The error term isn't Pareto distributed

I wish this article featured actual data or a simulation instead of these napkin drawings. Does a set of uniform or gaussian inputs to this interaction model produce power-law output distribution? Is it just because some nodes on the graph are highly connected?

I have a vague theory that if you can't root cause a failure, the system is too complex, but haven't given it much thought.