A human equipped with pencil and lots of paper can do a good job, but that's an awful lot more powerful than just a human.

A professional scientist falls more under "human equipped with pencil and lots of paper". Nobody said second-level learning was easy.

So, you're thinking that human abstraction ability derives from probable morphisms rather than certain morphisms over weaker classes?

Maybe? I think that first what happens is that we identify a probable morphism, and then we adjust the two theories at hand in order to eliminate the uncertainty from the morphism by pushing it "upwards" (into the imprecision of the higher-level model) and "downwards" (into the parameter dimensionality of the lower-level model).