A gestures is recognized properly in training mode but not when it is actually executed. Does training mode work differently? If so, how to make the actual execution similar to training mode?
You can try increasing the Match Probability in the Preferences tab if you have similar, but different gestures.
Also, note that each time you train, it adds another set of point patterns for that gesture. If you feel things are inconsistent, deleting and recreating the gesture will clear all point patterns for the gesture. So if you're having a conflict between two, delete them both and recreate to see if that helps.
The Match Precision can also be increased, this affects how many interpolation points there are made during the comparison.