[boost] [Containers] Should flat_* expose implementation vector?

Dear All, Should Ion's proposed flat_set, flat_map etc. expose the underlying vector that they use as their implementation? Pros: it lets the user efficiently load and save the content, and implement algorithms that cannot be done via the flat_* interface. Cons: it lets the user break the invariants. Const access to the impl is another possibility. Related to that: could the implementation type be a template parameter? Example: could one implement a flat_set on top of a stable_vector? (Is that useful?) And/or: should flat_* actually be adaptors that take a reference to their underlying implementation? Regards, Phil.