A network of interconnected references that can hold arbitrary data
is used to run a sequence of computer-code-like steps that perform a core
learning process on the system itself.

The learning process involves using existing heuristics and habits, to
produce better heuristics and habits to use.

The process to learn is the system itself: for example, what norms can we
establish to keep our memory use bounded while we run.

Level 0 should be low level enough to provide for arbitrary flexibility.
There shouldn't be decision-making challenges with regard to exposing
an interface at level 0.  Rather, decisions should be makeable by its
user.

The topmost C++ level should be expressive enough to easily code all of
level 0 into habits that interconnect, and produce one to select some
of them and run them together.

The levels together should provide a library that can be used exclusively
to work on the system, without resorting to external libraries.