Today while I was returning from a DM meeting I remembered your case...
One of the things that probably happens is that the model selects the best classifiers and then it has no more data to keep creating more groups. If you have, for instance, 10 cases, and your best classifier splits those 10 into two groups of 5, then the next field will split those groups in maybe 2 or 3 cases, and then there are no more records to keep on splitting. When the model stops splitting has to, I believe, with the "splitting rules". You may configure the model to keep on splitting, but there is one point where those splits won't be solid cases and rules, just punctual instances not relevant to the whole dataset and will most probably even vary a lot. You won't get those splits past the evaluation and validation phase. This is known as overfitting. ¿When do you know you have achieved this point? I can't answer this question. For me it has always been trial and error, sadly.
Maybe Tim or another senior (in terms of experience, not age 
) DMiner can confirm what I'm saying.
So how can you use those fields? Well, I remember the "drugA drug B etc" tutorial making a new field from two fields, Na / K. This field was a very good predictor, but Na and K, separately, weren't.
One intriguing subject for me has always been "what the model can do and can't do when it comes to fields". Apparently Decission Trees can't create new fields from the original fields, but I believe other models such as Neural Networks and Discriminant Analysis do this. What you need to know is what relations (such as X / Y) can and cannot be made by the model itself. This is an interesting subject indeed and it belong to data preparation, one of the most important steps in data mining.
I'm no expert, so I won't feel bad if someone corrects every word I have just said. Please feel free to do so!