Decision rules, guard banding, TUR and calibration intervals under ISO/IEC 17025

What a decision rule is, and why ISO/IEC 17025 requires one

A decision rule is the documented method you use to turn a measured value and its uncertainty into a pass or fail statement against a specification. Every measurement carries uncertainty, so a reading that sits close to a tolerance limit is genuinely ambiguous: the true value could be on either side of the line. The decision rule is the policy that resolves that ambiguity the same way every time, and it makes clear who carries the risk when a result lands in the gray zone.

ISO/IEC 17025:2017 ties this to two clauses. Clause 7.1.3 requires that, when the customer requests a statement of conformity to a specification or standard, the decision rule is clearly defined and, unless it is inherent in the requested specification, agreed with the customer. Clause 7.8.6 then governs what the statement of conformity must contain: it must identify which results the statement applies to, which specifications or limits are met or not met, and the decision rule applied, taking account of the level of risk associated with that rule.

In plain terms: you cannot stamp PASS on a certificate without first deciding, and recording, how you handle uncertainty near the limits. The underlying statistics come from the GUM (JCGM 100:2008, the Guide to the expression of uncertainty in measurement) and its conformity companion JCGM 106:2012, which formalizes how measurement uncertainty translates into the risk of a wrong pass/fail decision. ILAC-G8:2019 is the accreditation-body guidance that turns those ideas into the decision-rule language assessors look for.

Test uncertainty ratio (TUR): the 4:1 rule of thumb

The test uncertainty ratio answers a simple question: how much of the tolerance does my measurement uncertainty eat up? ANSI/NCSL Z540.3 defines TUR as the ratio of the span of the tolerance to twice the 95% expanded uncertainty of the measurement process used for the calibration:

TUR is not the older test accuracy ratio (TAR). TAR compares only the tolerance of the unit under test to the tolerance of the reference standard, ignoring the method, the environment, repeatability, and everything else that actually contributes to uncertainty. TUR uses the full uncertainty budget, so two labs with the same reference standard can have very different, and more honest, TURs. Where you see an old TAR claim, treat it as optimistic.

The 4:1 rule of thumb says that if your TUR is at least 4:1, the measurement is precise enough relative to the tolerance that simple acceptance is reasonable, the uncertainty consumes about a quarter of the tolerance band on each side. Z540.3 makes this concrete by capping the probability of false accept at 2% and treating a 4:1 TUR as the practical alternative when that probability cannot be computed directly. The 10:1 ratio is a legacy ideal from an era of much coarser standards; it is rarely attainable on modern, tight tolerances and is no longer the working target.

Guard banding: ILAC-G8, and who carries the risk

A guard band is an offset, written w, that moves the acceptance limit away from the specification limit. With guarded acceptance you only accept results that fall inside the tightened limit:

The most common, and the most assessor-friendly, choice is w = U, the 95% expanded uncertainty. ILAC-G8:2019 calls this binary statement with guard band: you set w equal to the expanded uncertainty so that a PASS means the result is inside the tolerance even after the uncertainty is subtracted. This is also known as stringent or 95% guarding, and it pushes the probability of a false accept down toward roughly 2.5% or lower at the limit.

Guard banding is fundamentally about choosing whose risk you protect:

There is no guard band that minimizes both at once, every step that protects the consumer costs the producer, and vice versa. That is why the decision rule must be agreed with the customer (clause 7.1.3): you are explicitly choosing a risk posture on their behalf. In safety-critical work, biasing toward low false-accept risk is usual; for production gauges with tight tolerances and high scrap cost, a smaller guard band that accepts more producer risk may be the right business call, as long as it is documented and agreed.

Two worked examples

Example 1: a 2:1 TUR with a w = U guard band

You are calibrating a digital multimeter on its 10 V DC point. The manufacturer's tolerance is 10.000 V ± 0.020 V, so the specification runs from 9.980 V to 10.020 V (a 0.040 V span). Your calibration process has a 95% expanded uncertainty of U = 0.010 V.

