What problem does it solve? Mathematical proofs assembled from multi-stage workflows can contain hidden gaps, unsupported scope upgrades, or counterexample-vulnerable claims that standard reviews miss. This Skill provides a final adversarial review pass that actively tries to break the weakest claims before a theorem candidate is accepted. ## Core Features & Use Cases - Adversarial Attack Surface Construction: Identifies the weakest theorem-level claims, assembly links, and step-level local units, then prioritizes obligations that would invalidate the theorem if false. - Stress-Case Testing: Tests claims against boundary regimes, degenerate cases, parameter extremes, and assumption-minimal examples, tracking convergence modes, explicit rate dependence, and notation provenance. - Repair Depth Localization: Classifies attacks as verified breaks, candidate counterexamples, or tentative concerns, and routes each blocking issue to the shallowest repair target (assembly, step, sketch, or idea). - Use Case: After assembling a LaTeX proof bundle from accepted proof steps, run this reviewer to verify the main theorem is self-contained, that no step exported an unproved invariant, and that simplified corollaries do not hide dropped dependencies. ## Quick Start Run the adversarial proof reviewer on the assembled LaTeX bundle and accepted step artifacts for perspective 1 idea 2 to produce proof_review_adversarial.md with a PASS or BLOCKING verdict.
