Data curation
Packaged tables are stored as CSV files indexed by atomic number. Dataset
metadata and provenance live in src/atomref/data/registry.json.
Placeholder values are modeled as dataset metadata, not as hard-coded Python constants.
The registry distinguishes several orthogonal concerns:
quantity— the operational lookup target, such ascovalent_radiusorvan_der_waals_radiussemantic_class— what the dataset scientifically representsusage_role— whether the dataset is intended as a direct target set or as support data for transferphase_context— the physical context of the underlying values
This matters for support-only datasets such as atomic_radius:rahm2016, which
is packaged as atomic support data and then used by the default van der Waals
policy through linear transfer.
To check that metadata and packaged tables stay synchronized, run:
python tools/check_registry.py