the label is the news: recovering 80% of a news reader’s score gain

On Agenthon's official Development board, a label-only numerical forecast recovers 80.3% of the observed gain from adding a language-model news reader to the same numerical base. Widening the 22 shock-family forecasts by 1.5, with zero documents and zero additional model calls, moves board loss from 0.9936 to 0.9701, against 0.9643 for the reader. The reader and widening gains have the same sign on 18 of 19 cards whose displayed reader score changes. A crossed document--instruction experiment locates a large scale effect in one prompt sentence. In two cached reply sets, combining the reader's full forecast mean with fixed numerical width scores above either original arm. The label's value follows its windows: the same base and widening lose 0.224 per origin on 232 public pre-shock origins. On a further frozen comparison of 81 windows on 79 cards, the numerical base outperforms full-document and excerpt adjustments on the 43 newly analyzed cards: the adjustments' two-tail gains against the base are -0.003808 and -0.008461. Among 12 presaved numerical rules, v3 selects center and width strengths separately and improves the complete 81-window cohort's mean on all five scores, including CRPS by +0.010048. Together, the label-only comparator, the reusable reader-center hybrid and the numerical rules give participants concrete ways to improve forecast distributions and measure what document reading adds.