What Is Jev? A Practical Guide to Decision Models for Marketing

Jev is a decision model from TypeSafe AI, released in September 2026. Instead of writing text, it reads a piece of context, such as an agency note or a support ticket, answers a question you define, and returns a probability for each option. For marketing teams, that makes a judgment classifier you can set up with plain-English questions, without training a model first.

This guide covers what Jev is, the five types of decision modeling it sits next to, which marketing workflows it fits and which it doesn't, and how to test it on your own data before you trust a single answer.

What is Jev?

Jev is a hosted model built for one job: answer a named question about some input and return typed results. TypeSafe describes it as "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." TypeSafe calls this category System One models. The name Jev is a nod to William Stanley Jevons, not an acronym.

You define the question and the allowed answers. Jev returns one of three kinds of result:

• Yes or no. Is the statement true for this input? Example: "Did this note report a tracking change?"
• Rating. Where does this input fall on a scale you define? Example: "Rate the risk of creative fatigue from 1 to 5."
• Choice. Which of these options fits best? Example: "Which cause is most likely: creative fatigue, stock constraint, tracking change, or inconclusive?" One documented setup supports up to 255 options.

For a choice question, you get back the selected answer, a probability for every option, and a confidence value. Nothing has to be parsed out of a paragraph.

Published details: Jev was in early access at launch. TypeSafe lists input pricing at $0.042 per million tokens, with output tokens free. TypeSafe also publishes latency of 70 to 500 milliseconds per call. These are the company's own figures, not independent benchmarks.

Is Jev an LLM?

No. A large language model generates text, and Jev does not. Thoughtworks puts the distinction this way: an LLM generates text where novelty and unpredictability are features, while Jev delivers only structured, probabilistic decisions.

Two cautions matter for marketers:

1. A valid format is not a correct answer. TypeSafe says Jev "can't hallucinate," which means the output always matches the schema you defined. It does not mean the answer is right. WWT makes the point directly: a structured answer doesn't magically make the answer correct, and independent testing has found performance differences depending on the task.
2. Speed and price claims need your own test. TypeSafe says it can't prove its pricing isn't subsidized, and expects its headline speedups to be "on the higher end of real world gains."

Five types of decision modeling, and when to use each

"Decision model" covers several different jobs. Most vendor pitches blur them together. Separate them before you choose a tool.

Google describes Meridian as built to answer three questions: historical ROI and contribution by channel, response curves by channel, and how to allocate future budget to maximize outcomes. It is designed to estimate the causal impact of marketing. Causal and optimization models answer those questions. A Jev-style model does not.

The useful way to hold all five at once: the decision model interprets the situation, the causal model estimates what an action will do, the optimizer chooses among feasible actions, and deterministic code enforces the rules. Each piece answers a different question, so check each one separately.

Where Jev works and where it doesn't

Here are the workflows marketing teams actually run, with a verdict for each and the reason.

The rule behind the ticks: Jev is useful when the input is text, the possible answers can be written in advance, and a wrong answer leads to a review, not a budget move. When any of those three is missing, pick another tool.

The CFO test: can you trust an answer like "58% tracking change"?

A CFO asks: spend went up 30% last month, and revenue stayed flat. Why?

That question is causal, so Jev can't answer it directly. What it can do is help you check the candidate causes quickly, if you give it the evidence. Here is how that plays out. The numbers are illustrative, not real Jev output.

Step 1: Turn the question into candidate causes. Write a choice question with mutually exclusive options:

• Tracking or measurement change
• Stock-out during peak weeks
• Creative fatigue
• Price or promotion change
• Platform over-claiming (the platform reports sales it didn't drive)
• Inconclusive

Step 2: Give it the evidence, not the conclusion. Pass in the week's agency notes, the checkout change log, and the inventory summary, as text. Don't pass in the revenue numbers and ask "why."

