Does good repeatability mean the results are correct?

No; a systematic offset can affect every result similarly.

Is repeatability always measured by a standard deviation?

Standard deviation is common for quantitative results, but the measure should suit the data and purpose.

Can a deterministic AI model have imperfect measurement repeatability?

Yes; changing image acquisition noise or other inputs can change its measurements even when its computation is deterministic for an identical input.

AI Glossary / Statistics and probability / Foundation

Repeatability

Definition

Repeatability is the closeness of agreement among repeated measurement results obtained under the same specified measurement conditions. In measurement science, those conditions typically include the same procedure, operator, measuring system and location over a short period. Repeatability describes a component of precision, not necessarily agreement with a true or reference value. Other disciplines sometimes use the word more broadly, so this entry states the measurement sense explicitly rather than treating all uses as interchangeable.

What it means

A measurement process can produce slightly different values even when the item and intended procedure stay unchanged. Repeatability asks how much such short-term variation occurs under controlled conditions. Small variation means good repeatability for those conditions, but every result could still share a calibration offset. Reproducibility examines agreement under specified changed conditions, such as different laboratories or operators. Neither term is informative without describing what stayed fixed, what changed and what quantity was measured.

How it works

Choose an item or suitable stable reference, define the measurement procedure and hold the relevant conditions as constant as practicable. Take repeated measurements over an appropriate short interval and record the circumstances. Estimate within-condition variation using a standard deviation or another justified precision measure. When measurements cover several items or sessions, use an analysis that respects the grouping rather than pooling unlike variation indiscriminately. Check for trends, resolution limits and dependence before interpreting a single summary as a stable process characteristic.

Three practical examples

Example 1: A technician repeatedly weighs the same reference object on one balance during a short session. Example 2: A factory tests how consistently one vision measurement system estimates the width of a stationary part under fixed illumination. Example 3: A laboratory assesses repeated readings from the same specimen using the same instrument and method. These exercises describe the specified measurement conditions. They do not by themselves show consistency across different specimens, operators, instruments or days.

Think of it this way

Think of placing several marks close together on a target. Their tight grouping represents repeatability, while their position relative to the center concerns accuracy and bias. The analogy also has a limit: a real measurement study must define the conditions behind the marks. A group made with one instrument in one minute cannot automatically describe performance after transport, recalibration or a change in operator.

Worked example

Suppose three measurements of a stable reference are 99.8, 100.0 and 100.2 units under the same conditions. Their mean is 100.0. The squared deviations sum to 0.04 + 0 + 0.04 = 0.08, so the sample variance is 0.08/(3 - 1) = 0.04 and the sample standard deviation is 0.20 units. If independent measurement errors with stable variance are a reasonable model, the estimated standard error of the mean is 0.20/square root of three, approximately 0.1155 units. Now add a calibration offset of two units to every reading. The mean becomes 102.0, but the standard deviation remains 0.20. Repeatability is unchanged even though agreement with a reference value of 100 has worsened. Three readings are far too few to establish a robust long-term precision characterization.

Illustrative case study

In this fictional case study, Maple Instruments prepares an AI-assisted dimension checker. Repeated measurements of one fixed part cluster closely, so a manager initially calls the system accurate. The metrology lead compares results with a reference and identifies an offset, then separately studies changes across operators and lighting setups. The report presents short-term repeatability, calibration findings and variation under changed conditions as separate results. It avoids claiming that a narrow spread in one controlled session guarantees reliable measurements throughout the factory.

Common misconceptions and mistakes

Common mistakes include equating repeatability with accuracy, assuming repeated identical software outputs establish measurement quality, and treating all repeated observations as independent. Another error is mixing day-to-day drift with within-session variation without identifying the components. A digital instrument's coarse rounding can make values look perfectly repeatable while hiding smaller fluctuations. In computational research, repeating code with the same data addresses a different question from measuring the same physical quantity repeatedly under metrological repeatability conditions.

Business use

Businesses use repeatability studies to judge whether measurement noise is small enough for inspection, monitoring or model-input requirements. A measurement system with poor repeatability can obscure genuine process changes and weaken labels used to train AI. Improvement may involve stabilizing fixtures, procedures or environmental conditions. Decisions should compare the estimated variation with the tolerance or task requirement while accounting for uncertainty in the precision estimate. Calibration, reproducibility and suitability across the operating range still require their own evidence.

Check your understanding

What happens to repeatability standard deviation when two units are added to every reading in the example?

Answer: It remains 0.20 units.

Why is the standard error of the mean smaller than the standard deviation of readings under the stated assumptions?

Answer: Averaging independent errors reduces variation of the mean.

Does one fixed-condition study establish performance across laboratories?

Answer: No; those changed conditions require a suitably designed reproducibility study.

Related terms

Bias
Calibration
Related terms include precision
measurement accuracy
reproducibility
measurement uncertainty
gauge repeatability and reproducibility
and intermediate precision

Learn next

Learn next how measurement variation is separated into within-session, between-day and between-operator components. Study the distinction between standard deviation and standard error, then examine uncertainty budgets that include calibration and other contributions. For AI inspection, test both image-acquisition variation and the model's sensitivity to it instead of evaluating repeated identical files alone.

Sources

Sources: NIST, Technical Note 1297 Appendix D1, Terminology, https://www.nist.gov/pml/nist-technical-note-1297/nist-tn-1297-appendix-d1-terminology; NIST/SEMATECH e-Handbook, Variability, https://itl.nist.gov/div898/handbook/mpc/section1/mpc114.htm; Analysis of repeatability, https://www.itl.nist.gov/div898/handbook/mpc/section4/mpc441.htm. The numerical measurements and fictional case are original examples, with independence stated only where required for the mean's standard error.

Previous: Reparameterization trick Next: Repeated game