Simulation is not a substitute for flight testing. It is the disciplined place to expose assumptions, integration errors, and unsafe operator workflows before hardware and people carry the cost.
Aviation AI should not be evaluated only by model accuracy. It must also be evaluated by context, uncertainty, operator workload, intervention timing, and the consequences of being wrong.
Multi-vehicle systems make coordination problems visible: interfaces, ownership, observability, deployment, recovery, and documentation become as important as the control logic itself.