> Well, one thing that this solution misses is virtual dispatch - HexadecimalEvaluator won't work for Addition(Literal(1),Literal(2)) for example, because Exten

> Well, one thing that this solution misses is virtual dispatch - HexadecimalEvaluator won't work for Addition(Literal(1),Literal(2)) for example, because ExtendedStringEvaluator doesn't know to call HexadecimalEvaluator back. Ah of course, I accidentally switched back into closed recursion rather than open recursion (that was ExpressionOpen). This indeed is exactly analogous to early binding vs late binding. You can't exactly write the same form of open recursion I wrote earlier because C# doesn't have higher-kinded generics (so for example you can't write interface Mappable<L> { public abstract L<B>Map<A>(Func<A, B> f, L<A> x) } which prevents you from writing record Fix<F>(F<Fix<F>> x) ), but that's a language-specific thing. If you had higher-kinded generics you could exactly model virtual dispatch with open recursion because they're the same thing. > That is, the Expression doesn't "know" any more that it is being visited, so you couldn't create a LoggingExpression let's say. On the other hand, the Visitor now has more control over the way it interprets expression subtypes. So even here we are doing double dispatch: we are dispatching manually based on the runtime type of `expression`, and then allowing the language to dispatch based on the runtime type of `this`. You can call it visitor, but at this point it really is the same as discriminated unions (indeed this is exactly analogous to the first iteration of code just without the last bit to ensure that combinations matched up). And you can exactly replicate `LoggingExpression` in the evaluator, there is no difference in expressivity. This shouldn't be surprising; after all the whole thing that kicked off all of this is that there is an isomorphism between discriminated unions and the visitor pattern. It should not be a surprise that it is possible to interpret one in the other. My point at the beginning of all this was to object to the statement that this was not an isomorphism. Anything you can do with the visitor pattern you can do with discriminated unions and vice versa. The only differences are in ancillary language support (e.g. virtual dispatch and higher-kinded generics). Hence in a very real real sense the visitor pattern is just discriminated unions and vice versa. The only preference for one or the two comes down to issues of code size and maintenance.