It is very, very hard to enforce behavior to an optimization system just with rewards / penalties and no explicit constraints. Which is why in manufacturing we

It is very, very hard to enforce behavior to an optimization system just with rewards / penalties and no explicit constraints. Which is why in manufacturing we use MPC, not RL (or use them within a system that can outright reject their recommendations if dangerous). There will be always cases that sacrificing one direction (operating rules) can improve the other one (profit).