In reference to this discussion on title ban for HNQ
An empirical approach would be far better than what we're doing right now imo. More specifically, a keyword should be banned for HNQ if the ratio of poor to good questions with the keyword is high. Given a large number of questions with a particular keyword it isn't that hard to reach some consensus on the bad-to-good ratio.
The problem with asking a small bunch of meta users about this is:
They don't have the patience or incentive to browse through tens or hundreds of questions to arrive at a bad-to-good ratio, they're going to be biased and use a smaller pool of data - the questions which they happen to remember or have googled.
The problem mentioned in 1 gets compounded since most users are infact probably look at overlapping data pools to arrive at their decision - googling a keyword turns up similar answers for everyone, remembering recent questions brings up common memories of the recent questions, and looking for high-voted questions again means wasted effort as everyone is looking at the same data to reach their decision.
Many poor questions are deleted and hence don't show up in searches. This significantly biases the data you have access to, unless you're a mod or frequent visitor of the delete queue.
What this site instead needs IMHO is a vote counter that can only be used by high rep users, since that's what you guys seem to be aiming for. You don't trust "average Joe"-driven voting to decide what is worthy of HNQ, you wish to give more weight to knowledgeable persons. Which is okay I guess, that's for another debate.
Since a coding solution is off the table, what would work is if the mods provided access to a decent-sized corpus of questions including deleted ones, and then distribute the task of filtering good and poor questions among high rep users. Or if that's too much effort, perhaps the mods themselves could do a bit of such filtering, arrive at good-to-bad ratios and take the decision themselves. I don't see how your pool of meta users with biased data is going to do a better job except when it comes to the most obvious ban worthy keywords (good-to-bad ratio nearing zero).