This puzzle is an example where letting us mutate just a few (like maybe 3) residues (of our choosing) might result in an even more stable fold. For example, residues 82-86 IRRRI - 2 oranges separated by 3 blues - make psipred think the final structure is a helix. They might make it fold as a helix too. But I'd be willing to bet it was designed as a sheet. Substituting an orange for the middle R would convince psipred that it was a sheet, and might make it actually fold into a sheet more reliably than the given sequence does.
You could give us the protein with a few mutations allowed as a separate puzzle, if you don't want to mess up the results of whatever experiment you are currently doing with us predicting foldit designs.
You could find any solutions that resemble the original design, run those mutated sequences through Rosetta, and compare the results using only backbone atoms to the original design, to see if you get better funnel-shaped graphs.