To me, that's looking at the wrong end of the processes, because in many cases that have been identified so far, it's the data that has demonstrated bias. If you're running an algorithm against biased data or an unrepresentative sample, then you're going to get erroneous or erratic results.
In some cases the algorithm is proprietary commercial technology, and I don't know if an auditor would have the capacity to evaluate what the algorithm is intended to do, or how they would evaluate working versus non-working and under what standard.