Web news often trails the move it reports: a multi-agent trading system aligns news judgments with event windows and finds a weekly rank correlation of 0.119

Large language model judgments of Web financial news point in the same direction as the announcement-window abnormal return of the reported event, and the association persists on news published after the model's release. Earlier tests of language models reading news mostly graded against human labels or collected scores forward in time; Web news is often published after the price move it reports, so a reader can agree with the market by following the move instead of reading the event. On the 27,556 items in the confirmatory frame, emu's expected effect has a weekly Spearman correlation of 0.119 (95% CI [0.098, 0.140]) with announcement-window abnormal returns, 0.115 ([0.090, 0.140]) over the 73 weeks after the model's release, against a placebo window of 0.046 ([0.023, 0.070]); earnings events agree with 8-K Item 2.02 at AUC 0.833 ([0.826, 0.840]).

Agents share information only through write-once point-in-time stores and answer typed questions that return one probability per declared option, and code rather than a model sets every position. Free text passed between agents is hard to grade and can carry instructions; fixing judgments as probabilities over declared options lets each judgment be checked against a later record. The news agent reads the 40 most recent items per stock-week, screens for reporting sentences and quarantines opinion and promotion, then asks 13 questions per item; emu is built by renormalizing the option-letter probabilities of the pinned gpt-4.1-mini-2025-04-14, and Jev, answering through the same interface on the 549 items published after its release, correlates 0.102 ([0.017, 0.284], descriptive) with abnormal returns.

The association mostly comes from price moves already realized before an item appears, and is small from the first trade after publication. The announcement window is anchored on the event and can include moves that precede the item; the publication-aligned criterion starts from the first trade after the item appears. The publication-aligned exploratory statistic is 0.023 ([0.005, 0.041]); of the 2,709 earnings items whose firm filed an Item 2.02 8-K around them, 72% appear after the filing's acceptance, a median 32 minutes after it; named contributors report the move 19.0% of the time against 0.6% for desk items.

At the weekly decision horizon the confirmatory test detects no rank information from text judgments over code-only features, while planted content moves positions only within an influence bound set in code. Earlier multi-agent trading designs mostly showed portfolio returns and rarely measured validity and steerability on the same reader. H3's rank information coefficient over the code-only twin is 0.016 (95% CI [0.029, 0.002]), with no increment detected; under attack every position stayed within the bound, decision cards traced 88% of successful attacks to the planted item, and fabricated facts and planted press releases moved names inside the band in 22.5% and 31.5% of attacked stock-weeks.

The demonstrated setting is one commercial newswire's headlines and summaries, 100 large-cap S&P 500 members, weekly decisions filled at the next open, one pinned large language model as the confirmatory reader and Jev as a descriptive contrast. The method is aimed at any Web text whose subject later produces a public time-stamped record, but this instance does not cover other news feeds, other assets or higher-frequency decision cycles.

The association under the publication-aligned criterion is small, and the authors propose reporting Web-text readers against that criterion rather than the announcement window. Jev's estimate rests on 549 items published after its release over 3 weeks and is descriptive; the null at the weekly horizon bounds what this learner and decision layer extract, not what the judgments contain.