According to CryptoSlate, a research preprint by Science Prediction Universe tested 13 agent configurations covering 12 models. Across 100 public experimental ranking tasks, the average Spearman rank correlation coefficient between available scored predictions was 0.52, and the average correlation coefficient with empirical results was 0.33. Following merged ranking, results for 73 tasks outperformed the average submission, though common errors may persist. The aforementioned correlation coefficients do not represent response accuracy; as coverage across tasks varies among configurations and the study did not establish a human prediction baseline, it cannot be inferred that these findings apply to the performance of all models or scientific prediction scenarios.