In an ad hoc network, trust systems usually rely on both local and remote evidence in order to build a trust relationship. Remote evidences are usually gathered and concatenated with each other so as to collaboratively infer a trust relation. However, the ever-changing topology and the high versatility of wireless links (coming from e.g., presence of interferences and noise) imply that evidence is partial. Hence the resulting trust relation varies from time to time, depending on e.g., the connectivity. To tackle this issue, we exploit the notion of confidence interval that corresponds to an interval estimate of a parameter (e.g., mean, standard deviation) that characterizes the evidence population. We herein consider this interval to indicate the reliability of the trust estimate and inflect the decision making of an intrusion detection system. More particularly, this detection system detects attacks threatening a routing protocol. This detector distinguishes itself by adopting a mechanism of punishment based on the level of harm produced by the intruder. In a nutshell, this harmfulness is measured by the number of misuse goals (e.g., route disruption, exhausting resources) that could be realized as a result of the intrusion. Performance evaluations of our intrusion detector along with the confidence measure have been also conducted.