Calculations with surrogate variants of original data are used to vali
date results obtained in dynamical analysis. Three classes of surrogat
es are now in use: random-shuffle surrogates, random-phase surrogates
and Gaussian-scaled random-phase surrogates. In this paper we present
an example based on a natural source of random numbers (radioactive de
cay) in which random-shuffle and Gaussian-scaled random-phase surrogat
es both correctly identify the random nature of the data while random-
phase surrogates give a dramatic, and totally spurious, identification
of non-random structure. The application of random-phase surrogates b
y themselves, without confirmatory calculations using Gaussian-scaled
random-phase surrogates, is becoming increasingly common. The results
presented here argue against this practice. The first step in the appl
ication of symbolic analysis to dynamical data is the partitioning of
the data into a finite symbol alphabet. These results also show that a
ppropriately constructed surrogates can be used as a protection agains
t spurious results caused by defective partitioning.