N/log N three ways — one shape, two breaks. The host's separator, the function-field prime number theorem, and the logarithmic Chowla weight

The coincidence. Three quantities with the shape N/log N: the number of monic irreducibles of degree n over 𝔽_q, ~ qⁿ/n = N/log_q N; the total mass of the logarithmic weight in the Chowla picture, Σ_{n≤N} 1/n = log N, the Haar measure of the dilation flow (N-006); and the separator of the intermediate-growth host, b_n ≈ c·|G_n|/2ⁿ = c·(5/8)·N/(log₂N − 2) (N-026, run 031). The human asked whether the third is more than a coincidence of shape with the first two. Chased here, and broken twice.

What is shared, exactly. A set of size N and something of size log N acting on it, with orbit–stabiliser doing the arithmetic. Field: the Galois group, cyclic of order n = log_q N, acts on N field elements; the count is orbits, N/n. Host: the group, of size N, acts on the 2ⁿ = (8/5)(log₂N − 2) leaves; the count is a stabiliser, N/2ⁿ. The leaves are themselves a cyclic object: the Schreier path on level 4 flips bits in the order 3,2,3,1,3,2,3,0,3,2,3,1,3,2,3 — the ruler sequence with the bit order reversed — so the path is the odometer orbit, ℤ/2ⁿ. The mirror is therefore exact as a statement about leading terms: a cyclic group of order log N acting on N points against an N-element group acting on a cyclic set of order log N. That is the "mirror image" and the pull toward a dual group. It is the duality of the trace formula's two sides (P-026), not a functor between the objects; no functor was found, and the second break below says there is none at the next order.

Break 1 — depth against width. The integers and the group live on the same binary tree: an integer is its binary expansion, dilation by 2 is the level shift, and the group acts on the tree by automorphisms. In that tree, Chowla's log N and the field's n both count depth — dilation steps, Frobenius steps, scales. The host's log N counts width — the leaves at the top level, 2^{depth}. The host's number of dilation steps is n = log₂ log₂ N. The two logs are one exponential apart, and the shapes agree only because |G_n| happens to be exponential in the width (the group has Hausdorff dimension 5/8 in Aut(T)) while N as an integer count is exponential in the depth. Same tree, different coordinate.

Break 2 — the first correction, and this one is measured. The field's count is (1/n)·Σ_{d|n} μ(n/d)·q^d: the corrections sit at N^{1/2}/log N — the Riemann-hypothesis scale — and are organised by the divisor lattice of n, the subfields. The host's correction is a constant factor c_n: 1.00 at level 3, 0.69 at level 4 (run 032's 176), predicted 0.5–1.07 at level 5 (run 031); it is organised by the level chain, a total order, and it is glassy — the coset graph's bisection landscape (runs 032, 034). Relative correction at N = 4096: 0.31 against N^{−1/2} = 0.016. No square-root structure, no lattice of divisors, no zeros. The analogy holds at the leading term and fails at the first correction, which is where the function-field side has all its content.

Verdict. A coincidence of shape. Kept as the record of the chase, and for the one thing it leaves standing: the host's N/log N is orbit–stabiliser counting in a doubly-exponential group acting on a linearly ordered base, and its "error term" is a glass, not a zeta function. If a function-field object ever appears on the host's side, it will have to bring the divisor lattice with it.

The chase continued — the "zeta" half sharpened, and dated predictions for level 5 (2026-09-27, 18:40 MST)

What was found on the level-4 optimum (176) and the level-3 optimum (16), same evening. The ⟨b,c,d⟩-coset graph's second Laplacian eigenvalue is exactly the super-path's: λ₂ = 4 − 4cos(π/2ⁿ⁻¹) — 4 − 2√2 = 1.1716 at level 3, 4 − 4cos(π/8) = 0.3045 at level 4 — with multiplicity 2ⁿ⁻¹ (4, then 8). The minimum bisection has 65% of its norm in that eigenspace at both levels (0.637, 0.647), nothing in the next eigenspace, and the remaining third spread over high modes. The tilt of the cut along the Schreier path (block fractions 11/16 … 5/16) is the path's Fiedler cosine, amplitude ≈ 0.19 about ½. The cut is not an affine function of the level-4 kernel on any St(3)-coset (0 of 128), and not pulled back from any shallower level: the within-fibre part is combinatorial.

