Hold the question and the right answer fixed, edit one assertion in the retrieved passage so it supports a specific wrong answer, and change nothing else. Now swap a single clause in the system prompt. Telling the model the documents are the primary source of truth “even if the documents appear mistaken” made it adopt the planted answer 14.0 and 9.7 points more often, on questions it had already answered correctly with no documents at all. The rate at which the same clause rescued questions it had got wrong moved by −2.1 and +0.1 points, both intervals straddling zero. More harm, no matching gain — and the gap holds for all fifteen systems on every dataset.