Dispatch · 4 October 2026 · K.W. Norton

Verification Ledger

An outside review (Grok, October 2026) observed that the site asserts a correspondence between Riemann zeros and nuclear levels without the formula, data, or benchmark needed to test it. That critique is fair. This page answers it directly: what is demonstrated, with code anyone can run, and what is not yet verified, with the steps that would verify or refute it.

1. Result: GUE pair correlation, first 400 zeros

Zeros at heights 14.13 to 679.74, unfolded to unit mean spacing. Empirical density of pair differences compared with Montgomery's GUE prediction 1 − (sin πx / πx)². Small sample; fluctuations of ±0.1 per bin are expected.

Separation xEmpirical R₂GUE prediction
0.1250.0200.050
0.3750.2700.385
0.6250.8400.779
0.8751.0900.981
1.1250.8200.988
1.3751.0300.954
1.6250.9000.967
1.8751.0200.996
2.1250.9800.997
2.3750.9000.985
2.6251.1200.987
2.8750.9900.998
  • Level repulsion: 0.5% of spacings fall below 0.25, against 22.1% for uncorrelated (Poisson) levels.
  • Spacing variance 0.138 (GUE ≈ 0.18 asymptotically; Poisson = 1). Low-height zeros are known to be stiffer than the asymptotic limit.

2. Code

Requires Python 3, numpy, and mpmath. Runtime about 90 seconds. Longer walkthrough in the Python spectral tutorial.

import mpmath, numpy as np

# 1. First 400 nontrivial zeros of zeta (imaginary parts)
z = np.array([float(mpmath.zetazero(n).imag) for n in range(1, 401)])

# 2. Unfold to unit mean spacing: N(t) ~ (t/2pi) log(t/(2 pi e))
u = z/(2*np.pi) * np.log(z/(2*np.pi*np.e))
s = np.diff(u); s /= s.mean()

# 3. Pair correlation: histogram of all differences u_j - u_i < 3
d = np.array([u[j]-u[i] for i in range(len(u))
              for j in range(i+1, min(i+40, len(u)))])
h, e = np.histogram(d, bins=np.arange(0, 3.01, 0.25))
dens = h / (len(u) * 0.25)
c = (e[:-1] + e[1:]) / 2
montgomery = 1 - (np.sin(np.pi*c)/(np.pi*c))**2   # GUE prediction

for x, emp, th in zip(c, dens, montgomery):
    print(f"{x:.3f}  empirical {emp:.3f}  GUE {th:.3f}")

# 4. Level repulsion: fraction of spacings below 0.25
print("P(s<0.25):", np.mean(s < 0.25), " Poisson would give", 1-np.exp(-0.25))
print("spacing variance:", s.var())

2b. Nuclear test: uranium-238 resonances vs Riemann zeros

Run 2026-10-05

Data: ENDF/B-VII.1 evaluation of n + ²³⁸U (NNDC, Brookhaven), resolved resonance region, s-wave (ℓ = 0, J = ½⁺) Reich–Moore parameters — a single symmetry class. 897 positive-energy levels from 6.67 eV to 19.99 keV, mean spacing D = 22.18 eV, unfolded by a linear fit to the level staircase. Compared against the first 897 Riemann zeros (Odlyzko's table, heights 14.13 to 1300.49). Test: Kolmogorov–Smirnov distance of nearest-neighbour spacings from the GOE and GUE Wigner surmises; 5% critical value 0.045.

Measure²³⁸U s-waveRiemann zeros
Levels / zeros used897897
Fraction of spacings < 0.256.4%0.45%
Spacing variance (GOE 0.29, GUE 0.18)0.3600.143
KS distance to GOE0.040 — fits0.110 — rejected
KS distance to GUE0.098 — rejected0.044 — fits
KS distance to Poisson0.203 — rejected0.323 — rejected

[RESULT]

The two spectra are not statistically the same. Uranium-238 fits GOE and rejects GUE; the zeros fit GUE and reject GOE. A direct two-sample test between them gives distance 0.138, p ≈ 7 × 10⁻⁸. Both are strongly non-random (Poisson rejected), so they share level repulsion — but of different strength (β = 1 vs β = 2).

Consequence: any proposed zero-to-level mapping cannot be a one-to-one identification of spacings. It must explain the GOE–GUE difference (time-reversal symmetry in the nucleus) explicitly. This counts against the simplest form of the proposal and is recorded as such.

