Guard banding explained with a worked example

Guard banding, written as one word as often as two, narrows the acceptance zone by a multiple of the measurement uncertainty so that a pass has a controlled probability of being wrong. Tolerance ±0.5 °C. Measurement uncertainty U = 0.12 °C (k = 2). With a guard band w = U, acceptance limits become ±0.38 °C. A measured error of 0.40 °C is within tolerance but outside the acceptance limits: under this rule it is not declared a pass.

A guard band is one decision rule, not the only one

Everything here is the arithmetic of a particular decision rule: narrow the acceptance zone, then apply simple binary acceptance inside it. The choice of w is not a technical detail hidden in a spreadsheet — it is the decision rule itself, and the standard requires that it be documented and agreed with the customer before the work, not chosen afterwards when the number is already known.

Simple acceptance with no guard band is also a decision rule, and a defensible one where the risk has been agreed; what is not defensible is having no stated rule and deciding case by case. The alternatives, and what has to appear on the certificate for each, are set out in the decision rule article.

If you want the numbers for your own tolerance and uncertainty rather than the ones below, the guard band calculator does the arithmetic in your browser and reports the false-accept probability at the acceptance limit.

Why 2.3 %

At a measured value exactly on the acceptance limit (0.38 °C), the true value has a roughly normal distribution centred there with standard deviation U/2 = 0.06 °C. The tolerance limit (0.50 °C) is two standard deviations away, so the probability that the true value exceeds it is 2.3 %. Inside the acceptance zone the probability is smaller.

Choosing w

• w = U: conservative, common in regulated environments.
• w = 0.5 U or another fraction: when false rejects are costly and the consequence of a false accept is modest.
• w > U: when a false accept is dangerous.

A second example, where it changes the answer

A reference thermometer is specified at ± 0.05 °C. The calibration returns an error of 0.041 °C with U = 0.020 °C at k = 2. Simple acceptance passes it comfortably — 0.041 is inside 0.050. Under a guard band of w = U the acceptance limit is ± 0.030 °C, and 0.041 is outside it: the instrument fails.

That is not the rule being harsh. It is the rule surfacing the fact that the measurement interval runs from 0.021 °C to 0.061 °C, and the upper part of it is out of specification. Simple acceptance did not prove the instrument conformed; it declined to ask.

The σ in the second column is u_c, the combined standard uncertainty, which is U/2 when k = 2. That is the whole arithmetic behind the table: a guard band of one U puts the tolerance limit two standard uncertainties away from the acceptance limit, and two standard deviations of a normal distribution leave 2.275 % in the tail -- 2.3 % to the precision any of this supports.

You will often see 2.5 % quoted for this instead. That is the figure for k = 1.96, which gives exactly 95 % coverage; at k = 2, which is what certificates almost always use, the tail is 2.275 %. The difference never changes a decision, but quoting the k = 1.96 number beside a k = 2 uncertainty is the kind of small inconsistency this audience notices.

Guard bands work in both directions

Everything above narrows the acceptance zone to protect against false accepts. The mirror image exists and is occasionally the right choice: widening the acceptance zone by w so that instruments close to the limit are passed rather than rejected, accepting a higher false-accept risk in exchange for fewer false rejects. ILAC-G8 permits it where the consequence of a false accept is low and the cost of rejecting good instruments is high — but it has to be stated just as explicitly, because a reader will otherwise assume the conventional direction.

What the certificate has to say

• The tolerance or specification the conformity is assessed against.
• The expanded uncertainty and its coverage factor — a guard band quoted without the U it was built from cannot be checked.
• The guard band width, as a multiple of U or as an absolute value.
• The resulting acceptance limits, spelled out. Do not make the reader compute them.
• The verdict, and for non-binary schemes, which band it fell in.

A guard band cannot rescue a bad uncertainty ratio. At a TUR of 1:1 the guard band consumes the entire tolerance and nothing can be declared conforming — which is the honest answer, and a signal to improve the measurement rather than the paperwork.