Partial Identification for Mechanism-Uncertain User Simulation

Stateful LLM user simulators must specify how latent user state changes after an agent intervention. When data are concentrated under baseline policies, several cognitive--affective transition mechanisms can fit the same observations yet predict opposite effects for a new policy. We formulate this residual uncertainty as partial identification. Observationally compatible mechanisms induce a set of policy contrasts rather than a point ranking; for finite discrete trajectory summaries, we give a distribution-free outer confidence set and transfer its coverage to the policy contrast. For pathway-separable affine contrasts we derive exact endpoints and a transparent decomposition of interval width. A finite-state social-engineering construction shows a sign-changing interval for stabilization order, while delay plus independent verification remains positive within the declared class. The results are model-conditional: they do not establish human intervention effects without human interventional evidence and a defensible mechanism class.