These links are derived by comparing normalised Sanskrit surfaces. They show that two collections carry the same wording. They are not, by themselves, a demonstrated line of transmission between them, and the direction stored on an edge is a property of how it was built.
Atharvaveda is one of four collections held here. Cross-corpus comparison bottoms out at the weakest shared surface, because a Devanagari text and a romanised text share no code points.
How these were derived2
Parallelism is symmetric and this graph stores it as a directed edge, so these relations are traversed in both directions and each row states which way the stored edge ran. All 1,684 REUSES_TEXT_FROM edges point Samaveda-to-Rigveda, and 1,421 Rigvedic passages have reuse edges only inbound; following the arrow would show those 1,421 verses no reuse at all.
measured
SHARES_ENTITY_VOCABULARY_WITH is not textual parallelism. It means two passages name some of the same registry concepts and says nothing about shared wording, so it is reported with relation_kind=ENTITY_VOCABULARY_OVERLAP and is excluded from textual_total.