Science One Framework: Verifiable Autonomous Research
Google Research's Chain-of-Evidence architecture binds an autonomous research system's claims to retrieved sources, executed code, evaluator outputs, and reproducible scores as the work is produced.
Can chain-of-evidence guarantees extend from code-scored benchmarks to open-ended research where correctness depends on judgment, physical experiments, or contested literature?
Science One treats verifiability as an architectural constraint rather than a final proofreading step. Its Chain-of-Evidence model requires every important claim to point to supporting evidence and requires that the evidence genuinely support the claim.
The prototype grounds references through retrieved papers, explores candidate solutions in isolated branches, preserves evaluator outputs in a read-only record, and builds a structured claim map before rendering a manuscript. A companion CoE Audit independently reruns scores, checks task constraints, verifies references, and compares the written method with the executed code.
Reported evidence
Across 75 generated papers on five systems-optimization tasks, Google reports no phantom references for Science One, verified scores, and stronger method–code alignment than the tested baselines. The system also matched or exceeded human-expert performance on the five internal tasks and produced valid results on MLE-Bench and Parameter Golf. The authors label it an experimental prototype, not a production tool.
ASI relevance
Automated research compounds only when accepted improvements are real. Evidence captured during generation can reduce the chance that polished prose launders a broken experiment into the archive. The hard extension is from code-scored tasks to science whose evaluators are incomplete, expensive, or social.