QM Bench

The Deming funnel: why adjusting a stable process makes it worse

W. Edwards Deming dropped marbles through a funnel onto a target and tried four ways of “correcting” the funnel after each drop. All four panels below get exactly the same random noise, so the only difference is the rule.

0 / 400 drops, same random noise for all four rules
Rule 1: leave the funnel alone
No adjustmentσ × 1.00

The funnel stays above the target. Only common-cause variation: the best result possible.

Rule 2: compensate from the last position
Adjust by the errorσ × 0.00

After each drop, move the funnel by minus the error from its current position. Variance doubles (σ × 1.41).

Rule 3: compensate from the target
Set opposite the errorσ × 0.00

Place the funnel at the target minus the last error. The pattern swings back and forth and grows.

Rule 4: aim at the last result
Follow the last pointσ × 0.00

Put the funnel over the last marble. A random walk: the process drifts away without limit.

Drawn at a smaller scale: it would leave the target.

What it teaches

Rule 1 is the only one that does not add variation. Every other rule reacts to common-cause variation as if it were a signal; this is called tampering or over-adjustment. Rule 2 increases the standard deviation by about 41 % (the variance doubles), rule 3 makes the process oscillate with growing amplitude, and rule 4 turns it into a random walk that drifts away.

Where you see it on the shop floor

  • An operator changes the machine offset after every measured part (rule 2).
  • Each new batch is set up “to compensate” the last batch’s error (rule 3).
  • A worker copies the last part instead of the master sample, or each training is given by the last trainee (rule 4).

The remedy is a control chart: adjust only when it shows a special cause. See it in the SPC simulator and test yourself in the SPC quiz.