# Feature Reproduction

No pretrained embeddings, model checkpoints, CondenGT-specific target tables,
opaque feature caches, or pre-generated missing-count views are included. The
missing-count views are deterministic transformations of
`data/raw/phase_diagram_original_scale.csv`.

1. `python scripts/prepare_missing_splits.py` partitions rows by the number of
   NaN values across `group_11`–`group_44` without changing the source rows.
The generated views are intentionally excluded from `metadata/file_inventory.csv`.
A changed pandas version can alter CSV byte formatting; validate the total row
count (280,778), fields, and value equality when regenerating.