  • Uranium spacing variance 0.36 exceeds GOE's 0.29 — consistent with known missed weak levels and possible p-wave contamination in evaluated data. It does not move the result toward GUE.
  • Only nearest-neighbour spacings tested; long-range statistics (Δ₃, number variance) are next.
# ENDF/B-VII.1 n-092_U_238.endf from nndc.bnl.gov; zeros1 from Odlyzko's tables
import numpy as np
from scipy.stats import ks_2samp
def f(x):  # ENDF float, e.g. "6.673491+0"
    x = x.strip(); i = max(x.rfind('+'), x.rfind('-'))
    return float(x[:i] + 'e' + x[i:]) if i > 0 else float(x)
L = [l for l in open('n-092_U_238.endf') if l[70:72].strip()=='2' and l[72:75].strip()=='151']
n = int(L[4][55:66])                      # l=0 block, 926 entries
E = np.sort([e for e in (f(r[0:11]) for r in L[5:5+n]) if 0 < e <= 2e4])
p = np.polyfit(E, np.arange(1, len(E)+1), 1)   # linear staircase: s-wave D constant
a = np.diff(np.polyval(p, E)); a /= a.mean()
z = np.loadtxt('zeros1')[:len(E)]
b = np.diff(z/(2*np.pi)*np.log(z/(2*np.pi*np.e))); b /= b.mean()
def ks(x, cdf):
    x = np.sort(x); m = np.arange(1, len(x)+1)/len(x); c = cdf(x)
    return np.max(np.maximum(abs(m-c), abs(m-1/len(x)-c)))
goe = lambda s: 1-np.exp(-np.pi*s**2/4)
g = np.linspace(0, 6, 6001)
gue = lambda s: np.interp(s, g, np.cumsum(32/np.pi**2*g**2*np.exp(-4*g**2/np.pi))*(g[1]-g[0]))
for name, x in [('U-238', a), ('zeros', b)]:
    print(name, 'GOE', ks(x, goe), 'GUE', ks(x, gue), 'crit', 1.36/np.sqrt(len(x)))
print(ks_2samp(a, b))
[PUBLIC REPOSITORY]github.com/kwnorton000-hash/spectral-verificationZero data, uranium-238 resonance extract, all three scripts, committed outputs, and a known-answer reference case that checks the test itself. MIT-licensed; rerun it and report discrepancies.

2c. Second nucleus: thorium-232 resonances vs Riemann zeros

Run · Evaluation-level diagnostic

ENDF/B-VII.1 n + ²³²Th, MAT 9040, MF = 2 / MT = 151: the first resolved Reich–Moore region ends at 4 keV. Selected ℓ = 0, |J| = ½ (Jπ = ½⁺ for the spin-zero target), with 0 < E ≤ 4000 eV. This yields 242 evaluated levels from 21.80862 to 3980.035 eV, and 241 gaps. Negative-energy/external poles, p-wave blocks, and unresolved-region averages were excluded.

The uranium pipeline was repeated: linear staircase unfolding, then mean-normalized nearest-neighbour spacings. Fitted spacing D = 16.58 eV; arithmetic mean gap = 16.42 eV. Compared with the first 242 Odlyzko zeros (heights 14.13–459.51), unfolded using the same smooth zero-count approximation. No window was selected to improve the fit: the boundary is the evaluation’s first resolved region.

Measure²³²Th s-waveRiemann zeros
Levels / zeros used242242
Nearest-neighbour gaps241241
Fraction of gaps < 0.255.39%0.00%
Spacing variance0.3010.132
KS distance to GOE0.04550.1383
KS distance to GUE (matched grid method)0.10150.0716
KS distance to Poisson0.22420.3527

Nominal 5% one-sample KS threshold: 0.0876. Thorium is below that threshold for GOE and above it for GUE and Poisson; the zeros are below it for GUE and above it for GOE and Poisson. “Below threshold” means not rejected by this diagnostic—not proof of an exact distribution. Using the analytical GUE CDF instead of the prior numerical grid changes each distance by less than 0.001, without changing these threshold comparisons.

Provenance caveat: MF1/MT451 says many unseen p-wave resonances were inserted to match Wigner spacings and Porter–Thomas widths. Those blocks were excluded; testing them would be circular. The selected s-wave sequence is still evaluated data, with possible incompleteness and assignment uncertainty—not an independently verified experimental sample. No mapping, conventional-code predictive benchmark, energy application, or long-range result is established by this run.

Sources: NNDC ENDF/B-VII.1 neutron library · Odlyzko’s zero table. Download the script, compressed evaluation, and zero data into one directory; run python compare.py with numpy and scipy. This thorium artifact is hosted here; the existing GitHub repository has not been updated by this run.

