Challenging, but solvable
The best zero-shot scores are 44.0% on Visual (GPT-5.5) and 49.3% on Tabular (GPT-5). Four of 13 models stay within three points of chance. People reach 94.8% with one minute per question. Heading (T4) and compass direction (T8), which depend most on how the two views are oriented, are the hardest.