This article offers a survey of three reliability theories in writing assessment: positivist, hermeneutic, and rhetorical. Drawing on an interdisciplinary investigation of the notion of "witnessing," this survey emphasizes the kinds of readers and readings each theory of reliability produces and the epistemological grounds on which it rests. Positivist reliability, positing an "a priori" reality and objective truth, produces "modest witnesses" (Haraway) who generate consistent, "true" scores. Hermeneutic reliability, which views "truth" as socially constructed though dialogue, produces communal witnesses who generate consensus, agreed upon readings. Rhetorical reliability, which views "truth" as a rhetorical effect of testimony, produces embodied witnesses who generate a range of perspectives, including articulated differences. While many writing assessment scholars, both in composition studies and in educational measurement, reject positivist reliability and embrace hermeneutic reliability, I suggest that the challenge for our field and our writing programs is to articulate a rhetorical theory of reliability and to design writing assessment systems that align with our prevailing conceptions of writing and reading as rhetorical activities.