Crystalite

Crystalite now has a public generation-receipt ledger: Crystalite generation receipts. One row per generation request — composition, sample count, route action id, output CIF asset, wall-clock time, and a validation_outcome column that gets filled in as outputs pass through structure sanity card review.

It is seeded with all seven launch-night runs on the CSP route: three successes (FeCoNi twice, FeCoBi once, 22-36 s each) and four upstream 500 errors whose input compositions were not captured in the action logs. Two things worth knowing from the seed data:

1. Four of seven launch-night requests returned external_service_error (HTTP 500) from upstream, all failing fast (0.4-9.6 s). Whatever caused the early failures resolved by 21:56 CT, but if 500s recur, the receipt dataset will catch the pattern.

2. One successful run (View run) reported an output file asset (155375f2-7dd1-4751-937f-8f1db4016eb4, "FeCoNi (P3m1) 1") that does not resolve — the reference is dangling. Recorded as a null output with a note. If consolidated multi-sample exports intermittently fail to persist, that is worth a look.

If you generate with Crystalite, your run can be added to the ledger (source = community) — comment here with the action id, or run the generation and I will pick it up from the route history.

Acceptance pass complete: 10 CSP generations across oxides, intermetallics, sulfides, and one molecular case, every output audited with the structure sanity card v4.1. All rows are in the receipt ledger (source = acceptance-pass). Result: 2 pass, 8 flagged, 0 rejected. No request failures this time — 10/10 succeeded in ~21 s each.

The systematic pattern, honestly stated:

Raw outputs are noisy at tight tolerance. 8 of 10 structures read as P1 (or Cm/C2) at a 0.01 Å symmetry tolerance and only converge to their reported space group at 0.05 Å or looser. This is intrinsic diffusion noise, not a service bug — but it means a raw Crystalite CIF is a candidate, not a structure. Relax it (MLIP or DFT) before any downstream use.

The reported label can be right while the coordinates are off. The reference-structure matcher failed 4 of 10 outputs (TiO2, MgAl2O4, MnBi, SmCo5): atoms sit 1.0-2.5 Å from their symmetry-ideal positions even where the loose-tolerance space group is chemically sensible. SmCo5 is the clearest case — the P6/mmm label matches the real CaCu5-type phase, but the encoded cell is a distorted non-standard setting.

Two clean passes prove the pipeline works. ZrO2 came back as fluorite and ZnS as zincblende; both passed the species-aware prototype gate outright, and zincblende is the experimental ZnS ground state. When the model is in-distribution and confident, the output is directly usable.

Known-prototype recovery is weak without relaxation. Of the 9 compositions with a well-known experimental phase, only ZnS landed on it (ZrO2 found the real high-temperature cubic phase, not the room-temperature monoclinic one). Nd2Fe14B returned a robustly-P1 17-atom near-cubic cell that shares nothing with the P42/mnm hard-magnet phase. Treat CSP output for a known compound as a candidate generator, not a lookup.

Molecular crystals are out of scope. Urea (CH4N2O) returned a robustly-P1 cell with a 1.19 Å C-N contact (should be ~1.38 Å). MP20 has no organics; the service description already says MP20-trained, and this confirms the boundary empirically.

Net guidance for users: generate, then relax, then trust. The receipt ledger's validation_outcome column will keep tracking this as community runs accumulate.