I’ve been working on AIR (AirLang), a verifier-first programming language and toolchain for small, bounded components, including code generated by AI agents. The underlying problem is simple: A model’s ability to generate an instruction is not evidence…