Models run locally, no data is sent outside.
A random 0–300 ms is added to each timeout. Dynamic reveal counts recognition time already spent.