I did try to rule out question age as a major source of error in the model by training on a sample of ~1 million questions all from ~2012. But it didn’t do any

I did try to rule out question age as a major source of error in the model by training on a sample of ~1 million questions all from ~2012. But it didn’t do any better on those. And (theoretically) duplicates are deleted by the moderators.