For the people have made things on early access already, got 3 questions.

For the people have made things on early access already, got 3 questions. Can Jev score dynamically generated candidate sets well, or does it perform best when the option space is relatively stable and known ahead of time? For systems with evolving state over time, do you recommend treating Jev as stateless and keeping memory externally, or have you found better patterns for temporal decision-making? How does Jev behave out of distribution? When it’s unsure, are the confidence scores calibrated well enough to use as a routing signal to a larger model?