Video models recover 18.1% of frequent events

Gemini 3.6 Flash recovered only 18.1% of true events in the high-count, high-frequency regime of a 2,190-video test, with just 0.2% of final counts correct. The trace-grounded study found that higher frame sampling lifted bouncing-ball answer accuracy from 19.6% to 29.3%, yet the reported event sequence matched ground truth only 3.7% of the time. Builders evaluating video systems should score timestamped event traces rather than final answers alone, especially before trusting the temporal fluency promised by new multimodal video architectures.