Error vs uncertainty: the measurement words people mix up

A supplier's inspection report lands on the quality engineer's desk: "Bore diameter 25.012 mm, accuracy plus or minus 0.005, error within tolerance." The metrologist down the hall reads it twice and puts the kettle on. Three measurement words in one line, not one of them used the way the writer thought, and the word the report needed, uncertainty, is missing entirely. In everyday English these words are interchangeable. On a calibration certificate they are not, and the difference decides which parts ship. Here are the pairs that trip people most, and how the people who measure for a living keep them apart.

A difference versus a doubt

The error of a measurement is a signed difference: the measured value minus a reference value for the measurand, the exact quantity you set out to measure. Read 25.012 when the bore is really 25.010 and the error is plus 0.002. The catch is in that word really. The reference only stands in for the true value, which no measurement can reveal, so nobody knows the error exactly either.

Uncertainty is the working answer to that problem. It is never signed and never zero: a figure for how widely the real value could plausibly sit around your reading, given everything you know. A complete measurement result carries both halves, "25.012 mm, plus or minus 0.003 mm," and to a metrologist a number without its uncertainty is half a sentence.

When an error is known well enough, you remove it with a correction: the lab found the gauge reads 0.002 high, so subtract 0.002. What remains is not zero error. It is the uncertainty of the correction. "We corrected for it" and "it's gone" are different claims, and auditors know the difference.

The dartboard

Picture a dartboard. Darts in a close cluster near the rim are precise but not accurate. Darts scattered evenly around the bullseye are on target on average, and sloppy one by one. Precision is how closely repeated readings agree with each other, whatever the true value. Trueness is how close their average lands to the reference. Accuracy covers both at once: how close a single reading comes to the truth.

The strict point that annoys newcomers is that accuracy is a quality, not a quantity. One gauge can be more accurate than another, but a measured value does not come with "an accuracy of 0.005." That number is an uncertainty, a manufacturer's spec or a tolerance, and it matters which.

A process engineer calls a narrow spread tight, and a tight cluster can still be in the wrong place. Neither word means resolution, the smallest change a display can show. A six-digit meter with a wandering zero only looks precise, and its last digit, counted in counts, says nothing about whether the first five are right.

Two kinds of wrong

Every error has two parts. Random error changes from one reading to the next and shrinks when you average. Systematic error stays put or moves predictably, and averaging does nothing for it. Take a hundred readings with a worn caliper and you get a beautifully stable average of the wrong number.

Bias is the estimate of that systematic part: the average reading minus the reference value. A bias that shows up as a nonzero reading at zero input is an offset, the thing operators zero out at the start of a shift. A bias that creeps over weeks is drift, which is why labs keep a check standard and measure it on a schedule. On a production floor the same idea goes by golden unit or golden wafer: a known part that should read the same every time, and raises eyebrows when it does not.

Same part, different hands

Repeatability is one person, one gauge, one method, measuring the same part again and again in a short window. Reproducibility is what happens when you change something on purpose: another operator, another shift, another lab. The second number is always the bigger one, because it carries the first inside it, and it is the one that matters when two factories argue about whether a part is good.

The floor test for both is the gauge R&R, part of a wider MSA. Several operators measure several parts several times, and the study splits the scatter three ways: the parts themselves, the gauge's own repeatability and the operators' reproducibility. The headline is the P/T ratio, measurement spread as a fraction of the tolerance. "Thirty percent P/T" means the gauge eats almost a third of the tolerance before the parts get a say.

Three kinds of limit

Tolerance is what the drawing allows. Uncertainty is what you know about one measurement. Confusing them is how good parts get scrapped and bad ones shipped. The four to one rule keeps them in proportion: measurement uncertainty should be no more than a quarter of the tolerance it checks, a ratio formally called the TUR. Its cousin, the ten-to-one rule, is about resolution: the gauge should read in steps no coarser than a tenth of the tolerance. When a reading lands near the edge, a guard band pulls the acceptance limits inside the tolerance, so a part has to clear the line by more than its own uncertainty.

Two other limits get mixed up even more often. Specification limits come from the customer or the designer: outside them, the product is OOS and nonconforming. Control limits come from the process's own history and mark where routine variation ends. A point beyond them is OOC, which means something changed, not that anything is bad. A process can be in control and out of spec, or out of control and making perfect parts.

Sigma, k and the interval words

Uncertainty borrows the vocabulary of statistics and bends it slightly. A standard deviation describes the scatter of a set of readings. A standard uncertainty is any uncertainty component expressed at that one-standard-deviation size, whether it came from scatter (a Type A evaluation) or from a certificate or data sheet (a Type B evaluation). Combine them in an uncertainty budget, multiply by a coverage factor of two, and you have an expanded uncertainty, which the lab calls two sigma and the certificate calls about 95 percent coverage.

The intervals have three names that are not interchangeable. A coverage interval brackets the plausible values of one measurand. A confidence interval brackets an unknown parameter, such as a process mean. A tolerance interval brackets a stated share of all the individual parts. Ask for the wrong one in a meeting and you will get a correct answer to a question nobody asked.

Five pairs to keep straight

1. Error and uncertainty: a signed difference you can never know exactly, versus a stated spread you can.
2. Accuracy and precision: close to the truth, versus close to each other.
3. Tolerance and uncertainty: what the drawing allows, versus what the measurement knows.
4. Spec limits and control limits: the customer's line, versus the process's own.
5. Calibration and adjustment: comparing an instrument against a standard and writing down what you found, versus turning the screw. A calibration with no adjustment is still a calibration, and the as found data is often the most useful line on the certificate.

Get those five right and the supplier's report rewrites itself: "25.012 mm, expanded uncertainty 0.003 mm at k = 2, within tolerance." The metrologist can drink the tea while it is hot.