--- title: "Jev: Bounded semantic judgments for messy text" description: "Jev is a managed API from TypeSafe that performs bounded semantic judgments on messy text, returning probabilities instead of binary answers or generated text. It supports three question types: Choice selects from developer-defined options, Noul provides a probability for yes/no queries, and Score rates on an ordered scale such as frustration level. Each decision uses application state as input and returns confidence values, allowing code to apply business thresholds rather than treating the model’s output as permission. Jev is not self-hostable; it runs as a closed API in TypeSafe’s US West Coast data center, introducing network latency and data transfer concerns for teams outside that region. The article uses a coffee shop metaphor to illustrate Jev’s role as an order taker making bounded judgments, contrasting it with generative models and exact computation, and noting alternatives like open models on vLLM." --- # Jev: Bounded semantic judgments for messy text [Red Hat](https://yomu.fyi/company/red-hat) · Maarten Vandeperre · Oct 20, 2026 **Type:** Explainer ## Summary Jev is a managed API from TypeSafe that performs bounded semantic judgments on messy text, returning probabilities instead of binary answers or generated text. It supports three question types: Choice selects from developer-defined options, Noul provides a probability for yes/no queries, and Score rates on an ordered scale such as frustration level. Each decision uses application state as input and returns confidence values, allowing code to apply business thresholds rather than treating the model’s output as permission. Jev is not self-hostable; it runs as a closed API in TypeSafe’s US West Coast data center, introducing network latency and data transfer concerns for teams outside that region. The article uses a coffee shop metaphor to illustrate Jev’s role as an order taker making bounded judgments, contrasting it with generative models and exact computation, and noting alternatives like open models on vLLM. ## Context The article presents Jev as a solution for bounded semantic judgments, a category distinct from exact database queries and generative text output. Using a coffee shop analogy, it explains that some customer messages require classifying the appropriate team, detecting whether a refund is explicitly requested, or rating frustration level. These judgments are not well handled by keyword rules or open-ended LLMs. Jev provides a structured decision service that returns probabilities for predefined questions, allowing application code to set thresholds and initiate follow-up actions. ## Approach / What changed Jev processes an application state object (plain string or JSON) and a set of predefined questions of three types: Choice (select from developer-defined options, up to 255), Noul (yes/no probability), and Score (ordered scale with 2-10 levels). Each question returns a probability or weighted score, along with a confidence value. The model is a specialized inference system that does not generate text; it produces typed, probabilistic judgments. The official Jev model is run only on TypeSafe’s managed API, not available for self-hosting. The article also notes alternatives like open models on vLLM for self-hosted deployments. ## Takeaways - Jev provides three bounded question types: Choice selects from developer-supplied options, Noul returns a probability for yes/no questions, and Score rates on an ordered scale. Each returns probabilities and confidence values. - Jev decisions are probabilities, not binary answers; application code interprets the probability and applies its own business threshold before taking action. The model does not authorize actions directly. - Jev is a managed API hosted exclusively on TypeSafe’s US West Coast infrastructure, not available for self-hosting or download. Teams requiring local processing or data sovereignty must consider alternatives like open models on vLLM. ## FAQ ### Can I self-host Jev? No, Jev is a managed API from TypeSafe. The official model cannot be downloaded or served on your own infrastructure. You must send requests to TypeSafe's hosted endpoint, currently on the US West Coast. ### Does Jev return binary yes/no answers? No, Jev's Noul type returns a probability between 0 and 1, not a binary true/false. A value of 0.5 means the model considers yes and no equally likely. The application code must interpret the probability and apply its own threshold. ### How does Jev differ from a generative LLM? Jev is designed for bounded semantic judgments, returning structured probabilities and choices rather than generating text. It takes application state and predefined questions as input, whereas a generative LLM would compose free-form responses. Jev is more like a classifier than a chat model. **Tags:** [AI](https://yomu.fyi/topic/ai), [Decision-Making](https://yomu.fyi/topic/decision-making), [Jev](https://yomu.fyi/topic/jev), [Semantic Judgment](https://yomu.fyi/topic/semantic-judgment) - Source: [Red Hat](https://developers.redhat.com/articles/2026/10/20/jev-bounded-semantic-judgments-messy-text) - Source URL: https://developers.redhat.com/articles/2026/10/20/jev-bounded-semantic-judgments-messy-text - Ingested by Yomu: 2026-10-20T15:00:17.530Z [Read original post](https://developers.redhat.com/articles/2026/10/20/jev-bounded-semantic-judgments-messy-text)