It Passed Every Test. It Can Still Fail.Here’s the 125-year history of checking medicine, the 8 problems no lab can fully solve, and the 4-box experiment that could finally tell us which differences matter.

Before there was quality control, there was a blood type

Before anyone understood blood groups, a transfusion could help one person and seriously harm the next. In 1901, Karl Landsteiner published his discovery of human blood groups, which explained why: blood from different people could react, and the red cells would clump together.

Notice what that means. The blood wasn’t contaminated. Nobody made it wrong. It was simply wrong for that person.

A treatment can be good in itself and incompatible with the person receiving it.

Everything in this essay descends from that sentence. Blood typing and compatibility testing were the first versions of a move personalized medicine now makes constantly: measure something about the individual before deciding what to give them.

One piece of vocabulary before we go on. Strictly speaking, quality control means testing materials and products. Quality assurance means designing and overseeing the whole process so errors don’t happen in the first place. You need both, and the history keeps rediscovering why: every time testing alone fell short, the field had to push its checks further upstream.

What the factory can check today, and what a pass doesn’t tell you

As of October 2026, manufacturers can measure many properties of a personalized treatment precisely. Approved CAR T therapies are made under extensive controls. Personalized cancer vaccines are at an earlier stage: clinical trials have demonstrated working manufacturing workflows, and regulators are still writing guidance for their unusual problems. In August 2026, FDA published draft guidance on potency assessment for active immunotherapy products, including patient-specific ones.

Below is the kind of panel a batch gets checked against. Every row is a real category of test. Now flip it over.

Every right-hand line is the same gap in a different costume. The tests describe the bag. The question is about the bag inside a person. So let’s go through the specific ways that gap opens up. There are eight.

A dish is not a body

“Potency” means biological activity relevant to what the product is supposed to do. For CAR T cells, potency testing can include killing target cells in a dish, measuring the signaling molecules cells release when they meet a target, and measuring how many cells were successfully modified. FDA discusses using complementary methods, because each captures a different slice of function.

But think about what the cells actually have to do. Survive the infusion. Multiply. Find the cancer. Kill it. Find more. Keep going for weeks or months. A short experiment in a dish sees the first kill.

Here’s a hypothetical example that shows what gets lost. Two batches both kill 80% of their targets in the first round. Run the test again with fresh cancer cells, and again.

So make the test more realistic? Every step toward realism has a price. The patient’s own cancer cells may be unavailable or hard to grow. Longer experiments delay a treatment for someone who may not have time. A more elaborate assay may give different answers in different labs.

The engineering problem is choosing a fast, repeatable test that captures enough of the biology to support a reliable release decision.

“Enough” is doing all the work in that sentence. Hold that thought. It’s the whole second half of this essay.

“Same factory, different file” is only half true

Personalized mRNA cancer vaccines have a different raw material: information. The patient’s tumor is sequenced, targets are chosen, and the factory writes those targets into RNA. It’s tempting to picture the factory as a printer. Same machine, new document.

But a new sequence is a new material. RNA’s length and sequence affect how it folds, and that can influence its stability and how it behaves during manufacturing. The UK regulator’s draft guidance asks developers to show control across the expected range of sequence lengths, structures, and compositions, rather than assuming every personalized sequence behaves the same.

There are byproducts, too. The enzyme commonly used to make RNA can produce double-stranded RNA impurities that affect immune stimulation and protein production. Researchers have engineered enzymes that make fewer of them, which shows the problem is solvable, but only by changing the manufacturing technology itself. And since you can’t make and fully characterize every future patient’s vaccine in advance, manufacturers need a demonstrated operating range, rules for flagging unusual sequences, and extra investigation when a design falls outside it.

Then comes the harder problem. A vaccine can be manufactured flawlessly and still encode weak targets. Each chosen mutation has to clear a gauntlet: be present in the relevant cancer cells, come from a gene that’s actually expressed, produce a suitable fragment, get displayed on the cell’s surface by the patient’s own immune machinery, and then be recognized by T cells that mount a useful response.

That historical benchmark is why predicted targets need careful interpretation. (If you want the full story of how a tumor becomes a target list, it’s in Your Tumor Is Now a File.)

So why not experimentally verify every target before manufacturing? Because those experiments need suitable patient material, time, and sensitive assays, and a negative result doesn’t settle the question. FDA’s August 2026 draft guidance acknowledges that developing a separate biological release assay for every personalized product may be impractical. For some personalized products, it proposes that a qualified manufacturing process, a scientifically justified target-prediction pipeline, and appropriate physical and chemical measurements could support release. That’s a proposed framework for justified assurance. It is not proof that every predicted target will work.

Improve the factory and you may have changed the medicine

A new way to grow cells, a different reagent, an automated instrument, a second manufacturing site. Each can raise capacity. Each can also change something about the product that matters biologically. So manufacturers run comparability studies, to show that material made after a change is close enough to the material behind the existing clinical evidence.

Here’s the trap. Patient A’s cells were made with the old process. Patient B’s cells were made with the new one. If B’s product looks different, is it the process, or is it B? Splitting one person’s starting material between both processes isolates the change, but it consumes material that is scarce by definition.

The direction regulators are taking is to build evidence around a whole manufacturing platform, shared methods and shared knowledge, while still checking each individual batch where differences could affect performance.

Four findings that change what “good cells” means

The most promising pattern across recent research isn’t a magic marker. It’s a method: find a candidate feature, deliberately change it, see which functions change, then check whether that predicts outcomes in new patients. There’s substantial support for the pieces. There is no established winner that solves the whole problem.

Every quality measurement that matters has to climb that ladder, and every rung needs its own evidence. A measurement can still be useful for quality control without directly causing success, as long as you know when it stays trustworthy, especially after the manufacturing process changes.

Five more moves that turn a score into an explanation

The four boxes are move one. The rest build on it.

For personalized mRNA vaccines, the same logic means separating two experiments that are easy to blur. Hold the encoded cancer targets constant while varying RNA design or formulation, to study manufacturing and delivery. Separately, hold the platform constant while varying target selection. Then trace expression, display on the cell surface, immune recognition, and outcomes. Otherwise a bad target choice can be mistaken for a manufacturing defect, or the other way around.

I’d start small: one CAR T product, one disease. The first milestone is showing the system can detect known functional defects and tell their causes apart. The next is predicting relevant outcomes in an independent group of patients, followed forward in time. Only after that would you test whether decisions based on it actually improve treatment.