I don't want to bet on this either way, but I'd like to note that Leaderboard scores are computer using 30% of the data, whereas Accuracy Threshold defined under Information > Evaluation is computed using total number of members (100%). This makes it even harder to beat 0.4 accuracy. In any case, Round 1 Milestone Prizes are coming soon. Good luck fellow competitors!

This is to prevent overfitting.

A good algorythm can be expected to get similar scores on both the entire dataset and the 30% subset.  A bad algorythm will overfit the subset, and score poorly on the holdout.

I was under the impression that the 30% is not included in the scoring set - based on the graphic at the bottom of this page:

http://www.heritagehealthprize.com/c/hhp/Data

My reading of that suggests that only the unknowable data is going to be used for prize scoring (I hope that is the case).