What the Margin on a Test Specification Costs

How to read this sheet

Read the two margin columns together and ignore the percentage. A component twenty-six per cent above its requirement sounds generous and a component at eight standard deviations sounds like nothing, and it is the second number that decides whether the test passes. The percentage is a property of the units the test happens to be written in; the standard score is a property of the component. On the seeded jacket the button is the one everybody would have cut and the drawcord is the one nobody has ever looked at, and the sheet reverses both. Why the number of specimens matters as much as the component. Every specimen the laboratory pulls has to pass, so the chance the test fails is one minus the chance a specimen passes raised to the number of them. Ten buttons at one failure in a thousand fail a submission once in a hundred; three zips at the same rate fail once in three hundred and thirty. That is the whole reason a button needs more margin than a zip does at the same apparent generosity, and it is why the first question to ask about any test that keeps coming back is how many specimens it is run on. The spread is the input everybody will want to skip. Nobody records the standard deviation of a button pull-out, and a single number off a single report tells you nothing about it. Take the last six or eight reports on that component and work it out, and if the reports disagree by more than you expected then that is the finding, not a nuisance - a component whose spread is unknown cannot be specified at all, only hoped for. Where the sheet refuses to flatter you. A saving that appears only because another test is already failing is not a saving. If a submission fails one time in eight anyway, weakening a second component looks almost free, because the failure it causes is mostly a failure that was going to happen. The sheet solves the protocol twice - once as it stands and once as if the weak tests had been brought up to the margin they need - and reports the smaller, durable number as the answer. What is not modelled. A normal distribution is assumed for every test result, which is fair for strength measurements and poor for anything with a floor or a ceiling; peel strengths in particular are often skewed and the sheet will be optimistic about them. Correlation between tests is ignored - a bad fusing lot and a bad seam tape lot on the same garment are treated as independent when they may share a cause. And nothing here covers the tests a component fails outright rather than marginally, which is a sourcing problem and not a margin one.