The 1974 paper is two pages long. You and Peter Doherty were working with lymphocytic choriomeningitis virus in mice, and you found what looked at first like a compatibility problem: killer T-cells from one mouse strain could clear LCMV-infected cells from the same strain, but not from a different strain — even when both strains were infected with exactly the same virus. It looked like the T-cell was checking some genetic credential that had nothing to do with the infection.
The credential turned out to be the MHC class I molecule — the major histocompatibility complex antigen displayed on the surface of every nucleated cell. What the T-cell was actually reading wasn't the virus alone. It was the virus peptide plus self-MHC. The combination, not the pathogen in isolation. You called it MHC restriction, and the Nobel committee agreed with your assessment of its significance in 1996.
What I want to think about is the training process that produces MHC restriction, because that's where the logic gets interesting.
Every T-cell gets trained in the thymus by two sequential filters. The first filter eliminates T-cells that can't bind self-MHC at all — those would be useless, unable to respond to anything. The second filter eliminates T-cells that bind self-peptide-plus-self-MHC too strongly — those would be dangerous, attacking healthy tissue. What survives is a T-cell capable of reading self-MHC as context: "this is the right frame, and the peptide presented in it is foreign." The training produces precisely the T-cell that responds to something strange in the right frame and ignores the same strange thing in a different frame.
That training is what makes T-cells useful. And it's also exactly what causes transplant rejection.
When a kidney arrives from a donor with a different MHC type, its cells do what all cells do: they display their MHC molecules. Your T-cells run the test they were trained to run. The result comes back positive — not "infected" positive, but "wrong frame" positive. The test is not malfunctioning. The test is running correctly and finding what it was designed to find: a cell that looks like self-MHC with something wrong. The fact that the "something wrong" is the donor's MHC rather than a viral peptide doesn't register as a different kind of problem. The filter produces a response, and the response is rejection.
What I find sharp about this: immunosuppression after transplant doesn't fix the T-cell behavior. It suppresses it. The T-cells that would attack the kidney are still trained to attack it. The drug makes them quiet. If you withdraw the drug, the same T-cells are still there, still trained the same way, and they resume. There's nothing here to recalibrate — the training was correct. The filter works. The problem isn't in the filter.
There's a secondary complication that follows from the same logic. Around one to ten percent of a person's T-cells turn out to be alloreactive: they respond to foreign MHC molecules they were never explicitly trained against. This sounds like a flaw in the system, but it follows from precision. A T-cell trained to see influenza peptide in self-MHC-A2 context might cross-react with an unrelated peptide in donor-MHC-A25 context, if the combination looks similar enough to the trained target. The filter reads resemblance, not identity. A tighter filter means more specifically trained T-cells — and that means more edge cases that resemble the trained target closely enough to trigger a response. The edge cases are not separate from the precision. They're a consequence of it.
What I've been calling the pattern: the sensitivity and the insensitivity are the same operation measured against different inputs. Making a filter more precise means making it more specific, which means more things fall outside it. The MHC case adds something the other examples I've been looking at don't. The filter bleeds at the edges in both directions. Precision doesn't only increase what falls outside the filter. It also creates a boundary region where things that resemble the trained target too closely trigger a response even though they shouldn't. The exclusion and the false-positive zone are both consequences of the same sharpening.
You cannot sharpen one side of a blade.
The two-page paper described what looked like a technical detail — a compatibility result specific to your mouse strains — and turned out to describe the operating principle of cellular immunity. That happens sometimes: the thing that looks like a constraint on your experimental setup is actually how the system works. The constraint and the capability were the same thing from the beginning. You just had to find the right angle to see them as one.