Ask a model whether it answered from the retrieved document or from what it already knew and you get a confident one-word reply. This paper tests what that reply is actually computed from. Delete the document from the cache after the answer is written and the reply never changes — 0 flips in twelve model-and-edit combinations — while the model can still quote a replacement document word for word 100% of the time. Copy the cached question-and-answer state across from a run that saw a different document, changing no visible text at all, and the reply follows the donor in 100% of sixteen tests. The self-report is a read-off of a cached trace, not a check against the evidence in front of it.