So..., If I take a lower stddev_multiplier, I will reduce the (single, globally valid) T, thus making point removal more aggressive. Also, mean_k should not be selected larger than the smallest expected clusters?

From another source it has been suggested, that the dataset may be too small (which in my humble opinion still would not warrant a contradiction in results). So here goes some more information: The training set contains 34 data points, the testing set the 4 points used in the example. For the 663 possible predictors it is guaranteed, that each column has 80% values >0 which are spectral counts from a proteomics experiment. Using either the svms from e1071 or ksvm I usually get accuracy values (i.e. the rate of correct predictions) of roughly 80% when doing a 10-fold cross validation.

This of course is only an example where I omitted the data (it is a large 38x664 object). For completenesses sake: The true value would be 'titino'.