The goal here is to convert all the encountered source from Scheme to Clojure as directly as possible while implementing the exercises in Clojure as well.
SICP tends to use recursion a lot more than seems strictly necessary in some places. However, it works well for Scheme because of its excellent tail call optimization, which currently isn't as far along in Clojure (because of limitations in the JVM.) So the challenge often seems to be avoiding the temptation of declaring and using recursive functions that rely on TCO which will either not compile in Clojure or cause the stack to blow (indeed, depending on the size of the data, even the Scheme code from SICP will blow the stack, too.)
What you end up with is, often, simpler than the original Scheme implementation and often gives Clojure an edge by using functions that are inherently parallelizable.