I think Aubrey's interpretation can be rescued if we adopt the following:

Inexacts represent intervals but also have a canonical value that lies
somewhere in the interval. When a operation is performed on inexacts,
the result is the inexact whose interval includes the mathematical value
obtained by performing the corresponding mathematical operation on the
canonical values of the arguments.

An example implementation would be with floating point numbers and all
operations with zero ULPs of error.

I'm not saying this is practical or desirable. I am also not saying that
it is the only valid reading of R5RS.

Regards,

Alan
--
Dr Alan Watson
Centro de Radioastronomía y Astrofísica
Universidad Astronómico Nacional de México