If you aren't willing to burn social capital there is a danger that you may be subject to positive test bias. So probably you should experiment in ways that you expect to make you do worse. If you're good at the coarse stuff and just want to make adjustments to the fine stuff, then this isn't much of a loss.

As an example of positive test bias, you might lump together two negative emotions as meaning "stay away," while in fact one of them requests help. If you always stay away, it's hard to learn this mistake. WrongBot's project is difficult because it mixes measurement of other people with intervention. Emotions, or at least facial expressions, are pretty simple and you probably could learn to label them from a book.