From the moment a message lands to the moment a signal fires.

More than 1,000 sources, live. We add new ones whenever we find ones worth ingesting, and we don't have to rebuild anything to do it. We're also building our own ingestion in parallel, so we depend less on third parties over time. Today: more than 2,000 messages a day flow through this layer.

News in isolation rarely tells you much. To trade on it well, you need a model of the recent history: what has been said, by whom, and whether the message in front of you is original information or a relay of something that came out five minutes ago. The system maintains that history, refines its view as it goes, and filters out recycled inputs. Most of our competition still does this part by hand. We have automated it.

The model takes what just came in, weighs it against what the market already knew, and outputs a structured judgement: how much this moves things, in which direction, with how much confidence. Interpretability is built in. We always know which features mattered, and by how much.

For each message that crosses the bar, the model outputs two numbers: an expected value and an expected return. Trading desks take it from there. What they do with the numbers is up to them.

Every input we see is stored and labeled with the outcome that followed. The result, after years, is a deeply annotated dataset that no one else has. Each new version of our model learns from it. The dataset and the model are coupled, and they get harder to replicate together every month.