This article is a focused examination of two Russian question particles, "razve" and "neuzheli". Previous literature characterizes them as markers of negative epistemic bias and positive contextual bias, roughly corresponding to English "really". Zooming in on subtle differences in their conventional meaning and conversational dynamics, I propose a novel conceptualization of negative bias in questions as a linguistic phenomenon sensitive to non-monotonic belief revision and make a case for examining contextual bias through the prism of research on evidence in language. I further show that this view makes correct empirical predictions not captured by the existing typology.