Step 3: Read the output. Suppose it returns:

Can you trust it? Not as an explanation. You can trust it as a ranking of what to check first. A 58% probability on tracking means "check tracking first," not "tracking broke." Four checks decide whether to act on it:

1. Calibration. On past cases, were the answers marked around 58% right about 58% of the time? If the 60% bucket is right far less often, don't read the score as a probability.
2. Evidence. Does the text actually mention a tracking change? If the output is high but the evidence is thin, the model may be guessing from the wording.
3. Independent confirmation. Pull the event logs. If purchase events dropped on the day the checkout changed, the finding holds. If the logs are clean, the 58% was a lead that didn't pan out.
4. Robustness. Rephrase the question and shuffle the option order. If the top answer flips, don't use it.

The decision rule: act on it, meaning hold spend increases and open an investigation, when the top answer clears your threshold and an independent check confirms it. Otherwise treat it as a lead.

What to tell the CFO: "The model ranks tracking first. We are checking the event logs. Spend increases are on hold until we confirm. If the logs confirm it, the revenue gap is partly measurement, and we'll run a lift test to size it." That answer is accurate because it separates the ranking from the causal estimate.

How to use a Jev-like decision model

This section is the playbook. Work through it in order.

1. Pick a scenario

Choose one judgment that someone reads every week, and check it against three filters:

• Is the input text? If the facts already sit in clean fields, use a rule.
• Can the answers be written in advance? If people argue about what the categories mean, fix that first.
• Does a wrong answer lead to a review or a budget move? Start with reviews. Only let the model move money after it has passed the checks below.

Good first scenarios: tracking-change notes, promo-end notes, cause sorting for flagged accounts, creative fatigue ratings.

2. Write the question and the answers

• One question per judgment. Don't bundle "cause" and "action" together.
• Answers are mutually exclusive, with an "inconclusive" or "unclear" option.
• Each answer is something a reviewer could confirm from the text or from another data source.

3. Gather the evidence and the independent checks

Decide what goes into the input (the note, the change log, the inventory summary) and what you'll check against afterward (event logs, inventory system, spend data). Jev can't verify facts. Your data can.

4. Label 100 to 200 past cases

Have two people label the same cases independently. Count how often they disagree. Where they disagree, the question is unclear. Fix the definition and label again before you test the model. Labeling is the most valuable week you'll spend on this.

5. Check the criteria

Run the model on the labeled set and check:

• Accuracy at the threshold you'll use. Track missed flags and false alarms at that threshold, not just overall accuracy.
• Calibration. Group predictions into buckets (0 to 20%, 20 to 40%, and so on) and compare each bucket with how often it was right. A confidence score you can't rely on is worse than no score.
• Robustness. Rephrase the question and shuffle the options. Count how often the top answer changes.
• Held-out period. Test on cases from a time period you didn't use to write the definitions.
• Enough cases. Small tests swing a lot. PayPal's team found that retraining on a later data release moved one model's hard-tier accuracy by 11.7 points. Don't draw conclusions from 20 cases.

6. Set the rule

Write down what each answer triggers, with thresholds. For example: if "tracking change" is above your threshold, hold spend increases on that account and open an investigation. The model changes what the team looks at. The rule changes where money goes.

7. Add what's missing

• Verification sources. The event logs, inventory system, and spend data that confirm or reject the answer.
• An audit log. For every case: the input, the output with probabilities, the rule that fired, and what happened next.
• An owner for each rule. Someone who reviews the rule and its thresholds on a fixed schedule.
• Exposure limits. A cap on how much spend any automated rule can move per campaign per day.

8. Run in shadow mode, then check weekly

Run for two to four weeks without changing spend. Each week, read a sample of the "no" answers, track the disagreement rate with human labels, and check whether misses cluster in one category. Re-run calibration every quarter, and again whenever the team changes how it writes notes, because the inputs drift.

What the research says

• Cisco's content-safety test found that Jev, used zero-shot, beat a 31-billion-parameter LLM judge on overall ranking on seven of eight benchmarks, but it trailed a purpose-trained classifier at strict false-positive budgets. Cisco sells a competing model and its labels were LLM-generated, so treat the numbers as directional.
• PayPal's team reported that an in-house typed judge cost about 1% of a frontier LLM judge on their workload. They also found stored LLM-judge labels matched a fresh consensus on only 26.9% of sessions where two judges agreed. Label quality mattered more than model choice.
• Anthropic's formal verification work makes the same point from a different field: a proof checker can confirm a program meets its specification, but it cannot tell you whether the specification matches reality. Your labeled set plays the role of the specification here.

