Warning - this is highly technical. If you are not scientifically orientated skip to the last line and ignore the proof I am presenting.
Printing and measuring are one system. Every dimension a printer produces gets checked by an instrument with its own error, and until you know what that instrument can actually resolve, you don't have a measurement — you have two numbers arguing with each other. That's what three calibration coupons, five print runs and three weeks turned into: not one clean result, but a chain of findings that kept turning out to be about the caliper instead of the part.
The pegs that wouldn't agree
The first coupon was simple: a plate of bores and a matching plate of pegs, printed to find press-fit clearance in PETG. It worked — the winning cell was unambiguous. 0.06mm enters and holds itself; 0.04mm won't enter under force; 0.10mm holds only by jamming askew. Clearance logged: 0.06mm.
Then the plate was printed a second time, and four nominally identical 5.00mm pegs stopped agreeing with each other. The first run measured 4.77 / 4.78 / 4.81 / 4.80mm — a 0.04mm spread. The second measured 5.00 / 4.90 / 4.83 / 4.80mm — a 0.20mm spread, five times wider, on identical geometry from the same slice and the same plate. It got logged as a real finding: something in the process was inconsistent.
Nobody asked what the caliper's own repeatability was. That question turned out to be the entire rest of this post.
The gear coupon that couldn't fail
The second coupon was twelve 20-tooth, module-1 gears in six meshed pairs, each pair set to a different backlash: 0.00, 0.05, 0.10, 0.15, 0.20, 0.30mm. The output it was built to produce was simple — the smallest backlash at which a printed pair still turns freely.
Every cell ran free. Including cell 1, which has zero designed backlash. A zero-backlash pair cannot physically turn freely. The read at the bench was right before any measurement backed it up: there's a looseness between the shaft and the cog, probably more than in the calibration numbers themselves.
The fixture's mounting posts were 5.00mm with 0.40mm of clearance per side, deliberately — the coupon was built on the reasoning that the bore fit must be decisively slacker than anything being measured, so a cell that won't turn can't be blamed on a tight bore. That reasoning is sound, and it's only half the problem: it guards against a false bind and says nothing about a false free.
Measured on the printed part: a 4.805mm mean post inside a 5.655mm mean bore is 0.850mm of play. At the gear's 20-degree pressure angle, that converts to 0.619mm of equivalent backlash — more than double the entire 0.00-0.30mm ladder the coupon was supposed to be testing. The gears weren't meshing at a calibrated backlash. They were rolling apart until the slack ran out.
The reasoning behind that 0.40mm clearance had a second problem: it was chosen because the bore fit was unknown at the time. By the time this coupon ran, the first one had already measured it. The premise for the slack had expired, and nothing had gone back to check.
The error was in the machine, not the design
Four posts, measured on two axes each: 4.75/4.86, 4.76/4.87, 4.72/4.87, 4.75/4.90mm. Each axis was repeatable on its own — 0.04mm spread — but every post was oval by roughly 0.13mm, in the same direction, and that direction lined up with the printer's own X and Y axes. The same effect was sitting in the first coupon's pegs and holes, unnoticed.
That single finding rippled backward through everything already logged. It cast doubt on the original peg spread — those readings were all taken on one axis, at whatever clock angle the caliper happened to land on, against a part that wasn't round. The "spread" may never have been part-to-part variation at all. It explained a fit that had been unaccountable until then: two identical printed spigots, one dropping into its socket cleanly and the other going in "very tight" — an oval spigot in an oval socket is a clocking-dependent fit, fine at one rotation and interfering at another. And it became a live candidate for a cracked part elsewhere in the shop: a circumferential crack around a socket bore, split along a layer line, with the hoop-tension signature ovality produces. The joint in question is a bayonet fit, which rotates the spigot through its tight orientation on every assembly rather than letting anyone find the easy one.
The fix that got withdrawn the same day
The obvious move was to tighten the fixture's bore clearance. It was proposed and dropped within the day, on a number that hadn't been computed yet: the cyclic part of the error doesn't depend on the fit at all. The post is oval by 0.130mm and the bore by 0.210mm, and as a gear turns, the two ovalities sweep against each other — a 0.130mm swing at 0.08mm of clearance per side, at 0.20mm, at the original 0.40mm, because the swing is set entirely by the ovality, not the nominal gap. That's worth 0.095mm of equivalent backlash, cycling twice per revolution, against a ladder whose step size is 0.05mm.
