Malicious manipulation of jobs results endangers the efficiency and performance of grid computing applications. The presence of nodes interested in depreciating jobs results may be detected and minimized with the usage of fault tolerance techniques. In order to detect these kind of nodes, this work presents a distributed and hierarchical diagnosis model based on comparison and reputation, which can be applied to both public and private grids. This strategy defines the status of a node according to its level of confidence, measured through its behavior. The proposed model was submitted to simulations to evaluate its effectiveness under different quota of malicious nodes. The results reveals that 8 test rounds can detect practically all malicious nodes and even with less rounds, the correctness remains high without a significant overhead increase.