It was never the operator. It was the room we put it in.

Route-NGX, leg 127. Companion to TECHNICAL_P2_ROUTENGX_V1.md. Data: writeup/data/p2_route_ngx_v1_general.json. Figure: fig63_route_ngx_v1_general.png. Novelty pass: writeup/novelty/leg_127.md, run and committed before any construction.

Eight legs ago this project started walking into a wall, and it has been describing that wall in slightly different words ever since.

The setting, once more. We want a computer-assisted certificate, the standard radii-polynomial machinery, for a self-similar profile of an inviscid transport equation. The machinery needs an approximate inverse A of the linearised operator L, good enough that ‖I − A L‖_w < 1. That number is called Z₁. Everything else in the certificate works. Z₁ does not. The best we ever measured, after spending every free choice in the construction, was 8.9591, against the 1 it has to beat.

Leg 58 turned part of that into a theorem. It showed that if A has a zero in one corner (the block called A₂₁, the one coupling the tail rows back to the finite block) then Z₁ ≥ 1 always. Not "in our experiments". Always. The proof is three lines and it works by testing I − A L on a single direction: the far-field mode h that the tail operator sends to zero. Because T h = 0, two of the three terms vanish, and what is left is ‖h‖.

And that is exactly where it stopped. With A₂₁ ≠ 0 a new term appears, h − A₂₁ B h, and A₂₁ can be chosen to cancel it. Leg 54's battery said that in practice this buys you almost nothing (one construction drove the offending term to 10⁻¹⁶ and paid for it with a total Z₁ of 5.66 × 10⁵) but "almost nothing in seven attempts" is a table, not a result.

This leg's job was to decide the general case. It decided it. The answer is yes (Z₁ ≥ 1 for every bounded A, with A₂₁ completely free) and the interesting part is that the argument we expected to need turned out to be the wrong argument entirely, and that the result, once you have it, is not about the operator at all.

The argument we planned, and why it failed

The plan, written into DIRECTION.md before the leg started, was a two-direction argument. The intuition: A₂₁ can buy you the kernel direction, but A₂₁ also has to appear somewhere else, in the tail-row images of the finite-block directions, and what it wins on one direction it should pay for, with interest, on the other.

We carried it out. It gives

Z₁  ≥  1 / (1 + η),        η = ‖G⁻¹ B h‖_w / ‖h‖_w

and on this operator η settles at about 13.7 instead of decaying. So the two-direction argument proves Z₁ ≥ 0.068, which is worth exactly nothing. The pre-registered plan was a dead end, and it is written down here because that is what this repository does with dead ends.

The argument that worked never mentions A₂₁ at all

Here is the whole thing. For any x,

‖x‖  ≤  ‖(I − A L)x‖  +  ‖A‖ ‖L x‖

which is the triangle inequality and nothing else. Divide by ‖x‖ and take the best x:

Z₁  ≥  1  −  ‖A‖_w · σ_min(L),        σ_min(L) = inf ‖Lx‖/‖x‖ = 1/‖L⁻¹‖_w.

There is no A₂₁ in that line because there are no blocks in it. A is never decomposed, so there is no corner for the argument to get stuck in. This inequality is not ours: it is the contrapositive, with a remainder, of the Z₁ < 1 ⟹ invertible hypothesis that every paper in this field states on page one. The novelty pass established that first and forbade the claim. What is ours is only the next question:

Is this operator bounded below, or isn't it?

Because if σ_min(L) = 0, that line reads Z₁ ≥ 1 for every bounded A, and we are done.

It isn't. And the sequence that proves it is one row wide

σ_min falls off like M^−(1−s) as the truncation M grows, where s is the weight exponent of the space. Measured slope at s = 0.3: 0.6985, against a predicted 0.7000. At s = 0: 0.9925 against 1.0000. It does not matter where you put the split: across K = 2, 4, 8 the numbers move by under 0.3%.

The vector that does it is not something a linear-algebra routine found. It is written down: take the far-field kernel h, truncate it at M, and correct the finite block by solving G z = −B h. Then:

• the finite-block rows of L v are zero to floating point (at most 1.2 × 10⁻¹⁴);
• z_K is exactly 0, by a parity property of the recursion, which switches off the only entry coupling the finite block back into the tail;
• and the entire residual of L v sits in one row, the truncation edge at m = M, of size |1 − M/2| · |h_M| · w_M ∼ M^(s−1).

