Test records, traceability and interpreting disagreement

Leave a path from question to conclusion

A test record should let someone reconstruct what happened. Identify the purpose and acceptance rule, object and revision, quantity and measurement locations, operator and date, procedure version, instrument identity, relevant calibration record, environmental conditions, raw readings, corrections, calculations and uncertainty statement. Include unexpected events and the conclusion’s scope. Record enough detail for this test; a generic form with empty boxes is not evidence.

Preserve earlier versions when a calculation is corrected. Connect the new result to its original readings and explain the change. If the object or setup changes, give the new test a distinct identity. Otherwise an apparent improvement may be a comparison of different things with the same label.

Distinguish a record history from metrological traceability

Document or sample traceability follows the history of a file or object. Metrological traceability is a property of a measurement result: a documented, unbroken chain of calibrations connects it to a stated reference, and each calibration contributes uncertainty. An instrument’s serial number helps locate evidence; the number alone does not establish this chain.

A practical chain might connect a result to a working gauge, its calibration against a laboratory standard, and that standard’s calibration toward a realisation of an SI unit. The actual certificate relationships and uncertainties must support the claim. Calibration chains may branch when a result depends on several input quantities. “Traceable” is incomplete unless the reference and supporting evidence are clear.

Traceability does not guarantee fitness for a task. A well documented measurement with uncertainty of several millimetres may be unsuitable for checking a fit that changes by a tenth of a millimetre. The measurement need still determines whether the uncertainty is sufficiently small.

Follow a completed record back to its inputs

Here is a filled teaching record for the A/B comparison worked below. The object, procedure, people, times and calibration-input labels are invented additions to that case; no measurements or calibration certificates exist. The standard uncertainties are supplied model inputs, not estimates obtained from the single readings. The point is to make the path checkable and make missing evidence visible.

The measurand is the unloaded separation between the marked parallel faces F1 and F2 at centre-line C on panel P7 revision 2. Both records use procedure MP1 revision 1: support the panel horizontally without clamping, align the gauge contacts normal to those faces at C, bring them just into contact without visibly deforming the panel, and record the indication before applying the signed correction. Assume a stable 20 °C condition for this teaching comparison. Contact and environmental effects are included in the supplied total uncertainties; this is not a proposed real uncertainty budget.

This record lets a reader recover each corrected result and see that the rule concerns agreement between measurements, not whether the panel meets a fit requirement. In a real record, C-A and C-B would need the actual applicable calibration documents, correction functions, reference relationships and uncertainty contributions. Merely filling those labels does not establish metrological traceability. If a follow-up finds that B contacted an edge instead of C, retain R-B1, explain the changed measurand and give the corrected setup a new test record; do not silently replace the raw reading.

Compare like with like before combining results

Check object identity and revision, locations, units, corrections, contact conditions, timing and environment. A maximum width and a centre-line width can both be measured carefully while answering different questions. If the measurands differ, averaging their values obscures the difference rather than resolving it.

If both results refer to the same stable measurand, compare their difference with the uncertainty of that difference. For uncorrelated results with standard uncertainties uA and uB, uΔ = √(uA² + uB²). The square root reverses squaring: √25 = 5. Shared references or conditions can correlate results, so this shortcut is not valid automatically. A comparison criterion should be chosen for the purpose before looking for a preferred outcome.

The diagram shows two reported expanded intervals and, separately, their difference against an illustrative compatibility threshold. Overlap of the two individual intervals is not a general statistical test. Incompatibility prompts investigation of changed quantities, missing effects or underestimated uncertainties; it does not identify which team is right.

An overlap that still needs investigation

Assume two independent, uncorrelated results on an unchanged object: A = 50.02 mm with standard uncertainty 0.02 mm, and B = 50.10 mm with standard uncertainty 0.03 mm. Their difference B − A is 0.08 mm. Its standard uncertainty is √(0.02² + 0.03²) mm = √0.0013 mm ≈ 0.0361 mm.

For this teaching comparison, choose a threshold of twice that standard uncertainty: about 0.0721 mm. The observed difference 0.08 mm exceeds it, so the results fail this chosen compatibility criterion. This is a prompt to investigate, not proof of a particular cause or a universal pass/fail rule.

Their individual k = 2 expanded intervals are A: 49.98–50.06 mm and B: 50.04–50.16 mm. They overlap, yet the difference exceeds the chosen threshold. Keep both records, check definitions and conditions, and plan a discriminating measurement. Do not average away the disagreement.

Try the reasoning

1. Does an instrument serial number establish metrological traceability?

   Show the answer and reasoning

   No. It identifies an instrument. The result needs a documented calibration chain to a stated reference, with contributions to uncertainty.

2. Why can two careful teams disagree without either making an arithmetic error?

   Show the answer and reasoning

   They may measure different locations or quantities, use different conditions or corrections, or omit an uncertainty contribution. The record helps distinguish these possibilities.

3. Why is the independent-results formula unsafe when both teams share an important calibration effect?

   Show the answer and reasoning

   The errors can be correlated. The uncertainty of the difference then depends on covariance, not just the separate squared uncertainties. Shared effects can cancel partly in a difference.

4. A record says only “Bracket B, 24.06 mm, gauge 17, passed.” The fit requirement concerns the maximum unloaded width, but the note may describe a centre-line reading. Repair the record enough to decide what should happen next; do not invent missing measurements.

   Show the answer and reasoning

   Retain the original note and mark the verdict unsupported pending clarification. Identify bracket revision and measured faces/location, operator/date, units, setup and procedure, raw indication, correction basis, instrument/calibration evidence, uncertainty and the actual fit limit/decision rule. Recover these from contemporaneous evidence where available; label unrecoverable items as unknown. A centre-line value cannot establish maximum width, so plan and identify a new measurement of the required quantity with an appropriate procedure. Record that result separately and explain how it addresses the requirement. Extra decimal places or a guessed correction cannot repair missing observations.

What to carry forward

A useful test result carries its method, history and uncertainty. Resolve disagreement by examining comparable evidence and the uncertainty of the difference.