Given a base model, we build the lineage DAG (parent → child) and take the base plus all
descendants as the family. Each model reports only its incremental training delta;
the family total sums the subtree with deduplication for merged nodes.
Coverage shows what fraction disclosed a dia_report footprint block —
totals are a lower bound when reports are missing.
Publisher pretraining estimates (e.g. Meta GHG tables) appear separately and are
not included in DIA rollup totals unless ingested as a dia_report.
Graph interaction inspired by
Model Atlas
(Horwitz et al., 2025).