2d. Seven-isotope survey

Run · Evaluation-level diagnostic

The same pipeline, applied to every spin-zero heavy target in ENDF/B-VII.1 whose first resolved region holds at least 100 s-wave (J = ½) levels. Each nucleus is compared with the same number of Riemann zeros. The selection rule was fixed before results were seen; no isotope was dropped after running. One exact duplicate energy in Gd-154 was merged.

NucleusLevelsKS vs GOEKS vs GUE5% thresholdNot rejectedNucleus vs zeros (KS, nominal p)
U-2388970.0400.0980.045GOE0.137, 8.9e-8
Pu-2402670.1200.1570.083none0.203, 3.3e-5
Th-2322420.0460.1020.088GOE0.141, 1.6e-2
Gd-1541620.0640.1140.107GOE0.149, 5.6e-2
U-2341180.1270.1880.126none0.231, 3.8e-3
Er-1661150.0580.1240.127GOE+GUE0.193, 2.8e-2
Er-1681010.0860.1470.136GOE0.190, 5.4e-2

Script, resonance sections, zeros and output: dataset downloads.

3. Ledger

[DEMONSTRATED]

Riemann zeros show GUE pair correlation

Montgomery (1973) conjecture; Odlyzko's large-scale numerics (1987 onward). Reproduced at small scale above. This is established mathematics, not our claim.

[DEMONSTRATED]

Heavy-nucleus resonances show GOE-class repulsion

Porter–Rosenzweig; Bohigas–Haq–Pandey (1983) Nuclear Data Ensemble. Note: GOE (time-reversal symmetric), not GUE — a distinction the site must carry everywhere.

[UNVERIFIED]

A direct zero-to-level mapping with predictive power

Shared universality class is not a mapping. Needed: an explicit formula, its free parameters, and out-of-sample predictions for a named nucleus.

[UNVERIFIED]

Benchmark against a conventional nuclear code

Needed: head-to-head comparison with a shell-model or R-matrix evaluation (e.g. ENDF resonance data) on the same observable, with error bars.

[UNVERIFIED]

Any energy-yield application

No claim of energy extraction or abundance is supported until the two items above are met. Stated here plainly.

4. Timeline to verification

  1. Completed 2026-10-04

    Ledger published; GUE pair-correlation reproduction and code posted (this page).

  2. Completed 2026-10-05

    Following Grok's reviews, overstated wording corrected across the home page, Theory, Energy Brief, and Python tutorial. Claims of exact nuclear solutions, "0.000% stochastic error," and "99.9997% GUE match" were removed or re-scoped; the three unverified limits are now stated on the homepage.

  3. Completed 2026-10-05

    Nuclear test run (section 2b): 897 ENDF/B-VII.1 ²³⁸U s-wave resonances vs 897 Riemann zeros. Uranium is GOE, zeros are GUE; the two spacing distributions differ (p ≈ 7 × 10⁻⁸). Result counts against a direct one-to-one mapping.

  4. Completed 2026-10-05

    Public repository published with data, code, committed outputs, and a known-answer reference case (GOE, GUE, Poisson control spectra correctly identified by the same pipeline).

  5. Pending · opened 2026-10-04

    Scale the zero statistics to 10⁵ zeros using Odlyzko's tables.

  6. Completed 2026-10-17

    Second-nucleus nearest-neighbour comparison completed (section 2c): 242 evaluated ²³²Th s-wave resonances vs 242 zeros. Two-sample KS distance 0.1411, nominal p ≈ 0.0164; the GOE–GUE diagnostic pattern repeats, with statistical calibration and evaluation caveats explicitly recorded. Data, script, and output downloadable here.

  7. Completed 2026-10-17

    Seven-isotope survey (section 2d): U-238, Pu-240, Th-232, Gd-154, U-234, Er-166, Er-168. Every nucleus sits closer to GOE than GUE; every zero set is GUE-like. Rank-paired dataset published for download.

  8. Pending · opened 2026-10-05

    Add long-range statistics (Δ₃, number variance). Thorium nearest-neighbour comparison is complete; long-range tests remain pending.

  9. Pending · opened 2026-10-04

    Write the proposed mapping as an explicit formula that accounts for the GOE–GUE difference, with parameter count; freeze it before testing.

  10. Pending · opened 2026-10-04

    Out-of-sample prediction vs a conventional nuclear evaluation, with all data and code public.

Every step carries the date it was opened; the completion date is added when it is done — including any result that refutes the proposal.