An AI change ships only when its claims are proven by real sandboxed execution. LLM proposes, code decides.