The Operator
Once the substrate (M, LM) is chosen, the closure-response operator is fully fixed by:
On a connected finite graph with inf σ(LM) = 0, the operator Cφ is strictly positive definite, its smallest eigenvalue is exactly φ−2, and the operator-norm of its inverse is
The continuum projection in one coordinate has a closed-form Green's function G(x) = (φ/2) exp(−|x|/φ) with decay scale φ. No shape parameter, no fitted threshold, no learned weight enters the operator at any level.
The Substrate: V600
The discrete substrate used by both empirical witnesses is the 600-cell graph V600: 120 unit vectors on the 3-sphere, generated by three canonical coordinate families — 8 axis vertices, 16 half-integer vertices, and 96 φ-mixed vertices — connected by short edges with inner product φ/2.
Computed from canonical generators
|V| = 120, |E| = 720, uniform degree 12 by H4 Coxeter transitivity. Antipodal symmetry s(−v) = 8 − s(v). 9-shell H3 partition {1, 12, 20, 12, 30, 12, 20, 12, 1}.
Nine eigenvalue classes in ℤ[φ]
Laplacian eigenvalue multiplicities {1, 4, 9, 16, 25, 36, 9, 16, 4} summing to 120. All values numerically resolve to closed-form expressions in the ring of integers of the golden field.
All graph and spectral facts are reproduced numerically by repro/verify_kernel.py from the canonical generators alone — no external data input. The polytope's choice itself is post-hoc motivated by the empirical landings; the construction inside it is theorem-level rigorous.
Discrete-to-Continuum Agreement
For a localised source f at any vertex, the discrete response ψ = Cφ−1f is shell-constant to machine precision: the within-shell variance across all 120 source choices is ∼ 2 × 10−15. The multiplicity-weighted per-vertex Pearson correlation between ψ(v) and the continuum kernel G(‖v − vsrc‖) is:
ρ = 0.976 on the unweighted graph Laplacian.
Numerical agreement between two independently-defined objects: a 120-dimensional discrete inverse and a continuum exponential kernel. Not a definitional identity.
Of three Laplacian variants tested on the same construction, the unweighted Laplacian ranks highest on the geometry-only criterion. The two φ-cocycle weighted variants — with edge weights ω(v) = φκ(v) via shell-grade exponents in φ-geometric and φ-arithmetic forms — score ρ = 0.888 and 0.884 respectively.
| Variant | Per-vertex ρ | Geometric ranking |
|---|---|---|
| Unweighted Laplacian | 0.976 | 1st |
| φ-cocycle, geometric edge weights | 0.888 | 2nd |
| φ-cocycle, arithmetic edge weights | 0.884 | 3rd |
This pure-geometry ranking reproduces — on a different test — the qualitative ranking established independently by the b-anomaly paper's data-χ2 comparison. Geometric and data criteria agree on which variant is the operator.
Two Domain-Disjoint Empirical Landings
The same fixed operator Cφ on the same fixed graph V600, with no shape-parameter retuning between regimes, appears as the shared response primitive beneath two empirical works covering qualitatively distinct physical settings:
The kernel κ(q²), as the shell-mean projection of Cφ−1f, describes the q² shape of the b→sμμ angular anomaly across five public datasets (LHCb 2015 K*, LHCb 2021 K*+, CMS 2025, LHCb 2025, LHCb 2015 Bs→φ). Sign uniformity holds in 5/5 datasets. AIC tie with a constant Wilson-coefficient shift inherited verbatim.
A recurrent self-model layer above the same Cφ on the same V600 — with one condition-dependent self-injection coupling and one substrate-pinned nonlinearity at the graph's average degree — passes 17/18 preregistered substrate/neuroscience correspondences (frozen 2026-04-18) at standard methodology, 18/18 after a documented N=20 deep-dive (thresholds unchanged), plus 6/6 literature-thresholded drug/sleep EEG signatures.
By design, the two witnesses share exactly the geometry-fixed operator: the same Cφ, substrate V600, and shift φ−2. Above that operator level, they share no fitted parameter, no threshold, no dataset, and the empirical claims are tested on disjoint physical domains — flavour physics versus cortical neuroscience.
Mapping Numerics to Claims
The paper applies an explicit claim-boundary discipline. Numerical results map to admissible language:
Spectrum, norm, correlation
The Laplacian spectrum, the operator-norm bound ‖Cφ−1‖ = φ2, the per-vertex correlation 0.976. Reproduced from canonical generators by a single ~470-line script in ~1 second.
Empirical landings
Sign uniformity in 5/5 b-anomaly datasets; aria-chess 17/18 standard plus 18/18 after the documented deep-dive with thresholds unchanged. Direction confirmed; not a uniqueness claim.
Forbidden language
"Derives the kernel." "Proves uniqueness." "Establishes selection." "Settles the underlying physics." None of these claims is made anywhere in the paper.
Stated Limitations
- The φ−2 shift is a design-level stability clamp, not derived from a closure functional or symmetry argument
- The 600-cell choice is post-hoc motivated by the empirical landings — not an a-priori derivation; alternative regular 4-polytopes (24-cell, 120-cell) are an explicit open ablation
- Kernel uniqueness is not claimed on either landing — the b-anomaly's free-width Gaussian alternative and AIC tie (wVFD = 0.348 vs wFREE_C9 = 0.652) are inherited verbatim
- The aria-chess Mode-B linearised-to-non-linear refit drift (+2.77 AIC units on LHCb 2025) is also inherited verbatim as the largest single methodological uncertainty in the passive-regime witness
- No selection theorem on the adaptive-closure-transport 4-tuple — the Lyapunov function on the reduced flow, edge-space decomposition under 2I-equivariance, and full reduced-flow convergence remain open
- No circuit-level model on the active-regime side — the aria-chess substrate-witness scope is imported verbatim, not relitigated
What This Is, And Is Not
One operator. Computed, not fitted. The same operator beneath two unrelated empirical settings, with no retuning between regimes.
verify_kernel.py. MIT licence.