Jev: A System One Model for Efficient AI Evaluation
Ayyaz@ayyazdev
Jev is a System One model that takes state as input and outputs typed choices, scores, or booleans with calibrated probabilities, avoiding string generation and performing parallel evaluations.
SignalDiogo Almeida (ex-OpenAI, RLHF) built it as a System One model: state in, typed Choice/Score/Boolean out with calibrated probabilities.
Pattern↑ Jev AI decision models
TypeSafe AI opened early access to Jev this week. Diogo Almeida (ex-OpenAI, RLHF) built it as a System One model: state in, typed Choice/Score/Boolean out with calibrated probabilities. No string generation.
Parallel eval of all questions in one shot. On Vercel AI Gateway as typesafe-ai/jev via AI SDK 7's experimental_evaluate. $0.042 per million input tokens, output free. They claim 70-500ms end to end.
Pranit Sharma at Vercel replaced Luna 5.6 with Jev on a command-safety classifier and got 5 to 18x faster with better accuracy (TechCrunch).
Agent loops keep doing tool pick, retry-or-stop, jailbreak checks with a frontier model. That always felt wasteful. A cheap decision model you can threshold on confidence is closer to how I'd wire those steps.