Some of the stuff generated by the volunteer transcribers is weak to say the least. I have thought about this problem for well over a year now and ruminated on potential solutions. One extreme solution is to have every transcription done twice, each time by a separate person. If the two transcriptions match then you can publish it live to the public. Otherwise the two versions would have to be reconciled. I have done some transcriptions for Family Search. This is the model they use and it seems to work for them. It may not work for Billion Graves as you may have problems at certain times of the year finding the volunteers to do the transcription once let alone twice.
Another approach would be to make each transcriber pass some sort qualification test to prove that they actually know how to transcribe according to your guidelines. This would help eliminated errors due to lack of knowledge about the requirements to transcribe information.