You send a state and a map of named questions. You get back one typed answer per question. That is the whole API.
The state is whatever your code already has — a ticket, a document, a JSON record. The questions are yours to name, and those names never reach the model, so each question has to say everything in its own instructions.
Jev never writes prose and never explains itself. It returns probabilities, and your code does the branching — which means the decision is always yours and always auditable.
