When a reasoning system encounters a contradiction like p&q=>false, it may try to eliminate it by retracting some belief that supports either p or q. This paper addresses the issue of how to determine which of the possible supporting beliefs to retract. The problem is studied in the context of a heterogenous distributed environment in which we can not make assumptions about how agents derive their beliefs. The presented model considers two properties of beliefs: the authority of the agents that provided the belief in the first place and the costs incurred upon other agents that already used the belief (directly or indirectly) for their own decision making or action, if the belief is to be retracted. The model is implemented in an information infrastructure we have created for enterprise integration.