Jev Makes Decisions, Not Text
Jev is TypeSafe AI’s first System One Model, a model class designed to produce fast, structured decisions that software can use directly. Its relevance is that it targets automation: instead of returning open-ended text that must be parsed and validated, Jev accepts unstructured state and returns typed, probabilistic outputs with confidence information.
The distinction begins with what the model produces. Conventional language models generate strings, which can serve many purposes but may contain hallucinations, refusals, or type errors. Software generally needs to interpret and validate those responses before using them. Jev gives up string generation in favor of structured values whose possible outputs and organization are defined in advance. TypeSafe describes this as a frontier-intelligence function call: unstructured input goes in, and typed decisions come out. Because the output must match its schema, the company states that Jev does not make type errors.
Jev is also designed differently at the sampling level. Traditional language models generate sequentially, producing one token at a time and conditioning each token on the previous one. Jev generates its outputs in parallel in a single query. TypeSafe reports end-to-end response times of 70 to 500 milliseconds for its System One-shaped queries, compared with a reported three to 329 seconds for the frontier models cited in its comparison. The company lists input pricing of $0.042 per million tokens and says output is currently free to meter. These figures apply to TypeSafe’s service and published comparisons, not to every possible workload.
The model’s training approach is called Reinforcement Learning for Calibrated Decisions, or RLCD. Rather than optimizing primarily for human-preferred writing or programmatically verifiable answers, it is intended to produce calibrated probabilities on System One tasks. Jev communicates confidence and uncertainty with each output; according to TypeSafe, higher confidence is intended to correspond to higher accuracy, while similar inputs should receive more consistent answers.
The proposed uses are decisions inside software: classification, routing, scoring, extraction, and branching where fixed rules are too rigid. TypeSafe also describes workflows that decompose complex tasks into independent questions, use probability-sensitive behavior, and then produce a discrete branch. Other stated applications include reducing large datasets into features, real-time software experiences, and checking or guarding the prompts and outputs of other language models.
TypeSafe’s workflow evaluations compare models using the same code-defined workflows and use reference probabilities from larger external models. The company reports that Jev performs strongly on these evaluations, while also explaining that the workflows were created by its capabilities team and that some bias may remain. Jev is available in early access.
Jev’s central design choice is to make software-consumable decisions rather than free-form text. Its structured outputs, parallel sampling, calibrated probabilities, and focus on type safety are presented as the basis for faster and more dependable AI-powered workflows.
Sources:
