It is in principle possible to use the architecture described here to build an interpreter mixing system using some of the ideas from [Jones, Sestoft And Søndergaard]. With our simple and regular program and language architecture, we can make this much simpler than the partial evaluator presented by [Jones, Sestoft And Søndergaard]. This results not in a full mixer function, but in a compilation system; for each node of the expression tree, we expand the node out using the operator definition closure for that node---an inlining of the operators. There are two notable restrictions on this operation: the procedure being compiled must not change its expression reflectively; and all the operator definitions must be in the same language as each other, since the resulting closure will be in the language used to define the operators, and each closure is allowed only one language. (This second restriction may be relaxed at the cost of having to merge several languages, with automatic renaming of operators where necessary.)
If operators are provided to escape entirely from the meta-tower system down to the real substrate system (in Platypus, this is Common Lisp), it should be possible to generate real compiled code from any executable procedure (that is, one using only grounded definitions) that does not reflect into its expression or interpreter, by repeated substitution of operators as described above, until operators shadowed by the real substrate are reached. At this point, code in the real substrate language may be substituted (possibly from the shadow definitions, but it may be necessary, or at least better, to have other code for this), and the results passed to the substrate's native compiler. The result of this compilation may then be put into the appropriate shadow maps, if it was generated from a closure used as a type-evaluator or as an operator.
Using these techniques, a program-tuning system could be written, that runs a program (possibly its caller) on an interpreter that does time- or call-counting- profiling, then arranges for new primitives to be installed, if possible, to make that program run faster.