This program tree shows an actual program generated by a test run on the tricky network that has been slightly simplified for human comprehension. This program takes the first approach of learning topography implicitly through evaluation. This has the disadvantage of being specific to the particular network used to evaluate it.
A potential way to generate programs which make use of the second, more general method is to evaluate with several networks simultaneously, summing the fitness for all of them. The toolkit has such capability, but time constraints did not allow us to explore this avenue vigorously.