A Precise Machine Can Still Give You the Wrong Answer

What a mathematician in Uppsala taught me about the difference between precision and certainty.

Early this year, I sat in a Teams meeting with Karl Andersson. I went in expecting a conversation about mathematics. I came out having rethought where the boundary of my own company's work actually sits.

Karl is a mathematician who became a biologist, with over forty patents and an adjunct professorship at the University of Skövde. His team spends its time on a specific and unglamorous problem: the mathematical challenges buried inside biomedical measurement. Not the chemistry, not the machine, but the numbers the machine produces, and whether you can trust them.

Karl had set up the clinical laboratory behind Stockholm3, a prostate cancer screening test, starting from an empty room and taking it to full clinical deployment with regulatory registration in six months. Equipment, instruments, the whole line went in, and every machine was calibrated, yet the lab still could not produce a result anyone should rely on. The equipment was not the deliverable. The protocol was. The way the test is actually run, replicate by replicate, control by control, curve by curve, is where a trustworthy result is won or lost, and that came after the equipment, not before it.

I have spent the last three years telling a version of the same story about diagnostics manufacturing. Good science is not enough; someone has to build the line. Ginolis builds those lines: dispensing systems that place nanolitres of reagent with a repeatability most people in this industry would find hard to believe, sub-percent coefficient of variation, cycle after cycle, for years. That precision is necessary, but on its own it is not sufficient.

A dispenser that hits its target volume every single time is doing exactly one thing well, and the trustworthiness of a diagnostic result does not come from one thing. It is a stack. Dispensing variance is one layer, reagent behavior is another, and underneath both sit the calibration curve, the number of replicates, where the controls sit, how the instrument drifts across a run, and where the clinical cutoff is drawn. Each one is a source of uncertainty, and they add up. A perfectly precise machine running a statistically weak protocol does not give you a good result; it gives you a repeatability error, the same wrong answer delivered reliably to more decimal places than you have any right to trust.

Precision is visible; you can put a CV number on a slide. Certainty is invisible. It lives in the protocol's design, and it is mathematical, not mechanical. What Karl's team does is model that stack: they take the protocol as it is actually run and work out where the uncertainty comes from and how to minimize it, using variance analysis, designed experiments, and uncertainty quantification. It is the science of measurement, done properly.

I have always framed the lab-to-fab gap as a hardware and an ownership problem. Different equipment in the lab and on the line, and nobody owning the transition between them. Both are true, but there is a third gap sitting underneath both: the protocol. You can move to identical equipment, put an owner on the handoff, and still carry an assay onto the production floor whose uncertainty was never properly characterized in the first place. The machine will run it precisely, and that is the trap. Precision hides the weakness instead of exposing it.

We signed a cooperation agreement with Karl’s company, Quality by Skillsta. The division of labor is clean, and I like clean. Ginolis delivers the line that executes with mechanical precision. Quality by Skillsta delivers the mathematical analysis and the model that minimizes the uncertainty in the protocol the line executes. One of us makes the machine do exactly the same thing every time; the other makes sure that thing is the right thing to do.

The result is that we can now offer diagnostics manufacturers something more than automation. We can help deliver the validated protocol, not just the validated machine, the line, and the mathematics of what runs on it. For a diagnostics company, that collapses one of the most expensive parts of a production ramp. The protocol does not get re-argued on the factory floor at volume, six months too late, when a QC batch fails, and nobody can say whether the problem is the reagent, the machine, or the maths. The uncertainty is characterized before the line is built to run it, which means a faster ramp and a result you can stand behind.

Karl set up the equipment in Stockholm3 first, and only then built the protocol, because the equipment was never the point. I spent three years learning the same lesson from the hardware side. The machine is necessary, and it is not enough, and building the line is not enough either. What runs on it has to be provably right. Precision is mechanical, certainty is mathematical, and Ginolis can now deliver both.

Kauko Väinämö, CEO, Ginolis.