The TOC YahooGroup has been talking about Deming's Funnel, and a smart person finally asked what it was. A quick google nets a wide variety of explanations. Tampering and effects diagnosis from the QualityAmerica knowledge center was a useful verbal description. And Yonatan Reshef at the U of Alberta has a nice numeric demonstration and simulator at his Deming's funnel experiment page.
In the end, the best explanation is that when you continually fiddle with a system on the basis of one sample (or an insignificant set), you ALWAYS increase the variation over the long term.
Proper adjustment comes from collecting a statistical sample of the behavior. Then you can decide whether you need to adjust your aim (the sample is consistently off-target), or the system (system is on-target, but the variation is undesirable).