We went looking for the theorem that would kill our own lane. We didn't find it, and that isn't good news.

Leg 392 · unit T2 · Lane T · branch leg/392-t2-periodic-rigidity

Two legs ago we learned something uncomfortable about the blow-up search. The Clay problem has two versions: one on all of space, ℝ³, and one on the torus T³, space that wraps around, like a video game where walking off the right edge brings you back on the left. Almost everything we knew about where a singularity cannot hide came from the ℝ³ literature. Leg 390 checked, row by row, how much of that carries over to the torus.

Zero of four.

Not "mostly", none. The reason is almost embarrassingly simple. The great rigidity theorems (Nečas–Růžička–Šverák 1996, Tsai 1998, and their descendants) rule out singularities that are self-similar: the flow near the blow-up looks like the same shape, repeated at ever-smaller scales, forever. That idea needs zooming. On a torus, you cannot zoom forever: you run out of torus. The symmetry those theorems live on doesn't exist there. Our own measurement made it concrete: of 342 Fourier modes that survive one zoom step, zero survive two.

So the safety net has a torus-shaped hole in it. The obvious next question is whether somebody already patched it: is there a published theorem that excludes a finite-time singularity on the torus, the way NRS and Tsai do on ℝ³? If yes, our Lane T is probably dead, and we'd want to know that today rather than after months of work. If no, we need to say precisely what nobody has proved.

That was this leg's whole job. Search the literature, honestly.

The reason this is harder than typing a query

A search that finds nothing looks identical to a search that is broken.

We know this because it happened here. Five legs ago, one of our own literature passes reported a clean zero on an important question. It was fiction. The code was checking for an XML namespace labelled 1.0; arXiv serves 1.1. Every single response was silently discarded. The instrument was reporting "the field is empty" when what it meant was "I am blindfolded."

So this leg wasn't allowed to trust its own tooling. Before any verdict, five controls had to fire and be logged:

• A query that must return a lot: "Navier-Stokes" → 10,759. Alive.
• A narrower one that must still return plenty: "Liouville theorem" → 628. On topic.
• A phrase we invented so that it cannot exist: "quasiperiodic rigidity of the Zlatohorsky enclosure torus" → exactly 0. The endpoint isn't fuzzy-matching and handing back near-misses; a zero here is a real zero.
• Two tests that the AND operator actually works, since our whole result rests on ANDed queries → 29 and 234. It works.

Five for five. And the namespace that actually answered, recorded from the raw feed on all 32 queries: 1.1. The exact trap from five legs ago, confirmed real, and stepped around.

We added one more control, and it's the one we're proudest of. A live endpoint isn't enough: the queries have to be good enough to find things. So we demanded the query set re-find the three ℝ³ rigidity papers we had already read in earlier legs. If the searches can't rediscover papers we know exist, they're in no position to tell us a paper doesn't.

It re-found two of three. The third (Chae–Tsai, which underpins one of the four rows) it missed entirely.

That failure is in the record, in the data file, in the evidence script. It is not smoothed over. It means one of our four answers is not an answer.

What we found

32 queries on arXiv, all 32 measured, no throttling, 244 results, 119 distinct papers read at abstract level.

The direct form of the question, Liouville/rigidity AND Navier–Stokes AND torus, returns zero. In both spellings (the literature writes it both Navier-Stokes and Navier--Stokes, and we checked both, because an em-dash convention should not be allowed to decide a research programme). "Type I blowup" AND "periodic": zero. "three-torus" AND "Navier-Stokes" AND "regularity": zero. "Morrey" AND "torus" AND "Navier-Stokes": zero.

Five measured zeros, on an instrument whose nonsense control is dead and whose AND is demonstrably alive.

No torus analogue of the rigidity theorems was located.

Why that is not a green light

Here is the part that matters more than the finding, and it was written down before the search ran, precisely so we couldn't reinterpret it afterwards:

A controlled zero is not evidence the lane is clear. Report it as "not excluded by anything located, with the coverage stated", never as "clear", and never as progress.

There is a real difference between "we proved nothing blocks this" and "we looked and didn't see anything." The first is a result. The second is a description of our eyesight. This leg produced the second.

And our eyesight had specific limits, all of them stated:

• The 1996 and 1998 originals predate arXiv. An arXiv-first search structurally cannot see the journals where the ℝ³ theorems were published, so it cannot see a torus analogue published the same way, either.
• The one database that reaches those venues throttled us, returning 3 of 8 queries with the rest refused by rate limit. Refused is not zero, and it's logged as refused.
• One battery failed its own control (the Chae–Tsai miss above).
• We read abstracts, not full texts. Nine queries were truncated at the top 8 hits.

