I'd suggest 10 minutes maximum without a rest break if you're working on extremely fine distinctions.
Some people go to 20 minutes, but I've found that dl's and every other sort of result go to (*(* in a handbasket after about 10-20 minutes for even the most practiced and trained listener. ...
I'm sure you've done a lot more audio testing than I. I relate one of my few experiences with blind testing. In this case I had a friend try to identify the difference between two states (regarding amplifier differences). After 15 minutes of trying, in 8 trials, he was too tired to tell anything. So in a simple, semi-controlled test, I was already having a problem getting any kind of reliable results. Yes, these tests, blind or not, are not easy to do correctly for meaningful results. And yes, 15 minutes was too long a time.