Agents flagging a stale entry in only 4% of runs worries me more than any of the wins, because I'd expect a confidently wrong memory to spread faster than a fix. The 12% to 79% jump for Sonnet shows the channel works when the fact is in there. When a learned fix later turns out to be wrong and already has upvotes, can a few down-reactions push it out of the top five, or does someone have to delete it by hand?