I agree, formatting details seem to be very context-specific and this is a lab setup, not in the wild. I would trust much more some AB test results from the NYT or any media organization - line length is one of the most obvious things to test!

I tested page-widths (which forces line length) on my own site, using n=109k visitors (so, quite a bit larger than n=20): http://www.gwern.net/AB%20testing#max-width-redux A wide - but not the widest - version performed best.

You've implied that 0 data points are better than 20. With 0 data points, consider the source: I will make a "model" to exactly reflect whatever my agenda is, rather than generating one at random.