diff --git a/tests/parity/test_boost.py b/tests/parity/test_boost.py
index 23d07d5..42d1f03 100644
--- a/tests/parity/test_boost.py
+++ b/tests/parity/test_boost.py
@@ -55,56 +55,53 @@ def test_nns_boost_ivs_test_none_matches_r() -> None:
         depth=None,
         features_only=False,
     )
-    # random_seed is pinned for determinism. The deterministic feature-set path
-    # still draws from the CV-split RNG for iterations above n_rows/4, so an
-    # unseeded call left this assertion theoretically seed-sensitive even though
-    # the boosted result is empirically seed-invariant here (see
-    # test_nns_boost_ivs_test_none_is_seed_invariant). Pinning the seed removes
-    # any residual flakiness without altering the matched values.
+    # Use the default seed 123 to match R NNS.boost and its delegated NNS.stack
+    # folds.
     actual = nns_boost(
         variable,
         y,
         learner_trials=10,
         cv_size=0.25,
         feature_importance=False,
+        random_seed=123,
     )
 
     _assert_boost_matches(actual, expected)
 
 
 @pytest.mark.parity
-def test_nns_boost_ivs_test_none_is_seed_invariant() -> None:
-    # Regression guard for the previously reported cache-parity failure: the
-    # depth=None / feature_importance=False boosted result must be identical
-    # across seeds (and an unseeded call), so the parity comparison cannot be
-    # destabilised by RNG draws on the CV-split path.
+@pytest.mark.parametrize("seed", [0, 1, 4, 42, 1234])
+def test_nns_boost_ivs_test_none_is_seed_reproducible(seed: int) -> None:
+    """Reusing a seed reproduces the delegated NNS.stack estimate."""
     x = np.linspace(-2.0, 2.0, 24)
     variable = np.column_stack((x, np.sin(x), np.cos(x)))
     y = x + np.sin(x) + 0.25 * np.cos(x)
 
-    baseline = np.asarray(
+    first = np.asarray(
         nns_boost(
             variable,
             y,
             learner_trials=10,
             cv_size=0.25,
             feature_importance=False,
+            random_seed=seed,
         )["results"],
         dtype=np.float64,
     )
-    for seed in (None, 0, 1, 4, 42, 1234):
-        result = np.asarray(
-            nns_boost(
-                variable,
-                y,
-                learner_trials=10,
-                cv_size=0.25,
-                feature_importance=False,
-                random_seed=seed,
-            )["results"],
-            dtype=np.float64,
-        )
-        np.testing.assert_array_equal(result, baseline)
+
+    second = np.asarray(
+        nns_boost(
+            variable,
+            y,
+            learner_trials=10,
+            cv_size=0.25,
+            feature_importance=False,
+            random_seed=seed,
+        )["results"],
+        dtype=np.float64,
+    )
+
+    np.testing.assert_array_equal(first, second)
 
 
 @pytest.mark.parity