Jev: An AI Model for Decision-Making Instead of Conversation

Most AI models we know today, such as ChatGPT or Claude, are designed to generate text. But Jev, introduced by TypeSafe AI, takes a different approach. Instead of producing sentences and paragraphs, its goal is to make fast, structured decisions for software systems. TypeSafe calls this category of models System One Models.

For example, instead of asking a language model, “Which department should handle this customer request?” and then analyzing its textual response, Jev can directly return a structured decision such as:

support: 0.82
sales: 0.14
spam: 0.04

An important feature is that Jev can provide a probability or confidence level alongside each decision. This allows an application to create simple rules, for example: if confidence is above 90%, take the action automatically; otherwise, send the case for human review.

TypeSafe says Jev is trained using a method called RLCD, or Reinforcement Learning for Calibrated Decisions. Unlike RLHF, which is commonly used to align language-model responses with human preferences, RLCD focuses on making decisions whose confidence scores are better calibrated.

The main advantage of Jev appears in tasks such as AI agents, request routing, classification, scoring, filtering, and automated workflow decisions. According to TypeSafe, Jev can perform these kinds of tasks much faster and more cheaply than using a full large language model for every decision.

Perhaps the most important idea behind Jev can be summarized like this:

LLMs are built to communicate with humans; Jev is built to make decisions inside software.

If this approach proves successful, future AI systems may increasingly use a combination of models: an LLM for reasoning, writing, and interacting with people, and models like Jev for handling thousands of small, fast, and inexpensive decisions behind the scenes.

One important distinction is worth noting: TypeSafe describes Jev as having “Zero Hallucinations,” but this does not mean its decisions are always correct. Rather, because Jev does not generate unrestricted free-form text, its outputs can be constrained to predefined types and formats. That reduces formatting and fabrication problems, while decision errors can still occur.