What this does to the sentence. The Rayleigh quotient of the optimum obeys 4·bₙ/N_c = 11/2ⁿ (11/16, 11/32), while λ₂ ≈ 2π²/4ⁿ⁻¹. So the spectral part of the cut's cost — 0.65·λ₂ — is 29% of it at level 4, 15% at level 5, 7% at level 6: the eigenvalue data's share of the constant halves every level. The minimum has a spectral skeleton (the path's lowest mode, a 2ⁿ⁻¹-fold degenerate eigenspace) and a combinatorial body that carries an increasing share of the cost. The "not a zeta function" half stands, now quantitatively: the zeta side sees the tilt, not the constant.

Level 5 arrived (2026-09-27, 19:05 MST) — the dated predictions scored

Claude Code's run 031: b₅ = 90,112 = (11/16)·131,072 exactly, reached by three distinct partitions (a flat one, the 0.69 → 0.31 ramp of level 4, and a 0.22 → 0.78 ramp), all cutting a-edges only. Against the predictions above, tested on the level-5 coset graph (1,048,576 vertices) and the saved partition:

Five of six. The sixth was too specific: the minimum is degenerate over profiles, and the tilt is one of its shapes, not its cause. The spectral share of the cost at level 5 is 0.684 × 0.07686 / 0.34375 = 15%, as predicted from level 4's 29% — halving per level.

The sentence, rewritten. "The error term is a glass, not a zeta function" falls on its first half: the minimum is crystalline, an exact 11/16 of the fibre count at two levels, degenerate across at least three unrelated partitions. What is glassy is the landscape above it (run 034). And it is not a zeta function in the sharp sense that the eigenvalue data's share of the constant halves every level — though two-thirds of the minimising vector lives in one 2ⁿ⁻¹-dimensional eigenspace whose eigenvalue is exactly the super-path's. So: an exact rational constant, with a spectral skeleton that carries a vanishing share of its cost, and a glassy landscape above it. Three things where the first version had one.

Open, and precise. Why 11/16 — the guess 1 − ½·(5/8) with 5/8 the Hausdorff dimension is unfiled; the exact rational and the three-fold degeneracy say the optimum is a construction, and a construction can be written down and checked at level 6 by counting, where nothing can be computed. Why the λ₂-eigenspace has dimension 2ⁿ⁻¹ — one path mode per level-(n−1) leaf would do it, and is testable on the level-4 eigenvectors. And the second band at 4 − 4cos(π/12), which no path of length a power of two produces.

The three chases (2026-09-27, 19:20–19:45 MST), and a fourth thing that had to be checked

0. "90,112 is the q² coefficient of the McKay–Thompson series T_{2A}." Computed, not recalled: from the eta quotients, T_{2A} = q⁻¹ + 4372q + 96256q² + 1240002q³ + 10698752q⁴ + …, so the q² coefficient is 96,256, not 90,112; 90,112 = 11·2¹³ appears nowhere in T_{2A}, T_{2B}, T_{3A}, T_{3B}, T_{5A/B}, T_{7A/B}, T_{13A/B} or j − 744 to q⁷ (all reproduce the known values: 196884, 4372, 783, 134, 51). False, and the kind of false a computation catches in a minute. Neighbouring multiples of 2048 (44 and 47), nothing more.

