A Thermometer, Not an Analyst: Lawful Uses of a Text Judge in a Live Trading Lab

A text model can sit next to a live trading system the way a thermometer sits next to a patient: it reads, it never prescribes. We follow one such model into a lab running an intraday strategy and a second, overnight strategy on one account. It is a small commercial judge, pinned to one version, that answers only typed questions about headlines. We describe six tests for a text signal: read a plain fact correctly against a real label, beat a simple count of its own headlines, show no memory of famous past events, give the same answer twice, always see the decision-time date, and check each forward verdict rule against random flags. They were assembled over three days, most of them after a failure, and we present them as the checklist we would now apply up front. We compare the judge with a general chatbot on blind rows, list what it may touch (stamps written after the close, counted forward) and what it may never touch (any number in the trading system), and report the plumbing failures that corrupted its readings. Its readings so far are small: one fact read well, one small effect below its bar, several that vanished; the fact was read under one setup, and its margin over a simple headline count is borderline. None of the six tests is new on its own; what is new is using them together on a text instrument beside a live system, and reporting every study it was used in, failures included. This is a paper about checking data and the instrument, not about profit: we make no claim about either strategy's returns. Working paper, also posted on SSRN (abstract 7589261). Abstract page with citation metadata: https://alexvidovich.com/papers/text-judge