ReviewForge: An Automated Architecture for Literature Retrieval, Screening, and Review-Ready Evidence Synthesis

Systematic reviews, scoping reviews, and other evidence syntheses are essential for cumulative science and evidence-based practice, but they are also labour-intensive and time-consuming. Even with reporting guidelines such as PRISMA 2020, the practical burden of multi-database searching, deduplication, screening, data extraction, and documentation remains a major bottleneck for researchers and review teams. In response, an ecosystem of computer-assisted tools has emerged, yet most tools address only one or two stages of the workflow, provide limited cross-source automation, or lack transparent provenance for review-ready outputs. This paper presents ReviewForge, a modular architecture for automating the end-to-end literature workflow from search to review-ready evidence tables. ReviewForge is designed to: (i) encode review protocols, search strategies, eligibility criteria, and extraction templates in a single declarative configuration file; (ii) orchestrate multi-source retrieval and import; (iii) perform robust deduplication and normalisation; (iv) support semi-automated title/abstract screening via rule-based filters and machine learning; (v) manage full-text acquisition and eligibility decisions; and (vi) generate structured outputs that can be pasted directly into narrative reviews or meta-analytic scripts. All processing steps are logged with complete provenance to support transparency and reproducibility in line with contemporary guidance on automation in evidence synthesis. Rather than evaluating specific machine-learning models, the present work focuses on the architectural design, module interfaces, and configuration semantics of ReviewForge. The goal is to provide a reference specification that review teams can implement, adapt, or extend, and to clarify how automation can be integrated into rigorous review conduct without obscuring decision paths or introducing undocumented bias.