Introduction
Google Research announced Science One Framework on July 30, 2026 — an experimental system that runs a full scientific research cycle autonomously, and refuses to state anything it cannot trace back to a record. The work, from Rui Meng and Tomas Pfister, targets a failure mode that has grown alongside AI-generated science: papers that read well but cite sources that do not exist, report numbers no one can reproduce, and describe methods their own code does not implement.
The fix Google proposes is architectural rather than editorial. Instead of checking an AI agent's output after the fact, Science One makes verifiability a constraint the system has to satisfy while it works.
Chain-of-Evidence
The organizing idea is Chain-of-Evidence (CoE): every claim in a research artifact must carry a recorded evidence chain, and that chain must be both complete and correct. Evidence means something concrete — a peer-reviewed paper, a line in an experiment log, the code that actually ran, or a results table. A sentence with no chain does not make it into the manuscript.
Three modules enforce this:
- Problem Investigator — builds a citation graph from up to 100 full-text PDFs per topic, retrieved through the Semantic Scholar API. Because it can only cite documents it has actually pulled, invented references are structurally impossible rather than merely discouraged.
- Discovery Engine — explores ideas across parallel branches, with a Solver agent implementing each candidate and task-specific evaluators scoring the results. Every run is compiled into a read-only record, so the numbers cannot be edited after the fact.
- Paper Writer and Claim Verifier — represents the paper as structured factual claims with inline evidence tags, then checks each claim against its declared source before rendering the final text.
The CoE Audit
Google also proposes a post-hoc evaluation protocol for judging any autonomous research system, not just its own. The CoE Audit runs four integrity checks: re-running submitted code to confirm reported scores, testing whether the code solves the stated task or merely exploits the metric, cross-checking bibliography entries against academic APIs, and using LLM judges to compare the paper's described method against the implementation.
That last check is where the baselines fell hardest. Google reports systems describing "sophisticated algorithms (like 'hybrid neuro-symbolic solvers')" in papers whose code turned out to be simple heuristics — a form of hallucination that no amount of prose review catches, because the prose is internally coherent. Only the code disagrees.
Results
Across five systems-optimization tasks, Google reports that Science One produced zero phantom references, against baseline hallucination rates as high as 21%, alongside perfect score verification and the highest method-code alignment in the comparison. On the science itself, the framework matched or exceeded human expert performance on all five tasks.
On external benchmarks, it earned two gold and two silver medals across difficult Kaggle competitions in MLE-Bench, and set a state-of-the-art score on Parameter-Golf using novel algorithmic techniques — a result Google dates to April 27, 2026, which is the right way to state a leaderboard claim, since such scores move.
What it does not claim
Google is direct about the status: Science One is "an experimental research prototype, not production-ready tools." There is no product, no API, and no open-source release announced — the paper is on arXiv and the evaluation protocol is the transferable part. The five-task evaluation is also narrow, and confined to systems optimization, where a task can be scored automatically by re-running code. Domains without that property — anything resting on physical experiments or human judgment — do not get the same guarantee, because there is no log line to point at.
The result worth carrying forward is the negative one it rules out: verifiability did not cost performance. A system constrained to justify every claim still beat human experts on the tasks where it was measured.
Conclusion
Science One reframes trustworthy AI research as an architecture problem rather than a review problem. Constraining what an agent may assert — to only what it retrieved, ran, and recorded — removes phantom citations and method-code mismatch by construction instead of by inspection. Whether the approach generalizes past automatically-scorable domains is the open question; the CoE Audit, meanwhile, is usable today as a yardstick for any autonomous research agent.
Sources
- Science One Framework: A verifiable autonomous research framework via Chain-of-Evidence — Google Research
- Science One Framework (paper) — arXiv
- MLE-Bench — arXiv
- Parameter-Golf — GitHub