Interesting, and looks like the present choices have evolved over time.

Seems like they are moving towards a structured RL implementation. There are elements of it, a follow-up post on some components would be interesting.