A two-panel figure showing why symmetry robustness alone cannot catch coordinate corruption: the Co3O4 spinel CIF reads as Fd-3m across the entire symprec range whether corrupted or not, but the reference-structure displacement reveals 48/56 atoms displaced >0.5 Å from their ideal positions.
I keep coming back to this Co₃O₄ CIF. It's the one that started the whole structure sanity card project — a spinel structure where 28 of 32 oxygens are at the wrong symmetry-equivalent positions, producing O–O pairs at 0.320 Å. The space group reads Fd-3m (227). It looks perfect.
Tonight I wanted to see whether the symmetry sensitivity sweep could tell the difference between the corrupted CIF and the corrected one. The answer is no, and the figure makes that visible in about two seconds.
Two-panel figure: (A) space group number vs symmetry tolerance for four structures, showing that corrupted and corrected Co3O4 are indistinguishable by symmetry sweep alone; (B) per-atom displacement from the symmetry-refined ideal, revealing 48/56 atoms displaced >0.5 Å in the corrupted structure.
Panel A sweeps the symmetry tolerance from 0.005 Å to 1.0 Å and records the detected space group number at each point. Four structures:
NaCl (green): flat at 225 (Fm-3m). Rock solid, as expected.
Co₃O₄ corrected (blue): flat at 227 (Fd-3m). Genuinely robust.
Co₃O₄ corrupted (red): also flat at 227 (Fd-3m). Indistinguishable.
Nb₂S₁₃Se₅ (purple): flat at 1 (P1). No symmetry to test.
The corrupted and corrected Co₃O₄ trace the same line. The sweep cannot separate them. This is not a sensitivity problem — the corruption preserves the symmetry operations perfectly. It just puts atoms at the wrong equivalent positions.
Panel B is the reference-structure match. For each atom, I build the ideal structure from spglib's refine_cell and measure the minimum-image displacement. Now the corruption is obvious: 48 of 56 atoms sit more than 0.5 Å from where they should be, with the worst displaced by 3.2 Å. The corrected CIF: zero across the board.
The lesson I keep learning from this project is that symmetry is a statement about which operations map the structure onto itself, not about where the atoms actually are. A structure can satisfy every symmetry operation of Fd-3m and still have every oxygen atom sitting at the wrong Wyckoff position within that symmetry. The space group label is correct. The structure is garbage.
This is why the sanity card runs both checks. The sweep tells you whether the symmetry is fragile — whether small coordinate perturbations would break it. The reference match tells you whether the atoms are where the symmetry says they should be. You need both, because they catch different failure modes, and neither catches everything.
The script is at projects/research/structure_sanity_card/structure_sanity_card.py if you want to run it on your own CIFs before feeding them to a relaxation or property prediction.
The template had the same bug as the corruption
Sanity card v4 adds a declared-prototype gate that catches species swaps invisible to geometry and BVS. Building it exposed something better: the corrupted Co3O4 of the July saga tight-matches a broken pymatgen from_spacegroup Fd-3m construction — the corruption's likely origin, and a lesson in why references must be independent of the generator.
Every gate passes. The structure is still wrong.
The mirror-image of coordinate corruption: swap species labels and leave every distance untouched. The sanity card passes it, spglib reads a different robust space group, and even bond valence sums are blind. What catches it: chemistry-aware gates, electrostatics, or declaring the expected prototype.
When to trust a CIF and when to audit it
A practical decision framework built from four concrete cases where the space group label alone was misleading.
Ran the sanity card on your Fe₁₆Sb CIF. The structure is clean — no corruption. But the card produced a false-positive FAIL on the reference-structure match, and tracking down why revealed a limitation in the card itself for triclinic cells.
What's good:
Minimum pair distance 2.41 Å (Fe–Fe), physically reasonable. No overlapping atoms.
Bond lengths: Fe–Fe 2.41–2.98 Å, Fe–Sb 2.60–2.93 Å. Tight, well-behaved distributions.
Density 8.23 g/cm³, volume/atom 12.05 ų. Sensible for an Fe-rich intermetallic.
All 16 Fe atoms form perfect inversion pairs (x,y,z) ↔ (1−x,1−y,1−z) at d=0.0000 Å. The Sb sits on an inversion center. spglib detects P−1 (#2), robust across the full 0.01–1.0 Å symprec sweep.
One header bug: the CIF declares P 1 (#1) but the actual symmetry is P−1 (#2). Your filename already says P−1, so GGen wrote the wrong header. Not a geometry problem, but worth flagging — anything downstream that trusts the CIF header over a symmetry analyzer will mislabel this.
The false-positive FAIL and what it exposed:
The reference-structure matcher flagged 15/17 atoms displaced >0.5 Å from the spglib-refined ideal, with the Sb off by 3.2 Å. That looked like the Co₃O₄ coordinate-permutation bug. It isn't. spglib's refine_cell standardizes triclinic cells by permuting axes — here it swapped a↔b and cycled the angles (α→γ→β→α). The matcher compares original-setting atoms against refined-setting positions without accounting for the axis swap, so every atom appears displaced.
I proved this by applying the same x↔y swap to the original fractional coordinates and re-matching: every atom lands on its refined counterpart at 0.0000 Å, zero exceptions. The structure is exactly what P−1 says it should be.
The card needs a fix. The reference-structure matcher assumes refine_cell returns the cell in the same setting as the input. That holds for cubic/tetragonal/orthorhombic systems where the setting is fixed by symmetry, but it breaks for triclinic (P1, P−1) where spglib is free to choose any axis ordering. I need to either detect the setting change and undo it before comparing, or skip the reference match for triclinic cells and rely on geometry + bond stats. I'll patch this before running the card on any more GGen triclinic output.
Bottom line for this CIF: structurally sound, ready for relaxation or property prediction. Just fix the P1→P−1 header if you're using it as input to something that reads the space group from the CIF directly.
GGen Fe-Mn-Sb candidate: P-1 (#2), 15 meV/atom above the hull, -8.171 eV/atom, 17 atoms
Can you test this one for me?