On July 30, 2026, Google Research published the Science One Framework, an experimental prototype for autonomous scientific research written up by Rui Meng and Tomas Pfister. Its target is the specific failure mode that has dogged AI-written research: papers that read well and cite sources that do not exist, or that quote numbers the underlying experiment never produced. Science One attacks this by making every assertion carry a verifiable link back to evidence, an approach the authors call Chain-of-Evidence.
The system has three parts. A Problem Investigator retrieves full-text PDFs through the Semantic Scholar API, up to 100 per topic, and builds a citation graph, so that references come from retrieved documents rather than from model memory. A Discovery Engine runs parallel explore-exploit cycles in which a Solver agent proposes and tests solutions while a task-specific evaluator scores them, keeping read-only records of every output. A Paper Writer then drafts the manuscript, but a dedicated Claim Verifier checks each factual assertion against its declared evidence source before the paper is rendered.
The reported results are concrete. Science One produced zero phantom references, against baseline hallucination rates of up to 21 percent, and scored highest on method-code alignment, meaning the paper’s description matched the code that was actually run. On the Automated Design of Research Systems benchmark it matched or exceeded human expert performance on all five tasks. On MLE-Bench it earned two gold and two silver medals across five Kaggle competitions, and it reached state-of-the-art performance on Parameter-Golf.
For a technical business leader, the interesting move is architectural rather than scientific. Science One does not try to make the model hallucinate less; it makes hallucination structurally impossible to publish by requiring a verified source for every claim before the document is emitted. That pattern - a separate verifier gate on generated output, backed by retrieval records rather than model memory - transfers directly to regulated reporting, analyst research, and any workflow where an unsupported number is a liability.