Tolerance stacks and the myth of the golden unit

Somewhere out there is a synthesizer built on a Tuesday morning by a technician having a good day, and it measures 0.004% better than the print on the box. Forums worship this unit. People make pilgrimages to buy it secondhand at four times list. This note is about why that unit is a statistical accident we specifically engineered away, and why the boring truth is better for you than the myth.

The golden unit is a real phenomenon in every industry with tolerances: the sample that happens to sit at the lucky center of every part distribution at once. It exists because tolerance is additive in the boring direction and multiplicative in the lucky direction. When a spec requires the stack of five parts to land inside a band, most units land in the middle, a few land near the edge, and a vanishing few land at the center of every distribution simultaneously. The golden unit is that last category, and there are maybe three of them per thousand. They are real. They are also the worst possible thing to design a product around.

The arithmetic of a stack

Here is the actual stack for a HELIOS-8 voice, simplified to the four terms that matter. The pitch you hear is produced by an oscillator whose frequency is set by a current that is set by a voltage that is set by a resistor network that is trimmed by a calibration table. Each stage contributes an error term, and the question every engineering review asks is: do these terms close inside the band?

Two ways to add these up, and the gap between them is the whole argument. The worst-case sum assumes every part is bad in the same direction, which is pessimism with a spreadsheet. The statistical sum assumes independent error sources, which is optimism with the same spreadsheet. Reality sits between them, and we ship to the number reality does not cross: the printed 0.02% is the worst-case sum rounded up, not the statistical sum rounded down. That one rounding decision is the difference between a spec we defend and a spec we apologize for.

What the last 0.01% costs

The first three decimals of pitch accuracy are bought with cheap money: better reference circuits, decent resistors, a sensible layout. The fourth decimal is bought with expensive money, and the exchange rate gets worse every step. Going from ±0.1% to ±0.02% roughly doubled our bill of materials. Going from ±0.02% to ±0.01% would have doubled it again, and the honest engineering answer is that nobody can hear the difference the second doubling buys, because 0.01% of one percent is a range smaller than the mistuning your own ears introduce when you sing along with your own patch.

So we stopped at the audibility line and spent the savings somewhere you can hear: the drive stage, the jack contacts, the panel. A tolerance budget is a budget. Money spent on an inaudible decimal is money taken away from a knob you will touch every day for a decade, and we think that trade is a theft from the user, even when it makes the spec sheet look heroic.

Why we measure per batch, not per era

Component suppliers revise parts. A resistor from 2024 and its 2026 replacement meet the same datasheet and measure differently in the tail. When we measure a sample at the start of production and stamp a spec for the year, we are publishing a promise about a future we cannot see. When we measure every batch and print the batch number next to the number, the promise stays attached to the thing you actually own. Your certificate has a batch date on it for this reason. The spec sheet on this website is the trailing twelve months across all batches, which is the least flattering number we can publish, and we publish it because it is the one that survives a subpoena.

The yield curve, because someone will ask

At ±0.05% we could ship 99.9% of units first-pass with no trim. At ±0.02% we trim every unit at the bench and still ship 99.7% first-pass. The trim takes ninety seconds per voice bank and is the single most expensive line in our production cost. It is also the line that makes the certificate meaningful. We have never considered removing it, and every new hire asks why, and the answer is that the certificate is the product.