To make their decisions, autonomous vehicles need to build a reliable representation of their environment. In the presence of sensors that are redundant, but not necessarily equivalent, that may get unreliable, unavailable or faulty, or that may get attacked, it is of fundamental importance to assess the plausibility of each information at hand. To this end, we propose a model that combines four criteria (relevance, trust, freshness and consistency) in order to assess the confidence in the value of a feature, and to select the values that are most plausible. We show that it enables to handle various difficult situations (attacks, failures, etc.), by maintaining a coherent scene at any time despite possibly major defects.