The tags can be sortable without enumeration. The problem is that the ordering will be different in every translation unit....

So, as I understand it, the compiler has to generate a unique typename for every class, and that has to be consistent over all translation units. It also seems to me that somewhere this must get hashed into an integral value or something so the compiler doesn't have to deal with strings as internal identifiers. Since we don't really care what enumerable value gets assigned to what system or to what tag, if we could somehow access the full typename or (even better) an integral hash of it, that would give us our sortable value automatically. For example, on g++ gives : FN5boost5units10length_tagEvE FN5boost5units2SI6systemEvE for typeid(length_tag()).name() typeid(SI::system()).name() If these could be hashed into a long int, that would be all we needed. Probably would involve very non-portable code, but it might be supportable on a few of the major compilers... I thought C++ name mangling was supposed to be standardized...

I would probably sort by dimension first to make checking the dimensions easier.

The way I'm envisioning it, any dimension typelist <S1,T11,P11>...<S1,T1N,P1N><S2,T21,P21>...<S2,T2M,P2M><S3,T31,P31>...<S3 ,T3L,P3L> would split into three typelists that are equivalent to the dimension<>::type typelist we construct now. These could then be treated separately, and appended afterwards. Since no inter-system conversion happens in the operators, I think that's all that would happen. This will definitely not happen before the review, but I think it could be done and would be interesting... Matthias