[jmares]

Data shows, as you say, that small scale performance is no indicator of large scale performance.

How then do you decide which projects are worth trying on the large scale?