Google Research presents Science One, an experimental autonomous research agent system built around Chain-of-Evidence (CoE), a verifiability protocol requiring that every claim in a generated research artifact carry a recorded evidence trail back to source data or code. The system is composed of a Problem Investigator that grounds literature review through the Semantic Scholar API, a Discovery Engine that explores candidate solutions across parallel branches, and a Paper Writer with a Claim Verifier that binds factual claims to workspace artifacts, audited via four checks covering score verification, specification violations, reference validation, and method-code alignment. Google reports the system achieved zero phantom references, versus baseline hallucination rates up to 21%, while maintaining state-of-the-art performance on benchmarks including MLE-Bench and Parameter-Golf.
