There is no rule of thumb, and therefore I need to dwell in more detail:
I usually assume that the email provider tells me: harbounce says there are persistent problems with a specific email address (i.e. it doesn’t exist) and softbounce says that there is only a temporary problem (i.e. the user’s disk quota has been exceeded) .
Assuming the provider, I would dump the email addresses using hardbounce from my list, saving them with softbounces. There may be several reasons for softbounce that may require user interaction with the email owner (i.e., freeing up some disk space on the server by deleting emails). Thus, you cannot predict the point in time when an email can be delivered to the user. But you can try to repeat based on a scaling sequence, such as a Fibonacci number and reset the address when a certain limit is reached.
By not trusting the provider, you can additionally count hardbounces for each address. This strategy has problems with the mail provider (sending false hard drives). I recently worked with a commercial mailing system. They used a counter for hardbounces and softbounces each. Since you usually do not send the newsletter every hour to the same address, there is enough time for the technical problems on the part of the supplier to be fixed. The default limit for hardbounces was 3, and the newsletter did not repeat after hardbounce. Addresses were deactivated only after reaching the global limit and reset after successful message delivery.