Auto-SLR: A Fault-Tolerant Architecture for LLM-Assisted Systematic Literature Screening

Systematic Literature Reviews (SLRs) require high-recall screening processes that are increasingly difficult to maintain manually due to the growing volume of scientific data. While tools like Rayyan and ASReview assist humans, fully automated pipelines using Large Language Models (LLMs) often suffer from brittleness. They frequently fail silently due to API instability, duplicate ingestion, or parsing errors. This paper presents Auto-SLR, a fault-tolerant reference architecture designed to operationalize LLM screening. We combine standard engineering practices, such as MD5 hashing for binary deduplication and an Automated Recovery Loop for handling complex PDFs, to ensure dataset completeness. In a validation case study of 345 agriculture-related papers, the architecture achieved a 100% decision rate. It successfully recovered 5 papers that failed initial parsing and provided auditable natural-language rationales for all inclusion and exclusion decisions.