Jev targets software decisions rather than conversation

TypeSafe AI’s Jev separates decision-making from text generation. The proprietary model evaluates application state against typed questions and returns structured answers in a parallel pass. Its outputs stay within a predefined answer space, eliminating schema and parsing failures without guaranteeing that a decision is correct.

Jev supports Choice for selecting a category, Score for grading against an ordered rubric, and Noul for returning the probability of a yes-or-no outcome. Its training targets calibrated probabilities, allowing applications to automate confident decisions and send uncertain cases to another model or a human reviewer. Potential uses include model routing, tool selection, retrieval checks and moderation queues.

TypeSafe reports latency of 70 to 500 milliseconds and pricing of $0.042 per million input tokens, with output billed at zero. Those performance claims come from vendor-selected workloads. Jev reaches roughly 68% accuracy on TypeSafe’s own benchmark, positioning its main advantage as speed and cost rather than superior accuracy. A fine-tuned small classifier may remain cheaper and equally accurate for stable tasks with sufficient labelled data.

Eden AI offers Jev through an alpha decisions endpoint alongside its model execution services, bypassing TypeSafe’s direct-access waitlist. However, request formats, model identifiers and pricing can change without a deprecation period. Jev has no published weights or self-hosted option, and confidence thresholds require validation against real application outcomes before consequential decisions are automated.