Most news feeds dump headlines on you.

Most news feeds dump headlines on you. Ours reads them first. Every headline that hits Tradeul is now typed in about 300 ms: what kind of catalyst it is, the detail, and whether it's confirmed or still a rumor. Here's how we built it. The raw material In the last two months our engine read 157,162 news items from 89 publishers, covering 12,621 tickers. Most of it arrives the way news always does: duplicated across feeds, full of noise, and often late. 1. Ingest Several feeds race for every story. The first copy wins and the rest are discarded. We read wires like PR Newswire, GlobeNewswire and Newsfile directly, racing their feeds against each other. 2. Clean Cross-source duplicates and translated reposts are dropped. Class-action notices get their own filter, so they never bury real catalysts. Tickers are pulled from the text itself, even when a release comes from a third party that only mentions the company. We measured that extractor on 29,290 releases: 95% recall. 3. Structure A definitive deal is not an FDA rejection, and neither is an ATM offering. So we built a catalyst taxonomy for traders: 17 types and 95 subtypes, from M&A · Definitive deal to Offering · ATM / shelf to FDA · CRL. Built for one-glance decisions. 4. Enter jev Every headline then goes to jev, a model from @typesafeai trained for calibrated, typed decisions. In a single call it returns the type, the detail and whether the story is confirmed or still a rumor, each with a probability. Median time: 296 ms. 5. Rumor or fact Traders price a report and a done deal very differently. jev reads how a headline is sourced and phrased, and flags what isn't confirmed yet. Since May it has flagged 1,259 headlines on our feed as rumors. In M&A, 4 out of every 10 were. 6. Precision over noise On a labeled benchmark of 395 headlines, jev picked the right type 97% of the time. When its confidence was 0.80 or higher, it got 354 out of 354 right. And when it isn't sure, the cell stays empty. A blank beats a wrong label. The result 17,096 headlines on our desk feed (OPN) are already typed, and every new one is typed live as it lands, with filters by type and by rumor. The News window is next. Every headline. Typed. In 300 ms. https://t.co/LkI6gJJZ53