Yes, the subset is really small (and, I admit, not yet sufficient for a metacircular bootstrap). The whole idea is to provide basic .NET reflection functionality and implement all the rest in Arc itself (exactly as MBase itself is unrolled from the very minimalistic core).
But, some fundamental Arc features are not compatible with this core at all: zap (needs a strange memory model, not quite possible to implement efficiently in .NET, and, any way, is too non-functional) and mutable strings (that is easy to fix, if someone really needs mutable strings, I am not quite convienced). Also, numerical tower must be implemented in the core compiler (at least a parser must recognise all the variations and compiler backend must emit all the literals correctly), but it does not seem to be strictly specified for an original Arc so far.
Next release (to be available soon) will contain more of the basic functionality and a consistent .NET bindings framework. This must be enough to grow the language from the core.
What we really need is a formal definition for Arc operational and denotational semantics.