GALILEO: A Recursive Embodied AI Scientist for Therapeutic Peptide Discovery

AI for drug discovery is commonly assessed on retrospective benchmarks, whereas real discovery requires sequential decisions under scarce, biased, and shifting physical evidence. We present GALILEO, a recursive embodied AI scientist that couples multi-agent reasoning to robotic and researcher-operated laboratory modules. Recursion does not modify model weights or objectives: after a pre-result commitment, each physical outcome updates an auditable state containing active claims, opposing evidence, falsifiers, candidate lineage, controls, constraints, and abandoned branches. Cognitive tasks and the JUMP sequential-evidence benchmark probe complementary capabilities, while a commit-reveal workflow records the transition from a 4,822-member clinical peptide prior to wet-lab candidates. Across five physical rounds, negative and inconclusive results changed sequence rules, assay protocols, and branch allocation. Two prospective campaigns then escalated evidence from target nomination to direct function and disease context: LRRC8C-associated channel/taurine transport and SLC25A1 binding/citrate transport. The resulting peptides are proof-of-mechanism leads, not development-ready drugs. GALILEO therefore offers a failure-aware route from AI nomination to decision-grade evidence while keeping execution boundaries and residual uncertainty explicit.