When we do experiments with low S/N, our experiments aren’t just underpowered — they’re unstable. Our data are both randomly blurred and irreproducibly biased.