Overview
- TypeSafe AI’s Jev is a model that returns calibrated probability distributions for typed questions instead of generating text, using three primitives: Choice (pick an option), Score (position on ordered levels), and Noul (yes/no probability).
- Several teams have replaced many chat-model judgment calls with Jev in agent control paths to save time and money, reporting answers in under a second and billing input at about $0.042 per million tokens with no output charges.
- Hands-on testing found concrete pitfalls: Score’s numeric value is a probability-weighted average so you must take the argmax of probabilities to pick a bucket, the provided confidence measures distribution peakedness not true uncertainty, and option lists act as part of the prompt so synonyms or omissions change outcomes.
- Practical integration patterns that emerged include giving Jev veto power only, routing low-confidence or edge cases to a chat-model fallback, and calibrating thresholds and option sets on your own data before automating actions.
- The change lowers per-decision cost enough that teams are automating many more small checks, so builders should expect higher decision volume, auditability needs, and the need to monitor calibration and failure modes in production.