diff --git a/models/rfd3/src/rfd3/engine.py b/models/rfd3/src/rfd3/engine.py
index 7d7bd5be..dec07b90 100644
--- a/models/rfd3/src/rfd3/engine.py
+++ b/models/rfd3/src/rfd3/engine.py
@@ -95,6 +95,15 @@ class RFD3Output:
     denoised_trajectory_stack: Optional[AtomArrayStack] = None
     noisy_trajectory_stack: Optional[AtomArrayStack] = None
 
+    @staticmethod
+    def _strip_atom_id(atoms):
+        """Strip ``atom_id`` so the CIF writer auto-generates unique
+        sequential IDs instead of using the duplicated values left by
+        virtual-atom padding."""
+        if "atom_id" in atoms.get_annotation_categories():
+            atoms.del_annotation("atom_id")
+        return atoms
+
     def dump(
         self,
         out_dir,
@@ -103,7 +112,7 @@ def dump(
         base_path = os.path.join(out_dir, self.example_id)
         base_path = Path(base_path).absolute()
         to_cif_file(
-            self.atom_array,
+            self._strip_atom_id(self.atom_array),
             base_path,
             file_type="cif.gz",
             include_entity_poly=False,
@@ -118,7 +127,7 @@ def dump(
         suffix = str(base_path)[-1]
         if self.denoised_trajectory_stack is not None:
             to_cif_file(
-                self.denoised_trajectory_stack,
+                self._strip_atom_id(self.denoised_trajectory_stack),
                 "_denoised_model_".join([prefix, suffix]),
                 file_type="cif.gz",
                 include_entity_poly=False,
@@ -126,7 +135,7 @@ def dump(
 
         if self.noisy_trajectory_stack is not None:
             to_cif_file(
-                self.noisy_trajectory_stack,
+                self._strip_atom_id(self.noisy_trajectory_stack),
                 "_noisy_model_".join([prefix, suffix]),
                 file_type="cif.gz",
                 include_entity_poly=False,
diff --git a/models/rfd3/src/rfd3/trainer/trainer_utils.py b/models/rfd3/src/rfd3/trainer/trainer_utils.py
index 59f43a1c..b088d298 100644
--- a/models/rfd3/src/rfd3/trainer/trainer_utils.py
+++ b/models/rfd3/src/rfd3/trainer/trainer_utils.py
@@ -189,7 +189,17 @@ def _cleanup_virtual_atoms_and_assign_atom_name_elements(
         atom_array.element,
     )
     atom_array.res_name[invalid_mask] = np.array(["UNK"] * sum(invalid_mask))
-    return atom_array[ret_mask]
+    result = atom_array[ret_mask]
+
+    # Strip atom_id annotation to prevent duplicate _atom_site.id values
+    # in CIF output. PadTokensWithVirtualAtoms copies the central atom (CB)
+    # including its atom_id; after virtual atoms are removed, sidechain atoms
+    # retain the duplicated atom_id. Removing it lets the CIF writer
+    # auto-generate unique sequential IDs.
+    if "atom_id" in result.get_annotation_categories():
+        result.del_annotation("atom_id")
+
+    return result
 
