Science One Framework: A verifiable autonomous research framework via Chain

July 30, 2026
Rui Meng, Research Scientist, and Tomas Pfister, Director, Google Cloud
Introducing the Science One Framework, an experimental research prototype designed to eliminate hallucinations by natively building verifiable evidence chains, and CoE Audit, an automated protocol to evaluate the integrity of AI-generated papers.
Large language models (LLMs) are increasingly being deployed not just as coding assistants but as autonomous agents capable of conducting end-to-end scientific research workflows. Recent systems (e.g., Sakana’s AI-Scientist, AutoResearchClaw, DeepScientist, AI-Researcher) can review literature, formulate hypotheses, execute experiments and write complete manuscripts that are comparable to human-authored papers. However, as the surface-level quality of these AI-generated manuscripts improves, a critical structural problem has emerged: verifiability. Because current autonomous research pipelines generate text iteratively, errors introduced at any stage are amplified. Some existing systems can generate non-existent citations, exhibit misalignments between the described methods and the actual code, and report experimental scores that are not fully reproducible from the provided code.
In our paper, we tackle this problem by introducing Chain-of-Evidence (CoE), a new verifiability framework for AI-driven research. We instantiate CoE with the Science One Framework, an autonomous research prototype that natively builds and maintains evidence chains, and the CoE Audit, a set of automated evaluation metrics that measures the integrity of AI-generated papers against their underlying code and evidence. Our results show that baseline systems hallucinate up to 21% of their references and frequently misalign their code and text, whereas the Science One Framework achieves zero phantom references and fully verifiable scores while achieving state-of-the-art performance on frontier benchmarks like MLE-Bench and Parameter-Golf.
The CoE is a conceptual framework that defines what makes a research artifact trustworthy, much as ACID defines what makes a database transaction reliable. Rather than prescribing how to build a research agent, the framework specifies the properties its outputs must have. It follows a single principle with two halves: every claim in a research artifact must carry a recorded evidence chain (completeness), and each chain must genuinely support the claim it is attached to (correctness). A claim may be a reference, a reported number, a method description, or a conclusion, that must link back to corresponding evidence, such as a peer-reviewed paper, an experimental log line, the code that actually ran, or the results table.
A hallucinated reference points to a paper that does not exist. An unreproducible score does not reappear when the code is re-run. A misdescribed method claims one algorithm in the paper while the code implements another. Each is a claim whose chain back to its evidence is broken; the CoE Audit makes these breaks measurable.
To demonstrate that verifiable AI research is possible without sacrificing problem-solving performance, we designed the Science One Framework. Unlike previous agents that generate a paper and try to link facts retroactively, the Science One Framework instantiates the CoE framework by construction through three main modules:
To rigorously evaluate the Science One Framework prototype against state-of-the-art baselines (e.g., Sakana AI's AI Scientist v2, AutoResearchClaw, DeepScientist, AI-Researcher), we developed the CoE Audit. This post-hoc evaluation protocol acts as an automated forensic reviewer, running four strict integrity checks on the generated artifacts (paper, solution, code and references):
We applied the CoE Audit to 75 papers generated across five systems-optimization tasks (Prism, Cloudcast, EPLB, LLM-SQL, and transaction scheduling) from the Automated Design of Research Systems (ADRS) benchmark.
The Science One Framework significantly outperformed existing baselines in verifiability. The CoE Audit applies the same independent protocol to every system, re-checking each reference against live scholarly databases, and under it, Science One Framework led on all four integrity checks. None of its references were phantom: every one pointed to a real, retrievable paper, compared to hallucination rates as high as 21% in baselines, because the Problem Investigator retrieves every reference rather than generating it from memory. It also achieved perfect score verification and the highest method-code alignment. In contrast, baseline systems frequently described sophisticated algorithms (like "hybrid neuro-symbolic solvers") when their submitted code was a simple, deterministic heuristic.
Crucially, implementing strict verifiability did not compromise the agent's scientific capabilities. The Science One Framework matched or exceeded human expert performance on all five ADRS tasks, achieving the best overall score among all systems on two of them (Cloudcast and EPLB).
To test its generalization, we deployed the Science One Framework on six highly complex external tasks:
As autonomous research systems scale to solve increasingly difficult scientific problems, solver quality alone will no longer be enough to differentiate them. What will separate their outputs is whether the resulting research can be trusted. Our findings demonstrate that verifiability must be treated as a first-class architectural constraint. By building evidence chains at the time a claim is produced rather than attempting to reconstruct grounding after the fact, The Science One Framework demonstrates that AI agents can produce rigorous, trustworthy and highly competitive scientific research. We hope the Chain-of-Evidence framework and its audit will serve as valuable tools for the community as we continue to build the next generation of AI scientists.
We would like to thank 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 for their valuable contributions to this work.
The Science One Framework is an experimental research prototype, not production-ready tools.

