Measurement decision rules

Illustrative comparison of three agreed rules

Assume a hypothetical upper requirement Y ≤ 100.0 units, measured value y = 99.0 units, and expanded uncertainty U = 2.0 units. The following are three different service agreements, not three simultaneous verdicts. Endpoints are included in the stated acceptance conditions.

1. Simple binary acceptance: accept when y ≤ 100.0. The result is accepted.
2. Guarded binary acceptance: accept when y + U ≤ 100.0. Here 101.0 ≤ 100.0 is false, so the result is rejected under that rule.
3. Three-outcome interval rule: accept if y + U ≤ 100.0; reject if y − U > 100.0; otherwise report unresolved conformity. The interval 97.0–101.0 crosses the limit, so this result is unresolved.

The third rule is defined solely for this example. Its intermediate label must not be assumed equivalent to a provider’s “conditional pass.” The measured result is unchanged in all three cases. The different dispositions come from different agreed treatments of uncertainty. These invented values illustrate logic, not a standard-prescribed limit or risk probability. Decision-error and acceptance-interval basis

Decision rule and scope

A measurement decision rule explains how measurement uncertainty is accounted for when declaring conformity with a requirement. It connects a result to a disposition such as pass, fail, or a defined conditional outcome. The specification establishes what values are permitted; the rule establishes how measured evidence is used to make the decision. Neither a calibration certificate nor a number inside a tolerance automatically identifies the rule. Global ACI decision-rule guidance, definitions

Global ACI-TECH-1-002 (G), version 1.0, has issue and implementation dates of 26 June 2026 on its cover. Its preamble and amendment table state that it reformats ILAC-G8:09/2019 without substantive content changes. Global ACI guidance, cover and amendment table

Establish the decision before measuring

First determine whether a conformity statement is needed. Where a regulation or governing standard prescribes the rule, use that basis. Otherwise, agree the specification, rule, and reporting outcomes with the customer before work starts. Check whether a method already incorporates uncertainty into its limits before adding another guard band. A rule based on a minimum uncertainty ratio also needs instructions for measurements that do not achieve that ratio. Global ACI guidance, sections 7–8

Choosing a rule involves consequences as well as mathematics. False acceptance releases a nonconforming item; false rejection withholds a conforming one. NIST’s Phillips and Krystek paper, published April 2014, describes uncertainty evaluation as technical work and decision-rule selection as a business decision involving product-specific costs. A stricter acceptance rule may reduce one kind of error while increasing another. Do not change the rule after seeing an inconvenient result. NIST conformity, decision rules, and risk analysis

Distinguish common rule families

An acceptance interval concerns permissible measured values; a tolerance interval concerns permissible values of the property. Keeping those intervals separate prevents “fails acceptance” from being mistaken for certainty that the item is physically outside specification. JCGM 106:2012, dated October 2012, develops these distinctions and the associated probability models. JCGM 106:2012, decision rules

Use the correct uncertainty and risk basis

Expanded uncertainty is U = k·u, where u is the combined standard uncertainty and k the coverage factor. Approximately 95% coverage for k = 2 depends on an applicable normal distribution and a sufficiently reliable standard-deviation estimate. An expanded interval is not an absolute error bound. Record whether the rule uses u, U, a percentile, a fixed margin, or another defined quantity. NIST expanded uncertainty and coverage

A probability for a particular measured item differs from an overall probability across a production population. Global risk depends on production and measurement distributions, not only on an uncertainty number. A quoted “2% false-accept risk” therefore needs a definition and model; it does not automatically describe every result near a limit. JCGM 106:2012, specific and global risks

Bounded laboratory examples

RMS Foundation states that its default conformity approach, unless another rule is specified or agreed, is simple acceptance without including uncertainty in the decision. It also specifies rounding to the decimal places required by the standard or technical specification before assessment, and includes a boundary where the permitted minimum/maximum is inclusive. This is a named laboratory’s policy. It demonstrates why rounding and endpoint handling must be obtained from the actual rule rather than guessed. RMS Foundation decision-rule policy

NI’s calibration-service explanation, updated September 20, 2024, describes a different implementation: guarded points can receive conditional pass/fail markers, identified in the certificate status column. Its stated trigger includes a TUR below 4:1 for double-sided limits and guard bands for single-sided limits. Those are provider-specific choices, not universal requirements of laboratory accreditation. Read the status legend and applicable service scope. NI calibration guard-band implementation

SATRA describes another approach for numerical results near requirements: within its uncertainty region it reports explanatory information without a pass/fail or performance-class statement. This illustrates that a useful report can communicate a result and its limitation without forcing a binary answer. It does not authorize substituting SATRA’s policy for a contractual binary rule. SATRA conformity-reporting approach

Make the decision reproducible

A practical decision record should identify the result and measurand, requirement and revision, units, uncertainty/coverage basis, formula or published rule, relevant limits, rounding stage, endpoint treatment, and outcome. Preserve the rule version alongside the report so a later reviewer can reconstruct the comparison. Define retesting and conditional-result handling in advance; selecting only a favorable repeat result changes the evidence rather than clarifying it.

For example, if an agreed calculation accepts y ≤ 100.0, a stored value of 100.04 and a displayed value of 100.0 can produce opposite outcomes depending on the specified rounding policy. This invented example explains why display formatting alone must not determine acceptance. RMS explicitly states its rounding policy and identifies its rule on certificates, showing how such ambiguity can be removed. RMS policy on assessment and reporting

Related references

Guard bands in conformity assessment · Inspection and quality documents