Sometimes historical dates written by humans are wrong - e.g. February 29 of a non leap year.
We should think about how to support these kind of dates that are technically invalid according to our calendar systems, as a different kind of ambiguous date with some margin of error. This might be something be a non-default that has to be opted in with a parameter, since supporting it by default could result in unexpected behavior.