Type Safe.AI launched its first model, Jev, a System One model that differs from LLMs. While LLMs predict the next token and generate text, System One models evaluate a state and produce structured answers. Jev uses RLCD (Reinforcement Learning for Calibrated Decisions) instead of RLHF, training the model to return decisions and probabilities.

The input is a state—context plus questions. It can be a simple message like "Is it possible for me to change my PIN number?", a JSON object with fields, or an array of messages. Best practice is to use an object with clear field names.

Questions are answered using primitives: Choice for predefined options (e.g., delivery, billing, account access), Score for ordered values (e.g., negative, neutral, positive), and Noul for yes/no. The model returns probabilities for all values, providing confidence levels.

Confidence is crucial for decision-making. For auto-replies, if the model classifies a request with high confidence, it can auto-respond; if uncertain, it can escalate. This makes Jev suited for classification and LLM-as-a-judge tasks.