• TUR = 0.040 / (2 × 0.010) = 2:1. That is below 4:1, so simple acceptance would leave an uncomfortable false-accept risk near the limits.
• Guard band w = U = 0.010 V. The guarded acceptance limits become 9.990 V and 10.010 V.
• Decision: a reading of 10.012 V is inside the spec (under 10.020 V) but outside the guarded limit (over 10.010 V), so under w = U guarded acceptance it does not earn a PASS, the uncertainty could put the true value over the line. A reading of 10.004 V is inside both, and passes.

Notice the cost: the usable acceptance window shrank from 0.040 V to 0.020 V, half the tolerance is now consumed by guard band. That is the honest price of a 2:1 TUR. The decision-rule statement on the certificate would read something like: Binary statement with guard band, w = U (95% expanded uncertainty); PASS only where the measured value lies within the tolerance reduced by the expanded uncertainty.

Example 2: an asymmetric specification

Not every tolerance is centered. Suppose a pressure gauge must read 100.0 kPa ^+0.8/_−0.3, an asymmetric tolerance with USL = 100.8 kPa and LSL = 99.7 kPa. Your process uncertainty is U = 0.15 kPa. The key point is that the guard band is applied to each limit independently, not symmetrically around nominal:

• Total span = 0.8 + 0.3 = 1.1 kPa, so TUR = 1.1 / (2 × 0.15) ≈ 3.7:1, just under 4:1.
• With w = U = 0.15 kPa, the upper guarded limit is 100.8 − 0.15 = 100.65 kPa; the lower guarded limit is 99.7 + 0.15 = 99.85 kPa.
• The accept window is 99.85 to 100.65 kPa, narrower on the tight (−0.3) side than on the loose (+0.8) side, exactly as it should be. A reading of 99.80 kPa is in spec but inside the guard band on the lower limit, so it does not pass.

The lesson: never collapse an asymmetric tolerance to a single ± figure before guard banding, you will mis-state the risk on the tighter side. Treat each limit on its own.

From decision rules to calibration interval planning

A decision rule controls the risk of one wrong call on calibration day. The calibration interval controls a different but related risk: how long an instrument can drift out of tolerance, unnoticed, before its next calibration catches it. Both are tuned to the same kind of target, an acceptable probability of using equipment that is actually out of tolerance, so they belong in one risk conversation, not two.

The bridge between them is your as-found data. Every time you calibrate, you record whether the instrument arrived in tolerance (as-found, before any adjustment). Aggregated across an instrument or a family of like instruments, this as-found in-tolerance rate is the measured reliability of your interval. If 96% of units come back in tolerance at a 12-month interval and your reliability target is 90%, you have headroom to lengthen the interval; if only 80% come back in tolerance, the interval is too long and is feeding bad equipment into production between calibrations.

NCSLI RP-1, Establishment and Adjustment of Calibration Intervals, is the standard reference here. It lays out several interval-analysis methods and gives decision trees for picking one based on inventory size, how much history you have, and your quality and budget priorities. The practical workflow it supports looks like this:

1. Set a risk-based starting interval. With no history, you cannot do reliability analysis yet, so you start from instrument type, usage rate, environment, and how critical the measurement is. A rugged, lightly used reference in a controlled lab earns a longer start than a portable gauge on a shop floor.
2. Pick a reliability target. A common end-of-period reliability target is 85% to 95% in tolerance, set higher for safety-critical or high-consequence measurements and where you want simple acceptance to stay defensible.
3. Collect as-found history over several calibration cycles, recording in-tolerance vs out-of-tolerance at receipt.
4. Adjust toward the target. Lengthen intervals for instruments comfortably beating the target, shorten them for instruments missing it, and investigate any instrument that fails repeatedly rather than just shortening blindly.

The connection back to decision rules is direct: a higher reliability target keeps more of your fleet inside tolerance, which keeps simple acceptance honest and reduces how often you are forced into aggressive guard banding on borderline instruments. Tight intervals and sound decision rules are two levers on the same false-accept risk.

Tools to do this

You do not need a spreadsheet of your own to run these numbers. We keep a small set of free, vendor-neutral calculators that run entirely in your browser, no signup and no data leaving your machine.

A common reason TUR comes out low is an underestimated uncertainty budget. If you work with temperature instruments, a dedicated thermocouple and RTD uncertainty calculator (on its way to our free tools page) builds the U₉₅ you would drop straight into the guard-band calculator above. Until then, the guard band calculator accepts whatever expanded uncertainty your budget produces.