How do I referee a paper?

Updated below We all know the critical importance of peer-review. It is the main form of quality control and feedback on research. However, to an early-stage researcher (like me), it can sometimes be a confusing system/process. Accordingly, there are several treatises on the scientific refereeing process that give guidance. Two (very different) examples from computer science -- this 1994 article by Parberry and a more recent one by Cormode -- offer great advice (though the latter might be a shade mischievous). Here, I'd like to solicit broader advice from the more experienced members of this community about the review process, with particular regard to the peculiarities of theoretical computer science. What are the main criteria for determining the significance of a paper's results? How do I judge whether a paper should be accepted to the conference/journal? Is it important to verify correctness? What are the main elements of a referee report, and which parts are most important? Is it always necessary to give a recommendation of (non)acceptance? What goes in the report and what goes solely to the editor? How does assessment for conferences differ from that in journals? How do reports for conferences differ from those in journals? (How on earth do I rate my "confidence" in my recommendation?) Should the journal version be significantly different from the conference paper? What if I don't understand the paper?...the proof? (Is it my fault or theirs?) What about typographical/grammatical mistakes? What if there are a lot of them? How much time should I spend on a report? How many reports a year am I expected to write? When is it acceptable to refuse a request to referee? Of course, any other relevant questions and answers on this topic are encouraged, since this is CW. This question is inspired by (stolen from) a similar post at MathOverflow. Update 15/02/2011: I am still very interested in getting more input on this question especially with regard to reviewing confere