Adversarial Technical Claims Reviewer

★ 7.5 · research

adversarial-claims-reviewer is a Claude Code skill that performs hostile verification of formal and quantitative claims in documents — math derivations, physics papers, statistical analyses, benchmark reports, and whitepapers. Every displayed equation and quantitative claim is assigned an ID (C1, C2, …) and classified into exactly one of four verdicts: VERIFIED, REFUTED, UNVERIFIABLE, or VACUOUS. Crucially, each claim is checked as named in the text — not a paraphrase or an adjacent statement — to prevent false claims from surviving review disguised as neighboring truths. Verification relies on deterministic methods first: SymPy/numpy scripts that exit nonzero on failure, numerical spot-checks at multiple parameter values, and dimensional analysis. The final report follows a structured template — verdict counts headline, the single most damaging finding leads, and per-claim justifications include script paths or explicit counterexamples. The skill is designed for anyone who needs to know whether the math in a paper, whitepaper, or benchmark report actually holds up.