diff --git a/packages/populace-build/tests/test_us_fiscal_refresh_builder.py b/packages/populace-build/tests/test_us_fiscal_refresh_builder.py
index 4da6614..eb01b50 100644
--- a/packages/populace-build/tests/test_us_fiscal_refresh_builder.py
+++ b/packages/populace-build/tests/test_us_fiscal_refresh_builder.py
@@ -221,9 +221,22 @@ def gate(diagnostics, *, targets):
     return gate
 
 
+@pytest.mark.parametrize(
+    ("sparse_selection_arm", "expected_cap_ratio"),
+    [
+        pytest.param(False, 0.12, id="dense-arm"),
+        # The leak regression: the sparse rmloss100 script ALSO passes
+        # --dense-default-dataset (an export-mode flag), so the arm
+        # discriminator must be the frozen-selection identity — a sparse-arm
+        # reconcile gates at 0.10 even in dense export mode.
+        pytest.param(True, 0.10, id="sparse-arm-dense-export-mode"),
+    ],
+)
 def test_ssi_reconciliation_returns_fresh_pair_the_stale_gate_would_reject(
     monkeypatch,
     small_frame,
+    sparse_selection_arm,
+    expected_cap_ratio,
 ) -> None:
     """The fresh pair is published where the stale pair the old loop gated fails.
 
@@ -347,6 +360,7 @@ def to_target_set(self):
         initial_result,
         target_specs,
         dense_default_dataset=True,
+        sparse_selection_arm=sparse_selection_arm,
         seed=3,
         epochs=5,
         learning_rate=0.01,
@@ -372,6 +386,9 @@ def to_target_set(self):
     )
     record = reconciliation.compilation["ssi_take_up_reconciliation"]
     assert record["exit_policy"] == "fresh_pair_under_returned_weights"
+    assert (
+        record["ssi_swap_delta"]["national_swap_sanity_cap_ratio"] == expected_cap_ratio
+    )
     assert [entry["pass"] for entry in record["pass_history"]] == list(
         range(1, len(record["pass_history"]) + 1)
     )
diff --git a/tools/build_us_fiscal_refresh_release.py b/tools/build_us_fiscal_refresh_release.py
index 0652b2b..f8910f0 100644
--- a/tools/build_us_fiscal_refresh_release.py
+++ b/tools/build_us_fiscal_refresh_release.py
@@ -293,6 +293,11 @@
 #: while remaining the right bar for sparse (populace#447 adjudication,
 #: 2026-07-18; trajectory recorded in the reconciliation pass_history). Both
 #: caps converge back to one number when #424 restores the candidate universe.
+#: ARM IDENTITY: the dense arm is the run WITHOUT a frozen selection manifest.
+#: ``--dense-default-dataset`` is an EXPORT-MODE flag that the sparse
+#: rmloss100 script also sets, and must never be used as the arm
+#: discriminator (a 32605c9-era sparse rerun gated at 0.12 through exactly
+#: that confusion; caught by its recorded cap ratio before anything shipped).
 SSI_TAKE_UP_SWAP_SANITY_CAP_RATIO = 0.10
 SSI_TAKE_UP_SWAP_SANITY_CAP_RATIO_DENSE = 0.12
 
@@ -5356,6 +5361,7 @@ def _reconcile_ssi_take_up_and_refit(
     target_specs: tuple,
     *,
     dense_default_dataset: bool,
+    sparse_selection_arm: bool = True,
     seed: int,
     epochs: int,
     learning_rate: float,
@@ -5523,9 +5529,9 @@ def _reconcile_ssi_take_up_and_refit(
             maximum_microsim_batch_size=maximum_microsim_batch_size,
             selected_support=selected_support,
             sanity_cap_ratio=(
-                SSI_TAKE_UP_SWAP_SANITY_CAP_RATIO_DENSE
-                if dense_default_dataset
-                else SSI_TAKE_UP_SWAP_SANITY_CAP_RATIO
+                SSI_TAKE_UP_SWAP_SANITY_CAP_RATIO
+                if sparse_selection_arm
+                else SSI_TAKE_UP_SWAP_SANITY_CAP_RATIO_DENSE
             ),
         )
         pass_history.append(
@@ -8725,6 +8731,7 @@ def main() -> None:
         result,
         target_specs,
         dense_default_dataset=bool(args.dense_default_dataset),
+        sparse_selection_arm=args.selection_source_manifest is not None,
         seed=args.seed,
         epochs=args.epochs,
         learning_rate=args.learning_rate,