I would like to point out that, with a primitive procedural interface,
static type inference becomes hard.  Indeed, if record types are values
constructed at runtime, it is hard to see how much useful type inference can be
done without some kind of partial evaluation.

In a previous post, I suggested dropping the syntactic layers.  Another
possibility, more type-friendly, would be to drop everything but
(a simplicication of) the explicit-naming syntactic layer.

Cheers
Andre