Inlay

We then build a probabilistic simulator of human persuadability. It compares an unstructured LLM target, a structure-conditioned LLM target, and a Bayesian-network target with explicit latent belief-state updates each turn.