Since ML outputs are usually "fuzzy", how does one check that the output is actually correct when using a technique like this in the field?