Jev: TypeSafe’s System One Model (and Why Brands Care)

For a decade, “AI for marketing” mostly meant models that write: captions, emails, landing pages, endless variants of the same paragraph. Then, in mid-September 2026, TypeSafe AI shipped something that does the opposite.

Jev does not generate a single word.

Built by Diogo Almeida (ex-OpenAI) after years in stealth on a training approach they call RLCD, Jev is TypeSafe’s first System One model — what some people simply call a decision model. You hand it a slice of the world as text (an ad, a page, a ticket, a game state) and a list of bounded questions. It returns typed judgments: which option, where on a scale, how likely a yes/no is true — with probabilities and confidence, in tens to hundreds of milliseconds, at a cost that makes thousands of calls feel cheap. Input is billed. Output is free.

So it will not draft your UGC script. It will tell you, at category scale, which hooks deserve the script.

That is why it moved past research Twitter so quickly. Within a week of launch, builders had it playing Doom at about 10 decisions a second, routing browser agents, tearing down hundreds of live ads for cents, and showing up on OpenRouter. The stunts were different. The thesis was the same: most AI products generate language. Most software — and most growth work — needs judgments.

Would this video stop the scroll? Is the hook curiosity or confession? Does the ad match the landing page? Should we scale or kill? Who does ChatGPT cite instead of us? Those questions were always there. They were just too expensive to ask on every creative and every page.

This piece explains what Jev actually is, how Choice, Score, and Noul work, which growth jobs open up when judgment becomes abundant, and how a growth engine like Frevana can use it as a decision layer — without pretending a decision model replaces creative taste.

What Jev is (and is not)

Jev does not write. No ads, no emails, no summaries, no chain-of-thought essays. You send it some text (the “state”) and a set of typed questions. It returns typed answers with probabilities and a confidence score.

TypeSafe’s framing: unstructured state in, typed probabilistic decisions out.

Simon Willison’s shorthand: a frontier-intelligence function call.

A useful plain version: a very smart switch statement with 2026-level judgment.

It is named after the Jevons paradox: when something gets more efficient, consumption often rises. Make decisions cheap and fast, and software will make more of them. “System One” nods to Kahneman: fast, intuitive judgment, not slow deliberative writing.

The three primitives

One API call can attach many questions to one state. Independent questions run in parallel, so asking 20 things about an ad is closer in latency to asking one than to chaining 20 LLM calls.

Speed and cost

Public numbers that keep showing up:

• Roughly 70–500 ms end to end per call
• About $0.042 per million input tokens
• Output is free (no token tax for the answer)
• TypeSafe’s own workflow evals put Jev near mid-tier LLM agreement accuracy at a tiny fraction of the cost and latency

That combination is why “hundreds of ads in under a minute for pocket change” posts travel. The unit economics finally match map-reduce over a corpus — creatives, pages, prompts, tickets — instead of “ask ChatGPT about this one thing.”

What Jev cannot do

Be blunt with your team:

• It cannot generate copy, code, or explanations
• Input today is text only (a string, JSON, or a list of strings) — not raw video or images
• A Choice is a closed set, on the order of hundreds of options, not an open world
• “No hallucination” in TypeSafe’s sense means type-safe output, not “always correct.” It can still pick the wrong valid option
• Confidence is not magic. You still need evals on your data

If your job is “write the UGC script,” keep an LLM. If your job is “score 120 scripts on stop-scroll,” use Jev.

How a call works

STATE                 +    QUESTIONS              →    TYPED ANSWERS
(ad as JSON)               Choice / Score / Noul       probs + confidence

Example state for a support ticket, in the style of TypeSafe’s quickstarts: a complaint email. Example questions: a Choice for department (billing, technical, sales), and a Noul for whether it is urgent.

The same shape works for growth:

• State is one video ad turned into text: on-screen hook, spoken caption, format tags, offer, CTA, landing claim
• Questions are a stop-scroll score, a hook-type choice, and a Noul such as “would a cold scroller stop in 3 seconds?”
• Your code thresholds on confidence, then ranks, kills, briefs, or spends

The pattern that showed up after launch is sometimes called Jev Engineering:

LLM proposes. Jev decides. Code acts.

That split is the whole game for brands.

Useful Jev use cases with Frevana

These are the jobs where decision models pay off. The pattern is a corpus, a schema, and a place to act on the answers.

Score category ads and learn what to make next

Collect recent video or static ads in a category. Turn each into text state. Ask Jev for stop-scroll likelihood, hook type, format, trust, and offer clarity. Roll up winning hooks, winning formats, and top ads. Brands use that playbook to brief the next shoot, instead of guessing from five favorites on a mood board.

Run a focus-group-style stop or scroll pass at scale

Same corpus, many buyer personas: for each ad and persona, ask stop or scroll. Thousands of judgments in seconds. Useful when you want a directional read on attention before you spend on testing.

Audit a DTC site for conversion

Crawl key pages: home, collections, PDPs, FAQ, shipping. Score path-to-cart, CTA clarity, offer clarity, trust, and whether claims stay consistent. You get a conversion scorecard and prioritized fixes, instead of a long narrative audit nobody ships.

Catch ad and landing mismatches

For each creative, compare the ad promise with the landing hero, offer, and CTA. Flag mismatches before you scale a winner that breaks on the page.

Kill weak creatives before they hit the auction

When you can generate dozens of variants, use Jev to gate fatigue risk, claim risk, and cold-traffic readiness. Auto-promote high-confidence SCALE calls. Hold or kill the rest.

Check AI visibility across many prompts

For shopper questions across ChatGPT, Gemini, Perplexity, and similar: who gets cited, why you lost, and how ready your URL is. Classify and score first. Only then write the pages that close the gaps.

Batch-score ecommerce listings

Shopify or Amazon PDPs as state: conversion readiness, intent cluster, and whether title, image, and bullets agree. Fix the weak rows in bulk.

Route budgets and next actions

Scale, hold, cut, or refresh creative from campaign state, with confidence gates so only clear calls run on their own.

Keep agents cheap inside the loop

Model routing, “is the agent stuck?”, “is this tool call risky?” — the same primitive brands can use whenever growth software is agentic, not only in one-off research.

For a brand marketer, the transferable lesson is:

• Growth work is already a pile of bounded questions
• Those questions are usually asked too rarely because they were expensive
• If you can turn creatives and pages into state, you can ask them continuously

The growth jobs Jev is naturally good at

Anything that needs a paragraph of reasoning, original storytelling, or open-ended strategy still wants a generative model, and a human.

Bottom line

Jev helps brands grow when growth is reframed as throughput of good decisions:

• More creatives scored than a focus group can watch
• More pages audited than a typical site pass covers
• More kill and scale gates before money hits the auction
• More citation and listing checks than a human can click

It will not replace your creative director. It replaces the fake choice between “gut feel on ten ads” and “an LLM essay on every ad.”

LLMs write. Jev decides. Frevana acts.

That is the model. Frevana is one place to put the last step: connect ads, storefronts, and AI visibility, then ship the fixes. The first step is understanding Jev on its own terms.

Further reading

• Introducing System One Models and Jev — Diogo Almeida, TypeSafe
• What is Jev?
• Jev introduces a new shape of LLM — Simon Willison, September 21, 2026