diff --git a/changelog.d/9808-element-shape-transfer-gate.md b/changelog.d/9808-element-shape-transfer-gate.md
new file mode 100644
index 0000000000..b7d3b19b3f
--- /dev/null
+++ b/changelog.d/9808-element-shape-transfer-gate.md
@@ -0,0 +1,25 @@
+**A relocated array whose element-shape proof does not exist no longer takes
+the side table to find that out** (#9792).
+
+`transfer_element_shape` runs from `layout_transfer` for every relocated
+array — growth forwarding, a copying minor, an old-gen defrag. It already
+computes `had_bit` for free from two header words it has read anyway, and
+then took `ELEMENT_SHAPES`' `RefCell` and hashed both addresses regardless,
+for two removes that on the overwhelmingly common path remove nothing. It now
+returns when neither address advertises a proof.
+
+The gate has exactly one safe shape and both halves are pinned by
+`a_transfer_skips_the_table_only_when_neither_address_advertises_a_proof`.
+Skipping on `!had_bit` alone would be wrong: a destination that still
+advertises a proof describes storage the move has just replaced, so that case
+keeps the full fail-closed path.
+
+What skipping leaves behind is a record at an address whose bit is clear, and
+that state was already part of the design rather than new: the bit is the sole
+authority for a read (`element_shape_proof` returns `None` before touching the
+table, and `note_element_store` is gated on the same bit), and `establish`
+draws every identity from `ELEMENT_SHAPE_PROOF_SEQ` rather than from whatever
+record sits at the address — precisely so a survivor cannot donate its epoch
+to the next array proven there. `prune_dead_element_shape_owners` drops it on
+the next collection, the same footprint-only guarantee a fail-closed transfer
+already relied on. The test asserts both defences directly.
diff --git a/crates/perry-runtime/src/array/element_shape.rs b/crates/perry-runtime/src/array/element_shape.rs
index 0158dbad33..29f0f245cc 100644
--- a/crates/perry-runtime/src/array/element_shape.rs
+++ b/crates/perry-runtime/src/array/element_shape.rs
@@ -672,6 +672,25 @@ pub(crate) fn transfer_element_shape(old_user: usize, new_user: usize) {
         // exactly the versioning a consumer guards on.
         let had_bit = array_gc_header(old_user as *const ArrayHeader)
             .is_some_and(|old_header| header_has_bit(old_header));
+        // #9792: neither address advertises a proof, so there is nothing to
+        // move and nothing to fail closed about — the `clear_bit` below would
+        // clear a bit that is already clear. Skipping is what the siblings in
+        // `gc::layout_tables` do with their emptiness flag, decided here from
+        // the header words this function has already read rather than from a
+        // side-table probe.
+        //
+        // What it leaves behind is a record at an address whose bit is clear,
+        // and that state is already part of the design: the bit is the sole
+        // authority for a read (`element_shape_proof` returns `None` without
+        // touching the table), and `establish` draws every identity from
+        // `ELEMENT_SHAPE_PROOF_SEQ` rather than from whatever record sits at
+        // the address, precisely so a survivor cannot donate its epoch to the
+        // next array established there. `prune_dead_element_shape_owners`
+        // drops it on the next collection, which is the same footprint-only
+        // guarantee a fail-closed transfer already relied on.
+        if !had_bit && !header_has_bit(new_header) {
+            return;
+        }
         let moved = ELEMENT_SHAPES.with(|m| {
             let mut map = m.borrow_mut();
             map.remove(&new_user);
@@ -795,6 +814,16 @@ pub(crate) unsafe fn test_element_shape_bit_set(arr: *const ArrayHeader) -> bool
     array_gc_header(arr).is_some_and(|header| header_has_bit(header))
 }
 
+/// Clear the advertising bit and leave the record in the table — the survivor
+/// state a skipped [`transfer_element_shape`] produces, and the one the
+/// bit-is-authority rule has to hold up under.
+#[cfg(test)]
+pub(crate) unsafe fn test_clear_element_shape_bit_only(arr: *mut ArrayHeader) {
+    if let Some(header) = array_gc_header(arr) {
+        clear_bit(header);
+    }
+}
+
 #[cfg(test)]
 #[path = "element_shape_tests.rs"]
 mod tests;
diff --git a/crates/perry-runtime/src/array/element_shape_tests.rs b/crates/perry-runtime/src/array/element_shape_tests.rs
index 80199a91a0..fa93b4f76c 100644
--- a/crates/perry-runtime/src/array/element_shape_tests.rs
+++ b/crates/perry-runtime/src/array/element_shape_tests.rs
@@ -541,6 +541,71 @@ fn a_fail_closed_transfer_leaves_no_record_for_the_next_array_to_inherit() {
     );
 }
 
+#[test]
+fn a_transfer_skips_the_table_only_when_neither_address_advertises_a_proof() {
+    // #9792: `transfer_element_shape` runs for every relocated array, and its
+    // `had_bit` verdict is free — two header words it has already read. When
+    // neither address advertises a proof it takes the side table anyway, for
+    // two removes that can only remove what no reader could reach. The gate
+    // that skips that has exactly ONE safe shape, and this pins both halves.
+    //
+    // Arm 1 — neither bit set: skipping is correct, and the record it leaves
+    // behind stays unreadable and cannot donate its identity.
+    // Arm 2 — the DESTINATION still advertises one: skipping would leave a
+    // live proof describing storage that has just been overwritten, so the
+    // gate must NOT fire on `!had_bit` alone.
+    let _serialized = test_serialize();
+
+    // Arm 1.
+    let src = built_from_pushes(CLASS_A, 2);
+    let dst = built_from_pushes(CLASS_A, 2);
+    let survivor = proof(dst).expect("proven").epoch;
+    unsafe { clear_element_shape(src) };
+    unsafe { test_clear_element_shape_bit_only(dst) };
+    assert!(
+        test_element_shape_record_exists(dst as usize),
+        "the fixture must actually leave a record behind the cleared bit"
+    );
+
+    transfer_element_shape(src as usize, dst as usize);
+
+    assert!(
+        proof(dst).is_none(),
+        "the bit is the sole authority for a read, so a record behind a \
+         cleared bit must not read as a proof"
+    );
+    let reproven = unsafe { ensure_element_shape(dst) }.expect("still homogeneous");
+    assert_ne!(
+        reproven.epoch, survivor,
+        "establishing draws a fresh identity from ELEMENT_SHAPE_PROOF_SEQ, so \
+         a survivor record can never donate its epoch"
+    );
+
+    // Arm 2: source proves nothing, destination still advertises a proof.
+    let src2 = built_from_pushes(CLASS_A, 2);
+    let dst2 = built_from_pushes(CLASS_A, 2);
+    unsafe { clear_element_shape(src2) };
+    assert!(
+        unsafe { test_element_shape_bit_set(dst2) },
+        "the fixture must leave the destination advertising a proof"
+    );
+
+    transfer_element_shape(src2 as usize, dst2 as usize);
+
+    unsafe {
+        assert!(
+            !test_element_shape_bit_set(dst2),
+            "a transfer whose source proved nothing must still fail the \
+             destination closed — gating on `!had_bit` alone would leave a \
+             live proof over storage the move has just replaced"
+        );
+    }
+    assert!(
+        !test_element_shape_record_exists(dst2 as usize),
+        "and it must take the destination's record with it"
+    );
+}
+
 // ---------------------------------------------------------------------------
 // Lifecycle hooks — what stops a recycled address inheriting a stale identity
 // ---------------------------------------------------------------------------