Robustness by Design: Interface Contracts for AI Control in High-Stakes Physical Systems

Fusion power plants will demand robustness and safety beyond today’s research machines. Data-driven methods promise new capabilities but introduce risks in plasma control. We argue that explicit interface contracts between physics-driven, data-driven, and heuristic modules are essential. Such contracts define domains of applicability, accountability, and fallback mechanisms, ensuring reliability in heterogeneous control systems. Using a data-driven confinement state identification module as illustration, we argue this principle extends to all high-stakes physical domains: robust and safe AI control depends as much on interface design as on algorithms and models.