Below is the dLogZeta distributions for both signal (prompt photons) and background (all QCD).  As expected, the signal distribution is concentrated at dLogZeta ~ 0 with the background distribution spread to much smaller values.  In particular, the background distribution features a "shoulder" at small values of dLogZeta while the signal distribution exponentially falls.

There are two features in the background distribution that warrant additional discussion.  The values with dLogZeta substantially larger than 0 are mostly energy depositions where only a single strip has any significant energy and that energy is still extremely small.  These low energy, single strip events strain the algorithm a bit (as it was tuned to full EM showers) and produce abnormally large values dLogZeta.  

Then there's the peak at dLogZeta = 0 which matches the signal distribution: these are in fact QCD events featuring SMD shapes entirely consistent with a single shower.  Electrons and hadronic showers are irreducible backgrounds, as are neutral pions that decay parallel to the strips.  This latter effect can be seen by looking at the eta and phi planes together.  Many events consistent with a single shower in the eta plane (dLogZeta = 0) show double shower structure in the phi plane (dLogZphi < 0) and vice versa.  Note also that the events with dLogZeta >> 0 or dLogZphi >> 0 are also confined to a single plane, as expected from an event with small energy deposition.

Lastly there are neutral pions whose decay photons cannot be isolated.  This effect should worsen at higher energies, and indeed the dLogZeta spectrum concentrates towards the single shower peak as the energy of the candidates is increased.

Finally a word on the shower shape model itself.  In order to avoid awkward geometry effects, the showers are not considered as distributions in eta or phi but rather absolute position along the eta/phi planes (in other words, z for the eta strips and r * phi for the phi strips).  With the strip widths already normalized by the use of absolute position, the location of the shower is set to 0 for each event by subtracting the mean strip position from each strip.  

Each shower is modeled with a heavy-tailed Cauchy distribution parameterized by the total energy, the full width at half max (FWHM), and the position.  Priors restrict each parameter to reasonable values, avoiding artifacts such as infinitely narrow or wide showers,

Energy: Gamma distribution with mean 15 and standard deviation 15.

FWHM: Gamma distribution with mean 2.25 and standard deviation 1.75.

Position: Gaussian distribution with mean 0 and standard deviation 7.

Remember that I make no claim that the shower model is perfect.  The model captures the gross features of reality and admits proper model comparison which then becomes a single variable input to the rest of the analysis no different then cluster energies or isolation sums.  Any implicit dependence of the analysis on data/simulation discrepancies is constrained in the signal extraction systematic bounds.