Most language models respond to prompts by generating text. A new arXiv paper describes an alternative: a System-1 architecture that answers typed questions about a state with calibrated probabilities. The model, introduced through TypeSafe's Jev, is designed for narrow decision-making rather than open-ended generation.

One decision type described in the abstract is Choice, where the model selects among options expressed in natural language. Because the outputs are probabilities, the system is intended to be calibrated, meaning its reported confidence should align with actual performance. The paper suggests that such narrow decision models could offer a more structured and reliable interface for specific tasks.

The paper is available on arXiv and has not yet been peer-reviewed. It focuses on a System-1 style of inference, which may complement larger System-2 models that handle broader reasoning.