In the absence of an authoritative explanation, I'll try to reconstruct the reasoning that led to audit I observed.
I assume that ways to justify the failure for known good and bad audits when these contradict original voting of reviewer "mirror" each other.
Reviewer who fails "known bad" audit despite their prior positive voting on the item is probably supposed to reason as follows:
It doesn't matter that your original voting for this item was opposite, because there is sufficient evidence that it was wrong. Accept the audit failure, learn that your original assessment was wrong and move on.
Those failed "known good" audit despite their prior negative voting are likely expected to reason the same way, exactly as written in above statement.
This "symmetry", in turn, should probably be based on the assumption that automated selection of items for known good audits is equally reliable as that for known bad ones - note how sufficient evidence is used to justify selection against original reviewer decision in both cases.
Voting close vs leave open, as well as voting up vs down used to select audit items for both kinds of items at first sight feels close enough to assume equal reliability of the selection.
But actually, there is a bit of disbalance related to the fact that voting down requires higher reputation than voting up, suggesting that on average voters down (who feed into known-bad audits) can be expected to have stronger understanding of site quality norms.
Worth keeping in mind that voting to close questions requires even higher reputation and that voting answer down carries a rep penalty, as opposed to totally free upvoting.
Above differences in "voting weight" used to select different kinds audits are probably ignored for the sake of simplicity.
It is worth noting that per my observations, selection of known bad audits actually involves more than close and down votes. Per my recollection, known bad audits I've seen were all picked from deleted posts.
Given that in order to be deleted, post has to go through quite stringent process, this would make a much more reliable selection than that based solely on close/down votes.
Actually, if "known bad" items are selected from posts deleted by moderators / 10K users, one can even argue that these audits passed through human verification before being presented to reviewer. If this is the case, it would be very important to take into account that selection of "known good" audits doesn't even come close to anything like that.