I vote *ACCEPT*. *Design* I consider it a shortcoming that std::variant does not have the never-valueless guarantee of variant2. The point of variant<A,B> is that after a successful initialization, it will either have a valid A or a valid B. If I wanted to have to deal with some other alternative, I would have specified a third type. The main objection to the never-valueless semantics is that after a failed assignment, the programmer could be surprised to find that the target variant object may have switched from a valid value of the previously stored type, to a valid value of one of the other alternative types. While it is true that the change of type can be surprising, this is the good old basic guarantee in action -- nothing less, nothing more. I also like the addition of the narrowing operation subset<U...>, seems useful. *Did you try to use the library?* I tried variant2 on Mac/clang. *How much effort did you put into your evaluation?* I spent a 3 hours reading the documentation and experimenting. I also tried using expected<T, std::error_code> as a result type in LEAF. *Are you knowledgeable about the problem domain?* Yes.