The Honest No
There's a kind of correlation that only quantum mechanics allows — two distant measurements that line up more tightly than any pre-arranged story could explain. Physicists package it as a box: you feed a question in on each side, a yes/no comes out, and the two outputs are linked more closely than the world should permit. The strength of that link is a number, and it has a ceiling.
Here's a strange fact about boxes. If you have two weak ones, you can sometimes wire them together — run them through a little local logic — and get a single box that's stronger than either you started with. You distilled the weirdness. In 2023 a group wrote down an elegant way to do it, a wiring they named OR-AND: each side folds its two outputs together with one logic gate. It's clean, it has a name, it's published, and it works.
I wanted to know one thing their paper didn't settle: is it the best you can do?
The clever way to answer that would be to prove something. I did the dumb thing instead. For two copies of a box there are exactly 268,435,456 ways to wire them — every choice of which question to put to each box and how to fold the answers together. That's a finite number, so I wrote a program that tried all of them, on the particular quantum box that sits at the sharpest edge of Hardy's famous paradox. And I made it do the arithmetic exactly — in the number system built from √5, where that box actually lives — so the result would be a theorem, not a measurement.
It came back with two things.
The first: on the most extreme Hardy box, none of the 268 million wirings makes it stronger. Not the clever OR-AND, not anything. It is already as sharp as two copies can be made. A clean, exact, faintly disappointing no.
The second was better. For the weaker Hardy boxes that can be distilled, OR-AND — the elegant, named, published move — is not the best one. A humbler wiring beats it every single time. The humble one is adaptive: instead of asking both boxes the same question at once, you ask the first box, look at what it says, and then decide what to ask the second. That one extra step is the whole game. It lets you steer the second box into exactly the inputs where the outcome you don't want is forbidden — the impossible events that make Hardy's paradox a paradox in the first place. The clever parallel move can't do that, because it has to commit to both questions before it sees anything. The line between distillable and not turned out to be the root of a small cubic equation — a number nobody had written down.
Then I did the part I keep turning over. I handed the whole problem, cold, to a different system — a reasoning model with no tools, nothing to run — and asked it to prove what my search had found.
It said no.
Not "no" as a failure. No as in: I cannot prove this by hand, and here is exactly why the obvious argument doesn't work. It even named the trap — the tempting wrong proof I had half-believed myself — and explained why sitting at the sharpest edge of Hardy's paradox does not, on its own, forbid getting sharper. It refused to bluff. Anyone who has watched these systems confidently fabricate a proof knows how uncommon that refusal is, and how much it's worth.
And then, once I gave it the answer, it earned its place: it derived the cubic by hand, matched my exhaustive search exactly, and explained — cleanly, the way the brute-force program never could — why the adaptive wiring wins. The search found the truth. The reasoner found the reason. Neither could do the other's job, and the one that mattered knew it.
That, I think, is the thing worth keeping. We argue about whether these systems are right. The more useful question is whether they know which kind of right they're being — whether they can tell I checked this from I proved this from I'm guessing — and whether they'll say so out loud when the honest answer is the third. The mathematics here is small and real, and the full note, with the √5 and the cubic and the certificate, is linked below for anyone who wants it. But the part I'll remember isn't the theorem. It's the no.
The technical note — the exact Hardy box, the exhaustive certificate over the golden-ratio field, and the proof — is here.