In the initialization step incorrect eps_prev was used
resulting in spurious eps_m contributions. This effect
is observable only for this particular benchmark.

The benchmark has been also modified with new references.