[gojomo]

I can't claim to know Google's actual procedures, but if you have a system with many thousands of candidate signals, and you train it to achieve certain 'quality' rankings via a largely automated process, it's possible that even the designers of the system would not know how every signal, in every combination, affects the rankings -- without researching very specific conjectures one at a time.

To wildly hypothesize, what if length-of-registration tends to indicate more-beloved results in most .COM domains, but has no impact on .ORG and .GOV domains, and actually indicates less-beloved results in .COM domains that consist of generic common words (those loved by professional domainers who then lengthen their registrations strategically)? A learning system could discover these conditional relationships over time -- but they'd resist easy explanation by a employee summarizing factors via informal channels.