Lesson 1 · ~10 minutes

Jev: the smart `if` statement

What Jev is, its three question types, and how to wire one decision into a TypeScript product.

Your win for today: look at any product feature and say whether a decision belongs to Jev or to an LLM, then write the Jev call and the threshold code for it.

0 · Four words first

📦stateThe data you hand Jev, such as a ticket or a message.

❓questionA named, typed ask with a fixed set of possible answers.

📊probabilitiesHow likely each allowed answer is.

🎯confidenceWhether there's one clear winner or the weight is spread out.

1 · What it is

Jev is a model from TypeSafe AI. It returns typed decisions, not generated text.¹ TypeSafe calls it a System One model: it's built for fast, automatic judgments that your software makes thousands of times, not for long chains of reasoning.¹ You decide the possible answers before you call it. The policy that acts on the answer stays in your code.²

Watch one call run, or scrub through it. Press ▶, use the chapter buttons, or turn on 🔇 → 🔊 for narration:

The numbers come from the request/response example in Firecrawl's guide.¹ Paused, the final frame works as a static diagram.

	Jev	Typical LLM
Output	Typed answer + probabilities	Free text (you parse it)
Latency	70–500 ms, most ~100 ms²	Seconds
Price	$0.042 / M input tokens, output free²	$0.20–$10 / M¹

2 · Three question types

Which option from my list? Returns the winning option, a probability for each option, and confidence.

Where on my rubric? Returns a position on the scale, plus probabilities and confidence.

Is this statement true? Returns a single probability from 0 to 1.

3 · The call, annotated

import { choice, TypeSafeClient } from '@typesafe-ai/sdk'
const client = new TypeSafeClient()

const { answers } = await client.systemOne({ 1
  state: { ticket: 'Export button crashes in Safari' }, 2
  questions: {
    category: choice('What kind of ticket?', { 3
      bug_report: 'Something is broken',
      feature_request: 'New feature request',
    }),
  },
})
// answers.category.choice      → 'bug_report' 4
// answers.category.confidence  → 0.95

systemOne makes one call. The raw endpoint is POST https://api.typesafe.ai/v1/systemone.²
The state: only what the decision needs.
The key (category) names the answer. The object lists the only answers Jev is allowed to return, and each description is read literally.²
A typed result. There's no parsing, and an answer outside your list is impossible.²

4 · Confidence → bands

A probability belongs to one option. Confidence describes the whole distribution.³ Both answers below have a winner, but only one can be trusted to act on alone. Use the toggle:

The 0.61 / 0.35 split is Copes' example: "returns" wins, but the ticket really touches two teams.³ Toggle to an illustrative clear-cut ticket and watch the weight move to one option as confidence rises.

Your code turns confidence into an action using two lines:³

Where 0.42 lands: below the confirm line, so a human triages it.

const ASSIGN = { act: 0.8, confirm: 0.5 }

const c = department.confidence
if (c < ASSIGN.confirm) return { action: 'triage' }                 // human handles
if (c < ASSIGN.act)     return { action: 'suggest', team: department.choice }
return { action: 'assign', team: department.choice }               // automatic

Don't copy the 0.8 / 0.5 numbers. Run Jev over labeled examples, measure accuracy in each confidence bucket, and set the lines where accuracy meets the error cost you can live with (for example, 90%+ for refunds).³

Zoom out: where the decision lives

TypeSafe's own design guide sums this up as "keeping code in control and giving System One narrow, structured decisions."⁵

5 · Jev or LLM?

Also prefer an LLM or a human for rare, high-stakes calls. Keep the state under 64k tokens and trim it.¹ Jev can't generate text and struggles with arithmetic and dates.¹²

In AI dev workflows, Jev works well as the fast gate around an LLM agent. It can decide which tool or path to take, whether an output passes a check, or whether to escalate. The LLM does the generating.

Practice: from memory

A · Place the marker

The zones are hidden. Remember where the lines are, pick the band, and repeat a few times.

B · Fill the faded example

This is the same call as in section 3, with parts blanked out. Fill it in without scrolling up.

const { answers } = await client.({
  : { ticket },
  : {
    team: ('Which team?', { billing: '…', technical: '…' }),
  },
})
if (answers.team. < 0.5) return { action: '' }

C · Quick quiz

1. Route each support ticket to billing, technical, or sales.

2. "This comment is spam." True or false?

3. Rate how severe a bug report is, from cosmetic to outage.

4. confidence measures the…

5. You need to know whether a cart total is over $500. How do you get that answer?

6. Your agent needs to draft a reply to the customer. Which model should do that?

🏆 Milestone challenge: Pick Jev vs LLM + Write solid calls

Pick one real feature from your product or agent workflow. Write the following and paste it back to me in chat:

The questions object: at least one choice and one noul.
The state you'll send, trimmed to only what the decision needs.
Your band lines, and what happens in each band. A rough sketch of your own band diagram is welcome too.

I'll review it like a code review. If it passes, I'll mark the milestones mastered on your dashboard.

Primary source: read the official TypeSafe docs next, starting with the Quick Start. Then read Flavio Copes on confidence, which covers setting thresholds in production.

Access note: direct sign-ups have been paused at times (one source reports a pause on 22 Sep 2026).¹ Jev is also available through the Vercel AI Gateway as typesafe-ai/jev, and through OpenRouter.

Stuck or curious? Ask me follow-up questions in the terminal. I'm your teacher for this, and nothing is too basic to ask.

Jev: the smart if statement