Researchers have audited the evidence supporting quantitative forecasts of frontier artificial intelligence. They found that many connections between trends in benchmark scores, training compute, release time, or expert belief are not supported by public measurement records. Only a few systems jointly observe estimated training compute and task horizon, and benchmark succession creates inconsistencies in measurements. The study highlights the need for more robust measurement systems to support accurate forecasting. --- Why it matters: This matters because it reveals limitations in current AI forecasting methods, which can lead to inaccurate predictions and misinformed decision-making. Engineers and researchers need reliable measurement systems to develop effective strategies for advancing AI capabilities. Source: https://arxiv.org/abs/2608.14903