Airguard

AirGuard: AI is writing the code that drives robots next to airplanes. AirGuard proves it's safe before it runs. PROBLEM: Driverless baggage tractors already operate beside aircraft at Singapore Changi and Dubai DWC (trials at Schiphol and Stuttgart). Their software is updated often, increasingly by AI, and nobody independently verifies an update before it drives next to jet engines, pushbacks and ground crew. Reading the code is not enough: hazards only appear when the code runs next to a real aircraft. WHAT IT DOES: AirGuard is an independent safety gate. Every controller update is red-teamed in a digital twin of the real San Francisco airport (OpenStreetMap geometry) and judged against the FAA's own rules (AC 150/5210-20A). Unsafe updates are blocked, rolled back, and ops get a Twilio call and SMS. Safe updates are deployed. HOW: Agents propose, Python decides. - Meta Muse Spark writes the update from a plain-English request ("make the route 20% faster") and describes violation frames. - SubQ red-teams it: one long-context call reads the full FAA rulebook, the SFO map, the code and every previous 0.1 s simulation trace without summarising (up to 676k tokens; about 4M in total), then designs harder scenarios each round. - SubQ Snap reads about 370k tokens of raw traces in seconds to triage severity and filter unrealistic tests. - A deterministic Python judge makes every PASS/BLOCK decision. Verdicts are hashed and reproducible offline. - A self-improving loop grows the test library and adopts a new version only if it catches more unsafe code on held-out tests without more false blocks. RESULTS: - Meta's live "20% faster" update was BLOCKED: it raised fog speed over the limit and drove into a 777's jet blast. - SubQ found real bugs the hand-written tests missed: a tractor run over by a pushback, one stuck in jet blast when a beacon lit, a 3.47 m crew near-miss, and a flaw in the already-approved controller. - Mutation testing: unsafe code caught rose from 56% to 89% with SubQ's scenarios, with 0 false blocks. A diff-only AI review wrongly rated the safe fix unsafe. - The fixed controller passed 4 live SubQ rounds and was deployed.