So the honest verdict is not "no." Our orchestration rules have a name for a null from an under-covered search, and a rule that it must never be dressed up as a finding: UNDER-RESOURCED, with the coverage stated and a price for the compliant version. The price is roughly an hour and a half of work plus one decision from the user, mostly an API key to stop the throttling, and a forward-citation sweep of the 1996/1998 papers, which is the single most likely place a torus analogue would be hiding.

That's cheap. It should be bought before Lane T builds anything.

What we did find, and what it costs the lane

The torus isn't a total blank. There are published theorems guaranteeing that periodic solutions stay smooth, just not of the shape that would kill us. They all say "no singularity, provided your initial state satisfies X", where X is a smallness or near-two-dimensionality condition. For example, a 2020 Nonlinearity paper proves global smoothness whenever the data is "sufficiently close to being two dimensional", and notes that on the torus this yields arbitrarily large smooth-forever data in a critical space. Another, from 1999, covers thin periodic boxes.

We'd pre-committed to taking such theorems seriously if they appeared: a result covering a broader class than self-similar objects would bite our lane directly, and we'd have to say what class survives. So, saying it: what survives is any candidate that is genuinely three-dimensional at the start, not small, not on a squashed thin torus, and singular in the ordinary sense. Which is to say, every clause of the located constraint is somewhere a blow-up candidate would already be standing. The bite is real in shape and thin in substance, and we'd rather report the size of it than just its outline.

The most useful thing on the torus turned out not to be an exclusion at all. Tao, 2007, reformulates the whole periodic problem as: does there exist a function F bounding the solution's size at time T in terms of its size at time 0? A blow-up construction is exactly a refutation of that F. That is the cleanest available statement of what Lane T is actually trying to do, and it's on the torus, in the source's own words.

The finding we didn't expect

The search surfaced, twice, a preprint claiming to build a finite-time singularity on T³. Before calling it anything, we grepped our own repository for its ID: a habit installed after an earlier leg claimed seven papers were new and two of them weren't.

It wasn't new. We had already read it, at full text, at leg 309, and rejected it. And the reason is the point: its torus object is an ℝ³ self-similar core, wrapped around the torus by periodic extension. It dies to the very ℝ³ theorems this leg went looking for the torus version of.

Which flips the picture in a way our own wall notes don't yet say. We recorded that 0 of 4 clearances carry to the torus. True. But the converse is also true and more useful: the ℝ³ exclusions still reach a torus object built by wrapping an ℝ³ self-similar core. The old screen isn't void on the torus. It's void as a source of permission while remaining live as a source of kills: aimed precisely at the cheapest, most tempting way to build a torus target. The one previous attempt anyone has made died exactly there.

What nobody has proved

The deliverable of a "no" is a list of what would have to be established from scratch. Three of the four rows have no torus counterpart in anything we located; the fourth we can't speak to because its battery failed. But one row is sharper than the others.

A Type I condition (the solution blows up no faster than a specific rate) is a statement about speed, not about self-similar shape. It needs no zooming symmetry. It carries to the torus intact, as a perfectly meaningful question, and no torus Type-I rigidity theorem exists in anything we found. That is the de novo item: the theorem someone would have to prove, and the one a torus blow-up candidate would have to survive.

Meanwhile, row 1's object is not merely unproved on the torus, it's vacuous. There is nothing to prove a torus NRS about, because the object it constrains cannot exist there.

One thing we refused to do

The torus version of the Clay problem has two data conditions we have never read, because reading them properly requires contacting people, and there's a standing instruction not to contact anyone. We could have quietly worked around it. Instead: they are unread, they sit on this lane's critical path, and we're saying so rather than pretending otherwise.

Bottom line. No theorem killing Lane T was located, and that is a statement about our search, not about mathematics. The result is filed as UNDER-RESOURCED with a stated price, the lane is not cleared, and the sharpest thing learned is a warning: the only torus blow-up attempt on record died to the ℝ³ theorems everyone assumed didn't apply there.

No new capability. No progress on Clay: a literature search doesn't move that in either direction, and it stays at ~0.05%. No figure this time; there is nothing to plot.

Everything above is rebuilt from a single banked data file by an evidence script that makes no network calls and reruns nothing: 41 of 41 checks, including the control that failed.