diff --git a/packages/populace-build/src/populace/build/us_runtime/voluntary_filing.py b/packages/populace-build/src/populace/build/us_runtime/voluntary_filing.py
index 2e5854c3..ed5c4d39 100644
--- a/packages/populace-build/src/populace/build/us_runtime/voluntary_filing.py
+++ b/packages/populace-build/src/populace/build/us_runtime/voluntary_filing.py
@@ -745,15 +745,30 @@ def _source_receiver_rows(
             .groupby(rows["_source_id"])
             .sum()
         )
-        if not asec_counts.eq(1).all():
-            bad = asec_counts.index[~asec_counts.eq(1)].tolist()
+        if asec_counts.gt(1).any():
+            bad = asec_counts.index[asec_counts.gt(1)].tolist()
             raise ValueError(
-                "US voluntary-filing support source units require exactly one "
-                f"ASEC row; invalid source unit(s) {bad[:5]}."
+                "US voluntary-filing support source units carry duplicated "
+                f"ASEC rows; invalid source unit(s) {bad[:5]}."
             )
-        source_rows = rows.loc[
-            rows["_support_channel"].eq(_BASE_ASEC_SUPPORT_CHANNEL)
-        ].copy()
+        # Prefer each unit's ASEC row, but a frozen-support selection may
+        # legitimately keep only a unit's PUF clone (the L0-survivor case the
+        # SSI reporter lineage also handles — Build M's certified 57,240
+        # selection does exactly this). Clones carry the unit's source
+        # predictors, so the surviving row predicts identically; pick it
+        # deterministically by channel then tax-unit id.
+        ordered_rows = rows.copy()
+        ordered_rows["_asec_rank"] = (
+            ~ordered_rows["_support_channel"].eq(_BASE_ASEC_SUPPORT_CHANNEL)
+        ).astype(int)
+        source_rows = (
+            ordered_rows.sort_values(
+                ["_source_id", "_asec_rank", "_support_channel", "_tax_unit_id"],
+                kind="stable",
+            )
+            .drop_duplicates("_source_id", keep="first")
+            .drop(columns="_asec_rank")
+        )
     else:
         if rows["_source_id"].duplicated().any():
             duplicates = rows.loc[
diff --git a/packages/populace-build/tests/test_us_voluntary_filing.py b/packages/populace-build/tests/test_us_voluntary_filing.py
index 1311bb72..a5ec9041 100644
--- a/packages/populace-build/tests/test_us_voluntary_filing.py
+++ b/packages/populace-build/tests/test_us_voluntary_filing.py
@@ -593,6 +593,53 @@ def fit(
     assert (by_source.nunique() == 1).all()
 
 
+def test_puf_only_survivor_units_predict_from_the_surviving_clone(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """A unit whose ASEC row was dropped by selection still predicts once.
+
+    Build M's sparse run died here: the certified frozen-support selection
+    keeps only the PUF clone for some source units (the L0-survivor case the
+    SSI reporter lineage already handles), and the receiver demanded exactly
+    one ASEC row per unit. The surviving clone carries the unit's source
+    predictors, so it serves as the prediction row; duplicated ASEC rows
+    remain a hard error.
+    """
+
+    expanded = clone_us_frame_for_puf_support(_frame(10))
+    person = expanded.table("person")
+    tax_unit = expanded.table("tax_unit")
+    dropped_source = tax_unit["tax_unit_source_id"].iloc[0]
+    dropped_units = tax_unit.loc[
+        tax_unit["tax_unit_source_id"].eq(dropped_source)
+        & tax_unit["tax_unit_support_channel"].eq("asec"),
+        "tax_unit_id",
+    ]
+    dropped = person["person_tax_unit_id"].isin(dropped_units)
+    sparse = expanded.select(~dropped.to_numpy())
+
+    class FakeFitted:
+        def predict(self, receiver: pd.DataFrame) -> pd.DataFrame:
+            return pd.DataFrame(
+                {_OUTPUT: np.ones(len(receiver), dtype=bool)},
+                index=receiver.index,
+            )
+
+    class FakeQRF:
+        def __init__(self, **_kwargs: object) -> None:
+            pass
+
+        def fit(self, *_args: object, **_kwargs: object) -> FakeFitted:
+            return FakeFitted()
+
+    monkeypatch.setattr(module, "QRF", FakeQRF)
+    predicted = impute_us_voluntary_filing(sparse, _donor(), seed=17)
+    survivors = sparse.table("tax_unit")["tax_unit_source_id"].eq(dropped_source)
+    assert survivors.any()
+    assert len(predicted) == len(sparse.table("tax_unit"))
+    assert predicted[survivors.to_numpy()].all()
+
+
 def test_real_qrf_recomputation_is_deterministic() -> None:
     frame = _frame(14)
     donor = _donor(120)