ResearchLoop: An Evidence-Gated Control Plane for AI-Assisted Research

A new technical report introduces ResearchLoop, an evidence-gated control plane designed to make AI-assisted computational research more auditable by treating every claim and piece of evidence as durable project state [1][2]. The system, described in a paper submitted on 27 May 2026, aims to address a growing tension in AI-assisted science: the same tools that accelerate ideation, implementation, evaluation, and writing also make it easier to state claims than to verify them [1][2]. ResearchLoop responds by formalizing the research lifecycle into a set of persistent, versioned objects. These include research questions, task contracts, evidence objects, claim ledgers, closeouts, and paper bindings, all realized as a repository-backed runtime [1][2]. The technical report provides the complete protocol specification, including the state model, transition rules, a claim-admission algorithm, and an insight-compounding mechanism [1][2]. The authors document an experimental record spanning nine distinct versions, from V0 through V9 [1][2]. That record includes a self-hosting case study, a controlled task-suite study with component ablations, a mathematical olympiad evaluation, and a supplementary SciCode boundary experiment evaluated with the official generated-code harness [1][2]. All artifacts, manifests, and verification reports are preserved in the project repository, the paper states [1][2]. By binding claims directly to evidence objects and maintaining a claim ledger, ResearchLoop creates a structure where the provenance of every assertion can be traced back through the system's state transitions [1][2]. The approach does not eliminate the risk of error, but it shifts the burden from post-hoc review to continuous, machine-readable auditability during the research process itself [1][2]. The work appears as a standalone technical report on arXiv, with no accompanying conference or journal publication noted at the time of submission [1].