An incoming support message needs a destination, a priority, and someone to handle it. Jev evaluates information against questions and possible answers you define, so it can help sort that queue. Define those decisions, use their probabilities to route work, and check the results before automating.
01Choose the answer type
TypeSafe calls Jev a System One model: a model built for fast, focused judgments that software can use directly. You supply the state, meaning the information to evaluate, and a schema, meaning the structure of the questions and answers. Jev returns values for code to use. A separate generative model writes replies or summaries. TypeSafe introduction.
| Type | Use it for | Read the result as |
|---|---|---|
| Choice | Which team should handle this message? | One of your named options, with probabilities and confidence |
| Score | How much does this issue disrupt the customer’s work? | A position on ordered levels you describe, with probabilities and confidence |
| Noul | Does this message request a refund? | The predicted probability of yes, from 0 to 1 |
A Choice needs options that separate the possible answers. Add other when the list might miss a case. A Score needs descriptions for each level. Three levels are numbered 0, 1, and 2; the returned score can fall between them.
With Noul, 0.5 means yes and no have equal probability. It does not mean medium urgency or medium lead quality. Define a scale with Score when you want to measure degree.
02Make the questions precise
Give each question one judgment to make. Automation creator Nate Herk’s I Tested Jev on 12 Real Use Cases. My Honest Thoughts. demonstrates this with separate questions for billing messages, sponsorship inquiries, urgency, and sponsor fit. Each answer becomes a usable column; a broad request to analyze an inbox leaves the desired output undefined.
Write the full question in instructions; the question ID is only for your code. Questions in one request run independently against the same state. Combine the answers afterward. If one answer is needed to fetch new information, make a second request with that information. Question design.
This original request follows the direct API’s documented format; it has not been tested against the live service. Send it to POST https://api.typesafe.ai/v1/systemone with Authorization: Bearer YOUR_API_KEY and Content-Type: application/json headers, or recreate the questions in TypeSafe’s playground.
{
"model": "jev-latest",
"state": "The dashboard stopped loading this morning. Our team cannot process orders, and we have no workaround.",
"questions": {
"department": {
"type": "choice",
"instructions": "Which team should handle the main issue in this message?",
"criteria": {
"billing": "Charges, invoices, or payment collection",
"technical": "Broken features, outages, or integration failures",
"other": "A request outside billing and technical support"
}
},
"impact": {
"type": "score",
"instructions": "How much does the reported issue disrupt work?",
"criteria": [
"Work continues normally",
"Work is disrupted but a workaround exists",
"Work is blocked and no workaround is available"
]
},
"refund_requested": {
"type": "noul",
"instructions": "Does the message explicitly request money back?"
}
}
}
As of September 19, Vercel also offers Jev through AI Gateway, using experimental_evaluate and model ID typesafe-ai/jev. Its example uses boolean for a yes-or-no question. Follow the documentation for the interface you choose; the request above is for TypeSafe’s direct API.
03Keep decisions and actions apart
Separate how well a request fits your service from how certain the model is about that judgment. In Jev is HERE. How to use it, podcast host Greg Isenberg and guest Ryan Vogel explore business routing through a design agency’s contact-form leads. An uncertain lead may need clarification. Quietly discarding it could lose a customer.
Choice and Score return confidence, derived from how their probabilities are distributed. It is a separate field from the winning option’s probability. Noul has no separate confidence field. Confidence documentation.
Our routing recommendation:
| Result | Next step |
|---|---|
| Clear fit for a tested category | Route to the corresponding queue |
| Ambiguous fit or missing information | Keep the item visible for review or clarification |
| A reply or detailed explanation is needed | Pass the source material to a person or generative model |
Choose thresholds from errors on your own examples. Keep permissions and execution in code: selecting a refund category must not itself issue a refund. Assume your agent uses every tool it has covers that boundary.
Herk applies the same division to YouTube comments: classify the pile, then send selected comments to a model that can write responses. Build an agent that finds your customers uses the same rule: call a model at the steps that need judgment.
04What the different demos teach
Fast decisions also fit inside interactive software. AI creator Matthew Berman’s We need to talk about Jev… contributes a simulated town where Jev chooses character reactions. Alongside the business examples, it shows several ways to divide the work:
| Source and example | What to take from it |
|---|---|
| Herk, meeting analysis, 10:57 | Ask whether a call names next steps, owners, and deadlines. Count the answers to find a recurring process problem. Your questions determine what the analysis can reveal. |
| Vogel with Isenberg, old leads, 18:03 | Apply the same classification to historical inbox records to find overlooked inquiries. A useful workflow can start with the backlog. |
| Vogel with Isenberg, service matching, 21:29 | Their proposed local-services business matches a request against supplied providers. Our implementation advice: fetch current coverage and availability first; calculate any quote from actual pricing rules. |
| Vogel, clips, 24:02, and Herk, 11:58 | Give Jev text passages prepared by transcription and candidate-selection tools. Judge whether each passage makes sense by itself and has a useful opening, then let editing software cut the clips. |
| Berman, simulated town, 6:24 | His coding model builds the world; Jev selects character reactions. A short response time lets decisions happen during interaction, while application code moves the characters. |
Herk’s browser extension labels posts as they enter view. Berman also discusses routing a prompt to another model. Both put a small decision before a larger action. Four shapes an AI workflow can take explains how those branches fit alongside steps that run in parallel.
05Check speed and accuracy
Herk reports that grouping more work into requests and running requests in parallel cut one email run from about 70 seconds to 6 seconds. That change is part of the application. Compare models with the same inputs, questions, parallel request count, and retry settings before attributing the whole difference to Jev.
TypeSafe’s launch post lists $0.042 per million input tokens and free output. The service still charges for input. Its largest speed and cost ratios come from its own workflow evaluations, and the company says those gains are toward the high end. Launch evidence and caveats.
The same post explains that its zero-hallucination figure means the output always matches the defined structure. A valid category can still be the wrong category. Berman’s broader reliability language should be read with that limit.
The demonstrations also expose limits. Vogel says his Bitcoin experiment performs poorly; Herk calls his trading demo unvetted paper trading. Berman describes a chess win caused by the opponent running out of time despite stronger play. These examples support testing speed under a deadline, not assuming that quick decisions are good decisions.
06Test one queue before routing it
“don't just plug in Jev and trust what it says automatically.”
Nate Herk, “I Tested Jev on 12 Real Use Cases. My Honest Thoughts.” · 12:52Use the method in Score your agent on known good answers. Track wrong routes, missed urgent items, and review volume separately. A system that sends everything to review may avoid automatic mistakes while saving no work. Include review and fallback costs in the comparison.
This original exercise turns the three videos into a testable workflow:
Repeat the comparison after changing criteria or models. Recheck your model choices each release makes that maintenance routine.