# T638 numerical evidence

Run: `20261031T115431Z` (build). This is build evidence, not an independent
critique acceptance. The table remains an unpublished draft.

The governing instructions are archived in
`agents/runs/20261031T115431Z-build-skills/manifest.json`. Both local and live
skills have SHA-256
`a2b28d9953264d7b0ca62663785c2952190c5be6d6d3597fd695d9a54d606278`.
The attached `skill-manifest.json` reproduces the manifest. Sources and their
SHA-256 digests are recorded in `sources.json`; their downloaded HTML is
preserved in the build-work directory.

## Quantity and precision

The 63 stored values are the raw moments alpha_n(d), with dimension d=2,...,10
and order n=2,...,8. Every value has 100 significant decimal digits. Decimal
strings carry NumberDB's uncertainty of one unit in the last written place.
The rigour is **heuristic (agreement-checked)**, because no rigorous quadrature
error bound was established.

`compute.py` computed all moments at 150 and 200 working decimal digits with
separate interval subdivisions. The maximum absolute change was
1.25358926129e-151; the maximum relative change was 5.89815000221e-151.
The generator requires relative agreement better than 1e-120 before returning
100 digits. `moments.json` retains both working-precision results and all seven
phase integrals in each dimension. `compute.log` retains the full run output.

The reciprocal phase factorization is expanded at infinity. If w=1/z and
J_k=integral(theta(t)t^k)/pi, its reciprocal is
w/sqrt(d) times exp(sum((-1)^k J_k w^(k+1)))/(1+w). Expanding its logarithm
and then its exponential gives the explicitly implemented coefficient
recurrence. Comparing with equation (56) in d'Eon and McCormick fixes the
sign and factor two in alpha_n. No float or rounded parameter enters this
calculation.

## Independent numerical checks

`checks.py` does not import the generator. It uses the Fock integral, equation
(52) of d'Eon and McCormick, and direct integration of mu^n H_d(mu)G_d(mu).
Let K_d(k) = 2F1(1/2,1;d/2;-k^2) and
B_d(k)=(1-K_d(k))(1+k^2)/k^2. Subtracting the exactly integrable rod kernel
and substituting k=tan(v)/mu gives

H_d(mu)=(1+mu) exp[-integral_0^(pi/2) log B_d(tan(v)/mu) dv / pi].

A transformed hypergeometric expression avoids small-k cancellation; an
independent recurrence for K_d is used at larger k. The outer substitution
mu=sin(phi) removes the Flatland inverse-square-root endpoint singularity.
Composite Gauss-Legendre rules of degrees 48 and 80, with independently tightened
inner adaptive quadrature tolerances, evaluate all nine moments together.

Before trusting that quadrature, the nonzero control integral of sin(v) on
[0,pi/2] was compared with 1 (error 1.11e-16), H_2(1) with
sqrt(2) exp(2 Catalan/pi), and H_3(1) with Kawabata's 2.907810529078606.
The latter two agree at double precision. Known lower moments alpha_0=2 and
alpha_1=2/sqrt(d) are tested in all nine dimensions.

Across all 63 higher moments the maximum absolute error is 7.55e-15. The lower
moment controls have maximum error 1.05e-14. The maximum quadrature-refinement
change is 8.22e-15. These independent computations check the mathematics to
roughly 13 decimal digits; they do not certify the last 87 stored digits.

Additional checks:

- The Busbridge recurrence through total order eight holds in all dimensions.
- Each second moment agrees with sqrt(d) alpha_2/2 = z_0(d) in the live T637
  document archived before this build; no published companion was edited.
- The exact alpha_2(2)=sqrt(2)(1/2+1/pi) and alpha_2(4)=2/pi formulas agree.
- All three available nontrivial conservative 3D moments in Kawabata Table 6
  agree within 6.90e-16 of the published fifteen-decimal values.
- The elementary dispersion recurrence agrees with direct hypergeometric
  evaluation on nine test arguments in each of nine dimensions, to a maximum
  phase difference 1.43e-100 at 100 working digits.
- The literal Programs snippet was executed and agreed with the computed value.

## Dry run and database read-back

The repository's `dry_run.py` and `check.py` were used unchanged. All 63 entries
passed the type, finite-value, prose and exact-in-disguise checks. The longest
written entry is 103 characters; the measured entries block is 8.8 KB.

After filling, `verify(sample=None)` reported **63/63 matched, 0 differing,
0 missing, 0 extra**. The independent Fock integration and all identities were
then rerun against values fetched from the database. The maximum Busbridge
residual was 3.49e-100; the maximum discrepancy with T637 was 1.07e-100; the
largest low-dimensional closed-form discrepancy was 1.34e-100, consistent
with rounding the stored operands to 100 significant digits. The direct
quadrature discrepancies were unchanged.

Audit fixes only shortened the definition, removed two low-reach tags and added
an installation command to the generator docstring. A final API read confirmed
that the Numbers and parameter order were unchanged. The final audit is clean.
See `audit-final.json`, `verify-stored.log`, `stored-document.json`,
`build-fingerprint.json` and `numerical-evidence.json` for machine-readable
results and source hashes. The fingerprint is a build aid; the critic must
obtain and judge its own fingerprint and argument-review sidecar.

Versions: SageMath 10.9, Python 3.12.3, numberdb 0.1.10, mpmath 1.4.1,
SciPy 1.17.1 and NumPy 2.4.4. No rigorous endpoint or quadrature enclosure is
claimed.
