BARRED

BARRED: Evidence-Gated Security Fix Review AI coding assistants can produce convincing security justifications even when the underlying code remains vulnerable. BARRED addresses this trust gap by making security claims pass through deterministic evidence validation rather than accepting model-generated explanations at face value. Combining Meta Muse models with adversarial multi-agent reasoning and AST-based verification, BARRED extracts and validates structural relationships between security guards and potentially vulnerable operations. Every review produces one of three outcomes: PASS, when required structural evidence passes deterministic validation within the supported proof domain; BLOCKED, when a proposed security claim fails evidence validation; or UNRESOLVED, when the available evidence is insufficient to establish a reliable conclusion. The hackathon implementation adds deterministic AST candidate extraction, mathematical receipt validation, tri-state adjudication, and Meta Muse model integrations. BARRED makes the boundary between what an AI proposes and what deterministic verification can establish explicit.