Science One Framework: A verifiable autonomous research framework via Chain-of-Evidence

Google Research has introduced the Science One Framework, an autonomous research prototype designed to eliminate hallucinations by natively building verifiable evidence chains, along with the CoE Audit, an automated protocol to evaluate the integrity of AI-generated papers. The framework is based on Chain-of-Evidence (CoE), a new verifiability concept that requires every claim in a research artifact to have a recorded evidence chain (completeness) and each chain to genuinely support the claim (correctness). The Science One Framework instantiates CoE through three main modules: an Idea Explorer, a Problem Investigator, and a Solution Developer. The CoE Audit applies four strict integrity checks: reference verification, score verification, code reproducibility, and method-code alignment. In evaluations on 75 papers from the ADRS benchmark, Science One achieved zero phantom references compared to baseline hallucination rates up to 21%, perfect score verification, and the highest method-code alignment, while matching or exceeding human expert performance on all five ADRS tasks. The framework was also tested on six external complex tasks including MLE-Bench and Parameter-Golf, achieving state-of-the-art results.