The paper studies how on-line inferring stability of a potential control-loop consisting of an uncertain plant interconnected in feedback with a candidate controller using plant 110 pairs recorded while the plant is possibly driven by a different controller. In such a context, a convenient tool to work with is to resort to the conceptual entity of a virtual reference (VR). The adopted approach consists of embedding, in the so-called unfalsified adaptive switching control schemes based on VR, a family of nominal models pairwise associated with the given candidate controllers. The result is that the supervised switching mechanism can moderate the chance that destabilizing controllers be switched-on and, hence, reduce both the magnitude and time durations of "learning" transients after start-up, while, in contrast with preexisting multi-model based methods, stability in-the-large is guaranteed under the minimal conceivable assumption that a stabilizing candidate controller exist. (C) 2009 Elsevier Ltd. All rights reserved.