IsnadGuard: Detecting Fabricated Chains of Narration in Hadith Transmission Networks

The isnad is the ordered chain of narrators through whom a hadith is transmitted. Hadith science has robust methodologies for evaluating reports based on narrator reliability and chain continuity. However, computational work has largely treated authenticity assessment as text classification over the report content, or matn. We study a focused task: detecting and localizing fabricated chains from narrator-transmission structure alone. Using Sanadset 650K, we construct a directed narrator graph and generate corruptions grounded in classical defect typologies: narrator substitution (tas-heef fi al-isnad) and chain splicing (idraj al-sanad). We propose IsnadGuard, a multiple-instance graph model that scores each transition using local statistics and narrator embeddings, aggregates transition scores with noisy-OR, and is trained jointly for chain classification, contrastive ranking, and edge localization. On 43,539 held-out chains, IsnadGuard improves AUROC from 0.727 to 0.836 and Hit@2 from 0.838 to 0.916 over an edge-frequency baseline.