SocietyBench caps social forecasting at 75

SocietyBench’s strongest frontier model scored 75.0 out of 100 across 125 counterfactual social-event prediction points, only halfway through the headroom above its 50-point trivial anchor, according to the August 4 preprint. The benchmark anonymized entities and shifted dates across five completed events to frustrate recall; three agent frameworks failed to improve their shared base model, and per-event gaps reached 21.4 points. Builders should pair that result with the five-task harness swing found in coding-agent evaluation: scaffolding is not evidence of better judgment, and one event is too narrow to validate a forecasting system.