> Looks like our anomaly detector is doing a decent job.
I don't agree. Calling ~5% of your dataset anomalies purely by construction always seemed like a crappy way of doing things to me. Look At those stars towards the end of the curve, they don't look like anomalies to me.