A number moved, so a meeting happened. Conversion dipped on Tuesday, so the funnel copy changed by Friday. Velocity fell one sprint, so the process gained a ceremony. Response times crept up for a week, so someone reorganised the on-call rota. Each correction was reasonable. Together they may be the main reason your numbers will not sit still.
W. Edwards Deming, the statistician whose quality methods reshaped postwar Japanese manufacturing, had a blunt name for this: tampering. His teaching demonstration used a funnel on a stand, dropping marbles at a target. Left alone, the drops scatter in a stable cluster — that scatter is the process talking, and it is normal. Then the demonstrator starts helping: after each drop, he moves the funnel to compensate for the last miss. The cluster does not tighten. It roughly doubles in spread, and under some correction rules it walks off the table entirely.
The mechanism is simple enough to hurt. A stable process varies on its own. If you adjust in response to ordinary variation, you are adding your adjustment on top of the noise — chasing each wobble with a new starting point, which manufactures the next wobble. The corrections compound. Then the widened swings get read as further proof that constant intervention is needed, and the loop closes.
Two kinds of movement, two different jobs
The way out is a question that has to be asked before any correction: is this movement the routine chatter of a stable system, or a signal that something specific happened? His tool for telling them apart is unglamorous — plot the number over time, weeks of it, and learn its normal range before reacting to any single point inside that range.
The two answers assign opposite jobs. A genuine signal — the number jumps outside its historical band, or a real event maps to the change — justifies finding the specific cause. Ordinary variation justifies no reaction at all in the moment; if the normal range itself is unacceptable, the work is redesigning the system that produces it, which is a project, not a Friday fix. Adjusting week to week does neither job. It is the cost of both with the benefit of neither.
Notice what this does to meetings. The weekly review stops being a court where the latest number stands trial and someone must announce a countermeasure. Most weeks, the honest finding is: within normal range, no action. That sentence feels like negligence in most cultures. It is the discipline, and it buys the thing tampering spends: a chart quiet enough that a real signal shows.
Watch out for
The trap on the other side is real: a stable chart can lull you through a genuine shift, and Deming's own late-career tone — everything is the system, individual effort is noise — overcorrects. Some Tuesdays the dip is the first day of the new reality. The safeguard is that a signal earns investigation of a specific cause, and the investigation either finds one or it does not. What the method forbids is not vigilance; it is correction without diagnosis.
Be honest, too, about why tampering persists: it is performed for an audience. A leader who adjusts something every week is visibly managing. Doing nothing because the variation is normal looks passive precisely to the people deciding your next role. That incentive problem is outside statistics, and pretending the chart alone will win the argument is its own naivety.
Answer this next
Pick the number your team reacts to most. Do you actually know its normal weekly range — and how many of the last five corrections were responses to movement inside it?
