An auditable evidence compiler for large language model-assisted systematic reviews

The article presents a framework for linking large language model (LLM)-assisted evidence processing to a versioned and auditable synthesis output. The research team used LLM agents to interpret sources and extract data, with deterministic code enforcing statistical rules and researchers resolving ambiguities. In a retrospective analysis of one registered prognostic review, 50 root-cause errors were documented, of which 12 changed results before correction. Forty-six errors were resolved and four remained as disclosed limitations. The corpus contained 454 reports, 445 studies, 441 cohort entities and 421 dependence clusters. All 39 source records reached a terminal state and two implementations agreed in 1,217 numerical comparisons. The framework enables inspection of synthesized evidence together with its provenance, statistical meaning and correction history.