Citation-Enforced Retrieval-Augmented Generation for Classical Arabic Sources: Honest-Unverified Accounting in the Athar Research Assistant

Generative language models produce fluent scholarly prose that cannot be verified, which disqualifies them for research assistance over classical Arabic sources. This article reports Athar, a system in which every generated claim must resolve to a retrieved corpus passage or be flagged as unverified. A curated corpus of 323 classical Arabic works (7,723,880 lines), imported from OpenITI and audited for authenticity, is served through lexical BM25 retrieval at page granularity with segment-level citation verification. A deliberately conservative classifier separates uncited claims from methodological framing, and unverified counts are reported for each deliverable. In a single live production run, ten uncited segments resolved into nine framing segments and one genuine claim, which the classifier is built so that it cannot exempt. A dense re-ranking stage demotes wrong-sense homograph hits in the recorded regression fixtures but is disabled in production, because measurement on the deployment host showed the encoder needs roughly 9GB on a 16GB machine already holding a 7GB index. Marginal generation cost is US$0.05 to 0.075 per request. Existing work on hallucination in Islamic and Arabic content is framed as detection, correction, or benchmarking; what we report instead is an architecture that refuses to treat an unresolvable claim as grounded and carries its failure counts into the product surface.