing",
+        "10% double glazing 2002 or later, 90% single glazing",  # exactly at threshold
+        "5% double glazing pre-2002, 95% single glazing",
+    ],
+)
+def test_dominant_single_glazing_resolves_to_single(description: str) -> None:
+    # The bug: the LLM flattened a >= 90%-single dwelling onto its *minority* double
+    # type. Single glazing is era-free, so a dominant-single split is fully
+    # determined — the guard must claim it (SINGLE) rather than defer to the LLM.
+
+    # Act
+    result = glazing_mix_guard(description)
+
+    # Assert
+    assert result is GlazingType.SINGLE
+
+
 @pytest.mark.parametrize(
     "description",
     [
         "100% double glazing 2002 or later",  # uniform — one type
-        "96% double glazing 2002 or later, 4% single glazing",  # >= 90% -> uniform
+        "96% double glazing 2002 or later, 4% single glazing",  # >= 90% double -> era ambiguity
         "80% double glazing 2002 or later, 20% double glazing pre-2002",  # same base type
         "some double glazing and a bit of single",  # unparseable — no percentages
         "",