Worse, tightening past a certain point poisons the reading outright. At 0.08mm per side, the worst-case clearance comes out to +0.035mm — a hard tight spot, twice per turn. The coupon's own protocol scores a tight spot at one angular position as a failed pair. The fixture would have manufactured the exact failure it existed to detect.
A coupon designed at the bench
The next coupon came from questions asked with calipers in hand, not from the model. Two perpendicular readings can't see an oval — at 45 degrees to its own axes, a 0.13mm oval reads 23.365 / 23.365mm, indistinguishable from round; an earlier reading of 23.28 / 23.29mm had settled nothing. So the new coupon carries eight index ticks, at 0, 45, 90 and 135 degrees, 0.80mm wide, engraved on the top face and stopping 0.60mm short of the rim so the jaws always land on undisturbed circle. Small measuring flats were considered and rejected — they'd have made a cleaner reading, but the question was how the printer renders a curve, not how well a flat measures.
Jaw pressure splays under hand pressure, so how deep a part sits changes the reading — every object on the new coupon is the same 4.00mm thickness for that reason, with a start/finish dot pair (fixed jaw on START) so the same two physical points get measured every time. And where the old ladder doubled at each step — 5, 10, 20, 40, 80mm — the new one steps 3, 6, 9, 12mm: enough to see whether an error scales, which doubling never could. Finally, a loose small circle tips over under its own jaws, so each size got exactly one raised boss on its own tile — something to hold, nothing in the way of the calipers at any angle.
The instrument was the limit
Repeated readings of the same axis, on the same part, re-seated between each attempt:
| Object | Spread (mm) |
|---|---|
| tile_5 | 0.280 |
| tile_3 / tile_6 | 0.210 |
| ring_40 | 0.180 |
| disc_12 | 0.170 |
| tile_9 | 0.150 |
| disc_20 | 0.130 |
| ring_80 | 0.120 |
| Median | 0.160 |
The ovality this coupon was built to catch is 0.130mm. The measurement is bigger than the thing it measures. The experiment can't answer its own question — and that was only visible because this was the first coupon in the whole project to repeat a reading at all. The verdict written at the bench, before any of this was tallied, was blunt: this is really hard to measure and see, and I do not 100% trust any of these numbers. It was correct. It's just quantified now.
Two more honest details belong in the record. The seating method that was supposed to make every object comparable worked cleanly on essentially one tile out of ten — the rest were either too small to use the caliper's own base or, at 40mm, about a millimetre too large for it. And the coupon was printed in grey, with engraved marks that are nearly invisible in grey. Grey wasn't the actual mistake — white would have been worse — the mistake was designing engraved index marks and never asking how they'd be read, when an earlier coupon had already recorded that silver-on-silver can't be read at all. The fix costs nothing: ink the grooves with a marker and wipe the surface clean.
What did survive the noise: every one of the ten objects came out undersize, by a mean of 0.164mm. Whether that error is constant or scales with size is still open — a least-squares fit gives -0.00072mm per mm of size, against -0.0036mm/mm for a purely proportional model and zero for a purely constant one. The data can't choose between them yet.
Checking the checker
Four separate "undersize" findings, across three coupons, all read with the same caliper — and nothing had ever checked whether that caliper reads zero correctly.
A standard circular coin is specified by the issuing authority. The coin that I used measures in at 23.00 x 2.20mm, and that number is already load-bearing elsewhere in the project — it's what a token design elsewhere is built on, so the check and the design agree by construction rather than by coincidence. If the caliper reads close to 23.00mm, the printer really is undersize, and 0.164mm is a real compensation number to build into future prints. If it reads closer to 22.84mm, the caliper itself is low, and four separate findings collapse into a single instrument error.
It's not a perfect reference — a circulated coin is worn, and a bimetallic one has a ring and a centre disc that can differ slightly — but it's good to roughly 0.05mm, which is more than enough to tell 0.16mm apart from zero, which is all it's being asked to do. It's worth several readings, since the same repeatability problem applies to a coin as to a printed part, and it's worth checking the 2.20mm thickness too, since that's a different axis measured against a different part of the jaw.
That check hasn't been run yet. Three coupons, five print runs and three weeks produced exactly one number that can currently be trusted without qualification — how repeatable this particular caliper actually is — and it turns out that was the number everything else depended on. You cannot calibrate a printer with an uncalibrated instrument. The shape of it keeps repeating: a check that was true when it ran, and stopped being true afterward, with no repetition to prove the point.