diff --git a/src/ascend/transform/auto_schedule/buffer_alias_analysis.h b/src/ascend/transform/auto_schedule/buffer_alias_analysis.h
index bb7e8c987d..df0a2e65ad 100644
--- a/src/ascend/transform/auto_schedule/buffer_alias_analysis.h
+++ b/src/ascend/transform/auto_schedule/buffer_alias_analysis.h
@@ -1245,11 +1245,17 @@ class BufferAliasAnalyzerImpl {
       IRStructure *child = access.task;
       while (child != nullptr && child->GetParent() != parent_scope)
         child = child->GetParent();
-      if (child == nullptr || !child->IsControl())
+      // An access outside this scope is a boundary of the child lifetime.
+      // Every access inside this scope, however, must be covered by the
+      // selected child. Dropping a direct or single-trip sibling access can
+      // hide the writer of a value that survives across parent iterations.
+      if (child == nullptr)
         continue;
+      if (!child->IsControl())
+        return nullptr;
       auto *candidate = static_cast<ControlNode *>(child);
       if (IsGuaranteedSingleTrip(candidate))
-        continue;
+        return nullptr;
       if (control == nullptr) {
         control = candidate;
       } else if (control != candidate) {
@@ -1297,11 +1303,12 @@ class BufferAliasAnalyzerImpl {
     }
 
     bool parent_live_through = child_result.live_through;
-    if (child_result.promote_live_through_as_live_in) {
+    if (child_result.promote_live_through_as_live_in &&
+        parent_scope == nullptr) {
       // The storage spans the whole child control, so it cannot alias inside
-      // that child. At the parent level the same lifetime is a finite prefix
-      // ending at the child's last read and may alias a later sibling
-      // generation.
+      // that child. Only a non-periodic parent can reduce it to a finite
+      // prefix ending at the child's last read. A loop parent needs the
+      // incoming value again on its next iteration.
       parent_live_through = false;
       child_result.live_out = false;
     }
@@ -1368,7 +1375,7 @@ class BufferAliasAnalyzerImpl {
       }
     }
     child_result.live_through = parent_live_through;
-    child_result.promote_live_through_as_live_in = false;
+    child_result.promote_live_through_as_live_in &= parent_scope != nullptr;
     child_result.generations = std::move(generations);
     child_result.summary_control = control;
     return child_result;
diff --git a/testing/ascend/auto_schedule/test_buffer_lifetimes.py b/testing/ascend/auto_schedule/test_buffer_lifetimes.py
index 486a11cb32..1c0226fe93 100644
--- a/testing/ascend/auto_schedule/test_buffer_lifetimes.py
+++ b/testing/ascend/auto_schedule/test_buffer_lifetimes.py
@@ -967,5 +967,74 @@ def test_l0_address_reuse_includes_the_bound_sf_lifetime(sticky):
     assert sa.data not in aliases and sb.data not in aliases
 
 
+@pytest.mark.parametrize("reader_depth", [1, 2])
+def test_preloaded_nested_reader_stays_live_across_parent_iterations(reader_depth):
+    @T.prim_func
+    def main(
+        x: T.Tensor((1,), "int32"),
+        y: T.Tensor((4,), "int32"),
+        out_x: T.Tensor((4, 2, 2), "int32"),
+        out_y: T.Tensor((4,), "int32"),
+    ):
+        with T.Kernel(1):
+            a = T.alloc_shared((8,), "int32")
+            b = T.alloc_shared((8,), "int32")
+            a[0] = x[0]
+            for k in T.Pipelined(4, num_stages=1):
+                for j in T.serial(2):
+                    if reader_depth == 2:
+                        for i in T.serial(2):
+                            out_x[k, j, i] = a[0]
+                    else:
+                        out_x[k, j, 0] = a[0]
+                b[0] = y[k]
+                out_y[k] = b[0]
+
+    # b(k) must not overwrite the preload read again by a child of k + 1.
+    assert ("a", "b") not in _insert_sync_alias_pairs(main)
+
+
 if __name__ == "__main__":
     tilelang.testing.main()
+
+
+def _make_parent_writer_nested_reader_program(versions: int, inner_extent: int, single_trip_writer: bool):
+    @T.prim_func
+    def main(
+        x: T.Tensor((8,), "int32"),
+        y: T.Tensor((8,), "int32"),
+        out_x: T.Tensor((8, inner_extent), "int32"),
+        out_y: T.Tensor((8,), "int32"),
+    ):
+        with T.Kernel(1):
+            a = T.alloc_shared((8,), "int32")
+            b = T.alloc_shared((8,), "int32")
+            T.annotate_buffer_versions({a: versions})
+            for k in T.Pipelined(8, num_stages=5, annotations={"enable_offset": True}):
+                with T.Stage(0):
+                    if single_trip_writer:
+                        for _ in T.serial(1):
+                            with T.Stage(0):
+                                a[0] = x[k]
+                    else:
+                        a[0] = x[k]
+                with T.Stage(1):
+                    for j in T.serial(inner_extent):
+                        with T.Stage(0):
+                            out_x[k, j] = a[0]
+                with T.Stage(4):
+                    b[0] = y[k]
+                    out_y[k] = b[0]
+
+    return main
+
+
+@pytest.mark.parametrize("single_trip_writer", [False, True], ids=["direct-writer", "single-trip-writer"])
+@pytest.mark.parametrize("versions", [2, 3])
+@pytest.mark.parametrize("inner_extent", [1, 2], ids=["single-trip", "multi-trip"])
+def test_parent_writer_and_nested_reader_keep_periodic_lifetime(versions, inner_extent, single_trip_writer):
+    # At physical step 6, a(6) is waiting for its stage-1 reader when b(2)
+    # writes. Whole-allocation reuse is unsafe even though a(k)'s last read
+    # precedes b(k)'s first write in the same logical iteration.
+    program = _make_parent_writer_nested_reader_program(versions, inner_extent, single_trip_writer)
+    assert ("a", "b") not in _insert_sync_alias_pairs(program)