The jevchat tool allows users to interact with the Jev model via a chat interface. Unlike general-purpose LLMs, Jev is categorized as a "System One" model, designed to prioritize speed and intuition for structured decision-making over the deliberate reasoning used to generate prose.
Jevchat simulates text generation by treating symbols as decisions for the Jev model
PLUS ULTRA by Amenoyomi
The model operates as a classifier that returns predefined "typed" decisions—specifically Choice, Score, or Noul—accompanied by probability distributions. This allows Jev to function as a high-speed, cost-effective "smart function call" that avoids the computational overhead and potential hallucinations often associated with free-form text generation.
jevchat achieves pseudo-generation by repeatedly querying Jev's Choice API to determine which symbol from a given alphabet should come next. The tool samples from the returned normalized distribution, appends the character to the reply, and repeats the process until a stop symbol is drawn.
Performance varies significantly based on the presentation of options. In symbol mode, Jev must internally append the option to the text before evaluating it. In hypothesis mode, the options provided are the resulting finished strings, which allows the model to rank completed text—the primary function for which decision models are built. This method roughly triples the top-1 accuracy and doubles the probability mass on the correct symbol.
While Jev is individually cheaper per decision than frontier models, the generation loop used by jevchat is computationally impractical. Because the system requires a separate API request for every symbol produced, the aggregate cost is prohibitive for standard use.