Does it work beyond the example?

A system gets 99% of its training photos right but fails at night. Which test shows this?

The team wants to know if it works outside what it has already seen.