Reputation mechanisms are being used to increase the reliability and performance of virtual societies. Different decentralized reputation models have been proposed based on interactions among agents. Each system agent evaluates and stores the reputations of the agents with whom they have interacted and can also testify to other agents about such reputations. The main disadvantages of such approaches when applied to open large-scale multi-agent systems are the difficulty of establishing strong links between the agents, the sometimes infeasible witness search process, the fact of the reputation is being influenced by the point of view of another agent and the fact that the agents cannot be willing to testify and collaborate with possible competitive agents. In this work we propose a hybrid reputation system with centralized and decentralized characteristics to overcome such problems. The reputations are provided by the system agents themselves but also by centralized subsystems that can be easily reached by any agent and can supply reliable reputations of any agent based on testimonies about undesired agent's behavior, characterized by the violation of system norms. Such centralized subsystems are instances of the proposed framework.