I have noticed that the progress meter becomes wildly inaccurate when Rosetta@home processes work units that cause it to shut down early because it has created enough decoys (typically 99) to force it to terminate early rather than create too many decoys for the server. I believe that a more accurate progress meter would be generated if the progress was defined as the maximum of the original CPU time-based progress meter or the number that is generated by dividing the number of decoys completed by the number of decoys allowed before an early shutdown is triggered.

Oops, I forgot that I posted this idea over three years ago.