A relevant problem when dealing with community detection of real-world data is how to test the partition discovered by a model. Generative models allow to define a likelihood for each parameters' configuration, and usually one selects the one that maximizes it. But what if two configurations have slightly different values of likelihood but are associated to very different partitions? It seems unreasonable to simply rule out one of the two. To tackle this problem it would be relevant to analyze the entire energy landscape and derive phase diagrams to quantitatively measure how certain or uncertain the model predictions are.