A story, and an invoice

Picture Kestrel Pay, a made-up payments app. Every support message, about two million a month, is read by a large language model that answers, in a friendly paragraph: "This looks like a billing issue, the customer seems frustrated, I would treat it as high priority." Then a bit of code tries to read that paragraph to work out which queue the ticket goes to. Sometimes the parsing breaks. Sometimes "high priority" turns into a 3 when the team meant a 4.

The engineer looking at the invoice notices something odd. The model is writing a whole paragraph to make a decision that fits in three words: billing, urgent, at risk. And the writing is the expensive part.

That observation is the whole reason typed-decision models exist.

The one-sentence version

A typed-decision model takes some text and a list of questions, and returns values: a category, a score, or a yes/no probability. It does not write sentences.

Jev, from a company called TypeSafe AI, is the first commercial one. Its makers call the category System One models (launch post), after the fast, gut-reaction thinking in Daniel Kahneman's Thinking, Fast and Slow. An LLM is closer to System Two: slow, wordy, deliberate.

   LLM
   ticket text ──► [ model writes a paragraph ] ──► your code parses it ──► queue
                     (slow, priced by output tokens)   (can break)

   Typed-decision model
   ticket text + questions ──► [ model returns values ] ──► queue
                     (fast, output is free at list price)   (always the shape you asked for)

Why anyone should care

Four things, in order of how much they matter:

1. The answer is always the right shape. No parsing paragraphs. Your code branches on queue == "billing".
2. It is fast. The vendor says 70 to 500 milliseconds; independent write-ups measured a median around a quarter of a second (BestHub, Flowtivity). Fast enough for a user waiting on a screen.
3. It is cheap per decision, at least at TypeSafe's list price. Chapter 2 does the sums properly.
4. It says how sure it is. That lets you send only the unsure cases to an expensive model. (Chapter 6 explains why you must not take that "sure" at face value.)

The name is a small joke. It nods to the economist William Stanley Jevons, whose paradox says that when something gets cheaper to use, people use more of it. If asking a question of a model costs almost nothing, you start asking it about everything.

How well proven is it?

Here is the honest part, and the reason for this course's shape. Every headline claim, checked against what independent people found in the first ten days. Filter by label. Open the ones that surprise you.

If you take one thing from this chapter: "0 hallucinations" and "0 type errors" describe the shape of the answer, not whether the answer is right. A perfectly formatted wrong decision is still a wrong decision, and in a payments or healthcare system that is the one that matters.

What this course promises

How the chapters fit

  0 Start here ─► 1 What a typed decision is ─► 2 Does it save money?
                                                       │
        ┌──────────────────────────────────────────────┤
        ▼                    ▼                         ▼
  3 Triage cascade   4 Guardrails         5 Decompose, filter, route
        └──────────────────────┬───────────────────────┘
                               ▼
                  6 Trusting the numbers ─► 7 When NOT to use it
                               │
                  8 Jev vs Laya ─► 9 Your projects ─► 10 Study guide

It leans on the harness ideas from Agentic AI Harness Patterns: the approval gate, the circuit breaker, the shadow evaluation harness, and the fallback ladder. A typed-decision model does not replace those. It plugs into them.

Sources

All pages read on 25 September 2026. Figures were pulled through a summarising fetch tool, so check any number on the linked page before you quote it.

• TypeSafe AI, Introducing System One Models & Jev. Vendor source: architecture, pricing, benchmark claims and stated caveats.
• The Register, TypeSafe AI debuts model for machines that plays Doom, 16 Sept 2026. Press coverage with a sceptical note on "hallucination-free".
• Cloudflare, Jev model page. API shape, context limit, availability label.
• xbill, Jev after eight days of independent tests, DEV Community, 24 Sept 2026. Independent synthesis of several studies.
• BestHub, Laya vs Jev and Flowtivity, Laya, benchmarked honestly. Latency measurements.
• jevtypesafeai.com, Pricing. A reseller that states it is independent of TypeSafe AI.