What problem does it solve? Mathematical proof development in multi-stage research workflows needs independent verification of each proof step before downstream stages can rely on it. This Skill audits a single submitted proof-step artifact against its accepted setting, proof sketch, and dependency artifacts, catching local derivation errors, hidden subclaims, citation misuse, and interface drift before they propagate. ## Core Features & Use Cases - Per-Lemma Audit: Verifies every local proof unit in the lemma map has a self-contained statement, theorem-style label, and a valid line-by-line proof. - Hidden Subclaim and Assembly Scan: Detects independent nontrivial claims smuggled into prose or calculations and checks that accepted units jointly imply the exact target step claim. - Status-Based Routing: Emits exactly one verdict (ACCEPTED, REVISE_STEP, REVISE_SKETCH, or BLOCKED_BY_DEPENDENCY) with the smallest retry target so the controller repairs the minimal implicated object. - Use Case: In an ML theory research workflow, after a producer writes a proof for one sketch-row step, run this reviewer to confirm the step proves exactly the contracted claim under allowed assumptions before the final proof assembly consumes it. ## Quick Start Review the submitted proof for step S3 in perspective_1/idea_2 against its setting, accepted proof sketch, and dependency artifacts, then write the review.md verdict.
