I think double-checking results would bring the most benefit, especially if done by a native speaker of the target language. They don't have to understand the source language, just point out that a sentence makes no sense or uses awkward grammar.
There is a long standing tradition at a major customer of my company, unrelated to this system, that any worker anywhere on the line can hit a Big Red Button which will stop the line if they detect a quality issue. Hitting the Big Red Button has immediate costs: it delays production ($$$), it incurs overtime expenses to get the lost production back ($$$), it throws your supply lines into chaos ($$), it causes you to miss targets with your customers and pay out contractual damages ($$$$$$), etc, etc.
Hitting the Big Red Button on a software project is cheaper, but not free. (For example, the Big Red Button plan for customer acceptance testing might sound like "Nobody goes home if there is an unknown risk to delivery", and a translation error anywhere means there is an unknown risk to delivery until you have verified that whatever process introduced the error was not similarly out of control elsewhere.)
1) Similar to 99Designs, etc, everyone could vote on quality, so that could give you some confidence in the text.
2) Translator users could accumulate karma and reviews.
3) Translators' previous work would be public for inspection, and you could see comments and votes on previous work.
4) The requester could seek higher quality by offering a higher price.