Tag
This paper identifies a non-composition principle in AI benchmark evaluation: support for adjacent projections does not automatically warrant their composition. It proposes a projectibility audit to diagnose unsupported joins in benchmark-to-use arguments, with a legal-research case study and simulations.