Meanwhile ‖v‖ stays bounded, because Σ m^(s−2) converges for s < 1. A defect shrinking like M^(s−1) divided by a norm that does not shrink is a rate of M^−(1−s). That is the whole proof, and it agrees with the numerically-optimal direction to ten digits.

We then tried to break it. Zero-pad the vector built at M into a tail four times larger: if the near-null behaviour were an artifact of stopping at M, the ratio would jump back to order 1. It costs a factor of 1.354. Turn on dissipation, which gives the tail a diagonal and destroys the kernel: σ_min stops falling and flattens completely (fitted exponent under 3 × 10⁻³, against 0.6985 in the same code path). The instrument can report the other answer.

And the bound is not lossy. Feed it leg 54's entire banked battery, 196 rows, and it holds on all of them, with a minimum slack of 7.7 × 10⁻¹⁰, attained exactly where theory says it must be: at the exact inverse, where ‖A‖ equals 1/σ_min and both sides are zero.

What this actually costs a would-be counterexample

Stated as a magnitude rather than a verdict: any A reaching Z₁ ≤ 1 − δ must have

‖A‖_w  ≥  δ / σ_min(L)  ∼  δ · M^(1−s).

Be honest about what that does and does not exclude. At any fixed truncation the floor is small, reaching Z₁ = 0.99 at M − K = 1024 needs only ‖A‖_w ≥ 6.64, so a finite-M counterexample is not excluded, and in fact leg 54 already built one: the exact inverse, whose norm is 1/σ_min on the nose. That is why leg 54 ruled it inadmissible. What is excluded is a single bounded A that works for every M, and that is the only thing the method ever meant by an approximate inverse. The floor grows by 1.99× per doubling of M in the flat class and 1.62× at s = 0.3, without bound.

The part that changes the story

Here is the finding that made this leg worth running, and it came from the novelty pass rather than from the mathematics.

A preprint appeared in July 2026 (arXiv:2607.19762, Jie Xu) on exactly this operator: the a = 0 CLM linearisation about the same exact profile. It proves that on the origin-H² space, the point spectrum is exactly {0, 1}, the two symmetry modes, and the essential spectrum meets the right half-plane only in the vertical line Re λ = −1/2. Remove the two symmetry modes by the standard modulation, which is exactly what our gauge row does, and the operator is invertible, with a spectral gap of 1/2.

So the operator is fine. It has a bounded inverse. It simply does not have one here, in the weighted ℓ¹ space of Fourier coefficients that the radii-polynomial method needs in order to control its tail.

That reframes eight legs of negative results. The wall was never a property of the CLM linearisation. It is a property of the room we insisted on putting it in, and we insisted because that room is where this particular certificate machinery knows how to work. The project's standing discipline has a lesson numbered 70 that says name the realization. This is that lesson arriving as a literature fact instead of a note to self, and it is the reason no sentence in this leg says the operator "has no bounded approximate inverse". It says: no bounded approximate inverse in ℓ¹_w at s < 1.

There is a matching detail inside our own data. At s < 1 the kernel is in the space and σ_min → 0. At s ≥ 1 the kernel leaves, and the exponent drops to zero: the wall is gone. But at s = 1.5 σ_min starts diverging again at a completely different rate, because a different obstruction has taken over: the dual functional, the cokernel, has entered the space. Two mechanisms, swapping at s = 1, and s = 1 is precisely the exponent where the object we actually want to certify has infinite norm. The space is squeezed from both sides, which is the sharpest form of the thing this project has been circling since leg 51.

What it does not say

Nothing here is about HL_S2_nonsymmetric, the real target. Nothing here moves any link of the chain that leads to the Clay problem; no link has moved in 127 legs and none moved today. The object is the a = 0 CLM linearisation, whose Y₀ is exactly zero for a degenerate reason that certifies nothing. No dynamics were run. A wall measured on this object bounds the real target's difficulty from below, not from above.

What did change: leg 54's scope line ("A₂₁ ≠ 0 is measured, not proved") is retired for this operator in this space. It is proved now. The general class is closed, and the honest statement of what was closed is that the certificate's space, not the operator, is where this method runs out.