Facts Only

* Rui Meng and Tomas Pfister introduced the Science One Framework and CoE Audit on July 30, 2026.
* The issue addressed is the verifiability of AI-generated research pipelines where iterative text generation amplifies errors.
* Chain-of-Evidence (CoE) is a framework defining trustworthy research artifacts, requiring every claim to have a recorded evidence chain for completeness and correctness.
* The Science One Framework instantiates CoE by building evidence chains natively through three modules.
* The CoE Audit is an automated protocol that evaluates the integrity of AI-generated papers against their code and evidence.
* Baseline systems exhibited up to 21% hallucination rates for references and exhibited misalignment between code and text.
* The Science One Framework achieved zero phantom references and perfect score verification.
* The framework maintained state-of-the-art performance on MLE-Bench and Parameter-Golf benchmarks.
* The framework achieved the highest method-code alignment.
* The prototype was evaluated across 75 papers from five systems-optimization tasks in the ADRS benchmark.
* The Science One Framework matched or exceeded human expert performance on all five ADRS tasks.
* The system was tested on six complex external tasks.

Executive Summary

A new framework, Chain-of-Evidence (CoE), and an evaluation protocol, CoE Audit, are introduced to address the verifiability of research generated by Large Language Models (LLMs) acting as autonomous research agents. The core problem stems from existing AI systems generating iterative text where errors amplify, leading to hallucinations such as non-existent citations or misalignment between described methods and actual code. The Science One Framework instantiates CoE by design, creating evidence chains natively within the research process, and includes CoE Audit to automatically evaluate the integrity of AI-generated papers against their underlying code and evidence. Experimental results show that while baseline systems hallucinate up to 21% of references, the Science One Framework eliminates phantom references and achieves perfect score verification while maintaining state-of-the-art performance on benchmarks. The framework posits that trustworthiness requires every claim in a research artifact to be linked to verifiable evidence, ensuring completeness and correctness across all claims.

Full Take

The development of the Chain-of-Evidence framework suggests a necessary shift from post-hoc fact-checking to intrinsically verifiable generation in AI research, treating verifiability as an architectural constraint rather than an afterthought. The pattern where sophisticated capabilities (state-of-the-art performance) can be achieved alongside structural integrity (verifiability) implies that the barrier to deploying autonomous agents in high-stakes scientific domains is often not computational power but epistemology—knowing what constitutes reliable knowledge.
The contrast between baseline hallucination rates and the framework's zero phantom references highlights a fundamental gap: current AI research pipelines prioritize generative fluency over grounding. The failure mode observed where systems describe complex methods unsupported by the actual code points to a systemic weakness in symbolic reasoning within LLMs when operating autonomously. This suggests that moving from mere pattern matching to genuine scientific agency requires encoding logical constraints into the system architecture itself, mirroring established principles like ACID for database reliability.
The implication is that future AI science cannot be assessed solely on output quality; it must incorporate an audit layer that enforces accountability at the point of knowledge creation. The framework demands a paradigm change where the focus shifts from maximizing superficial performance metrics to guaranteeing verifiable provenance. This forces a reckoning with what we value in scientific progress: speed and fluency, or truth and traceability.
Bridge questions: If verifiability is treated as a first-class constraint, how can this framework be generalized across modalities beyond text-based research artifacts? What new metrics are necessary to evaluate the "correctness" of complex methodological chains that involve multi-step experimental reasoning rather than simple reference checks? What are the long-term risks if evaluation focuses only on internal evidence versus external, real-world scientific validation?

Sentinel — Human

Confidence

This text exhibits the highly structured, technical complexity and focused argumentation consistent with human-authored academic research communication, focusing on defining a novel framework for AI verifiability.

Signals Detected

Moderate sentence length variance; exhibits clear academic, dense prose typical of a research paper abstract/introduction.

High thematic focus with logical progression from problem identification to proposed solution and results; demonstrates expert-level synthesis.

Strong internal logic linking the framework (CoE) to the prototype (Science One) and the evaluation protocol (CoE Audit); no obvious template matching.

Specific, technical claims regarding hallucination rates (21%) and performance metrics are presented in a highly structured, empirical manner, suggesting grounding in specific experimental results.

Human Indicators

The text employs complex, layered argumentation where concepts (CoE, ACID) are explicitly defined and then applied to a novel system prototype. The voice shifts appropriately between conceptual framing and quantitative reporting, characteristic of academic writing.

The conclusion deliberately frames the work as an 'experimental research prototype' rather than a finished product, demonstrating nuance typical of scientific communication.