RFD3 output cif contains duplicates in id coulmn

I apologize that I don't know the setting to make rfd3 fully deterministic; however, I believe this should be fairly reproducible.

Run Command:

rfd3 design out_dir=/path/to/output inputs=/path/to/glob.json dump_trajectories=False prevalidate_inputs=True diffusion_batch_size=1 seed=42

with glob.json:

{
  "glob": {
    "length": "100-100",
    "is_non_loopy":true
    }
}

My output json:

{
    "diffused_index_map": {},
    "metrics": {
        "max_ca_deviation": 0.08025217056274414,
        "n_chainbreaks": 0,
        "n_clashing.interresidue_clashes_w_sidechain": 0,
        "n_clashing.interresidue_clashes_w_backbone": 0,
        "non_loop_fraction": 0.87,
        "loop_fraction": 0.13,
        "helix_fraction": 0.87,
        "sheet_fraction": 0.0,
        "num_ss_elements": 4,
        "radius_of_gyration": 13.636231638805786,
        "alanine_content": 0.14,
        "glycine_content": 0.01,
        "num_residues": 100,
        "diffused_com": [
            1.206663727760315,
            0.28238099813461304,
            -2.9324238300323486
        ]
    },
    "specification": {
        "length": "100-100",
        "extra": {
            "example": "glob",
            "task_name": "glob_glob",
            "sampled_contig": "100",
            "num_tokens_in": 100,
            "num_residues_in": 100,
            "num_chains": 1,
            "num_atoms": 500,
            "num_residues": 101,
            "example_id": "glob_glob_0"
        },
        "is_non_loopy": true
    },
    "ckpt_path": "/home/tristin/downloads/large_language_models/foundry/rfd3_latest.ckpt",
    "seed": 42
}

Now the output structure (after gunzip) seems to vary very slightly, but I'm pretty consistently duplicates in id 19:

...
ATOM C CB  . ASP A 0 4   . 4   ASP A CB  19  3.6175138405647456e-12 1.0 ? False -10.776768   -2.932267    4.327731     1
ATOM C CG  . ASP A 0 4   . 4   ASP A CG  19  3.6175138405647456e-12 1.0 ? False -11.970276   -2.7417085   5.2360744    1
ATOM O OD1 . ASP A 0 4   . 4   ASP A OD1 19  3.6175138405647456e-12 1.0 ? False -12.903551   -2.0365722   4.8578053    1
ATOM O OD2 . ASP A 0 4   . 4   ASP A OD2 19  3.6175138405647456e-12 1.0 ? False -11.906276   -3.3094902   6.332546     1
...

While pymol isn't bothered by this cif2pdb doesn't seem to like it -- I guess it's not proper cif format. I can renumber easily enough, but I figured I should note the issue in case you think it's something rfd3 should be doing.

Start by reproducing the provided rfd3 design command with the shown glob.json and inspect the generated structure after gunzip. Trace how residue or atom IDs are assigned in the RFD3 output path; done means the CIF contains no duplicate IDs and can be converted by cif2pdb.