Limitations

These are current implementation limits, not planned behavior.

Storage and data model

• SQLite and PostgreSQL are supported backends; SQLite is the sole portable archive format.
• SQLite has no server, writer coordination, or access control. PostgreSQL provides those natively; STOMS adds no authorization layer, connection pool, or backup service of its own. See PostgreSQL operational ledger for what STOMS adds on top of PostgreSQL's own concurrency guarantees.
• Structure, evaluation, and catalog iteration is cursor-backed. Trajectory, stored-query, and aggregation iteration materializes matching rows first.
• Structures, evaluations, trajectories, and aggregation membership cannot be updated or deleted. Tags are append-only.

Querying and selection

• Queries cannot inspect numerical property values, atom counts, array contents, or scientific/geometric similarity.
• STOMS does not infer scientific equivalence between method specifications.
• MethodSpec.model_id is not persisted, hashed, or matched by SQLite. Do not use it as a durable evaluator identity; use method_identifier and parameters.
• DatasetQuery(require_evaluation=True) does not itself select an evaluator; strict selection still requires an evaluator criterion.

Export and VASP

• extxyz export performs no unit conversion. It exports standard energy/forces, optional node energy, custom scalar values, and custom arrays with atom-count first dimension; other custom values are omitted.
• The VASP adapter parses vasprun.xml and selected sidecar metadata. It does not create structures, ingest records, or preserve general VASP artifacts.
• A missing POTCAR can fall back to XML titles. A present POTCAR with no parseable datasets is an error.
• XML recovery depends on pymatgen support. Versions without exception_on_bad_xml cannot request a distinct recovery parse.