Potentially not the case; if you allow that some fraction of the upvoters are not aware of the 12-months-matched and are upvoting based on 1-month-matched, the discounting might disappear.

While they might not be aware of 12-months-matched, 1000/mo should still rank significantly higher than 2000, and it doesn't.

This screams karma = log(dollar amount) to me.

Me, too; it coheres with what I know about how humans cope with large numbers. I had tried a log regression, but there's just not enough data...