When tool output contradicts the prompt, models follow recency over stated source reliability.

A synthetic benchmark constructs conflicts where exactly one evidence source matches ground truth, independently varying modality, recency, stated reliability, and provenance. Across open-weight instruction-tuned models the arbitration is systematic: distinct text-versus-number preferences, temporal recency weighted more consistently than explicit reliability cues, and over-trust of external forecasts contradicting direct in-context evidence. arXiv An out-of-date but recent-looking tool response quietly outranks a correct fact you put in the prompt. Labeling your sources "authoritative" does less than you think.