Schwarz turns SMT failures into local repair tasks and verifies 91.5% of SV-COMP ReachSafety against CPAchecker's 60.1%

arXiv·medium signalXBlueskyLinkedInCopy link

Agentic verification loops usually see only a coarse verifier error, timeout or unknown, so the model cannot tell whether the spec is wrong, a lemma is missing, or the obligation needs a different theory view. Schwarz exposes program-point snapshots of checked facts, lets the agent propose local lemmas, and applies theory-aware solver policies for numeric, quantified, memory and floating-point obligations. Evaluated on 1,475 tasks for C and Rust/Verus, it solves 95.2% of 475 recent agentic-verification benchmarks and 91.5% of 1,000 SV-COMP 2026 ReachSafety tasks averaging 1,427 lines, versus 60.1% for CPAchecker.