 
 def _readout_seq_from_struc(
diff --git a/models/rfd3/tests/test_cif_duplicate_ids.py b/models/rfd3/tests/test_cif_duplicate_ids.py
new file mode 100644
index 00000000..80419830
--- /dev/null
+++ b/models/rfd3/tests/test_cif_duplicate_ids.py
@@ -0,0 +1,104 @@
+"""Regression tests for duplicate CIF atom_id (#148).
+
+PadTokensWithVirtualAtoms copies the central atom (CB) including its
+atom_id annotation.  After virtual atoms are removed, sidechain atoms
+retain CB's duplicated atom_id.  The CIF writer uses these values for
+_atom_site.id, producing duplicate IDs that violate the mmCIF spec.
+"""
+
+import tempfile
+from pathlib import Path
+
+import numpy as np
+from atomworks.io.utils.io_utils import to_cif_file
+from biotite.structure import AtomArray
+from rfd3.engine import RFD3Output
+from rfd3.trainer.trainer_utils import (
+    _cleanup_virtual_atoms_and_assign_atom_name_elements,
+)
+
+
+def _make_alanine_array_with_duplicate_atom_ids(n_residues=4):
+    """Build a protein AtomArray where every atom in each residue carries
+    the same atom_id as CB — the corruption caused by virtual-atom padding."""
+    names = ["N", "CA", "C", "O", "CB"]
+    n_atoms = n_residues * len(names)
+    atoms = AtomArray(n_atoms)
+
+    for i in range(n_residues):
+        for j, name in enumerate(names):
+            idx = i * len(names) + j
+            atoms.chain_id[idx] = "A"
+            atoms.res_id[idx] = i + 1
+            atoms.res_name[idx] = "ALA"
+            atoms.atom_name[idx] = name
+            atoms.element[idx] = name[0]
+            atoms.coord[idx] = [float(i * 4), float(j * 1.5), 0.0]
+
+    # Simulate the bug: every atom in a residue shares CB's atom_id
+    atom_ids = np.array(
+        [i * len(names) + 4 for i in range(n_residues) for _ in range(len(names))]
+    )
+    atoms.set_annotation("atom_id", atom_ids)
+
+    # Annotations required by _cleanup_virtual_atoms_and_assign_atom_name_elements
+    atoms.set_annotation(
+        "is_motif_atom_with_fixed_seq", np.ones(n_atoms, dtype=bool)
+    )
+    atoms.set_annotation(
+        "is_motif_atom_unindexed", np.zeros(n_atoms, dtype=bool)
+    )
+    atoms.set_annotation("gt_atom_name", atoms.atom_name.copy())
+
+    return atoms
+
+
+def test_cleanup_strips_atom_id():
+    """_cleanup_virtual_atoms_and_assign_atom_name_elements must remove
+    the atom_id annotation so the CIF writer generates fresh IDs."""
+    atoms = _make_alanine_array_with_duplicate_atom_ids()
+
+    # Precondition: atom_id exists and has duplicates
+    assert "atom_id" in atoms.get_annotation_categories()
+    assert len(set(atoms.atom_id)) < len(atoms.atom_id)
+
+    result = _cleanup_virtual_atoms_and_assign_atom_name_elements(atoms)
+    assert "atom_id" not in result.get_annotation_categories()
+
+
+def test_cif_output_has_unique_ids():
+    """CIF _atom_site.id values must be unique after _strip_atom_id."""
+    atoms = _make_alanine_array_with_duplicate_atom_ids()
+    atoms = RFD3Output._strip_atom_id(atoms)
+
+    with tempfile.TemporaryDirectory() as tmpdir:
+        out_path = Path(tmpdir) / "test"
+        to_cif_file(atoms, out_path, file_type="cif")
+
+        cif_text = Path(f"{out_path}.cif").read_text()
+
+        # Parse _atom_site loop to find the id column
+        lines = cif_text.splitlines()
+        col_names = []
+        data_start = None
+        for i, line in enumerate(lines):
+            if line.strip().startswith("_atom_site."):
+                col_names.append(line.strip().split()[0])
+            elif col_names and not line.strip().startswith("_") and line.strip():
+                data_start = i
+                break
+
+        assert "_atom_site.id" in col_names, "missing _atom_site.id column"
+        id_col = col_names.index("_atom_site.id")
+
+        ids = []
+        for line in lines[data_start:]:
+            stripped = line.strip()
+            if not stripped or stripped.startswith("#") or stripped.startswith("loop_"):
+                break
+            parts = stripped.split()
+            if len(parts) > id_col:
+                ids.append(parts[id_col])
+
+        assert len(ids) > 0, "no atom records found in CIF output"
+        assert len(ids) == len(set(ids)), f"duplicate _atom_site.id values: {ids}"