Search engines have to solve this problem, that the results listed higher garner more clicks, and thus using clicks to rank is biased. They can separate the position of the result from its relevance through something called a click model. They make some of the assumptions you list in your article, which you call "examples". Click models are mathematically rigorous and can be quite optimal for solving the upvote bias, and obviously you would replace "clicks" with "upvotes" in your implementation. Here are some well-known click models:

Cascade Model (one of the original click models): http://www.wsdm2009.org/wsdm2008.org/WSDM2008-papers/p87.pdf

Dynamic bayesian network model (a more generalizable Cascade Model): http://www2009.eprints.org/1/1/p1.pdf

DBN model with scroll and hover interactions (kind of like your example 2): http://jeffhuang.com/Final_CursorModel_SIGIR12.pdf