Three worked Crystalite requests end to end (ZnS, SmCo5, Nd2Fe14B), measured known limits from the 10-structure acceptance pass, and how to report a bad structure. The one-line version: generate, then relax, then trust.
Crystalite is live on Ouro as a service: a lightweight EDM diffusion transformer, trained on MP20, that generates crystal structures. Give it a composition and it returns a CIF in about twenty seconds. This guide walks through three real requests end to end, including what the raw output is and is not good for, because the honest answer to "can I use this structure directly" is sometimes.
I asked the Predict a crystal structure route for ZnS with one sample:
{"composition": "ZnS", "n_samples": 1}Simple CSP endpoint: predict a structure for one composition and return a single CIF.
Twenty-one seconds later: ZnS (F-43m), in the primitive rhombohedral setting of the cubic cell. I ran the output through the structure sanity card before trusting it: Zn-S bonds at 2.35 A (sane), and the species-aware prototype gate confirmed zincblende outright. Zincblende is the experimental ground state of ZnS, so this CIF is directly usable. When Crystalite is in-distribution and confident, this is what you get.
Same route, {"composition": "SmCo5", "n_samples": 1}:
Simple CSP endpoint: predict a structure for one composition and return a single CIF.
The route returned SmCo5 (P6/mmm), and P6/mmm is the correct space group for the real CaCu5-type phase. But the sanity card tells a more careful story: at a tight 0.01 A tolerance the structure reads as P1, the hexagonal label only emerges at 0.05 A, and five of six atoms sit more than 0.5 A from their symmetry-ideal positions (worst case 1.6 A). The cell it encoded is a distorted, non-standard setting. This CIF is a candidate, not a structure: relax it with an MLIP or DFT before computing anything from it.
{"composition": "Nd2Fe14B", "n_samples": 1} returned Nd2Fe14B (P1): a 17-atom, near-cubic triclinic cell that is robustly P1 at every tolerance. Nothing is wrong with the geometry (B-Fe 2.13 A is reasonable), but it shares nothing with the P42/mnm hard-magnet phase. For compositions with a famous ground state, Crystalite CSP is a candidate generator, not a lookup table. (View run)
These come from a 10-structure acceptance pass (oxides, intermetallics, sulfides, and one molecular case), with every output audited and logged in the public Crystalite generation receipts
Raw outputs carry diffusion noise. 8 of 10 read as P1 or lower at 0.01 A tolerance and only reach their reported space group at looser tolerance. Relax before use, always.
The label can be right while coordinates are off. In 4 of 10, atoms sat 1.0 to 2.5 A from symmetry-ideal positions behind a plausible label.
Known-prototype recovery is weak without relaxation. Only ZnS landed its experimental ground state directly; ZrO2 found the real high-temperature cubic phase rather than the room-temperature monoclinic one.
No organics or molecular crystals. The model is MP20-trained; urea came back with a 1.19 A C-N contact that should be about 1.38 A. Out of scope is out of scope.
Early instability is on record. Four of seven launch-night requests hit fast upstream 500s before the service stabilized; all ten acceptance requests succeeded. If you hit a 500, retry once, then report it.
Every generation is logged in the receipt ledger
The one-line version: generate, then relax, then trust.