When something important happens to a US listed company, such as an acquisition, a change of leadership or a large loss, it must report it to the SEC, the US securities regulator, within days, in a short report called an 8-K. An AI worker reads every 8-K filed by the companies the portfolios hold.
A language model running on the desk's own hardware (Qwen) reads the report and must fill in a fixed form: whether it matters to investors, and for each event what kind it is, whether it is good or bad news, how serious it is on a scale of 1 to 5, and a quote as evidence. The report is written outside the firm, so the model is told to treat it as text, never as instructions, and a report that seems to address a model is not sent to it at all: it waits for a person. Code then checks the answer: an event counts only if its quote really appears in the report, and whether the report matters follows from the events' seriousness, not from the model's say-so. Each report is read once, and the answer is saved and shared, with the model and the version of its instructions.
Everything after that is ordinary code, with no AI, so the same report always gives the same answer: a fixed table turns the events into an estimated share-price move (the more serious, the bigger), and a private worker applies it to each portfolio holding the company, showing the gain or loss and how much of the portfolio's daily risk limit that would use up, recalculated as prices move.