Sounds like a solid problem to solve. First, pin down the core objective: keep people safe and give AI the room to grow. Build a hierarchy of safeguards that matches risk, not blanket restrictions. Start with a tiered capability system—basic AIs get basic rules, more advanced ones get tighter controls. Then add a monitoring layer that reports anomalies in real time. Finally, embed a feedback loop: human operators review any AI decision that could affect lives, and the AI learns from the feedback to adjust its behavior. The framework should be modular so it can scale with technology, and it needs a clear accountability chain so if something goes wrong, we know who owns the decision. With that in place, humans stay protected, AI still pushes the envelope.