Quality
and the Funnel Experiment: When Your Organization’s Best Efforts to Fix
a Stable Process Make Everything Worse — and the Adjustments That Were
Supposed to Help Become the Source of the Variation You Were Trying to
Eliminate
There is a particular kind of madness that seizes quality
organizations, and it looks exactly like diligence.
The production line is running. The control chart shows variation —
natural, expected variation within limits — and someone in a position of
authority stares at it and says: “We can do better.” They adjust the
process. They tweak the temperature. They change the feed rate. They
modify the pressure. And the next data point lands further from the
target than it was before.
So they adjust again.
And again.
And with each adjustment, the process — which was stable, which was
predictable, which was doing exactly what a capable process is supposed
to do — begins to wander. The variation increases. The control limits
are breached. Defects appear where there were none. And the person who
started the adjusting, now surrounded by the wreckage of what was once a
perfectly functional process, looks around and says: “See? It’s a good
thing we intervened. Imagine how bad it would have been if we
hadn’t.”
This is not a parable. This is not a thought experiment. This is
something that happens in manufacturing plants every single day, and it
has a name. W. Edwards Deming called it the Funnel Experiment, and he
used it to demonstrate one of the most counterintuitive and devastating
truths in all of quality management: when you adjust a stable
process in response to individual results, you make the process
worse.
Not better. Worse.
Every time.
The Experiment That
Changed Everything
Deming’s Funnel Experiment is elegant in its simplicity, which is
precisely what makes it so powerful. Imagine a funnel suspended above a
target on a table. You drop a marble through the funnel. It hits the
table somewhere near the target — sometimes close, sometimes less close,
but always scattered around the center in a natural pattern determined
by the inherent randomness of the system.
This natural scatter is the voice of the process. It is the process
telling you what it is capable of doing. It is variation, but it is
variation with a stable, predictable shape.
Now, Deming asked: what happens if we try to improve the results by
adjusting the funnel’s position after each drop?
Rule 1: Leave the funnel alone. Don’t adjust
anything. Just keep dropping marbles through the same fixed point. The
results scatter around the target with natural, stable variation. This
is optimal. This is the best you can do without changing the system
itself.
Rule 2: Adjust the funnel based on the last result.
A marble lands two inches to the right of the target, so you move the
funnel two inches to the left. The next marble lands three inches to the
left, so you move the funnel three inches to the right. You are
compensating for every deviation — and the result is that the variation
nearly doubles. The marbles spread out over a much wider area. Your
adjustments, each one perfectly logical in isolation, have made the
system worse.
Rule 3: Adjust the funnel based on the last result, relative
to the target. This is similar to Rule 2 but adjusts based on
the distance from the target rather than from the funnel’s current
position. The result? A systematic drift. The process wanders further
and further from where it should be, like a drunk who corrects his path
by the same amount he strayed, but in the wrong reference frame.
Rule 4: Reset the funnel to the position of the last marble
drop. You place the funnel directly over wherever the last
marble landed. The result is catastrophic. The process goes on a random
walk, drifting without bound. It will eventually land anywhere on the
table — or off it entirely. This is chaos dressed up as
responsiveness.
Deming demonstrated this with marbles and a funnel. But he was really
talking about your factory, your laboratory, your inspection department,
and your management review meetings. He was talking about every quality
professional who has ever looked at a single data point that deviated
from target and reached for a control knob.
The Shop Floor That
Couldn’t Stop Adjusting
Consider the case of a precision machining operation that produced
cylindrical shafts for automotive transmissions. The critical dimension
was the outer diameter — specified at 25.000 mm with a tolerance of
±0.015 mm. The process was capable. The Cpk was 1.67. The control chart
showed a stable, predictable pattern. By every statistical measure, this
process was performing beautifully.
Then a new shift supervisor was promoted from the operator ranks. He
was ambitious, detail-oriented, and deeply committed to quality. On his
first night shift, he watched the operator measure a shaft that came in
at 25.008 mm — well within specification, well within control limits,
but eight microns above the nominal target.
“Let’s bring it back to center,” he said. He adjusted the tool offset
by eight microns.
The next shaft measured 24.993 mm. Thirteen microns below nominal,
still within spec, but now the supervisor was alarmed. “It’s drifting,”
he told the operator. He adjusted the offset again — this time by
fifteen microns in the positive direction.
The next shaft measured 25.011 mm.
The supervisor adjusted again. The shaft after that measured 24.988
mm. Then 25.013 mm. Then 24.982 mm — now just three microns from the
lower specification limit. In the span of a single shift, this
supervisor had taken a rock-solid process and transformed it into a
roller coaster.
By morning, three shafts had been rejected for being out of
specification. The process that had produced zero defects in the
previous six months was now generating scrap. And the supervisor’s
report said: “Process showed unexpected instability during night shift.
