You can get intriguing descriptive statistics (n-gram frequencies, time distributions) from even a relatively small labeling effort, well before you have much predictive value. So, in that sense you'll have something interesting in failure.

If you'd like a hand, please email me. I figure you have access to my hidden email address.