Predicting downstream capabilities of frontier AI models with scale remains challenging due to the complex transformations involved in computing metrics like accuracy and Brier score. This paper investigates why these predictions are elusive, revealing that downstream metrics depend on comparing the correct choice against a small set of incorrect choices, making accurate predictions difficult. The study shows that these metrics are computed through a sequence of transformations that degrade the statistical relationship between performance and scale. By analyzing multiple model families and benchmarks, the research demonstrates that the probability mass on incorrect choices fluctuates with scale, affecting predictability. The findings suggest that while pretraining scaling laws are more predictable, downstream capabilities require understanding how probability mass on both correct and incorrect choices changes with scale. The paper emphasizes the importance of designing evaluations that account for these factors to reliably track the progression of frontier AI capabilities.Predicting downstream capabilities of frontier AI models with scale remains challenging due to the complex transformations involved in computing metrics like accuracy and Brier score. This paper investigates why these predictions are elusive, revealing that downstream metrics depend on comparing the correct choice against a small set of incorrect choices, making accurate predictions difficult. The study shows that these metrics are computed through a sequence of transformations that degrade the statistical relationship between performance and scale. By analyzing multiple model families and benchmarks, the research demonstrates that the probability mass on incorrect choices fluctuates with scale, affecting predictability. The findings suggest that while pretraining scaling laws are more predictable, downstream capabilities require understanding how probability mass on both correct and incorrect choices changes with scale. The paper emphasizes the importance of designing evaluations that account for these factors to reliably track the progression of frontier AI capabilities.