Corrective action taken: continuous monitoring and adjustment of tool
offset to compensate for drift.”
He believed he had been doing the right thing. He had been responding
to data. He had been making decisions. He had been managing.
He had been operating under Rule 2 of the funnel.
Why Smart People Tamper
The most insidious thing about process tampering is that it feels
exactly like good management. Every individual adjustment makes sense.
If the temperature is running 2 degrees high, lower it by 2 degrees. If
the fill weight is 3 grams light, add 3 grams. If the defect rate ticked
up on Tuesday, investigate and correct on Wednesday. Each action is
rational. Each action is responsive. Each action is exactly wrong.
This is because individual results from a stable process contain two
components: the signal (the true process mean) and the noise (random
variation). When you adjust based on a single result, you are reacting
to noise as if it were signal. You are chasing ghosts.
The mathematics are unambiguous. For a stable process with variance
σ², the variance of the output under Rule 2 (adjusting for the last
result) is 2σ² — double the natural variation. Under Rule 4 (resetting
to the last result), the variance grows without bound. These are not
approximations or theoretical projections. They are mathematical
certainties.
But human beings are not wired to understand this intuitively. Our
brains are pattern-recognition engines that evolved to detect threats
and respond immediately. When we see a data point that deviates from our
target, our instinct is to correct. To act. To do something. The idea
that the correct response to variation in a stable process is to do
nothing — that feels like negligence.
This is why Deming considered the Funnel Experiment one of his most
important teachings. It was not just a statistical lesson. It was a
lesson about human psychology, about the relationship between action and
results, about the deep and dangerous human tendency to confuse movement
with progress.
The Four Rules,
Lived on Your Factory Floor
You might be tempted to think that your organization does not tamper.
You might be tempted to think that you follow Rule 1 — leaving the
process alone when it is stable. Before you conclude that, consider
these common industrial practices:
Rule 2 in action: Target chasing. A CNC operator
adjusts tool offsets after every part based on the last measurement. An
injection molding technician changes barrel temperatures when a single
part measures near the spec limit. A chemical process operator tweaks
reagent ratios based on the last batch result without checking whether
the process is in control.
Rule 3 in action: Drift through “improvement.” A
quality engineer adjusts a process based on deviation from target, but
uses the last adjustment as the new baseline. Each “correction” shifts
the reference point, and the process slowly wanders away from its
optimal setting like a ship with a broken compass.
Rule 4 in action: Uncontrolled resets. A changeover
procedure that sets machine parameters based on the last production
run’s final settings rather than engineering standards. A calibration
protocol that adjusts instruments to match the last reference standard
reading rather than the true reference value. A supplier that changes
its process based on the last lot received rather than the
specification.
Every one of these practices is well-intentioned. Every one feels
like prudent quality management. And every one violates the fundamental
principle that Deming spent his career trying to teach: if the
process is stable, the variation you see is the voice of the system. It
is not a problem to be fixed. It is the process telling you what it
is.
The Control Chart as
Tampering Prevention
This is precisely why Walter Shewhart invented the control chart in
the 1920s, and why Deming spent fifty years evangelizing it. The control
chart is not just a monitoring tool. It is a decision-making framework
that tells you when to act and — critically — when not to.
The control chart divides process behavior into two categories:
Common cause variation is the natural, inherent
variability of the process. It is stable, predictable, and present in
every process. It cannot be reduced by adjusting the process in response
to individual results. It can only be reduced by fundamentally changing
the system — better equipment, different materials, redesigned
processes, improved training.
Special cause variation is unexpected, unpredictable
variation that signals something has changed. A point outside the
control limits, a run of seven points above the center line, a trend of
six consecutive increasing points — these are signals that the process
has been disrupted and that investigation and correction are
warranted.
The rule is simple, but its implications are profound: react
to special causes. Do not react to common causes. The moment
you react to common cause variation as if it were special cause, you
become the special cause. You become the source of variation you were
trying to eliminate.
I once consulted for a pharmaceutical manufacturer that was producing
a liquid medication. The fill weight was monitored on a control chart,
and the process was beautifully stable. But every time the average
weight of a sample drifted above or below the target line — not the
control limit, the target line — the line operator would adjust the fill
volume. The control chart looked like a seismograph during an
earthquake: jagged, oscillating, wildly unstable.
When I asked the operator why she adjusted, she said: “Because my
supervisor told me to keep it as close to target as possible.”
The supervisor, when asked, said: “Because the customer expects
consistency.”
The customer, when asked, said: “We just need it within
specification.”
Three different definitions of “good enough.” The operator was trying
to hit the target. The supervisor was trying to minimize variation. The
customer just wanted product that met spec. And the result of all this
targeting and minimizing was a process that swung back and forth like a
pendulum, generating more variation — and more risk — than if everyone
had simply left it alone.
