If I had to venture a guess, it would be "because that's how...

If I had to venture a guess, it would be "because that's how they coded their demo". I imagine that they probably put something together that would by default display the top inferred item with the greatest accuracy based on the results, and the code allows the user to 'reclassify' an image that is identified incorrectly.

If you look at some of their examples, such as the MobileNet/ImageNet inference, you can see how they list the top inferred items based on their accuracy. That's the great thing about their MetaTF Development Environment. Developers can play around with the existing samples and do their own testing in preparation. Smart move to get people acclimated to it before receiving the processors.

Just remember that what you are seeing is a prototype demonstration for them to sell the technology and most likely not representative of a commercial product using the IP or Akida processors that would be a lot more rhobust in commercial applications.