System 2 Reasoning Requirement for Safer RAG Against Knowledge-Poisoning

To mitigate misinformation in retrieved documents, this approach restricts untrusted document access to agents capable of deliberative System 2 reasoning. This prevents models from being influenced by incorrect information they are otherwise capable of detecting, reducing the computational overhead of strict isolation.