A portable Agent Skill for paper review: it reconstructs a paper's real research question and checks, claim by claim, whether the authors' claims are backed by page-level evidence.
Takes an arXiv ID, DOI, or PDF and extracts full text with page markers; labels each claim supported, partial, or unsupported and assesses its overclaim risk.
Judges only from the paper itself and never bypasses paywalls; when figure extraction fails it is recorded as a limitation, not as a lack of support.
Reimplements Marginalia's review workflow at the behavioral level without copying upstream code.