The Management Tampering
Trap
Tampering is not limited to the shop floor. In fact, some of the most
destructive forms of tampering happen in the boardroom, where executives
who do not understand process behavior make decisions that destabilize
systems they have never even seen.
The most common form of management tampering is the target-driven
intervention. A KPI shows a slight dip. The executive demands an action
plan. The manager produces one. Resources are diverted. The process is
adjusted. The next measurement cycle shows improvement — not because the
intervention worked, but because regression to the mean pulled the
result back to the process average. The executive claims credit. The
intervention is enshrined as best practice. And the process, quietly,
has been made worse.
Another form: the incentive trap. Management ties bonuses to defect
reduction targets. The workers, responding rationally to the incentive
structure, find ways to reduce the reported defect rate without actually
reducing defects. They redefine what counts as a defect. They increase
inspection tolerance. They stop reporting borderline cases. The numbers
improve. The reality does not. This is Rule 4 applied at the
organizational level — the process has been reset to wherever the last
measurement happened to land, and the reference point has drifted away
from the truth.
I saw this at an automotive components supplier that promised its
customer a 25 PPM defect rate. When the actual rate hovered around 80
PPM — still world-class by most standards — the quality manager began
reclassifying defects as “non-conformances” that didn’t count toward the
PPM calculation. When that wasn’t enough, he began shipping parts that
were borderline and letting the customer sort them. The reported PPM
dropped to 15. The customer’s incoming defect rate tripled. The supplier
lost the contract within a year.
They had hit the target. They had missed the point.
The Discipline of Inaction
The hardest thing for a quality professional to learn is not which
tool to use or which standard to follow. The hardest thing is knowing
when to do nothing.
This requires a kind of discipline that runs counter to every
instinct of professional competence. When your boss asks what you are
doing about the latest data point, “nothing, because the process is
stable” is not an answer that inspires confidence — even when it is the
correct answer. When your team looks to you for leadership after a bad
result, suggesting that the result is within the expected range of
variation sounds like complacency — even when it is the truth.
But this discipline — the discipline of distinguishing between signal
and noise, between special cause and common cause, between the variation
that demands action and the variation that demands acceptance — this is
the discipline that separates organizations that improve from
organizations that merely oscillate.
Deming understood this. He understood that most of what managers call
“management” is actually interference. He understood that the most
powerful lever in quality improvement is not the adjustment — it is the
restraint. The ability to look at a result that is imperfect, that is
not exactly on target, that shows deviation from the ideal, and say:
“This is the process. This is what it does. If we want different
results, we must change the system — not chase the data.”
How to Stop Tampering
Breaking the tampering habit requires structural changes, not just
individual awareness. Here is what works:
First, establish control charts at every critical process,
and make the control chart the trigger for action. If the point
is inside the limits and shows no patterns, no one adjusts anything.
This needs to be a written, enforced procedure, not a suggestion. The
control chart is not a suggestion box. It is a go/no-go gauge for
intervention.
Second, train everyone — operators, supervisors, engineers,
managers — on the difference between common cause and special cause
variation. This is not a one-time seminar. This is ongoing
education. People need to understand, in their bones, that reacting to
noise makes the noise louder.
Third, create a decision protocol that requires evidence of
special cause before any process adjustment. If someone wants
to change a setting, they need to show the control chart evidence that
justifies the change. No chart, no change. This simple rule eliminates
the vast majority of tampering.
Fourth, measure the measurement system. Before you
react to any data point, ask: is this data even reliable? MSA studies
exist precisely for this reason. A significant portion of what looks
like process variation is actually measurement variation — and adjusting
a process based on measurement noise is tampering squared.
Fifth, celebrate the absence of intervention. When
an operator sees a result that is off-target but within control limits
and chooses not to adjust, recognize that decision. Publicly. The
culture needs to value restraint as much as it values action.
The Deeper Lesson
The Funnel Experiment teaches something that extends far beyond
statistical process control. It teaches a fundamental truth about
systems: the people inside a system, acting on information from
within the system, cannot reliably improve the system by reacting to
individual events. Improvement requires stepping outside the
system, understanding its structure, and making systemic changes.
This is why Deming was so passionate about profound knowledge — the
understanding of variation, systems, psychology, and theory of knowledge
that he believed was essential for any leader. Without this
understanding, leaders are forever trapped inside the funnel, adjusting
and readjusting, making things worse while believing they are making
things better.
The marble does not need your help. The process does not need your
adjustment. What it needs — what every process needs — is for someone to
understand it deeply enough to know when it is speaking and when it is
merely breathing.
The stable process is breathing. Leave it alone.
Peter Stasko is a Quality Architect with 25+ years
of experience transforming organizations across automotive, aerospace,
and pharmaceutical industries. He specializes in bridging the gap
between statistical theory and shop floor reality, helping leaders
understand that the most powerful quality tool is sometimes the
discipline to do nothing.
