While I think Dimension 10 has done a good review on the Q/A, so perhaps it can go live immediately, I am worried about the refereeing section. Here, you need a robot to link arxiv papers automatically to questions, allow hand-insertion of non-arxiv literature, and, most importantly, I firmly believe that you need separate voting on originality of content, and on general accuracy/quality separately.
The two evaluators need to be completely separate for good review. The originality is extremely important, in that, for instance, 't Hooft's 80s models of holography are not 100% accurate (because he is analyzing thermal black holes in 4d, not string theory objects), but they are extremely original and contain the seed of the main idea of holography very clearly. The mistakes in the papers are of the form of crazy semi-string actions with wrong sign, and other weirdness, caused by having a slightly off picture. They would have a high originality score, but a relatively low accuracy score, although it should still be positive, because the calculations are correct given the assumptions.
I suspect that the easiest way to do it in Q/A is to have two linked questions, with citations and originality discussed on one, and with accuracy discussed on the other, and where the upvote score is equal to the originality score times the accuracy (or, perhaps exp(originality/10) times accuracy, to avoid sign problems, or perhaps exp(originality/10|originality|^.8) to avoid exponentially exploding reputations for paper-authors.
Having two questions per paper allows the citation issues to be sorted out on one page, and the accuracy issues on another. It is important not to mix the two up, because the main comments you will get on papers will be nearly entirely devoid of physics content, and will be "cite me" answers by people who feel their work has been unfairly neglected. This sentiment can be expressed in the "originality" section.