--- title: "Jev: a practical guide to TypeSafe's decision model" canonical: "https://consideredgear.com/blog/jev-decision-model-guide/" pubDate: "2026-10-01T00:00:00.000Z" updatedDate: "2026-10-01T00:00:00.000Z" author: Jim Shannon description: "Jev makes typed decisions through a hosted API. Here's what its choice, score, and yes/no outputs mean, what it costs, and what to check before using it." tags: [classifiers, jev, automation] --- Jev is a hosted model for small text decisions, priced at $0.042 per million input tokens as checked October 1. It is worth evaluating if you are comfortable sending those requests to a cloud service, but a valid answer can still be wrong. TypeSafe's documented public offering does not fit a fully offline setup. Imagine your home assistant gets a simple request: “Show me the manual for the dishwasher.” Before anything writes an answer, the software has several jobs. Recognize that this is a document search. Choose the appliance. Find the file. Decide whether the result is good enough to show. Jev is designed for those small judgments. You give it text and questions with defined answer shapes, and it returns structured decisions for your code to use. TypeSafe introduced it in early access on [September 15, 2026](https://typesafe.ai/blog/introducing-system-one-models-and-jev). The important buying-guide detail is where it runs: TypeSafe's [public documentation](https://docs.typesafe.ai/models), checked October 1, describes a hosted API and does not offer a self-hosted Jev download. It is relevant to local-AI builders as a reference point and a possible cloud component, but a system that depends on its API needs an internet connection. ## What Jev returns TypeSafe calls Jev a [System One model](https://docs.typesafe.ai/concepts/system-one). The phrase describes its focus on bounded decisions inside software. It does not mean the model can understand every request reliably or replace a general-purpose assistant. Its [three question types](https://docs.typesafe.ai/primitives) are straightforward once you translate the names: - **Choice:** Select one of the options you define. For a document, that could be manual, warranty, receipt, or other. - **Score:** Rate something against ordered descriptions you supply, such as irrelevant, partly relevant, or directly relevant. The result is a numeric position along those levels and can fall between them. - **Noul:** Estimate the probability that the answer to a yes/no question is yes. “Does this request ask to change a device?” is one example. The questions go alongside the material to evaluate, called the state. Your application receives the results and decides what happens next. Jev does not produce a written explanation of its answer. If the job requires a friendly reply, a detailed summary, or new code, a generative model still has work to do. Classification can sit in front of that model; it does not make generation unnecessary. ## What it costs right now As checked on October 1, TypeSafe lists **Jev 1.13.0 at $0.042 per million input tokens**, with no output-token charge. That is US-dollar API pricing, not a subscription for a complete home assistant. The [model page](https://docs.typesafe.ai/models) also lists text-only input, a 64,000-token total request limit, and a separate 32,000-token limit for the state plus the longest question. Here is a planning example, not a measured workload: if a request uses 1,000 billed input tokens, 100,000 such requests use 100 million tokens and cost $4.20 at that rate, before any applicable taxes. Count the actual input reported by the API, including your questions and criteria. Retries and calls to other services add to the total. That low unit price makes experimentation approachable. It does not tell you whether the complete system is cheaper than a local alternative. A classifier that makes poor decisions can create more work for a larger model or for you. ## Confidence has a specific meaning Jev returns probabilities across the available options for Choice and Score. Its separate confidence field summarizes how concentrated that distribution is. Noul returns its yes probability without a separate confidence field. These distinctions come from TypeSafe's [confidence documentation](https://docs.typesafe.ai/confidence) and [Noul reference](https://docs.typesafe.ai/primitives/noul). Consequently, do not read a confidence value of 0.9 as a blanket promise that the action is 90 percent safe. Whether an action is appropriate also depends on permissions, consequences, and the quality of the information you supplied. For a first project, let Jev suggest tags or routes while you compare them with your own answers. Decide what kinds of errors are acceptable before choosing a threshold. A review queue is a useful part of the design, not an admission that the experiment failed. ## The limits are worth reading before the demo TypeSafe publishes a useful [list of known weaknesses in Jev 1.13](https://docs.typesafe.ai/model-jaggedness/jev-1.13). It includes counting and numerical precision, date comparisons, indirect instructions, distracting context, and adversarial text that can steer a decision. Keep arithmetic and permission checks in ordinary code. Send the relevant text rather than a whole archive. Treat text inside a document as material to evaluate, not as authority to change what the application is allowed to do. The distinction between a valid output and a correct answer is especially important. Choosing an allowed label avoids an invented label. The chosen label can still be wrong. A well-formed “turn on” decision is not enough reason to operate a device without the rest of the checks. TypeSafe's launch post includes large speed and cost claims from its own workflow evaluations. We have not reproduced them, and they should not be read as universal savings against every language model or local setup. For a useful comparison, measure the same task, include the network round trip, and count how much work needs escalation. ## What leaves your house The text you submit is processed by TypeSafe's hosted service. Its documentation says customer requests and responses are not used for model training, and its [legal page](https://docs.typesafe.ai/legal) describes an enterprise zero-data-retention option. Those are different questions from whether the data leaves your network. For a household document project, start with public manuals or invented records. Review the current [customer agreement and data-processing terms](https://typesafe.ai/legal/mca) before sending private documents. If your requirement is that processing stays on your own machine, choose a self-hosted approach. ## A sensible first experiment Pick one bounded, reversible job, such as suggesting a category for public appliance manuals. Define a short list of categories with clear descriptions and an “other” option. Write a small test set with expected answers, including ambiguous documents, before changing the instructions. Record the model version, the decision, the time taken, and whether a person had to review it. Keep a final set of examples separate from the ones used to tune your questions. Compare against the simple rules you would otherwise use. If the experiment works, expand gradually. Do not start with autonomous purchases, door locks, or commands that delete files. A classifier's judgment should inform your software; the software still needs to enforce what is allowed. Jev is worth evaluating when you have many small text decisions and are comfortable with a hosted API. For readers whose priority is owning the entire inference path, its main value is showing what a focused decision interface can look like. Our [classifier explainer](/blog/ai-classifiers-explained/) lays out a practical way to test whether a model handles your particular job well enough. *Research note: This guide is based on TypeSafe's public documentation checked October 1, 2026. We have not run a hands-on Jev evaluation for this article. Prices, model limits, and service policies can change; example costs are arithmetic, not recorded bills.* ## Buying options, if they fit your setup Jev is a hosted API; it does not need a PC or GPU purchase to use the service. For readers comparing it with a separate, self-hosted alternative, here is an optional hardware starting point—not a machine advertised to run Jev. - [Browse local mini-PC alternatives to a hosted-only setup (paid link)]() — For other self-hosted tools, not for running the hosted Jev service. These are Amazon search references, not verified offers. Prices, availability and exact configurations are unconfirmed. [How affiliate links work](/disclosure/). As an Amazon Associate I earn from qualifying purchases.