TypeSafe Introduces Jev: Typed Decision AI with 20 Use Cases

Within modern artificial intelligence agent loops, thousands of minor judgments occur constantly, ranging from model selection and command safety checks to passage relevance and completion verification. To address this structural bottleneck, TypeSafe has launched Jev, a system designed specifically to handle these micro-decisions rather than engaging in general conversation, code generation, or summarization. Jev takes unstructured state inputs and returns typed decisions accompanied by calibrated probability metrics.

The operational workflow begins by submitting a state object, formatted as text or JSON, alongside a dictionary of typed questions. TypeSafe documentation outlines three core primitives tailored for developers building automated agent pipelines.

Every question is evaluated in parallel against the identical state within a single request. TypeSafe trains Jev utilizing Reinforcement Learning for Calibrated Decisions (RLCD), ensuring that higher confidence levels reliably track higher accuracy rates. Furthermore, the Choice function supports up to 255 distinct options in a single pass.

The primary performance claims indicate that Jev operates 193.6x faster and 444.6x cheaper. According to TypeSafe's launch disclosure, these figures represent the higher end of real-world performance gains, utilizing GPT-6 Astra and Fable 5.1 as reference baseline answers.

Stock photo for illustration only, not from the actual event

Open-model performance figures are self-reported by individual projects on custom harnesses and should be viewed directionally. The cleanest head-to-head evaluation comes from OpenRouter's Banking77 test, where Jev scored 3.3 points lower in accuracy than Claude Opus 5, while operating at a median speed 13 times faster and at approximately 1/22nd of the cost.