Capture metadata + the R-IDs this paper touches. LLM-assisted summary and applicability scoring run in L1.
Pick every invariant this paper meaningfully interacts with. Conflicts surface in L1 evaluation.