Have you looked a bit at replicode?: http://wiki.humanobs.org/public:replicode:replicode-main Or dylan?: http://opendylan.org/index.html Maybe you can grab an idea or two from them as well.

How does this compare? What makes it specially useful for reasoning newer ideas. I've read some of the introduction about it using the PAC model, but can you give an example of how it goes about using that?

I like the idea of the language, I think it follows with Global Workspace Theory: https://en.wikipedia.org/wiki/Global_Workspace_Theory and Global Neuronal Workspace: https://en.wikipedia.org/wiki/Dehaene%E2%80%93Changeux_model

Say you compile 2 machine learning algorithm implementations. If the reverse compiler gives you an interface with a choice between say those 2 algorithms and a % accuracy rate with a visual representation this definitely sounds like a working idea of a global workspace theory. As it abstracts the complicated parts into simplistic choices that are selectable to be do-able in linear time with some statistics.