Jev

### Long description
Jev is an artificial intelligence model that evaluates questions about a state and returns structured results that code can use directly. It belongs to a class of models that its developer calls System One, oriented towards decision-making within software rather than generating text for people. It does not generate free text: it receives unstructured information, such as text or the state of a program, and returns values in a predefined format.

Each query consists of a state and one or more questions of three types: Choice, which selects an option from a list; Score, which rates the state against a rubric; and Noul, which estimates the probability that a statement is true. Choice and Score return probabilities and a confidence level, which makes it possible to set thresholds so that the software acts autonomously or refers the case for human review. Noul does not include a separate confidence level.

Questions, even of different types, are evaluated in parallel and in isolation against the same state, in a single request. The possible outputs are defined in advance, so the result conforms to the specified schema and does not require interpreting free text. The model is based on a proprietary architecture, a parallel sampler and a training method called reinforcement learning for calibrated decisions (RLCD).

Query design starts from atomic questions, each about one specific aspect, whose results are combined through logic in the code. Uses include classifying, routing, scoring and extracting information, branching processes where hand-written logic is rigid, transforming large volumes of data into attributes, and evaluating or filtering the inputs and outputs of other models.

Access is through an API, with documentation and a console for developers. Jev is developed by TypeSafe AI, an AI lab focused on intelligence infrastructure for automation.