TypeSafe Jev
Answers typed questions about content you supply — yes/no, pick-one, rate-against-levels — with a probability on every option instead of a paragraph to parse.
Where TypeSafe Jev lands
Composite scores out of 10, from benchmark runs where every provider faces the identical task corpus. Rank is within that capability's field. Full method: how we measure.
TypeSafe Jev sits outside the web-access benchmark — the search/scrape/crawl corpora don't exercise a model inference API, so this page carries catalog facts without a scorecard.
The honest pitch
Reach for it when your code needs a decision it can branch on, not prose it has to read: which queue a ticket belongs in, whether a retrieved passage actually answers the question, how severe a policy breach looks, whether a message is a jailbreak attempt. Every answer carries the probability of each option and its own confidence, so you can act on the clear ones and send the rest to a person. The content is read once however many questions ride along, which makes a batch of twenty far cheaper than twenty calls.
Look elsewhere when you want the model to write something — there is no generated text, no streaming, no tool use and no conversation, so a summary or a reply belongs on an inference route instead. It reads text only: images, audio and PDFs have to be turned into text or fields first. And the whole request caps at 64k tokens, with 32k for the content plus the longest single question, so a full archive in one shot is the wrong shape.
About TypeSafe Jev
TypeSafe’s Jev is built for the decision a program makes, not the sentence a person reads. You send the thing to judge — a support message, a retrieved passage, a record, a chat log — and a map of questions you name yourself. Each question is one of three kinds. A yes/no question comes back as a number between 0 and 1: how likely the answer is yes. A pick-one question comes back as the option chosen plus the probability of every option you offered. A rating question, scored against levels you describe in your own words, comes back as a weighted value plus the spread across those levels.
Two of the three also return a confidence, and it is worth treating as a separate signal from the probability. The answer tells you what the model decided; the confidence tells you whether to trust that decision on its own. Thresholding it is how a pipeline routes the uncertain cases to a human while letting the clear ones through untouched — the practical difference between a classifier you have to babysit and one you can leave running.
The billing shape rewards asking more at once. Only input tokens are charged and output is free, and the content you send is read a single time no matter how many questions accompany it. Twenty questions about one document cost roughly what one question about it costs; twenty separate calls pay for that document twenty times. Answers come back keyed by the ids you chose, so fanning out speculative questions and ignoring the ones that turn out irrelevant is a normal way to use it rather than waste.
Requests name a model version, and the reply always states the concrete version that answered — so a request sent to a moving alias still tells you exactly what produced the result. Pin the version once your thresholds are tuned against it.
Reaching it through NativePort
A fixed route. The gateway applies this route's documented request policy, injects the upstream credential and returns TypeSafe Jev's response, metered against your balance at TypeSafe Jev's real published usage price — request and response fields sit in the endpoint reference.
curl -X POST https://api.nativeport.ai/typesafe/v1/systemone \ -H "Authorization: Bearer $NATIVEPORT_API_KEY" \ -H "Content-Type: application/json" \ --data '{"state": "Help! My payouts have been failing for 3 days.", "model": "jev-latest", "questions": {"is_urgent": {"type": "noul", "instructions": "Does this convey urgency?"}}}'