# Guard bands in conformity assessment

Separating specification and acceptance limits, selecting uncertainty-aware decision rules, and interpreting false-accept and false-reject risks.

Reviewed: 2026-11-01

Tags: inspection, metrology, conformity-assessment

An inward guard band moves the acceptance limit while leaving the drawing tolerance in place. For upper limit T and band width w, the acceptance limit is A = T − w. [ILAC-G8:09/2019, section 4](https://www.iecex.com/dmsdocument/3485/)

In the example, a tolerance of 9.900–10.100 mm and an agreed inward band w = U = 0.020 mm give an inclusive acceptance interval of **9.920–10.080 mm**. A reading of 10.090 mm is accepted by simple acceptance and rejected by this guarded binary rule. With U reduced to 0.005 mm and w still equal to U, the same reading is accepted.

Choose the decision rule before measuring. The consequences of accepting nonconforming items and rejecting conforming items inform that choice. [NIST decision-rule and risk-analysis paper](https://www.nist.gov/publications/assessment-conformity-decision-rules-and-risk-analysis)

## Keep the quantities separate

| Quantity | Meaning | Common mistake |
| --- | --- | --- |
| Tolerance interval | Values permitted by the specification | Shrinking the drawing tolerance instead of documenting acceptance limits |
| Acceptance interval | Measured values accepted under the rule | Assuming it always equals the tolerance interval |
| Standard uncertainty, u | Uncertainty expressed as a standard deviation | Substituting it for an expanded uncertainty without conversion |
| Expanded uncertainty, U | Standard uncertainty multiplied by coverage factor k | Assuming k = 2 always establishes exactly 95% coverage |
| Guard-band width, w | Separation of corresponding limits | Assuming w must equal U in every application |

NIST explains that **U = k·u** and that associating coverage with an expanded interval requires justified distributional assumptions. With an applicable normal distribution and a reliable standard-deviation estimate, k = 2 gives approximately 95% coverage. It does not mean the error is known to be bounded absolutely by ±U. State the uncertainty quantity and coverage basis used in the decision. [NIST expanded uncertainty and coverage factors](https://physics.nist.gov/cuu/Uncertainty/coverage.html)

## Specify the rule and its outcomes

Simple acceptance uses the tolerance limits as acceptance limits: w = 0. Guarded acceptance restricts the accepted interval. ILAC-G8:09/2019 describes both binary outcomes and non-binary reporting, and permits a guard band based on **w = rU** rather than requiring r = 1. A binary rule can reject a value still inside the tolerance interval. A non-binary rule can instead distinguish conditional outcomes near a boundary. Its terminology and boundary treatment must be explicit. [ILAC-G8:09/2019, sections 4–5](https://www.iecex.com/dmsdocument/3485/)

SATRA provides a concrete alternative to forcing every result into pass/fail. In its published testing approach, values within an uncertainty region around a requirement receive explanatory notes without a pass/fail or class statement. This is SATRA’s reporting policy, not a universal instruction to label every overlapping interval “indeterminate.” Use the rule required or agreed for the actual examination. [SATRA conformity-reporting approach](https://www.satra.com/spotlight/article.php?id=546)

## Understand which risk is being controlled

Consumer’s risk concerns false acceptance; producer’s risk concerns false rejection. **JCGM 106:2012**, dated October 2012 on its cover, distinguishes risk for a particular measured item from global risk over a production process. Global calculations depend on the distribution of produced items as well as the measurement system. A claimed “2% risk” is incomplete unless the risk definition, model, and conditions are identified. [JCGM 106:2012, sections 8–9](https://www.bipm.org/documents/20126/2071204/JCGM_106_2012_E.pdf)

For fixed production and measurement conditions, moving acceptance limits inward reduces false acceptance while increasing false rejection. Guarded rejection moves a decision threshold outward to reduce false rejection; it serves a different objective. Neither arrangement removes all decision risk. Avoid interpreting a coverage probability as a guaranteed fraction of defective items among shipments. [JCGM 106:2012, guarded decisions and global risks](https://www.bipm.org/documents/20126/2071204/JCGM_106_2012_E.pdf)

## Illustrative dimensional example

Assume a hypothetical part specification is **9.900–10.100 mm**, the expanded measurement uncertainty is **U = 0.020 mm**, and the agreed binary rule uses **w = U** at both ends. For this example, endpoints are included in acceptance.

The acceptance interval is therefore:

**[9.900 + 0.020, 10.100 − 0.020] = [9.920, 10.080] mm**.

| Measured value | Simple acceptance | Agreed guarded acceptance |
| --- | --- | --- |
| 10.070 mm | Accept | Accept |
| 10.090 mm | Accept | Reject under this rule |
| 10.110 mm | Reject | Reject |

For 10.090 mm, the reported expanded interval is **10.070–10.110 mm**. Its overlap with the specification does not supply the decision rule; the chosen rule supplies the disposition. The rejected result does not prove with certainty that the part’s value is outside tolerance. These numbers are invented to demonstrate arithmetic, not a prescribed dimensional-inspection rule. [Guarded-acceptance construction](https://www.iecex.com/dmsdocument/3485/)

If improved measurement reduced U to **0.005 mm** while retaining w = U, the acceptance interval would become **9.905–10.095 mm**. The same hypothetical measured value, 10.090 mm, would now be accepted. This illustrates the operational benefit of reducing uncertainty; it does not justify discarding repeat results until one passes.

## Bounded service examples and implementation

NI’s calibration-service page describes simple binary acceptance by default, with guard bands applied to double-sided test points when its TUR is below 4:1 and to single-sided limits. Guarded points use non-binary reporting, including conditional pass/fail markers. This is a named provider’s implementation; the 4:1 threshold is not a universal guard-band requirement. Read the actual certificate’s rule and status legend. [NI guard-band implementation](https://www.ni.com/en/shop/services/hardware/calibration-services/calibration-services-guard-band-implementation.html)

Transcat similarly distinguishes “Fails Acceptance” against guarded limits from “Out Of Tolerance” against original limits in its service description. That distinction is useful when interpreting records: a supplier’s status label should be read alongside the limits it references. Do not translate every guarded rejection into an unqualified physical nonconformance claim. [Transcat guard-band reporting example](https://www.transcat.com/calibration-resources/white-papers/guard-banding-101)

For a usable decision record, retain the measurand, specification and revision, units, corrected result, uncertainty and coverage basis, guard-band calculation, acceptance limits, endpoint/rounding policy, rule reference, and disposition. Agree how conditional results, retesting, and departures are handled before testing. Keep uncertainty evaluation technically valid, then choose the risk allocation appropriate to the requirement and its consequences. [NIST decision-rule selection basis](https://www.nist.gov/publications/assessment-conformity-decision-rules-and-risk-analysis)

## Related reference

[Inspection and quality documents](../../sourcing/documents/index.md)