extended dynamic sampling documentation?

The documetns/articles on the web are pretty high level. anyone know of any documents that go into more detail. When I look at a tkprof output I see somewhat different dynamic sampling queries when the optimizer level is at 2 (this is 10.2).

Case 1: Case statement in the query based off of a column in a join clause (between a GTT and a base table). The sample block clause is hard coded.

Case 2: No case statement. This query is also joining from a GTT to a base table. The sample block clause is populated with a bind variable.

I am hoping either oracle or someone has some more details of why dynamic sampling queries can be different?

--
http://www.freelists.org/webpage/oracle-l