the translation into CHRs would be something like:
class C a b | a -> b instance ctxt => C t1 t2
-->
C a b <=> infer_C a b, memo_C a b. (two roles)
infer_C a b <=> ctxt. (instance inference)
memo_C a b1, memo_C a b2 => b1=b2. (class FD)
so there'd be one preparatory inference rule to split all constraints occurring in the process into their two roles, one inference rule for each instance declaration, and one propagation rule for each FD. the CHR succeeds if there are no non-memo constraints left.
to avoid misunderstandings: this is not yet complete. it is sufficient for the example from the paper, because the FD-based improvements there are triggered by the two goal constraints, but the translation also needs to account for FD-related information taken from the instance declarations. claus