Measurement uncertainty explained

Every measurement result is an estimate. Measurement uncertainty is the parameter that describes the spread of values that could reasonably be attributed to the measurand, given everything known about the measurement. It is not the error — the error is unknown — and it is not a tolerance.

Building an uncertainty budget

• List every influence: reference standard, drift, resolution, repeatability, environment, method.
• Quantify each as a standard uncertainty. Type A evaluations use statistics on repeated readings; Type B evaluations use other information such as certificates and specifications, with an assumed distribution (normal, rectangular, triangular).
• Combine the standard uncertainties. For independent inputs the combined standard uncertainty is the root sum of squares.
• Multiply by a coverage factor k to obtain the expanded uncertainty. k = 2 corresponds to approximately 95 % coverage for a normal distribution.

Reading an uncertainty statement

'U = 0.12 °C, k = 2' means the laboratory believes the true value lies within ±0.12 °C of the reported value with roughly 95 % confidence. Without the coverage factor the number is ambiguous. Without knowing whether the instrument's own contribution was included, the number cannot be applied to your measurements.

Why it is the model's job to explain, not to calculate

Combining components is simple arithmetic, but it is exactly the kind of arithmetic a language model gets wrong quietly. In ValiTrac the combination and expansion are performed by a deterministic calculation engine and returned to the model as tool results; the model explains the result and cites the evidence, and any answer that involved a calculation is flagged for human review.

A complete budget, end to end

The idea is easier to hold once you have seen one worked. This is a 0.1 °C data logger calibrated by comparison in a stirred bath at 5 °C.

The last column is the point

Resolution contributes 68 % of the variance here, more than everything else combined, because components are squared before they are added and the largest one therefore dominates disproportionately. Replacing the reference thermometer with one twice as good would move u_c from 0.0351 to about 0.0339 — a 3 % improvement, for real money. Reading the logger from its 0.01 °C export instead of its 0.1 °C display would roughly halve the total, for nothing.

That is what a budget is for. It is a diagnostic that tells you where to spend, and reading it only for the final number throws away most of its value.

Type A and Type B are about method, not quality

A common misreading is that Type A components are the good ones and Type B are guesses. The distinction is purely how the uncertainty was evaluated: Type A by statistical analysis of repeated observations, Type B by any other means — a certificate, a specification, a distribution assumed from physical reasoning. A certificate uncertainty from a national metrology institute is Type B and is the best-known number in most budgets. Both types combine identically.

Reading a statement someone else wrote

• U = 0.12 °C, k = 2 means the true value is believed to lie within ± 0.12 °C of the reported value with roughly 95 % confidence.
• Without k, the number is ambiguous by a factor of two or more and cannot be compared with anything.
• Check whether the instrument was adjusted. An as-left uncertainty says nothing about what the instrument was doing during the interval that just ended.
• Check the point. An uncertainty quoted at 0 °C does not apply at 200 °C, and certificates often quote the best one prominently.
• Compare it against your tolerance before anything else. A valid ± 0.5 °C certificate is useless for policing a ± 0.3 °C limit, however good the laboratory.