01 Choice
Pick one option.Classify a ticket as billing, technical, or sales. Receive a selected option, option probabilities, and confidence.
choice: "technical" Independent field guide
Jev AI,
how it works.
Jev is TypeSafe AI’s model for typed, structured decisions. Learn how Jev turns context and a constrained question into a Choice, Score, or Noul result, decide whether it fits your task, and design a first workflow with validation and human review.
STATE → DECISION
Input01
JevSystem One
Choice Which team should handle this?
Structured answer02
"technical"Model output Application actionRoute or review
Illustrative example · not a live API response
Typed decisions Choice · Score · Noul
Text in, structure out Keep your code in control
An independent perspective Examples, sources & limitations
01 / Start with the basics
What is Jev AI?
Jev AI is TypeSafe AI’s System One model for structured decisions. You provide context, called state, and typed questions. Jev returns choices, scores, or yes/no probabilities that your software can use directly. It does not write a reply or carry out an action. Your code validates the decision, applies thresholds, and decides what to do next.
Think of a customer message about a broken integration. Instead of asking a model to compose a support email, ask which team should receive the ticket. Jev chooses from the options you supplied; your application routes the ticket or sends it for human review.
TypeSafe AI · System One02 / How Jev works
From state to a controlled action.
A Jev request separates the model’s judgment from the application’s action. That boundary is the central idea: Jev answers a typed question about supplied text, while your software remains responsible for validation, thresholds, permissions, and side effects.
- 01
Define one narrow decision
Start with a question that has a clear operational meaning. Ask Jev which support queue fits a message, how strongly a passage matches a criterion, or whether a specific policy condition appears to be met. Avoid combining unrelated decisions in one label.
- 02
Supply only useful state
Send the text and context needed for that decision. Jev accepts structured text values, but more context is not automatically better. Remove unrelated records, keep untrusted content separate from instructions, and let retrieval code choose the evidence first.
- 03
Choose the answer type
Use Choice for a selection among explicit options, Score for ordered levels with written criteria, and Noul for the probability that a yes-or-no statement is true. The type should match the decision your application actually needs.
- 04
Validate before acting
Check that the Jev response has the expected type and range, record the model version, and route uncertain or invalid cases to review. Authentication, access control, calculations, database writes, and external actions remain conventional application code.
03 / Choice, Score, Noul
Three ways to ask.
Ask a narrow question, choose the right answer type, and let ordinary code handle the action. You can ask several questions about the same state in one request.
02 Score
Rate against clear levels.Define ordered levels such as calm, frustrated, and very angry. Read the score, level probabilities, and confidence.
score: number 03 Noul
Estimate a yes/no probability.Ask whether a message is urgent. The result is a probability between 0 and 1; your code chooses an action threshold.
noul: 0…1 04 / Task fit check
Is Jev right for this task?
Use this quick Jev fit check before writing an integration. It does not call the API or predict model accuracy; it tests whether your task has the structure needed for a controlled decision workflow.
05 / First Jev workflow
Build a useful first application.
A support-ticket router is a practical first Jev project because the input, decision, and application action can be separated clearly. Keep the first version small enough that every result can be inspected.
Write the routing rule
Define billing, technical, sales, and needs_review in plain language. State how to handle messages that mention several teams. The fallback matters because Jev can only choose from the answer structure you provide.
Create representative examples
Collect ordinary tickets, ambiguous cases, short messages, hostile instructions, and out-of-scope requests. Label the expected route independently. This set gives you evidence for comparing Jev outputs instead of judging a few memorable examples.
Send and inspect one request
Pass the customer message as state and ask a Choice question with your defined routes. Validate the returned option and inspect its probability distribution and confidence. Treat sample output in this guide as illustrative until you run Jev with your own access.
Add review and failure handling
Choose review rules from observed errors and business cost, not from a universal threshold. Handle timeouts, authentication errors, rate limits, missing answers, and version changes. Make queue writes idempotent so a retry cannot duplicate an action.
06 / Guides
One practical learning path.
01
Find a useful task
Six examples, from ticket classification to agent routing.
Read the guide 02Understand the cost
Input-token billing, with transparent cost examples.
Read the guide 03Make a first request
From an API key to a typed answer you can read.
Read the guide 04Know the boundaries
Failure modes, language support, and human review.
Read the guide07 / Jev FAQ
Questions developers ask about Jev.
Short answers to the main questions behind a Jev evaluation. Each answer links conceptually to the deeper guides on this site.
What is Jev?
Jev is TypeSafe AI’s System One model for structured decisions over text. A request contains state plus typed questions, and Jev returns Choice, Score, or Noul answers that application code can validate and use. Jev is designed for constrained judgments, not for generating prose or autonomously completing a workflow.
How is Jev different from a generative AI model?
A generative model produces open-ended text such as an email, explanation, or summary. Jev returns a constrained decision in a declared answer type. That makes integration simpler for routing, ranking, classification, and policy checks, but the typed format does not guarantee that the judgment is correct.
What do Choice, Score, and Noul mean in Jev?
Choice selects one supplied option and includes option probabilities and confidence. Score evaluates ordered, described levels and returns a score with related probabilities and confidence. Noul expresses the probability that the answer to a yes-or-no question is yes. Each supports a different application decision.
Can Jev take actions or generate a customer reply?
No. Jev returns a decision; your application decides what happens next. Use ordinary code for permissions, calculations, persistence, and external actions. Use a generative model when the required output is new prose. Keeping those responsibilities separate makes failures easier to detect and review.
How should I evaluate Jev before production use?
Build a labelled set from your own normal, ambiguous, adversarial, and out-of-scope inputs. Pin the Jev model version, record the exact questions, compare results with expected decisions, and review false positives and false negatives. Then set thresholds, fallbacks, and human-review rules for the cost of those errors.
Where Jev fits
Use Jev for narrow judgments inside a larger application: selecting a route, checking relevance, or assessing a criterion. Your application still owns retrieval, exact calculations, access control, and execution.
Use casesWhere to use another tool
Use a generative model to draft a reply, a parser to extract exact fields, and normal code for arithmetic. A well-formed output is not a guarantee of a correct judgment.
LimitationsUnderstand it. Build with it.
Start with one useful decision.
Read the short setup guide, send a small request, and inspect the result before putting it into a workflow.
Reviewed: · Applies to: Jev 1.13.0