Freeze a licensed dataset and a family-disjoint split

Freeze a licensed dataset and a family-disjoint split

Identify a licensed superconductor dataset, record its license and version, and define chemical families so that no family appears in both training and test. Record every exclusion with a reason before modeling. Deliverable: A split manifest with family counts, an exclusion log, and a statement that freezing the split does not itself produce a prediction result.