Large language models (LLMs) are advancing toward autonomous scientific research, yet verifiability remains a critical issue. Existing systems often introduce errors, generate non-existent citations, and exhibit misalignments between methods and code, leading to unreproducible results.

To address this, Google AI introduces the Chain-of-Evidence (Co E) framework, instantiated by the Science One Framework. This prototype natively builds and maintains evidence chains, ensuring every claim is supported by verifiable data. The Co E Audit provides automated metrics to measure the integrity of AI-generated papers against their underlying code and evidence.

Results show baseline systems hallucinate up to 21% of references, while Science One achieves zero phantom references and fully verifiable scores. It also matches or exceeds human expert performance on scientific tasks, demonstrating that verifiability can be integrated without sacrificing capability. The framework's modules ensure grounded references, systematic exploration, and claim verification, making AI-driven research more trustworthy.