This question is actually the topic of my masters research which will be published Summer of '15.
The top probability estimates are as follows in order of our likelihood:
1. Private for-profit company
2. State (Government/Military) laboratory
3. University laboratory
4. Open source project
5. Lone Wolf developer
These are based largely on where incentives for AGI implementation lie, where funding is most available and where activity is occurring.
I can imagine a situation where someone suspects that one of these projects listed above is "superintelligent". What would be the test of that?
The AGI community hasn't agreed on a test or metric, however the best answer I have found in a study is the work of José Hernández-Orallo and his "Anytime intelligence test" [1]. Practically though, I personally think we will know once enough people are convinced, a kind of "know it when I see it" [2] situation.
[1] http://www.sciencedirect.com/science/article/pii/S0004370210... [2] http://en.wikipedia.org/wiki/I_know_it_when_I_see_it