The pipeline problem: false claims are easy to train on

Language models do not distinguish truth from repetition, polish, or citation style on their own. Preventing fabricated statistics requires tracing claims across collection, training, retrieval, evaluation, and serving.