Sorry, I actually wrote the haiku finding logic in November and it's been running since then while we figured out the look of the tumblr and the moderation workflow.
There are a few awkwardness checks we do that disqualify some haikus. For instance, if the second or third line has a comma right near the front or back, or if there are month or title abbreviations, etc. But I'm always looking to refine it further.
According to the American Haiku Society (yes, it exists), the syllable count is actually less important than including a seasonal word and a "cut" between two different sets of imagery. But that's a little harder to teach a bit of hack code to do... so...
Yes, to clarify, I started with the base CMUdict for syllable counts, but I had the program keep track of any term misses it ran into. This way I could augment its vocabulary. It also helped me find some tokenization bugs and also try some rules for dealing with compound words like "unsportsmanlike"