TypeSafe Jev 1.13: a model for application decisions

What goes in and what comes out

You give a description of the situation and questions about the allowed results. You get typed answers with probabilities. The vendor presents Jev as a model that gives up generating free text in favour of such an interface.

A sample task could be choosing a department, deciding whether urgent escalation is needed and judging whether a request is complete. These are different questions about the same case. The application uses the answers in the process rules, for example it sends an ambiguous request to an operator.

This describes a proposed use. It is no result of a Jev test at Syntalith, and no proof that the model will handle a specific company queue correctly.

Three kinds of question and access through the API

choice picks from the given categories, noul returns the probability of a yes answer, and score places a case on a defined scale. For a request, these could be the department, whether it asks for a refund and the level of urgency. The model takes text, also inside a JSON object; it does not analyse images or recordings directly. It also gives no reasoning for its decision.

According to OpenRouter's documentation, an OpenRouter key is enough, with no separate TypeSafe account. Requests go to the Decisions API at POST /api/alpha/decisions or, compatible with the TypeSafe SDK, POST /api/v1/systemone. The key stays on the application server. Several independent questions about the same state can be sent together; the answers do not see each other.

In a pilot, we record the version identifier the API returns. An alias that points to the newest model can change the result, so after a version change we check the thresholds again on labelled data.

Why speed attracts interest

Many applications need no long statement. They have to recognise an intent quickly, check a condition or choose the next path. Jev fits this need with a different way of getting structured results.

TypeSafe publishes its own speed and cost comparisons. The authors describe the set of tasks used, the conditions and the limits of the comparison. We treat these results as the vendor's measurements that justify a trial. We do not translate their percentages into savings for a specific company.

In an application, the network path, the input size, validation and the time other services take also count. Shorter model computation will not always shorten the whole process by the same amount.

A correct format and a correct decision

Keep two kinds of error apart. The model can return an invalid structure that the application cannot read. It can also return one of the allowed categories, but the wrong one.

TypeSafe's claims of no type errors concern the structure of the answer. The category „logistics” can be a valid data type and a wrong choice for a payment complaint. Likewise, a high confidence level needs checking on company examples. That is why the project covers evaluation, a threshold for passing a case to a person and handling of missing data.

Can Jev run locally?

OpenRouter provides a hosted TypeSafe service. In the materials we checked, we found no Jev 1.13 weights to download and run yourself. In a project that needs private processing, we check the vendor's terms, location and availability. We do not assume that the Jev name means offline work.

There are separate local projects inspired by this idea, including Kev and TinyJev. They have their own weights, training and limits. We describe them in the comparison of local alternatives.

How Syntalith can help

We choose a decision in the process that can be judged clearly. We compare rules, a small model and a chosen general model; we include Jev where its availability and terms fit the requirements. We measure accuracy, the escalation rate and the cost of an accepted decision.

If you have a queue of requests or documents to classify, describe it to Syntalith. We can prepare a test, choose how to host it and tune a local model to the categories your company uses.

What it costs

The cost of Jev calls depends on the OpenRouter price list, so the scenario does not count it. What remains is the time to check decisions and the cost of the application.

Where to start

In the free review we choose one process and count the volume and the team's working time, and within 2 working days you receive a written summary with a price. We run the build as part of AI apps, and price maintenance separately.