Accuracy of Balanced Realization (Pole location not preserved) in case of a sequence of a Laguerre Basis

A sequence of ss-objects are defined (based upon Laguerre Basis, needed for an identification algorithm). Since numerical aspects are important, I intended to make use of balanced realization (often this has improved the numerical behaviour )

As the example Shows, the pole-Locations are not preserved. If I discretize my Systems, with small sample time, pole Locations of the balanced realization are consistent with the original ones.

Does anyone have an idea why balreal.m does not perform as I expect? Can I not expect balreal.m to raise a warning when the pole Locations of the returned ss-object are not the same as the original one?