@NathanMerrill Unfortunately, my answer there is no longer the score to beat. Ofri Raviv's Python entry comes in at 0.221. He's using a clever strategy wherein he generates a sparse hashtable with only the bins for titles where 11 is a terrible guess. If the hashtable misses, he returns 11.
The result is slightly better than my hashtable approach where the load factor is well over 5.
@PhiNotPi There really isn't any predictable correlation between problem titles and their scores. I looked briefly at length and keywords, and there's really nothing there to exploit. If you want to make a good book-judging contest, find some "cover" that has all kinds of rich, exploitable patterns in the data.
@PhiNotPi Anything with "titles" in it (as in: English) is probably going to wind up being a hard coding contest. Maybe something like "correlations between tags and vote counts" or something.
How about view counts and vote counts? Those would correlate very highly.