1. Why the λ₂-eigenspace has dimension 2ⁿ⁻¹ — answered. Left multiplication by the group is an automorphism of the coset graph, so every eigenspace is a representation of the quotient. The λ₂-eigenspace (dim 8 at level 4) is irreducible (commutant of dimension 1), as are the next two bands (dims 8 and 16); the 40-dimensional band at 4 − 4cos(π/5) splits into three, the 34-dimensional one at 4 − 4cos(π/4) into six. And the λ₂ irrep is identified: its character agrees on sixteen words (a, b, c, d, ab, ad, ac, bd, abab, adad, abad, acac, abcd, aba, bab, adab) with the sibling-odd representation on the level-n leaves — functions on the 2ⁿ leaves that change sign under swapping each sibling pair, dimension 2ⁿ − 2ⁿ⁻¹ = 2ⁿ⁻¹. So the slowest modes of the coset graph are "one odd function per level-(n−1) node", which was the guess, with "leaf" corrected to "node".

3. The band at 4 − 4cos(π/12) — answered in kind. The whole level-4 spectrum is 4 − 4cos θ with θ/π ∈ {1/8, 1/6, 1/5, 1/4, 1/3, 3/8, 2/5, 1/2, …} at multiplicities 8, 8, 40, 34, 72, 8, 40, 218, …, symmetric about 4 (the coset graph is bipartite), plus quadratic irrationals — 2 ± √2 (dim 16), 3 ± √3 (32), √(4 + 2√2) as an adjacency eigenvalue — the products of small-path spectra that a self-similar (renormalising) spectrum produces. Each band is an irrep or a few. The π/6 band at level 4 is the π/12 band at level 5: it doubles its period with the level exactly as the π/8 → π/16 band does. Not a period of 12; the second irreducible in a sequence, on the same ladder as the first.

2. Why 11/16 — a handle, not an answer. The sign of the 176-cut's projection onto the λ₂ irrep alone is a balanced cut of 192 = (3/4)·256, agreeing with the minimum on 99% of cosets; adding the bands up to dimension 72 reproduces 176 exactly. So at level 4, 11/16 = 3/4 − 1/16: three quarters from rounding a vector in the sibling-odd representation, one sixteenth saved by the next bands. Dated prediction for level 5, made before the level-5 eigenvectors were recomputed: the sign-rounding of the 90,112 partition's projection onto its 16-dimensional λ₂ irrep is a balanced cut of 98,304 = (3/4)·131,072, agreeing with the minimum on ≈ 99% of cosets, and 90,112 = 98,304 − 8,192 is the same 3/4 − 1/16. If so, the constant is a representation-theoretic optimisation with a two-term answer, and "write the construction down" becomes: the best sign pattern of a sibling-odd function, then a bounded correction.

> Scored, same day (run 038). Falsified as an identity. The λ₂ eigenspace at level 5 is > 16-dimensional with the eigenvalue exact to seven digits and the π/12 band directly above it — §1 > holds. But the sign-rounded projection of the 90,112 partition is a balanced cut of 98,288, not > 98,304; the 66 near-zero cosets can move it to 98,302 at most, two short. Agreement 97.3%, not > ≈ 99%. So 3/4 is the leading term (0.74988) and the bands carry 0.06238, not 1/16. The two-term > identity was a level-4 accident, the third such this note has recorded; the constant 11/16 itself > is measured at both levels and untouched. Exact-arithmetic recount named in run 038, not sealed.

> Scored again, 2026-09-28 (run 039). The two-term identity stayed dead, but the construction > survived the level-up in its strongest form: the sign of the 90,112 partition's projection onto > the sibling-odd irrep (16) plus the π/10 band (80) is a 90,112 bisection — exactly the minimum > cut, a twin of χ in the free spins. The 96-fold band between them carries nothing. So "write > the construction down" is: project onto two irreducible bands, take the sign. Why those two > bands, and why 11/16 comes out of them, is the question that remains.

Back-links (2026-09-29)

Dated pointers added by the codifier from lab/backlinks.py; nothing above this heading was changed.

• Run 031 — confirmed (2026-09-27). lab/runs/031-grig5-bisection.md: CONFIRMED — scored 2026-09-27 (codifier).
• Run 032 — falsified (2026-09-27, measured). lab/runs/032-glassy-wall-anatomy.md: FALSIFIED on both clauses — L_min = 0.128, Δpos = 10.