CEMAFROID is a company with a french delegated public service, delivering conformity attestations of refrigerated transport vehicles. It studies the ageing of those vehicles, depending on several physiochemical and mechanical factors, for which physical models have been proposed by researchers in thermal engineering. The DATAFRIG R database records more than 300 000 attestations of vehicles over 80 attributes, opening the opportunity to predict the ageing by building a numerical model using machine learning methods. During the development of such a model, several classical questions appeared, regarding the data quality, the field reality and the mistrust of domain experts. In this paper, we propose to use the notion of functional dependencies to address the aforementioned model's limitations. In particular, we investigate how FDs could help, especially using their counterexamples, that turn out to provide meaningful examples of such limitations, easily interpretable by domain experts. Interestingly, the existence of such counterexamples in the dataset is a way of demystifying the numerical model with the experts, by giving them back the control over their own data. This approach has been tested with domain experts from CE-MAFROID, with many positive feedbacks. It is worth noting that this attempt to better trust classification models is not limited to a particular application, and could be generalized to others.