Disclaimer. It is really helpful to make concrete proposals in order to advance the discussion. This is a tentative proposal for an archival mechanism that tries to address some issues brought up by Henry Cohn and others. I'm sure there are plenty of other issues we haven't discussed yet. So please don't think that this is an official solution, the discussion is just starting!
What gets archived? Archival units are individual revisions of individual posts in a manner similar to what is found here but with some additional functionality such as direct links to the "live" post on MathOverflow, links to other revisions of the post, links to the matching archived question or matching archived answers depending on whether the post is an answer or a question, and other information described below.
Rationale: This is preferable to archiving page snapshots since it focuses on the actual information cited. Imagine a situation where user A gives a brilliant answer to user B's excellent question where user C had previously given an answer with a subtle mistake. After realizing their error, user C promptly retracts their answer by deleting it. However, user B takes a snapshot of user A's answer before user C's deletion and proceeds to cite it in an influential paper. That snapshot keeps being cited by other authors and user C's subtle mistake suddenly becomes a very embarrassing one.
When is it archived? Archiving would be done after a delay of $N$ days. Archival permalinks links would still be available right after posting but the archived version would be marked as pending.
- Posts that are deleted during the waiting period would be discarded from the archive.
- Posts that are put on hold/migrated during the waiting period would have their waiting period extended by $M$ days.
Rationale: The archival delay is to avoid archiving spam and other inappropriate content and to give users a reasonable chance at early retraction. Reasonable delays are $N = 10$ and $M = 30$.
What about deletions? Once archived, a post is never deleted from the archive. However, archived posts can be annotated after some trigger events.
- When a post is deleted on MathOverflow, all of its archived revisions are marked as retracted in a clearly visible manner.
- When a post is closed or locked on MathOverflow, some appropriate mark may be added near the link to the post on MathOverflow but the archived version would not undergo major changes.
Rationale: Archiving is fundamentally a one-way process: archived contents should never be deleted (unless there was an error in the archiving process itself). On the other hand, it is necessary to respect the users and the community. Annotations are the best we can do in this context.
Deleting, closing and locking are very different processes. Deleting is a clear retraction and needs to be clearly annotated. Closing and locking have to do with the functionality of the site and don't seem to need much annotation.
What isn't archived? Post comments are never archived, neither are vote counts, view counts and similar ancillary post data.
Rationale: These primarily have to do with the functionality of the site. It is true that some of this content is worth archiving but it seems too difficult to sort. If a comment really does contribute significantly to the archival value of a post, it should be integrated into that post or promoted to a separate post.
How does this integrate with MathOverflow? A new citation button would be added below each post, separate from the share button. The button would give the archive permalink of the current version of the post (even if archiving is still pending) as well as properly formatted BibTeX and AMSref entries.
Rationale: Sharing and citing are fundamentally different; it is confusing to merge the two into one. The share button will still exist and it will give a link to the actual MathOverflow post. The cite button needs to be separate since it links to the archived MathOverflow post.
If archiving is still pending, it's the user's responsibility to ensure that the archiving process has completed before using the link in a citation. Since there is usually a very long delay before citations are published, this shouldn't be a problem in practice. In cases where the link needs to be immediately usable, the sharing link is probably more appropriate.