TypeSafe AI · Evaluation model

Jev

Turn context into decisions, with probabilities attached.

Classify text, check criteria or score a response. Jev answers your questions with typed results your code can use.

One input. Three ways to evaluate it.

Illustrative values · No API call

Shared input

“My subscription was charged twice this month. I can’t update my card, and this is my second message. Please help.”

Pass text or JSON as shared state. Ask several questions in one request; each is evaluated independently against that same input.

Which team should handle this?

Billinghighest-probability option

Billing: 82%
Technical: 13%
Account: 5%

Define a set of categories. Jev returns the most likely option and the probability of every option, adding up to 100%.

TypeSafe’s Choice guide ↗

Read the whole distribution.

The winning option is only part of the answer. A clear preference and a close call can both select “Billing.” Their distributions tell you how much ambiguity remains.

TypeSafe’s native Choice and Score answers also include a value derived from the distribution’s shape. It is distinct from an option’s probability and is not a guarantee of correctness.

How TypeSafe defines confidence ↗

Clear preference

Billing: 90%
Technical: 7%
Account: 3%

Close call

Billing: 36%
Technical: 34%
Account: 30%

Illustrative Choice distributions.

Catalog pricing

Input / 1M tokens: $0.0462
Output / 1M tokens: Free
Request minimum: $0

Finalized provider inference cost plus 10%. Verified zero-cost requests remain free. Credit holds are temporary; settled charges appear in account activity.

Evaluation endpoints

Primary HTTP API: POST; Boolean probabilities, selected choices and rubric scores. Choice and score answers include their full probability distributions.
Native TypeSafe format: POST; Uses Noul for boolean questions. Choice and score answers also return confidence; scores include a level legend.

Building with Jev on MiniRouter

Text and JSON in, typed answers out: State can be a string, object or array. These evaluation endpoints do not generate chat replies or accept image, audio or streaming requests.
One state, independent questions: Batch up to 64 questions. If a question needs a previous answer, make a separate call. Extra questions still use input tokens.
Your categories and scale: Up to 64 options per choice; 2–10 ordered levels per score. Describe each option clearly, and add an “other” option when categories are not exhaustive.
32 KiB normalized request limit: The upstream JSON must fit within 32,768 UTF-8 bytes, including routing fields MiniRouter adds. Leave room for those fields in your payload.

Availability follows the qualified catalog and your API key permissions. If the qualified route is unavailable, requests return 503 before a credit hold or provider call.

HTTP example

Evaluation HTTP request

curl --fail-with-body https://api.minirouter.sh/v1/evaluate \
  -H "Authorization: Bearer $MINIROUTER_API_KEY" \
  -H 'Content-Type: application/json' \
  --data '{
    "model": "typesafe-ai/jev",
    "state": "The subscription was charged twice.",
    "questions": {
      "billing": {
        "type": "boolean",
        "instructions": "Does this describe a billing problem?"
      }
    }
  }'

Use an explicit model and an integration API key. Disable automatic retries: a lost response may still incur cost. See the guide for question formats and limits.