TypeSafe just released Jev, what they call a "System One" model. The name is from Kahneman: fast, intuitive, no reasoning. You send it structured data and typed questions, and it returns probabilities and classifications in 70-500ms. No text generation. The output is a number or a label, not a paragraph.
The pitch: snap judgments at $0.042 per million input tokens. That's roughly $0.00007 per decision. I wanted to know if those snap judgments are worth anything, so I pointed it at live prediction markets and let it run for 24 hours.
The pipeline
I built a system that ingests news from 40 RSS feeds and a Bluesky firehose, matches each item against live prediction markets on Polymarket and Kalshi, and runs four "engines" on every news-market pair: Jev with market prices in the prompt, Jev without prices, a keyword matcher, and a random baseline.
For each pair, Jev answers five questions in parallel: Is this news relevant to the market's resolution? Which direction should the price move? How big is the expected move? Is it already priced in? What's the probability the market resolves YES?
All five questions, one API call, 175ms, $0.00007.
What Jev classified
After 24 hours, Jev flagged 373 news-market pairs as strongly relevant out of 19,740 total. That's a 1.9% hit rate. The random baseline flagged 3.1%, and the keyword matcher flagged 7.9%. Jev is selective — it reads the news text against the market's resolution criteria and only flags when it sees a real connection.
The classification was consistent, too. When I ran the same request through Jev twice, the relevance label matched 90% of the time. Stripping out the market price and re-running changed almost nothing — 96.8% agreement — which means Jev wasn't anchoring on the current price. It was reading the news.
So the sorting works. The question is whether the sorted pile tells you anything about what happens next.
What Jev predicted
For the pairs Jev called "strong" and gave a direction, I checked whether the market price actually moved that way within an hour.
None of the engines beat a coin flip. Jev at 44.1% was slightly worse than random.
Why prediction markets were the wrong test
Before calling this a failure of the model, look at the data it was trying to predict. Out of 320 Jev "strong" signals with a direction, 175 showed zero price change after an hour. The average absolute move was 0.6 cents.
Prediction market prices update when a human manually places an order, not when the news hits. The liquid markets ($200k+ volume) barely moved at all — 0-6 cents per day on average. The markets that did move had thin liquidity ($20-60k), meaning a single trade can swing the price. There's nothing to predict on a one-hour timescale because the signal propagation mechanism is a person deciding to open a brokerage tab, and that can take days.
This is a structural problem, not a model problem. No engine — not Jev, not a frontier LLM, not a human — could reliably predict which direction a prediction market price will move in the next hour, because most of the time it doesn't move at all.
What I took away
The classification layer did its job. Jev sorted 19,740 items into buckets consistently, cheaply, and fast. But I was testing it on the wrong outcome variable. Trying to predict prediction market price direction from news is like trying to predict when someone will check their email from the subject line — the subject might determine what they do when they read it, but it tells you nothing about when they'll open the app.
The model can read. It can sort. It can tell you whether a document is relevant to a question. It did all of that for $2.63 in 24 hours. I just asked it to also tell the future, and it can't do that.
All code and data are at github.com/noreff/jev-signal.