Introduced by Gottweis et al. (Nature 2026) as a hypothesis-ranking system, Co-Scientist was extended in August 2026 (arXiv 2608.26701) by Google DeepMind with Duke, Columbia, and Texas A&M into an execution-grounded system: separate agents handle evolutionary ideation, three-phase experiment execution scored by an LLM reward model, and manuscript writing audited against raw lab logs. Reliability modules perform Hallucination Clipping and route work through a two-layer research-safety gateway.