Great to see Jev in ReqLLM. I've been playing with the model from the other end and ended up with a small GenServer-shaped client:
github.com/dannote/jev
Jev answers typed questions with probabilities, never text. On the BEAM that's just a message. So you reply to Jev from a callback, its answer arrives at 𝚑𝚊𝚗𝚍𝚕𝚎_𝚊𝚗𝚜𝚠𝚎𝚛/𝟹, and routing is pattern matching with confidence thresholds as guards. Crashed calls become error messages under a 𝚃𝚊𝚜𝚔.𝚂𝚞𝚙𝚎𝚛𝚟𝚒𝚜𝚘𝚛, recursion is a clause that replies again.