FAQ: what can I do now?

Can I test Jev on my own data this week?

Yes, and you don't need code to start. Pull 100 past agency notes, write one yes/no question, have two people label the notes, and compare. TypeSafe is in early access, so request access at typesafe.ai before you plan a rollout.

My CFO asks why spend didn't turn into revenue. What do I send back?

Send a ranked list of candidate causes with the evidence you checked, not a single answer. Use the four checks in the CFO test above, and say which causes are confirmed, which are open, and what the next check is. Then size the gap with a lift test or a marketing mix model.

Which workflow should I automate first?

Pick the one where someone reads text every week and a wrong answer means a review, not a budget change. Tracking-change triage is the usual first choice.

Should I use Jev or a general LLM as the judge?

Run both on your labeled set. Compare the calibration and the false alarms at the threshold you'll use, not overall accuracy. The published tests are close on overall ranking and mixed at strict thresholds, so your own data decides.

My notes are inconsistent and my labels disagree. Is it worth starting?

Start with the labeling. Disagreement between labelers tells you the question is unclear. Fixing the definition is useful even if you never use a model.

Can Jev decide where the budget goes?

No. It can route and classify. Budget decisions need a causal estimate, an optimizer, and a policy you wrote down.

Want a second pair of eyes on your labels and thresholds?

If you'd like help designing the labeled set and the rules for one scenario, talk to us at niti.ai/talk-to-us.

References

• TypeSafe AI, "Introducing System One Models & Jev," September 15, 2026. https://typesafe.ai/blog/introducing-system-one-models-and-jev
• TechCrunch, "A new kind of AI model from a ChatGPT inventor is thrilling developers," September 18, 2026. https://techcrunch.com/2026/09/18/a-new-kind-of-ai-model-from-a-chatgpt-inventor-is-thrilling-developers
• Simon Willison, "Jev introduces a new shape of LLM: System One, aka Decision Models," September 21, 2026. https://simonwillison.net/2026/Sep/21/jev
• TechTarget, "Jev decision model touted as quicker, cheaper LLM alternative," September 22, 2026. https://www.techtarget.com/news/Jev-decision-model-touted-as-quicker-cheaper-LLM-alternative/856949/
• Thoughtworks, "Jev and the rise of decision models: what you need to know and why they matter," 2026. https://www.thoughtworks.com/en-us/insights/blog/generative-ai/jev-decision-models-what-need-know-why
• WWT, "The Rise of Decision Models," 2026. https://www.wwt.com/blog/the-rise-of-decision-models
• Braintrust, "What is Jev? A guide to AI that makes decisions," September 26, 2026. https://www.braintrust.dev/articles/what-is-jev
• Cisco Blogs, "Jev: Decision Models on Trial," October 5, 2026. https://blogs.cisco.com/ai/jev-decision-models-on-trial
• Weng et al., "Our Own Jev: Typed Decisions in One Forward Pass, from an In-House Judge to Open LoRA Models," arXiv:2610.29052, October 2026. https://arxiv.org/abs/2610.29052
• Google, Meridian documentation, "Rationale for causal inference and Bayesian modeling," 2026. https://developers.google.com/meridian/docs/causal-inference/rationale-for-causal-inference-and-bayesian-modeling
• Konitzer, "Media Measurement and the Assisted Own Goal," arXiv:2607.09608, June 2026. https://arxiv.org/html/2607.09608v1
• "When Offline Evaluation Misleads: A Diagnostic Protocol for Reward and Policy Selection in Delayed-Feedback Contextual Bandits," arXiv:2608.11560, August 2026. https://arxiv.org/abs/2608.11560
• Anthropic, research post on using Claude to formally verify open-source software, 2026. https://www.anthropic.com/research/formal-verification

Jev is a product of TypeSafe AI. This guide is independent analysis. Niti AI is not affiliated with TypeSafe AI.