Why this library exists and the bar every part of it holds to.
A computer scientist and an aerospace engineer should both be able to pick this up and solve a problem with it. Not because it is dumbed down, but because the surface is small and general: typed transforms, a graph that runs them, a trace you can read, evaluation, and feedback. The deep ideas (control theory, network flow, neural error/feedback) inform the design; they are never the price of admission.
If using a capability requires you to first learn the theory behind it, the capability is not finished.
The core stays domain-free on purpose. Cost, capacity, tokens, latency, reward: these are signals a caller records, not concepts baked into the primitive. A domain (payments, evals, infra, research) sits ON TOP as an example or a feature package; it never leaks into the core. This is what lets the same primitive serve unrelated fields without becoming any one field's tool.
There is an inverse relationship worth naming: the more domain knowledge someone has, the more likely they are to pass over a general tool they could have used, because they assume they need a specialist one. Deep expertise narrows the search. This library is built against that failure mode. It offers general capabilities that a specialist can reach for without leaving their domain, if they are willing to do a small investigation first.
We rely on two things from whoever uses this:
- They actually want to solve a problem (not collect a framework).
- They will do a small investigation to see how the primitive maps to their problem.
In return the library promises that the investigation is small and the value is real: a working graph, an inspectable trace, an honest comparison, a next step. The deliverable is always "I can see what happened and what to change," not a score.
Use deep domain knowledge to DESIGN the primitive. Expose simple capabilities the user understands immediately: run a graph, read the trace, compare two runs, find what to change. The theory is in the bones, not on the label.