Speaking tasks involving peer-to-peer candidate interaction are increasingly being incorporated into language proficiency assessments, in both large-scale international testing contexts, and in smaller-scale, for example course-related, ones. This growth in the popularity and use of paired and group orals has stimulated research, particularly into the types of discourse produced and the possible impact of candidate background factors on performance. However, despite the fact that the strongest argument for the validity of peer-to-peer assessment lies in the claim that such tasks allow for the assessment of a broader range of interactional skills than the more traditional interview-format tests do, there is surprisingly little research into the "judgments" that are made of such performances. The fact that raters, and rating criteria, are in a crucial mediating position between output and outcomes, warrants investigation into how raters construe the interaction in these tasks. Such investigations have the potential to inform the development of interaction-based rating scales and ensure that validity claims are moved beyond the content level to the construct level. This paper reports the findings of a verbal protocol study of teacher-raters viewing the paired test discourse of 17 beginner dyads in a university-based Spanish as a foreign language course. The findings indicate that the raters identified three interaction parameters: non-verbal interpersonal communication, interactive listening, and interactional management. The findings have implications for our understanding of the construct of effective interaction in paired candidate speaking tests, and for the development of appropriate rating scales. (Contains 2 figures and 3 footnotes.)