TypeSafe System One / Jev: Judgments + Probabilities, Not Chat
Inference Desk explains Jev’s binary, categorical, and scoring questions, then connects the request schema to application control flow. The examples cover independent questions over one state and confidence gates that send uncertain cases to another handler.
Original by Inference DeskGetting startedIntermediate5 min 5 secPublished
Independent binary, categorical, and scoring questions can share one state and return probabilities without conversational prose.
Question identifiers belong to the application; the instructions and criteria must describe what the model should judge.
Application code owns thresholds and fallbacks, including escalation of uncertain or risky decisions.
Worth knowing
Schema-constrained JSON and typed probability outputs do not guarantee correct or hallucination-free decisions; thresholds must be tuned to specific domain risks.