A customer returns 40 shafts you rejected, with their own report showing every one inside tolerance. Both measurements were made with calibrated equipment by competent people. Neither of you is wrong, and neither of you can prove the other is, because the parts sit within the measurement uncertainty of both systems.
Every measurement carries uncertainty. Near a tolerance limit, that uncertainty means some conforming parts get rejected and some non-conforming parts get accepted. Guard banding is the decision about which of those two errors you are willing to make, written down before the argument starts rather than after.
The default rule in ISO 14253-1 is that uncertainty works against the person making the claim. If you are the supplier claiming conformance, the uncertainty is subtracted from the tolerance zone, and your acceptance limits move inward. If you are the customer claiming non-conformance, you carry the same burden in the other direction.
Take a 50 mm H7 hole, tolerance 0 to +25 µm, checked with a plug gauge whose expanded uncertainty U is 3 µm:
| Limit | Drawing | Acceptance limit for proving conformance |
|---|---|---|
| Lower | 50.000 mm | 50.003 mm |
| Upper | 50.025 mm | 50.022 mm |
The conformance zone has shrunk from 25 µm to 19 µm, and 24% of the drawing tolerance is now unusable to you. That is the cost of your measurement uncertainty, and it is real whether or not anybody calculates it.
The ratio between the part tolerance and the measurement uncertainty determines how much of the tolerance survives.
| TUR (tolerance ÷ 2U) | Guard band each end | Tolerance left |
|---|---|---|
| 10:1 | 5% | 90% |
| 4:1 | 12.5% | 75% |
| 3:1 | 17% | 66% |
| 2:1 | 25% | 50% |
The traditional gauge maker's rule of 10:1 exists for this reason: at that ratio the guard band is small enough to ignore in most production work. At 4:1, the widely used minimum in automotive quality systems, you give up a quarter of the tolerance. Below 3:1, the measurement is doing more harm than good and the honest answer is better equipment.
Gauge tolerance eats into this before you start. A plug gauge made to a standard gauge tolerance already consumes part of the part tolerance, which is why gauge class matters on tight work and why buying a cheaper class to save money can cost more in scrapped good parts.
Simple acceptance. Accept anything inside the drawing limits, ignore uncertainty. Fast, and standard practice on loose tolerances. Both parties share the risk roughly equally, and it only works when the TUR is comfortable and both sides have agreed to it.
Guard banded acceptance. Pull acceptance limits inward by the guard band. You will scrap some good parts. You will not ship bad ones. Use this for safety-critical features, for customers with a history of disputes, and where a field failure costs more than the part.
Guard banded rejection. Push rejection limits outward, so you only reject when you are confident. Useful for incoming inspection where rejecting a good lot creates a shortage.
The choice is commercial as much as technical. Write it into the quality plan, name the rule, and state the uncertainty you assumed.
Operators do not calculate guard bands at the bench. They read a limit off a work instruction, so the arithmetic has to be done once, by someone who knows the gauge.
That last line settles most disputes before they start. A report saying "accepted per ISO 14253-1 guard banded acceptance, U = 3 µm, k = 2" is a different document from one that just says "pass", and it is the one that holds up in a supplier audit.
A NABL-traceable calibration report from us states the expanded uncertainty and the coverage factor, because a calibration that does not tell you the uncertainty has not given you what you need to make a defensible decision.