SCOREAgents & automation
Jev: Type-Safe Probabilistic Reasoning
Hazem Omier@hazemomier
Jev is a non-chat model that takes typed state and questions, returning calibrated probabilities in a single pass without generating text.
SignalYou pass state + typed questions (Noul / Choice / Score); it returns calibrated probabilities in one parallel pass — ~70–500ms, $0.042/MTok in, output free.
Pattern↑ Jev AI decision making
Jev (TypeSafe, System One): not a chat model. You pass state + typed questions (Noul / Choice / Score); it returns calibrated probabilities in one parallel pass — ~70–500ms, $0.042/MTok in, output free. No string generation, so no prose hallucination by construction.
Vercel AI SDK now wires it via experimental_evaluate (model typesafe-ai/jev).
For agent stacks: do you put a System One gate in front of the LLM (route / escalate / score), or keep asking a frontier model to decide whether to call itself?