Data and code for: A single reference-energy outlier nearly eliminates Co–O chemistry from a generative model's training distribution

Supporting data and analysis code for the manuscript submitted to Digital Discovery (RSC). The paper traces a cobalt-specific failure of conditional crystal generation back to one row ofthe public training data. Every number reported in the manuscript is recomputed from the filesin this deposit by a single script (`verify_all_numbers.py`), which prints a mismatch count. This deposit is self-contained except for the training archive itself, which is 186 MB andis redistributed by its original authors; its exact provenance and checksum are recorded in`alex_mp_20_snapshot.json` so that the identical snapshot can be retrieved.