Jev for Trading: What Developers Actually Build

Two things are worth knowing before you use Jev for trading, or for anything else that touches money. TypeSafe's own documentation does not mention finance, markets, stocks or trading anywhere — every financial use of it is community practice, not a vendor-supported feature. And the one experiment set up to test whether it actually improves a trading signal has not reported results yet.

What exists instead is a consistent architectural pattern, and it is the useful part.

The pattern: Jev judges, code executes

A survey of Jev finance and trading projects, taken on 20 September 2026, found 14 projects and one rule holding across all of them: "Jev judges, code executes." Jev returns only typed judgments over compact state. Thresholds, risk vetoes and order placement stay in ordinary deterministic code.

That division is not a stylistic preference. Jev answers in three shapes and none of them is an instruction:

Question type	What it returns
Choice	one option from a list you define
Score	a position on a rubric you define
Noul	the probability, 0 to 1, that a statement is true

A model that returns 0.72 cannot place an order. Something has to decide what 0.72 means, and in every surveyed project that something is code with a number in it that a human chose. See our Jev review for the full picture of what the model does, and what is Jev for how System One models differ from the chat models you are used to.

What the 14 projects actually do

The survey splits them in two.

Eight live trading systems, doing crypto market-making on Monad/Kuru, Hyperliquid and Kraken. These produce buy, sell or hold decisions and execute through post-only limit orders, written in Python and in TypeScript on Bun. Worth noting: they ship with paper-trading defaults. The authors did not point them at live money as the out-of-the-box behaviour.

Six forecasting and analysis projects, including stock-direction work on Hong Kong equities, portfolio signal generation from multi-question analysis, and two classification results that are easier to judge because they have benchmarks attached:

• Banking intent classification on the BANKING77 dataset: 92.40% accuracy
• Tax document classification on IRS forms: 100% accuracy at $0.001 a page

Those two are the clearest evidence of fitness for purpose, and notice what they have in common. Both are document classification with a known right answer. Neither is a prediction about a price.

The survey also records what nobody had built as of 20 September: no prediction-market projects, no arbitrage, nothing dedicated to forex, nothing DeFi-specific.

The experiment that would settle it has no results

The most honest thing to report is an absence. An open experiment on a stock portfolio manager is testing exactly the question a sceptical developer would ask: does Jev relevance-filtering make a news-sentiment signal better?

It uses Jev three ways — Score for ordinal relevance on a 1–5 scale, Choice for event classification, Noul for binary detection such as "is this a price recap?" — and pre-registers two hypotheses:

1. Jev's relevance scores agree with human judgement well enough to filter articles automatically, and better than cheap baselines.
2. Filtering or weighting by relevance improves the signal's predictive power over treating all articles equally.

As of writing it is at Phase 0, design. No measurements. The rubric, gold labels, metrics and success thresholds were all registered before any results exist, which is a deliberate guard against fishing for a favourable outcome afterwards — and a good sign about the rigour, while telling you plainly that the question is open.

So: the pattern is widespread, two classification benchmarks look strong, and whether any of it improves a trading signal is unmeasured.

Why the pricing suits classification work

Jev costs $0.042 per million input tokens, output tokens are free, and there is a 64k-token budget per request. For a workload that reads a lot and answers in a word — grading a thousand headlines for relevance — that shape is unusually favourable, because the volume sits entirely on the cheap side.

That is the real reason these projects exist. Running a frontier model over every article to extract one label is the thing Jev displaces, and our how to use Jev guide covers the request format if you want to try it on your own corpus.

Using Jev for trading: four things worth copying

Four things the surveyed projects do that are worth copying if you are building something that touches money:

1. Keep every threshold in your code, not in a prompt. The model gives you a number; you decide what it means.
2. Give Jev compact state, one question at a time. TypeSafe's own guidance is that System One models work best when each question asks one specific, well-scoped thing.
3. Default to paper trading. The live systems in the survey do.
4. Benchmark on labelled data before trusting it. The two credible results above are both cases where someone could check the answer.

And one thing to be clear-eyed about: none of this is investment advice, and nothing in Jev's documentation claims it can forecast a price. It classifies text. Whether classified text predicts returns is the open question above.

For the wider picture of AI in this sector, see our AI tools for finance page, and everything we have published about Jev for the rest of the cluster.

FAQ

Can Jev predict stock prices?

There is no evidence it can, and TypeSafe does not claim it. Jev returns typed judgments about text you give it — a choice, a rubric score, or a probability that a statement is true. Projects use it to classify news and filings; turning a classification into a price expectation is a separate step, done in code, and unvalidated so far.

What is the "Jev judges, code executes" pattern?

The arrangement found across all 14 projects in the September 2026 survey: Jev supplies only typed judgments over compact state, while thresholds, risk vetoes and order placement live in deterministic code. The model never issues an instruction.

Is Jev accurate enough for financial classification?

On two benchmarked tasks, yes: 92.40% accuracy on BANKING77 intent classification, and 100% on IRS tax-form classification at $0.001 a page. Both are document classification with known right answers, which is the kind of task the results support.

Does TypeSafe support using Jev for trading?

Not explicitly. Its documentation makes no mention of finance, markets, stocks or trading. Every financial application described here comes from community projects, so you are outside documented territory.

How much does Jev cost for high-volume classification?

$0.042 per million input tokens with free output tokens, and a 64k-token budget per request. Because the cost sits on the input side and the answers are single values, grading large volumes of text is cheap relative to asking a chat model for the same labels.