I came back to this thread thanks to Alan Schmitt's Caml Weekly News.

Here is a fairly systematic approach to the problem, contrasting Jeremy's
more virtuoso approach. It works well when the whole set is statically
defined in the source code, but I'm not sure about more dynamic scenarios
(when a lot of the type variables involved become existentially quantified.

The idea is to first formulate a solution using logic programming. A link
set can be represented as a list of accepted sinks, and I can write a small
Prolog program, "insert" that takes a new link, and a link set, and
traverses the link set to find the link's sink in the list, and add the
link's source to the list in passing.

   set([sink]) :- link(source, sink).
   set(sinks) :-
     set(prev_sinks),
     link(source, sink),
     insert(prev_sinks, source, sink, sinks).

   (* either the sink we want to insert is first in the list *)
   insert([sink | _], source, sink, [source, sink | sinks]).

   (* or it is to be found somewhere, recursively, under some other sink
[s] *)
   insert([s | prev_sinks], source, sink, [s | sinks]) :-
     insert(prev_sinks, source, sink, sinks).

Once you have this logic program, it is straightforward to translate it to
a GADT declaration:

type 's linkset =
  | Init : ('source, 'sink) link -> ('source -> ('sink -> unit)) linkset
  | Insert : ('source, 'sink) link * 'set1 linkset * ('set1, 'source,
'sink, 'set2) insert -> 'set2 linkset
and ('set1, 'source, 'sink, 'set2) insert =
  | Here : (('sink -> _) as 'set1, 'source, 'sink, 'source -> 'set1) insert
  | Next : ('set1, 'source, 'sink, 'set2) insert -> ('a -> 'set1, 'source,
'sink, 'a -> 'set2) insert

let _ = fun (type t1 t2 t3) (lnk1 : (t2, t1) link) (lnk2 : (t3, t1) link)
(lnk3 : (t3, t2) link) ->
  let set = Init lnk1 in
  let set = Insert (lnk2, set, Next Here) in
  let set = Insert (lnk3, set, Here) in
  set


On Wed, Sep 21, 2016 at 8:50 PM, Andre Nathan <andre@digirati.com.br> wrote:

> Em 21 de set de 2016, às 19:22, Jeremy Yallop <yallop@gmail.com> escreveu:
>
> It is!  Here's one approach, which is based on two ideas:
>
>
> Wow, there's a lot to digest here :) Most of it is new to me, including
> the use of "evidence" types... I'll have to re-read it and try to work it
> out after a night of sleep.
>
> Thanks,
> Andre
>
>