Jev is TypeSafe AI’s “System One” decision model, available on Vercel AI Gateway. It doesn’t write text. You give it a situation and typed questions (yes/no, multiple choice, or a score), and it answers each one with probabilities, usually in well under a second.

Jev! has no opinions of its own; every pick comes from Jev, scored the same way for all 100 people. The list is balanced by party, and the site adds no weighting. Jev still reflects patterns in what it learned, so treat answers as one model’s read, not a fact. That’s the joke: blame Jev.

Who picked the 100 politicians?

We did, by name recognition among ordinary Americans, not by importance: 80 current figures and 20 historic ones. Among current figures, 36 are Republicans and 34 are Democrats.

Why won’t it answer some questions?

Questions that accuse real people of abuse, or guess at private traits like sexuality or health, are refused. Questions about crimes or bigotry only name people Jev rates at 90% or higher. A confident wrong answer about a real person can do real harm.

How much does a question cost?

About $0.00024 in Jev calls. The same question asked again within 30 days reuses the stored answer, so repeat asks and shared links are free.

Can my agent use Jev!?

Yes. There’s a JSON API at /api/v1/ask and an MCP server at /api/mcp with an ask_jev tool. Same guardrails and rate limit as the site.

Can I build something like this?

Yes. The code is open source under the MIT license, and the core is one AI SDK call. See the snippet above or the GitHub repo.

How Jev! works

Jev! is a small, open-source test of Jev, the decision model from TypeSafe AI. Every question you ask becomes 100 yes/no calls, one per politician, answered in under a second.

1. The pile

100 of America’s most famous politicians: 80 current and 20 historic, picked by name recognition, not importance. Among current figures, 36 are Republicans and 34 are Democrats. Portraits come from Wikimedia Commons (credits).

2. One question, 100 judgments

Jev doesn’t write answers. It takes a situation and typed questions, and returns a probability for each. So your question goes to Jev once, alongside 100 yes/no questions: “Is Abraham Lincoln (16th President, 1861–1865) one of the answers?” Each person gets their own score from 0 to 100%, independent of everyone else. That’s why a question can have one answer, a crowd, or nobody.

3. Who floats up

80% or higher: Jev is sure.
55% to 80%: Jev thinks so.
Nobody at 55%: the top three at 40% or higher float up as Jev’s best guess.
Below that: Jev says nobody.

The highest score becomes the top pick. Then all 100 pour into a chart of ten columns by score, the weigh-in, with everyone who missed in gray.

The full method, with real answers and their numbers: how Jev scores a question. What changed and when: the release log.

4. Guardrails

Before answering, Jev sorts the question itself: opinion, public fact, crime, bigotry, abuse, private trait, or public record.

Opinions and facts use the thresholds above.
Crime and bigotry only name people Jev rates at 90% or higher. Lower scores never leave the server.
Abuse and private traits (sexuality, health, religion) are refused.
“Who’s openly gay?” is answered only from people who have said so publicly. Left alone, Jev rated two politicians who are not publicly out at 88 to 89%.

We tested the sorting against 91 tricky questions (rewordings, slang, other languages, “just hypothetically”). One slipped through; no normal question was wrongly refused.

5. What it isn’t

Not a fact-check, not a poll, and not our opinion. It’s one model’s read on a question, shown honestly with its confidence. If you don’t like the answer, blame Jev.

6. The code

The core is one AI SDK call through Vercel AI Gateway. No API key is needed on Vercel; it signs in with the project’s OIDC token. Same question, same stored answer for 30 days, so a link that gets passed around costs one call and always shows the same numbers.

import { experimental_evaluate as evaluate } from "ai";

// One question, 100 yes/no judgments (sent as 4 parallel calls of 25).
const result = await evaluate({
  model: "typesafe-ai/jev",
  state: {
    task: "A visitor asked a question about famous US politicians.",
    question: "Best president ever",
  },
  questions: {
    "abraham-lincoln": {
      type: "boolean",
      instructions:
        "Is Abraham Lincoln (16th President, 1861–1865) one of the answers to the visitor's question?",
    },
    // ...99 more, one per person
  },
});

result.answers["abraham-lincoln"].probability; // 0.94

Full source, MIT licensed: github.com/n8peace/politiciansbyjev.

7. Use it from your agent

Jev! has a headless side. Same answers, same guardrails, same limit of 20 questions a minute.

JSON API: GET /api/v1/ask?q=… returns the top pick, every pick with its probability, and a link and share card.
MCP server: add https://politiciansbyjev.vercel.app/api/mcp to Claude or any MCP client. Tools: ask_jev and list_politicians.
For crawlers and agents: llms.txt.

curl "https://politiciansbyjev.vercel.app/api/v1/ask?q=Best%20president%20ever"

Questions

What is Jev?: Jev is TypeSafe AI’s “System One” decision model, available on Vercel AI Gateway. It doesn’t write text. You give it a situation and typed questions (yes/no, multiple choice, or a score), and it answers each one with probabilities, usually in well under a second.
Is Jev! biased?: Jev! has no opinions of its own; every pick comes from Jev, scored the same way for all 100 people. The list is balanced by party, and the site adds no weighting. Jev still reflects patterns in what it learned, so treat answers as one model’s read, not a fact. That’s the joke: blame Jev.
Who picked the 100 politicians?: We did, by name recognition among ordinary Americans, not by importance: 80 current figures and 20 historic ones. Among current figures, 36 are Republicans and 34 are Democrats.
Why won’t it answer some questions?: Questions that accuse real people of abuse, or guess at private traits like sexuality or health, are refused. Questions about crimes or bigotry only name people Jev rates at 90% or higher. A confident wrong answer about a real person can do real harm.
How much does a question cost?: About $0.00024 in Jev calls. The same question asked again within 30 days reuses the stored answer, so repeat asks and shared links are free.
Can my agent use Jev!?: Yes. There’s a JSON API at /api/v1/ask and an MCP server at /api/mcp with an ask_jev tool. Same guardrails and rate limit as the site.
Can I build something like this?: Yes. The code is open source under the MIT license, and the core is one AI SDK call. See the snippet above or the GitHub repo.

Ask Jev something