How we build and why — methods, evaluations, and engineering notes in the open.

While building the Interaction Lite and Pro model paths, we found that efficiency depends less on always calling a stronger model and more on deciding when reasoning is actually worth spending.