Structured Judgment with Jev

This week, we’re diving into Jev, a new type of AI model available from TypeSafe.ai that was released a few weeks ago. They’re approaching AI from a different direction with a new class of models that are optimized for making structured decisions.

The core difference between Jev and other LLMs comes down to how we interact with it. Existing LLMs largely involve passing unstructured text with the ability to subsequently send additional messages to refine the output.

Jev adds structure into what it expects for inputs and outputs, and it only runs one iteration. Because Jev’s outputs are always the same expected shape, downstream applications can reliably work with the data without the fear of breaking due to LLM hallucinations.

The structure requires a little more planning and work upfront, but it’s worth it: Jev is over 100x faster and cheaper than frontier AI models.

Let’s go a bit deeper and unpack how Jev works.

• Each request consists of state and questions
• Each response has answers to every question asked
• There are three types of questions: choice, score, and noul

State is the context or data that the questions are targeting. This can be unstructured text, numbers, collections of things.

A choice question is a set of labels where Jev’s answer provides a probability on a per choice basis and an overall confidence. Think of this like how emails are classified as important, spam, or marketing content within Outlook and Gmail.

A score question is similar although it contains an ordered list of outcomes like a rubric or letter grade; Jev’s answer is similar to that of a choice question, plus it includes an average score. Think of this like a 360 review where each person gives you a score from 1 to 5 and you get an average score of 4.6.

A noul question provides context about what something looks like if it’s true or false, and Jev’s answer returns a probability of it being true. Think of this like figuring out whether someone agreed to receive more information about a service or product.

Because this sounds a bit esoteric, let’s explore a use case that most of us are familiar with: meeting notes. Practically all of my clients have adopted AI-generated transcripts and notes on their meeting platforms, which produce a bunch of unstructured text.

The transcript and notes are the state passed into the request. Additional pieces of state may include descriptions of each of the participants, their role, and details about why the meeting happened.

Several questions can be asked about the meeting that align to each of the question types that Jev supports.

• Choice: What type of meeting is this?
• Choice: What emotions were present?
• Score: Where does the sentiment rank from negative to positive?
• Score: Is there convergence or divergence present in the topic discussed?
• Noul: Is there a clear next action from this meeting?
• Noul: Was a decision made during the meeting?

Jev’s answers to these questions can be handled by automation tools to update systems, trigger escalations, or pass into AI to produce coaching and guidance. That’s the benefit of Jev’s structured judgment: it’s not just another summary of a meeting, rather a set of answers that systems can act on.

The same pattern shows up elsewhere, like assigning a score to a prospect, determining what stage a relationship has hit, or to decide whether to initiate an action. Even if you never touch Jev, the redesign is the same: decide the questions and shapes before sending data to AI.