Google Introduces Science One — A Framework for Verifiable AI Research

Google has unveiled Science One, an experimental research framework designed to eliminate hallucinated citations and unreproducible results in AI-driven science. The system is built around Chain of Evidence (CoE) — every claim, number, and algorithm description must trace back to a concrete source: a paper, a code fragment, an experiment log, or a results table.

The framework operates in three stages: it studies up to 100 full-text papers and builds a source graph; it simultaneously tests multiple hypotheses and implementations; and it generates the final paper while cross-checking each assertion against the original data. For auditing, Google developed CoE Audit, which reruns code, verifies citations, checks for task substitution, and compares method descriptions with actual implementations. In tests, Science One produced zero phantom references, while some existing research agents fabricated up to 21% of their sources. The system is currently an experimental prototype.

Science One Framework (Google Research)