NOVA ResearchOS: Evidence-Risk-Gated Collaboration and Validation-Gated Self-Evolution for Auditable Research Agents

Abstract

Abstract: Multi-agent research assistants can route tasks through specialized roles, but adding more agents does not by itself guarantee reliable scientific conclusions. We present NOVA ResearchOS, an implemented, event-sourced prototype that couples claim–evidence provenance with two distinct control loops: an intra-task decision that can summon a Reviewer after evidence has been retrieved, and an inter-task process that proposes policy changes from recorded failures and screens them before deployment. This is a draft version of the research paper (dated 20 September 2026). Code availability: https://github.com/qingfeng123456-cyber/NOVA_ResearchOS Keywords: multi-agent systems; scientific evidence; adaptive collaboration; self-evolution; auditability; evaluation gates.