GMD Task 4 composition-support audit - structures and public-checkpoint predictions

This record preserves the raw inputs and outputs of an audit of compositional generalisation in machine-learned interatomic potentials. It contains 4,800 deterministically selected molecular structures from GMD-26 Task 4 (Fragment Combination), including the structure table, ExtXYZ geometries, 4,800 predictions from the public MACE checkpoint, and 4,800 predictions from the public NequIP checkpoint. It also contains the raw integrity manifest, composition-span audit, matched-carbon- block analysis, compute record, source lock, and inference run records. The archive demonstrates that every Task 4 amino-acid target composition lies in the span of the diamine and dicarboxylic-acid support compositions. Task 4 is therefore a composition-span-preserving negative control for the atomic-reference offset observed in the earlier Fragment Duplication audit. No model training and no new DFT calculation were performed to create this archive. The modifications consist of deterministic frame selection, conversion to Parquet and ExtXYZ, public-checkpoint inference, statistical analysis, and deterministic packaging. Source attribution: selected geometries and reference labels are derived from GMD-26 by Amir Masoud Nourollah, Irtaza Khalid, Stefano Leoni, and Steven Schockaert, released under Creative Commons Attribution 4.0. The original work is “Benchmarking Compositional Generalisation for Learning Inter-atomic Potentials”, ICLR 2026: https://openreview.net/forum?id=WxTlAbRUE6 The archive contains ARCHIVE_METADATA.json with the generating Git commit and SHA-256 checksum of every member. The corresponding code and complete audit history are available at https://github.com/Sensiel/mlip-fragment-mechanism-audit