On referee-author communications Every time I referee a paper, I dream of a system which would allow me to ask the author a question without troubling the editors. It would save time for everyone involved, most importanly the referee; in my case it would probably half the refereeing time. 
Here is the system I envision: together with the paper, the journal would give the referee a link to a secure web form where the referee could type a question which would automatically be emailed to the corresponding author, who in turn could reply in a similar web form. Of course, the editors should have access to the logs of these exchanges, and the system should be set up so that the exchanges (if any) can only be initiated by the referee.
Question. Has this ever been implemented? Are there reasons this cannot (or should not) be implemented in mathematics journals? 
 A: I think nothing must be changed in the present procedure.
1) We send a paper to a journal when we think that the paper is ready for publication, if it is not, we don't send it. We spend more time on it, if necessary we left the paper 1 or 2 weeks in our drawer before sending it, just in case.
2) Once the paper is in the hands of the referee, the referee judges a work assumed to be complete, not a draft. If it is not accepted the first time but may be improved, it's referee's job to tell what to do and the journal board's job to be the intermediary. 
3) After one or two rounds the paper is definitely accepted or rejected. That's the way it works.
The referee is not a teacher, he is a peer and judges as a peer.
A: The review system for research projects in Italy (during the last few years; it is possibly going to change now) was based on comparing the opinions of two different anonymous referees. At some point the two referees have to exchange opinions with each other and reach an agreement, while remaining anonymous. In practice, this is achieved by channeling communications through a website based at a central location. I post my comments, the other referee can read them and post his comments, etc. Since we both know that our exchange is watched by some official at the Ministry, we feel some pressure as to not revealing our identity. I think this works pretty well, and should work also for peer review. 
Actually, several journals have already an electronic system for reports; at the moment it is a little awkward to use it as a communication tool, because of the many steps involved (the messages must be approved through the chain referee->editor->chief editor->editor->author or something like that), but it should be quite easy to make it faster and transform it into a supervised direct channel between referee and author.
A: Anyone using one of the suggestions based on "create a fake free email account" should watch out for the possibility that the free email provider will include your IP address in the message.  If it does, it is trivial for the recipient (here, the author) to determine which university owns that IP address, which would probably let them make a very good guess at your identity (check the faculty web page for Podunk State University and see who is the expert on reticulated splines).   So that would give only the illusion of anonymity.
For example, Yahoo Mail does include the sender's IP address.  Gmail apparently does not do so currently, or at least doesn't include it in any obvious way (it could be encoded in some way that is well known but not by me).
A: My opinion is based much more on an idealization of the process of refereeing, rather 
than the actual process itself, so this opinion may have little worth.  It appears to me 
to have much common sense, so I will share it anyway.  (This may all be in whatever
procedure or policy manual for referees the journal has.)
The goal of the process is to have someone spend resources putting ink to paper.  There
are other goals, such as scientific progress, but let us take the economic viewpoint.
The representative of the people spending these resources is the editorial team.  The team
is involved in the process of deciding how to arrange the ink.  Your questions and
eventual evaluation are key in this decision.  I suggest that you send a dated letter
(or email) which they can easily extract and forward to the author.  In short, they need
to be actively involved in the forwarding of such questions.
There are several techniques you can use which ease the process.  For one, leave any
reference to yourself out of the letter, so the editorial team can put such in if needed.
If you leave whitespace between the questions, you encourage the answerer to fit their
answer in that space.  Having the answerer meet the challenge of giving a concise answer
usually promotes proper answering, and adds value to the process.
Nowadays, many people use a computer to search for a substring.  Even so, use every
appropriate landmark to help locate the relevant text, as well as an appropriate string,
e.g. "Near the end of your second section, just before Lemma C, you have the phrase '...'.
What were you thinking when you included that phrase?" .
With the letter to the author, include remarks intended only for the editorial team,
saying what will happen if the questions get answered promptly, e.g. your referee report
will come six weeks earlier, or you anticipate a particular problem which may evaporate
if the right answers come, otherwise you will recommend the problem be fixed.  If the
editorial team knows explicitly how passing the letter on will benefit them, they will
be more motivated to pass it on and save time.
Arrange it so that you only need to send two or three such letters at most. This will
involve phrasing your questions properly and accumulating them so that this part of
the process is economically advantageous. Also, while becoming a referee,
 make sure you understand and agree with enough of the editorial process, so
that it will allow timely transmission of such questions and answers.
Gerhard "Ask Me About System Design" Paseman, 2011.01.02
A: I did exactly as others have suggested: created a fake email account (this was pre-Gmail; I think I used "sithmail") called [author name]reviewer, and corresponded that way.
Actually, I did this not to ask about the paper, but to strongly recommend to the author that he put his article on the arXiv before signing anything (and pointed out that the editor himself frequently posted to the arXiv). I didn't see any point in putting the editor into a difficult position, about whether he was allowed to suggest to the author that the author arXiv the article.
A: For many years now, 
I have refereed only non-anonymously.
It has not been uncommon that the authors of a paper I recommended with reservation—or, especially
those I rejected—to write me directly,
and we subsequently engage in an online (but off-journal) dialog to improve their paper,
often for another venue.
Almost uniformly these dialogs have been productively positive.
Easy for me to say, because of course I hold all the power strings.
Nevertheless, what Igor B. suggests seems to me would be a significant improvement 
on the current system.
My experience with the more one-sided communication I have experienced
predicts that it could work well.
A: How about this:


*

*First adopt Andy Putnam's suggestion in the comments above: The referee sets up a pseudonymous gmail account to correspond with the author.

*If the referee's anonymity matters, one would want to use a different  pseudonym for each new paper one referees.  Otherwise, even if one can't be identified, authors might figure out that the referee named Rollland is the same person as the referee of some other paper, also named Rollland.

*Some system (a web site) would be available to pick fake names from (and tag them as already taken, if they are) so that the referee's particular personal style of choosing various fake names would not get decrypted.  Maybe just 9-digit numbers or the like.

*If many referees do this and editors object, then their standard boilerplate requests to referee might include their policy that this is not to be done.
All this assumes refereeing continues to be single-blind.  The desirability of blind refereeing is debatable.  And it won't work with double-blind refereeing unless the editor assists somehow.
