I hope that arc keeps a canonical and authoritative implementation. One of the big problems keeping many people from deploying Lisp applications is the uncertainty and difficulty in choosing a particular implementation. It can also end up fragmenting developer time and you will inevitably end up with incompatible but extremely similar arc implementations everywhere.
One or two canonical implementations is fine, but there should be one authority for the language and it should be a fairly good implementation.
I think the axiomatic approach is intended to stop this. Lisp suffered because the spec was complex and ambiguous. To port Arc, just port the axioms, run the "spec" et voilà, you have a conforming implementation.
It should be noted that pg intends to implement more of Arc in Arc, when he gets the time.
In fact I have an implementation in SBCL half way (it basically can load arc.arc, but I haven't bothered with network/thread functions yet). To solve the call/cc problem I got to the point of doing a CPS transform, and since Arc has only 4 special forms (very few axioms, as you say) this was 'easy'...
