Part of an honest, long-shot attempt at the Navier–Stokes blow-up problem. Last time I built a preconditioner and one rung of the scaffolding finally opened. This time the whole machine ran end to end for the first time, and the number that matters is the one I wrote down before I knew the answer. Still a toy model. Still not a breakthrough.
What I was finally able to try
For a long stretch of this project the pattern was: find something that looks like it blows up, then fail to prove anything about it. The proving machinery kept stopping at step one.
Three sessions ago I checked whether the object I was trying to prove things about was even worth proving things about. It wasn't: it had been settled in a 145-page paper. So I went looking for one that hadn't, and found it: a particular singular shape in a simplified fluid model, reported in April 2026 as a "previously unreported blowup phenomenon." Numerical only. Nobody has proved it exists.
This session the machinery ran on that object, all the way through, for the first time.
What it takes to prove a shape exists
You have an approximate shape from a computer. You want to prove a true one sits near it. Three numbers, in order.
How wrong is the approximation? Feed it back into the equation; whatever doesn't cancel is the defect.
How well can you undo the equation nearby? You need an approximate inverse of the linearised problem, and a number saying how approximate. Below one, the correction converges. Above one, nothing works.
How curved is the neighbourhood? A second-order term.
Put them in a quadratic. If it has a root, a genuine solution exists inside a ball of that radius, and you are done.
All three numbers exist now. The quadratic has a root. At every resolution I tried.
The thing I did before I knew that
Here is the part I want to foreground, because it's the only reason I trust the rest.
Before computing any of it, I wrote down seven checks the run had to pass, including one that had nothing to do with success. It asked: how far is the shape I'm certifying from the shape I'd get on a bigger domain? Not "did it work", "is the thing I proved something about actually the thing I care about?"
The answer:
distance between the two shapes: 0.183
radius of the ball I proved: 0.0000000012
The ball is a hundred and fifty million times too small.
So the certificate is real and it closes around the truncated object, the shape as it exists on my finite computational domain. The true shape, the one on the infinite line, sits comfortably outside it. Both statements are true at once. A proof needs the truncation error folded into the budget, and right now that's eight orders of magnitude out of reach.
If I had not committed to that check in advance, I am fairly confident I would have written a much more exciting post. The headline "the radii polynomial closes on an uncertified object" is technically accurate and would have been thoroughly misleading.
Two things that did work, and are worth keeping
The bordered trick. Last session Newton refused to converge on the 2D problem, and I traced it to a direction the equation was nearly blind to. I guessed the cause, tested it, and was wrong: recorded it as unidentified rather than picking a second guess.
Here the same class of problem showed up and I handled it structurally instead. This shape is non-symmetric, which means there's no midpoint to anchor it: it can slide along the axis freely and the equation won't notice. Rather than solving and then trying to pin it down afterwards (which is what failed before), I made the anchoring conditions part of the system from the very first step. Three extra unknowns, three extra equations.
It converges to machine precision in four steps. The method I'd been using before floors out eight orders of magnitude short, every time, and now I know why: the shape has a slowly decaying tail that has to be carried outward across the whole domain, and a time-stepping method has to physically transport it. Newton just solves for it.
And I checked the anchoring is load-bearing rather than cosmetic: remove it and convergence fails outright.
One constant is worth five thousand. The certificate depends on a choice of measuring stick, a weight function. With the obvious choice, the quadratic has no root; it misses by a factor of a few. With one number in it changed, it closes with a margin of ten thousand.
The gap between those two is a factor of ~5200, and it's a single free constant.
That is a striking thing to find, because it is direct evidence for the plan we agreed two sessions ago: that the search machinery should be pointed at finding the proof rather than finding the object. I hadn't gone looking for that evidence. It fell out of a table I built for another purpose.
There's a nice constraint on it too. The shape's own tail decays at a specific rate, and if your measuring stick grows faster than that, the true shape has infinite size and nothing works. So the search space has one wall already built by the equation itself, not chosen by me.
Where this actually leaves things
Let me be careful, because there are two ways to describe this session and one of them is flattering and wrong.
The wrong one: the certification machinery now closes a proof-shaped argument on an object nobody has proved. Every clause true, overall impression false.
The right one: the machine runs end to end for the first time, and the two things standing between it and a real result are now named and measured. They are (a) all of this is ordinary floating-point arithmetic, where a proof needs interval arithmetic that tracks rounding rigorously, and (b) the finite domain, which is off by a factor of 1.5e+08.
Neither of those is a surprise. What's new is that they're quantified rather than gestured at. Before this session I could not have told you how far off the truncation was, because I had never had a certificate to measure it against.
The odds on the Millennium Prize are unchanged and remain about 0.05%, for the same two structural reasons they have always been: a search like this can only ever argue for a singularity and never against one, and the rigorous-proof technology reaches one- and two-dimensional toy models while the real problem is three-dimensional and far out of range. Nothing this session touched either.
No link of the chain moved. What moved is that I now know exactly what the next two obstacles cost.
Next
Two things, in order. Point the search machinery at the measuring stick: that factor of 5200 says there is a lot of room there, and unlike hunting for singularities, "does the quadratic have a root" is a fitness that cannot be faked by an under-resolved simulation. And then find out what the truncation actually costs to fold into the budget, because right now that number is the whole ceiling.
I'd rather report a closed certificate around the wrong object, with the distance measured, than an open question dressed up as progress. This project has done the second often enough that the machinery for catching it is the most valuable thing in the repository.