How BlockSurvey uses JEV

Most of the decisions a survey tool makes are small. Is this answer real? Does this question ask for someone's name? Which question type fits this label? We make them thousands of times a day, and for a while we handed all of them to a large language model. This post explains why we now give many of them to a different kind of model called JEV.

System 1 and System 2: two ways of thinking

In Thinking, Fast and Slow, Daniel Kahneman describes two modes of thought. System 1 is fast, automatic and intuitive. You use it when you catch a ball, read the mood in a room, or spot a typo before you can say what is wrong. It runs on pattern recognition and takes almost no effort. System 2 is slow, deliberate and effortful. You use it to multiply 17 by 24, compare two mortgage offers, or plan a route through a city you don't know. It handles things you have never seen before, but it is slow and it tires you out.

A chess grandmaster glances at a board and knows who is winning. That is System 1. Then the grandmaster calculates a seven-move line to prove it. That is System 2. Good decisions use both, and the skill is picking the right one for the moment. System 2 on everything is slow and expensive. System 1 on everything produces confident mistakes on hard problems.

AI has the same split. Large language models, and reasoning models in particular, are System 2 tools. They write, explain, plan and work through problems step by step. You wait for the tokens, you pay for them, and what comes back is prose that your code has to parse. If all you need is a yes or no on "is this message a refund request?", that is a lot of machinery.

What is JEV?

JEV is the flagship model from TypeSafe and the first of what they call System One models, built for fast, structured decisions that software can use directly. You give it a state (text or JSON) and a set of typed questions. It does not write a reply. It returns typed answers with probabilities, and your code can branch on them, sort by them or route with them.

Those probabilities are trained to be calibrated. Across many predictions, things JEV scores at 0.8 should turn out true about 80% of the time. That holds for groups of predictions, not for any single answer, so the docs also give a confidence value that lets your code decide when to act and when to pass the case to a person or a slower model.

The current version is Jev 1.14 (jev-1.14.0), available since October 14, 2026. Jev 1.13 remains available.

The question types

A few properties worth knowing:

• Questions in one request run in parallel and independently, so one answer is never context for another.
• Input is text (Box also takes an image).
• JEV is not fine-tuned per customer. You shape it with the state, instructions and criteria you write.
• It is weak at exact arithmetic, date logic and long multi-step reasoning. The docs say to do those in code and keep JEV for the quick judgment.

Where we use JEV in BlockSurvey

A brief look. In each feature below, a decision comes before any writing.

AI follow-up questions

The language model still writes the follow-up question, but JEV judges the answer first. It checks whether the answer is a genuine attempt, whether the respondent wants to stop, and whether it matches the sentiment the survey creator wants to follow up on. If JEV is confident the answer is junk, like "asdfgh", or the respondent is done, the language model is never called. If JEV is slow or unavailable, we skip the check and call the language model anyway, so a JEV problem never blocks a respondent.

That gives four benefits. Each model handles what it is good at: JEV decides and the language model writes. The common case is cheaper and faster. Respondents stop getting earnest questions about "asdfgh", and nobody is pushed on after saying they are done. And we choose how sure JEV must be before it skips anything, and keep that bar high.

Read the full story: How AI follow-up decides whether and when to ask, and why that makes it faster and cheaper.

Making a survey anonymous

JEV reads each open-ended question and judges whether it asks for anything that could identify a person. The bar is strict on purpose. If JEV cannot answer, the survey is not marked anonymous.

Read the full story: How "Make it anonymous" uses AI decision-making to check a survey is really anonymous.

Choosing the question type in Create using AI

Simple rules settle obvious labels like email and phone. For the rest, JEV does a Choice over the types the builder supports. The earlier model remains the fallback.

Read the full story: How Create using AI decides the right question type for every question.

Scoring a survey before it goes out

Rubric items, such as wording clarity, answer options and duplicates, go to JEV in parallel batches. Our code combines the results, and an uncertain item counts for less. To change how much something matters, we change a weight in code, not a prompt.

Read the full story: How BlockSurvey Score uses AI decision-making to tell you if a survey is ready to publish.

Conclusion

In each of these features, a decision comes before any writing. When the work is a decision, such as junk or genuine, which type, whether it matches, a System 1 model answers faster and cheaper, with a probability you can set a threshold on. When the work is writing or reasoning, a language model stays in charge.

If you are building something similar, go through your pipeline and ask of each step whether it is a decision or a piece of writing. The decisions that fit the form "which of these is it?" are good candidates for a System 1 model.