@danluu that is the difference. Fuzzing is just sending in garbage to see that everything is handled. PBT is about checking invariants. So the library is really about having some ergonomics to shape complicated input domains so you can check invariants on (shaped(X), f(shaped(X))). Do this well, and the invariants are the specification for your application. An LLM can discover an implementation, which is not as feasible with just fuzz generation alone.