An efficient evidence-based Automated Fact Checking system

Abstract

The rapid spread of fake news, accelerated by Generative AI, has outpaced traditional fact-checking, overwhelming journalists and verification platforms. Addressing this, we present Sanctuary, an efficient automated fact-checking system using moderately lightweight, open-source language models within a hybrid Retrieval-Augmented Generation framework that grounds its reasoning in retrieved evidence. Unlike approaches reliant on costly proprietary models or basic classifiers, Sanctuary delivers competitive accuracy and robust reasoning, verifying real-world claims in under 30 seconds. In the Fact Extraction and Verification 2025 competition, Sanctuary ranked 3rd, outperforming several systems and the baseline. We also introduce FactCellar, a dataset of real-world claims in realistic retrieval settings, enriched with source credibility and potential impact annotations. Experiments show these metadata substantially improve verification accuracy. Together, Sanctuary and FactCellar advance scalable, transparent fact-checking, offering professionals and everyday users practical tools to counter misinformation.