In a blog post published on the official Google Cloud blog by Research Scientist Rui Meng and Director Tomas Pfister, researchers introduced the Science One Framework. This is an experimental research prototype designed to completely eliminate the phenomenon of hallucinations in scientific research conducted autonomously by artificial intelligence, through the structured and natural construction of verifiable chains of evidence. Alongside this, the researchers introduced CoE Audit, an automated protocol for evaluating the integrity and reliability of AI-generated scientific papers against their source code and supporting evidence.
The Structural Challenge: The Reliability and Verifiability Problem in AI-Driven Research
Large language models (LLMs) are currently deployed not only as code-writing assistants but as autonomous agents capable of managing end-to-end scientific research workflows. Recently unveiled systems, such as Sakana AI's AI-Scientist (version v2), as well as AutoResearchClaw, DeepScientist, and AI-Researcher, are capable of reviewing scientific literature, formulating research hypotheses, executing actual experiments, and writing entire papers that present a quality similar to papers written by humans.
However, alongside the improvement in the superficial textual quality of these papers, a critical structural problem has emerged: verifiability. Because existing autonomous research pipelines generate text iteratively, errors introduced at any stage tend to replicate and amplify in subsequent stages. Many existing systems generate bibliographical references to sources that do not exist at all, present a clear discrepancy between the described work methodologies and the code actually executed, and report experimental results and scores that cannot be reproduced by running the attached code. The researchers propose a solution to this problem by introducing a new framework for verifying AI-based research called Chain-of-Evidence (CoE).
Chain-of-Evidence (CoE): A New Concept for Reliability in Research
The Chain-of-Evidence (CoE) is a conceptual framework that defines what makes a research artifact trustworthy, similar to how ACID principles define the reliability of database transactions. Instead of dictating how to build the research agent itself, the CoE framework defines the properties its outputs must satisfy. It is based on a single principle with two parts: every claim in a research artifact must carry a documented chain of evidence (completeness), and each chain must genuinely support the claim to which it is attached (correctness).
A claim can be a bibliographical reference, a reported number, a methodological description, or a conclusion, and it must link back to the appropriate evidence – such as a peer-reviewed paper, an experimental log line, the code that actually ran, or a results table. A hallucinated reference points to a paper that does not exist; an unreproducible score does not reappear when the code is run again; and an incorrectly described methodology presents a certain algorithm in the paper while the actual code implements another algorithm. In all these cases, the chain linking the claim to the evidence is broken. The CoE Audit protocol allows these breaks to be measured accurately.
Science One Framework Architecture
To prove that verifiable AI research is possible without compromising on performance and problem-solving capabilities, the researchers designed the Science One Framework. Unlike previous agents that generate a paper and try to link it to facts retroactively, this system implements CoE principles by construction through three core components:
-
The Problem Investigator (literature grounding): To prevent fictitious references, the system builds a citation graph using the Semantic Scholar API. It reads up to 100 full-text PDF papers for each topic and generates a structured research brief. Every reference in the final paper originates from this grounded API-based call, thereby completely avoiding reliance on model memory.
-
The Discovery Engine (parallel exploration and exploitation): The system systematically explores and exploits ideas across multiple parallel branches. In each isolated cycle, a Solver agent implements a solution and a task-specific evaluator assigns it a score. High-performing branches are iteratively refined, and all raw evaluator outputs are collected into a strict, read-only record.
-
The Paper Writer and Claim Verifier: Before generating the final paper, the system builds a structured representation of every factual claim with an inline evidence tag linking it to a specific item in the workspace. A dedicated Claim Verifier component checks each claim against its declared source of evidence. Claims that exceed the existing evidence are reconciled and reformulated conservatively, rather than deleted, in order to maintain complete alignment with what the work actually supports.
CoE Audit Protocol: The Four Integrity Tests
The researchers developed CoE Audit – a post-hoc evaluation protocol that serves as an automated forensic reviewer and runs four strict reliability tests on the generated artifacts (the paper, solution, code, and references):
- Score verification: Extracting the reported score in the paper and comparing it to a completely independent re-run of the submitted code.
- Specification violation: Checking the solution code to ensure it actually solves the task and does not exploit loopholes in evaluation metrics or read ground-truth answer files.
- Reference verification: Cross-checking every bibliographical entry against academic APIs to identify fictitious references.
- Method-code alignment: Using LLM judges to compare the methods section of the paper side-by-side with the actual code, to ensure that the text faithfully describes the implemented algorithm.
Test Results and Performance Achievements
The researchers applied the CoE Audit protocol to 75 papers produced by five different systems across five system-optimization tasks (Prism, Cloudcast, EPLB, LLM-SQL, and transaction scheduling) from the ADRS (Automated Design of Research Systems) benchmark. Under this rigorous protocol, the Science One Framework led in all four integrity checks: None of its references were fictitious (0% hallucinations compared to hallucination rates of up to 21% in baseline systems), it achieved perfect score verification, and the highest method-code alignment. In contrast, baseline systems often described complex algorithms in the paper (such as "hybrid neuro-symbolic solutions") while the code they actually submitted contained only simple, deterministic heuristics.
The system proved that implementing strict reliability constraints does not harm its scientific capabilities. Science One Framework matched or surpassed human expert performance on all five ADRS tasks, achieving the best overall score among all systems on two of them (Cloudcast and EPLB).
Additionally, the system was tested on six highly complex external tasks:
- MLE-Bench: Across five difficult Kaggle competitions (medical imaging, fine-grained recognition, and 3D perception), the system achieved two Gold Medals (including a winning score on a 3D object detection task, where baseline systems failed completely) and two Silver Medals.
- Parameter-Golf: In a competition to train live language models under strict hardware and file-size constraints, while baseline systems failed to submit a valid solution, Science One Framework met all constraints and achieved an industry-leading score (as of April 27, 2026). The system discovered genuine and innovative algorithmic techniques rather than just superficial hyperparameter tuning.
Summary and Limitations Statement
As autonomous research systems scale to solve more complex scientific problems, solution quality alone will no longer be enough to differentiate them. The dividing factor will be the degree of trust that can be placed in the research outputs. The Science One Framework project demonstrates that verifiability must be treated as a first-class architectural constraint.
The researchers expressed gratitude to several partners who contributed to this work: Bhavana Dalvi Mishra, Jiefeng Chen, Chun-Liang Li, Palash Goyal, Mihir Parmar, Yiwen Song, Yale Song, Raj Sinha, Parthasarathy Ranganathan, Burak Gokturk, and Jinsung Yoon.
In conclusion, the authors emphasize that Science One Framework is defined as an experimental research prototype only, and does not constitute a commercial or production-ready tool.