Toward a Brain-in-a-Vat Experiment for AI Agents: Questions, Formal Limits, and Experimental Design

This position paper formulates a limited experimental translation of the brain-in-a-vat sce nario for artificial agents. A test agent develops inside a generated world and receives its evidence through a controlled observation–action interface. The central object of study is the agent's construction, comparison, and testing of hypotheses about inaccessible generative condi tions. A policy-relative equivalence relation defines when external implementations produce the same accessible history laws. Exact equivalence fixes binary identification at chance under equal priors; approximate equivalence yields a total-variation ceiling. These results motivate a graded epistemic protocol and a factorial design over inherited priors, environmental traces, interven tion capacity, memory, and generator responsiveness. The proposal remains methodological and revisable. It supports experiments on meta-world reasoning while leaving consciousness, genuine epistemic birth, and metaphysical access unresolved.