Open-Source Jev on Apple Silicon and CUDA

The hardest part of production is making thousands of simple decisions reliably.

Should this ticket be escalated? Which tool should run? Does a retrieved document support the proposed action? Is the user asking for a refund or an explanation?

Many teams have been answering these questions by prompting a general-purpose LLM to return JSON. The model generates a label, perhaps a confidence score, and the orchestration code moves on.

Field like "confidence": 0.93 looks precise, yet you still have to measure separately whether that number tracks how often the model is right.

However what you often need is to give a model some shared context and a set of typed questions, then read probability distributions directly from its internal representations.

The model never has to generate prose.

Let’s have a look at that clean split between deciding and acting.

Why Put a Decision Model Inside an Agent?

Picture a support agent with access to billing, shipping, and returns tools. A user writes:

Shoes arrived two weeks late and in the wrong size. Also I see two charges on my card.

A generative agent can reason about the ticket and pick a tool. A decision service can instead return distributions over a fixed set of options plus an escalation signal, and your application applies its own policy to the…