A second moment closure (SMC) computation is compared in detail with t
he direct numerical simulation (DNS) data of Le et al. (1997) for the
backstep flow at Re = 5100 in an attempt to understand why the intensi
ty of the backflow and, consequently, the friction coefficient in the
recirculation bubble are under-estimated. The data show that this reci
rculation bubble is far from being laminar except in the very near wal
l layer. A novel ''differential a priori'' procedure was used, in whic
h the full transport equation for one isolated component of the Reynol
ds stress tensor was solved using DNS data as input. Conclusions are t
hen different from what would have been deduced by comparing a full si
mulation to a DNS. In particular, the epsilon-equation, usually blamed
for faults in model predictions, has been found to give excellent res
ults in this case. In fact, the main problem comes from the <(uv)over
bar>-equation which predicts a too high turbulent force. A modificatio
n, by including the gradients of mean flow in the transport model, has
then been attempted and has cured 50 percent of the backflow discrepa
ncy.