Version 1.0 reports 1,199 evaluation tasks based on a panel of 1,244 U.S. adults surveyed from June to September 2026; 239 tasks are public and 960 are held out. Each task asks a system to predict a distribution over fixed answer options, scored against ground truth using total variation distance.
In the published results, six language models score between 76.2 and 80.6 distribution-agreement points. Rehearsals reports 94.1 for its Engram 1.0 system. The source says base models ran through a Kaggle harness, while Engram was assessed using Rehearsals' in-house harness with the same case and metric setup. The authors also report that LLMs select the top answer correctly 41% of the time on product questions, versus 74% on brand-perception questions. These are results reported by the benchmark's organizers, not an independent replication.