10 Nov 2026 6 min read antifragility

Monday Myth : Both Gauges Agreed

Two gauges agreed to within a fraction of a micrometre. The shaft passed inspection. Then a housing arrived from another supplier, and the shaft rocked inside it. This week’s Monday Myth asks what a second measurement can tell us when both gauges share the same reference.

Inside a fitted case on a toolmaker’s bench lies a set of polished steel blocks. The smallest are easily mistaken for shims. Each has two carefully finished measuring faces, separated by a known distance. Bring two clean faces together, slide one across the other with a little pressure, and the blocks adhere firmly enough to lift as a stack. The operation is called wringing. With a modest set of blocks, a toolmaker can assemble hundreds of different lengths. A workshop can carry a small repertoire of dimensions from the inspection room to the bench, where a stack can set a comparator before a batch of parts goes through it.

The dimension engraved on a block is its nominal size. The calibration gives the small departure from that size under defined conditions, along with an estimate of the uncertainty. The distinction matters even when the block has suffered no wear. Take a 100-millimetre steel block from a room at 20 °C to one at 25 °C and let it come to temperature. Using a typical expansion coefficient for gauge-block steel, it grows by nearly six micrometres. That is an inconsequential change for a great deal of work. For a dimension with a tolerance of a few micrometres, a trip across the building can account for the entire margin.

The length that travels

A laboratory can calibrate a master block by interferometry, using light to measure its length. A workshop can then compare a working block against a calibrated reference of similar length. The comparator has a much smaller job than the interferometer. It needs to resolve the difference between the two blocks. If they are of similar material and have settled to the same temperature, much of the effect of thermal expansion cancels in that comparison. The workshop can transfer the known length with equipment suited to daily use, while the more elaborate work of establishing the reference stays in the laboratory.

This arrangement gives factories a way to make things separately and have them fit together later. A supplier can machine a shaft and another can make the housing it goes into, without either having the other’s part on the bench. The length travels through a chain of comparisons, each carrying some uncertainty. That uncertainty has practical consequences. A result can be traceable all the way back to a national standard and still leave too much doubt to decide whether a part near the tolerance limit should pass. The importance of a micrometre depends on what the part will have to do.

Now imagine a workshop with two comparators on different benches. One sets up the grinding operation and the other inspects the finished shafts. Both have been set from the same working block. If the block wears and comes to measure short, a shaft ground to the resulting reference can pass inspection against it. The operators can change shifts, repeat the readings and keep impeccable records. The two comparators may agree within a fraction of a micrometre. The error entered before either operator touched the part.

What the second bench can see

A second bench still earns its place. It can catch a slipping fixture, an error in recording the measurement or a fault in the first comparator. Separate operators make several mistakes less likely to survive. Their independence extends as far as the causes they do not share. The worn block sits upstream of both readings, so the most relevant question is whether any comparison in the chain can expose a change in the block itself. Comparing it with a separately maintained check standard can make that change visible while there is still time to establish which batches it may have affected.

The size of the workshop changes the economics. Give a dozen benches a common reference and the cost of maintaining a reliable dimension falls. Parts from different lines can come together with less adjustment, and a disagreement between benches becomes easier to investigate. If the reference drifts, however, the same arrangement carries the error to a dozen places. In the particular case of a shaft and a housing that both inherit a related error, a trial assembly can even appear satisfactory. The error may surface only when a replacement arrives from elsewhere. At that point the expensive question is how many parts made over the preceding weeks can still be used.

A useful check therefore depends on where the suspected error could enter. A second supplier using a different measuring instrument may still rely on the same flawed drawing. Two laboratories using a common definition of the metre can nevertheless provide genuinely distinct evidence about whether one particular block has worn. There is no advantage in asking everyone to invent a length of their own. The difficulty lies in sharing enough to make the parts interchangeable while retaining a comparison capable of revealing which shared assumptions have travelled too far.

The software that agreed with itself

Consider a company replacing the billing component of an old order system. The new component runs in parallel with the old one. Each night, a reconciliation job compares invoice totals, while acceptance tests replay a selection of past orders and compare the resulting invoices with archived outputs. Both forms of checking are worth doing. One can expose discrepancies in the live workload and the other makes past cases repeatable. Several weeks of agreement would give the company solid evidence that the new component preserves the behaviour represented by those comparisons.

Suppose both components receive an order feed that already drops a type of credit raised by customer service. In the old arrangement, a separate process applies those credits directly to customer accounts. The migration team has faithfully reproduced the invoice calculation, and the two totals agree because the omitted credits never enter either calculation. If the wider cutover also retires the separate process, customers can owe the wrong amounts while every invoice comparison continues to pass. The point at which the accounts diverge lies beyond the boundary drawn around the comparison.

A reconciliation against credits recorded by customer service and adjustments actually posted to customer accounts could expose that gap. It would have its own problems. A credit might be approved today and posted tomorrow, or applied to a balance spanning several invoices. Making that comparison meaningful takes knowledge of how the business handles credits, as well as patience with genuine timing differences. The nightly comparison between invoice engines can be more precise, easier to automate and considerably cheaper to run. It can also provide thousands of perfectly valid answers to a question that ends just before the failure begins.

Where the discrepancy goes

The consequences change once the same comparison reaches a steering committee. Engineering reports that the parallel run passed. Finance reports that the two invoice totals reconcile. The programme reports that both functions have signed off. Each statement may be accurate. By the time the decision reaches the committee, one comparison has acquired three institutional witnesses. A customer service supervisor who mentions the missing credits brings something harder to absorb. The evidence may consist of a few awkward cases, some still waiting to settle, with no clean denominator and no line of its own in the cutover pack.

Organisations pay for this asymmetry in a particular way. It is cheap for the next group to reuse a completed reconciliation, and the result arrives in the language the decision already uses. Following a handful of customer accounts across the boundary between billing and customer service takes time from people who have other work. If the investigation delays the cutover, the delay has a date, an owner and a cost. If it prevents months of manual adjustments, the saving belongs to a future that nobody will get to observe. The temptation to request one more signature instead has a perfectly intelligible economic basis.

By the time the new system goes live, even the historical invoices may have acquired a different status. During migration they served as records of what the previous system did. Once promoted into a regression suite, they begin to determine what subsequent releases are allowed to do. Suppose a later team decides to show the credit on the invoice itself so the customer can see it. The old examples then become obstacles to a deliberate change. The team can preserve every historical expectation by putting the credit back into a separate process and leaving the invoice alone. The accident of where the old system recorded a transaction has acquired the force of a requirement.

Back in the workshop, the customer sends over a housing from another supplier and a shaft from the disputed batch. The shaft slips in with enough play to feel it rock. There is a small, awkward silence when it passes again on the inspection bench, in front of everybody. The housing waits beside the comparator while someone goes to fetch the fitted case. Until that morning, there had been a perfectly ordinary explanation for why the readings on both benches agreed.