We present Prometheus, a domain-agnostic scientific reasoning engine based on Active Inference and the Free Energy Principle. Given a likelihood model mapping hidden mechanisms to observable experimental outcomes, Prometheus autonomously selects maximally informative experiments, performs exact Bayesian belief updating, and converges to correct mechanism identification with statistically significant efficiency gains over uninformed baselines. We validate the architecture across four scientifically distinct domains — reaction mechanism classification in organic chemistry, antimicrobial resistance gene identification in clinical microbiology, crystal structure determination in materials science, and a canonical coin identification task — using an identical, unmodified engine instance (verified by cryptographic hash).
Chouhan et al. (Sun,) studied this question.