In their original formulation, Bayes nets were a way to capture conditional independence properties of probabilistic models. That is, given any probabilistic model for P(Y|X1,X2,X2,...), there is a Bayes net that captures some of the conditional independence relations in your model. Bayes nets certainly cannot capture all possible conditional independence relations: undirected graphical models, for example, capture a different class of independence relations, while factor graphs capture a superset of the independence relations expressible by Bayes nets.

In this light, I'm not sure that your challenge makes sense. Bayes nets are a way of expressing a properties of probabilistic models, rather than a model unto themselves. "Bayes nets" alone is as meaningless a choice of model as "models expressible in Portuguese".

Perhaps a particular Bayes net together with a certain choice of conditional probability function for each arc and a certain choice of inference algorithm would constitute a model.