It’s actually kinda easy. Neural networks are just weirder than usual logic gate circuits. You can program them just the same and insert explicit controlled logic and deterministic behavior. To somebody who don’t know the details of LLM training, they wouldn’t be able to tell much of a difference. It will be packaged as a bundle of node weights and work with the same interfaces and all.

The reason that doesn’t work well if you try to insert strict logic into a traditional LLM despite the node properties being well known is because of how intricately interwoven and mutually dependent all the different parts of the network is (that’s why it’s a LARGE language model). You can’t just arbitrarily edit anything or insert more nodes or replace logic, you don’t know what you might break. It’s easier to place inserted logic outside of the LLM network and train the model to interact with it (“tool use”).