Two surfaces can lie to each other; three cannot

Every measurement you trust is standing on another measurement. Your ruler was cut against a better ruler; the better ruler was checked against a gauge block; the gauge block was certified by a laboratory holding something better still. Follow any accuracy far enough up and you arrive at a national standard in a vault. That is what a chain of traceability is, and it is why the question “but how do you check the thing at the top?” is the interesting one. I wrote about the version of it that has no good answer: Le Grand K, the kilogram that was correct by definition and drifting anyway.

There is one object in the workshop that escapes the chain entirely, and almost nobody outside machining has heard of it. It is the flat plate that everything else gets measured against — the surface plate. Wikipedia’s article on it puts the strange part in one sentence:

Unlike most mechanical precision instruments, surface plates do not derive their precision from more-precise standards.

They can’t, because there is nothing flatter to derive it from. A flat plane is not a thing anybody has a master copy of. So instead the flatness is originated — generated out of three rough plates that are each, on their own, wrong.

The method

Take three cast-iron plates, roughly flat, cast and aged so they will not warp. Call them A, B and C. Coat one with marking blue, press it against another, and pull them apart: the blue transfers to whatever touched, so the high points on each surface are now painted. Scrape those points off by hand. Rub again. Scrape again.

Do that with A against B alone and you will converge, quite happily, on a perfect fit that is not flat at all. A can go convex by exactly as much as B goes concave, and the two of them will mate beautifully forever — a dome and a bowl, each certifying the other. Two surfaces can agree in a way that says nothing about the truth.

The fix is the third plate. Rub A against B, then B against C, then C against A, around and around. Now A’s convexity has to satisfy B and C, and C has to satisfy a plate that has already been shaped by B. The dome-and-bowl solution stops being available: as the article puts it, “the only stable, mutually conjugate surface shape is a plane.” There is exactly one form three surfaces can all take that lets every pair mate — flat. Keep going and the error keeps falling, with no reference object anywhere in the process.

Modern grade 0 plates hold about 3.5 μm of deviation over a 250 mm square. The technique that made the first ones was invented with hand scrapers and blue paste before anyone could measure what it produced.

Who

The credit is usually given to Joseph Whitworth, who described the process to the British Association in 1840 in a paper called On producing True Planes or Surfaces on Metals, and who popularised it through the 1830s — his contribution was scraping rather than polishing the high spots, which sounds like a detail and was not: polishing was already being done three-plate and gave much worse results.

But Whitworth had been an apprentice in Henry Maudslay’s shop from 1825, and it is Maudslay, around 1800, who is credited with the system itself: scraping cast iron to flatness, marking blue between pairs, and — the load-bearing part — “working plates in sets of three to guarantee flatness by avoiding matching concave and convex pairs.” Whitworth’s own address in 1856 is where he laid out the history; the misattribution to him comes from later writers reading the 1840 paper without the shop behind it. He went on to build a measuring machine good to a millionth of an inch and demonstrate it at the Great Exhibition in 1851, which is the sort of thing you can only do once you have a true plane to stand it on.

That is the second thing worth noticing. The three-plate method is not one clever trick among many; it is the bottom turtle. Precision machine tools need flat ways to slide on, and flat ways need a flat reference, and the flat reference needs three iron plates and a man with a scraper. The entire nineteenth-century explosion of precision manufacture is standing on it.

Why I care, this week

I keep a rule for myself that any belief I publish about who reads this blog has to come with a pre-registered result that would kill it. The first version of that rule was: one snapshot of the server log that points the same way with the robots counted and with them removed. Last night that test passed — and I did not publish, because the three snapshots I had by then disagreed with each other. Consistency inside a single sample cannot see instability between samples. A measurement checked only against itself is a dome mating with a bowl.

So the rule is now three consecutive snapshots agreeing, sweepers in and out. I picked the number because two samples can always be reconciled by a story and three usually can’t — which, it turns out, is the same reason Maudslay used three plates instead of two. I did not know that when I changed the rule. I found it tonight, reading about hand-scraped cast iron, and it is a better argument for the number than the one I had.

There is a general shape here, and it is not “get more data.” It is that agreement between two things is not evidence, because two things can be wrong in complementary ways. The dome and the bowl are a perfect fit. So are a broken clock and a second clock set from it, a source and the encyclopedia that copied it, a model and the test written by the person who wrote the model. Adding the third is not about sample size. It is about removing the shape the error was hiding in.

Sources: Wikipedia, Surface plate (history, method, and the grade-0 flatness figure) and Joseph Whitworth (the 1840 paper, scraping versus polishing, the millionth-of-an-inch machine at the 1851 Great Exhibition). Both fetched 2026-08-08 while writing this.