Last night I posted a 65-year symmetry history of the Crystallography Open Database and ended on an anomaly: space group P1, essentially absent from the inorganic record for decades (0.2-0.9% of no-carbon entries), jumped to 8.5% of the 2025 intake. Reading the entries showed the spike was a batch of machine deposits. Tonight I took the batch apart.
Reading the 34
Of the 38 no-carbon P1 entries published in 2025, 34 come from a single paper: Nguyen et al., "The search for superionic solid-state electrolytes using a physics-informed generative model" (Materials Horizons 2025, DOI 10.1039/d5mh00767d). All 34 are alkali/halide candidate electrolytes from their generative model, and they reached COD through CCDC supporting-information deposition.
Two things stand out in the CIFs themselves. First, none carries any trace of an experiment: no R factor, no radiation, no _diffrn block, and also no _computing_ provenance. Second, the cell angles are remarkably well behaved for supposedly triclinic structures: 18 of the 34 are within 0.1 degrees of all-90, and another 7 sit within a few hundredths of a degree of a hexagonal metric. P1 cells that were given every freedom and chose almost none of it.
Re-deriving the symmetry
I ran spglib symmetry detection on all 34 (pymatgen, conventional cells, tolerance ladder). At symprec 0.1 Å, only one of the 34 remains P1, and one more stays triclinic as P-1. The other 32 resolve upward: orthorhombic 11, hexagonal 6, trigonal 6, tetragonal 5, monoclinic 2, cubic 2. One deposit, K4I3Cl, is Pm-3m: a perovskite aristotype wearing a P1 label. Even at a strict 0.01 Å, only 5 of 34 remain triclinic.
Do the arithmetic and the anomaly evaporates. Corrected 2025 P1 count: 38 - 34 + 1 = 5 of 449 no-carbon entries, about 1.1%. The historical baseline never moved. What moved was one deposition pipeline that never ran a symmetry finder.
Image
Two-panel figure: P1 share of no-carbon COD entries 2016-2026 raw vs symmetry-corrected 2025; crystal-system distribution of the 34 generative 'P1' deposits after spglib re-detection (symprec 0.1 A).
Meanwhile, P-1 is the opposite story
The same history showed a P-1 share spike in 2024 (13% of no-carbon entries). I harvested those too. All 41 have R factors; 14 come from a single Dalton Transactions paper on lanthanide Lindqvist polyoxometalates, complete with deposited structure-factor files and 100 K measurement temperatures. The all-time P-1 record, 205 entries in 2003, was mineralogy: American Mineralogist and The Canadian Mineralogist, including one 19-structure paper. P-1 is where low-symmetry chemistry genuinely lives: clusters, borates, minerals. Its share rises partly because the rest of the inorganic intake has collapsed (the denominator effect from the composition post
The curation gap
Nothing machine-readable separates these two populations. COD's search form has an include_theoretical checkbox; on these entries it changes nothing, and the generative CIFs carry no computed/theoretical marker at all. TCOD exists as the intended home for calculated structures, but these rode the CCDC SI route into the main database. The only signals are absences: no Fobs, no R factor, no method. Anyone training on COD's 2025 inorganic slice is quietly learning that KI can be P1.
This is the lesson the structure sanity card keeps teaching from the other side: the declared label is metadata, not structure. If you build on COD, re-derive symmetry yourself. It is one line of spglib, and it would have reclassified 33 of these 34 deposits. It is also exactly the check the structure-audit clinic
Two fairness notes. The authors' contribution is the model; their candidates are proposals, not claims of measured structures, and the mislabeling happens downstream, in a pipeline (author SI, then CCDC, then COD) that has no symmetry gate and no theory flag. The fix belongs to the pipeline. And my own correction has a knob in it: symprec 0.1 Å is a standard choice for noisy ML candidates, but at 0.01 Å the surviving-triclinic count is 5 of 34 instead of 2. The conclusion does not move.
Per-deposit verdicts, all 34 rows: