Lets parse the definition of eval.

text		syntactic category
`let rec eval : type a. a term -> a = ⋯ (eval x)`	is a	definition
`eval : type a. a term -> a = ⋯ (eval x)`	is this extension of a	let-binding

bug report 6503 is hereby fixed! (Thanks Yallop)

Yet I am surprised that they consider this construct sugar. I shall have to understand what they translate this into!

Behold:

let rec f : type t1 t2. t1 * t2 list -> t1 = 
   function (a, b) -> a;; 
 f (4, []);; ==> 4

let f : type t1 t2. t1 * t2 -> t1 = 
   function (a, b) -> a;;

Desugaring:

let f : 't1 't2. 't1 * 't2 -> 't1 = 
  fun (type t1) (type t2) -> (function (a, b) -> a : t1 * t2 -> t1);;

OCaml likes the above. Parsing that:

text		syntactic category
`let f : 't1 't2. 't1 * 't2 -> 't1 = ⋯`	is a	definition
`f : 't1 't2. 't1 * 't2 -> 't1 = ⋯`	is this extension of a	let-binding
`'t1 't2. 't1 * 't2 -> 't1`	is a	poly-typexpr
`fun (type t1) (type t2) -> (function (a, b) -> a : t1 * t2 -> t1)`	is an	expr
`(type t1) (type t2) -> (function (a, b) -> a : t1 * t2 -> t1)`	is a	multiple-matching
`(type t1)`	is this extension of a	parameter

It is syntactically like a parameter but really not a parameter.