Constraining LLMs Just Like Users

LLMs can't be trusted any more than the users interacting with them, which means they need the same kind of constraints developers already use for human input. Rather than piling more instructions into a system prompt hoping the AI behaves, developers can limit which output tokens the model can even choose from, making rule-breaking mathematically impossible. This same approach powers smarter interfaces like natural language date pickers, and it forms the backbone of safer tool-calling systems.