Routing customer support tickets based on LLM sentiment or classification prompts leads to catastrophic nondeterminism during incident spikes.
We present a hybrid state-machine model where language models extract structured entities, but transition decisions are executed by a verified statechart.