numerical-evidence.md

back to table · edit · history · where entries came from · files · download

4833 bytes, as of the version from 2026-11-19 21:01 (current). Recorded here, not run.

# Numerical evidence for T747

This is build evidence, not the independent scientific or argument-selection
verdict. Read together with the matching archived canonical skill manifest:
`../20261118T063041Z-build-skills/manifest.json`, also attached as
`skill-manifest.json`, SHA256
bba13c9ac69e01f4f550078b34f7201dbc4cb422f40ff6e2d7360bb70cd61fe8.
Complete matching bundle SHA256:
d6e85d49873fac5bc748d068b0288cf242b0f9aebf3a9a0fff70295019227420.

Definition: journal Guttmann–Rogers (3), denominator
`t-cos(theta1)-cos(theta2)-cos(theta3)+cos(theta1)*cos(theta2)*cos(theta3)`,
probability angular measure on [0,pi]^3, real t>2. Exact rational argument
identities are enclosed from QQ; there are no approximate input parameters.

Enclosure proof: journal (4) gives G(t)=B(4/t^2)/t. For each nonnegative integer
r, both hypergeometric coefficients a_r and binomial coefficients d_r start at
1 and have positive ratios
`a_r/a_(r-1)=(8r-7)(8r-3)/(64r^2)` and
`d_r/d_(r-1)=(4r-3)/(4r)`, each <=1. Consequently the three-factor convolution
has 0<=b_r<=binomial(r+2,2). For L=M+1, q=1-z, its omitted majorant is exactly

`S(z,M)=z^L/2*((L+1)*(L+2)/q+(2L+3)*z/q^2+z*(1+z)/q^3)`.

The identity follows by expanding binomial(L+k+2,2) and summing 1,k,k^2
against z^k. `checks.py` also compares it with a separately summed tail at
z=1/4 and 256/289. The generator uses M=2592 and 464-bit real Arb balls,
including convolution rounding and rational input enclosure. It returns the
partial sum divided by t, with symmetric error S/t. At the worst argument
17/8, S/t is below 4.008e-130. The largest observed ball radius is below
4.008e-130 and the minimum relative accuracy is 429 bits (a measured metric,
not the source of certification). 100 significant digits are then serialized.
This bound applies to all 66 selected arguments since their z<=256/289.

Executed checks:

- Existing dry_run.py and check.py: 66 values of native type RealBall; no
  exactness, suspicious-decimal or prose findings; imports initialized Sage.
  Approximate raw enclosure block 12.8 KB, longest 159 characters.
- Every serialized interval, after the upload and readback, contains independent
  800-bit evaluations of both journal (4) via Arb hypergeometric and the
  algebraically transformed elliptic formula from (7),(9). Oracle relative
  radii are required below 1e-150, with finiteness and positivity tested before
  containment. This is a different implementation with shared source identities.
- Every one of the 66 arguments is checked by independently integrating the
  third angle analytically and applying SciPy dblquad to
  `1/sqrt((t-2+(1-cos(x))*(1-cos(y)))*(t+2-(1+cos(x))*(1+cos(y))))`.
  Absolute observed differences are <=2.23e-16, with tolerances 2e-12.
  These are non-certified double-precision checks, not 100-digit verification
  of the defining triple integral. Exact sample: the entire selected set in
  argument-selection.md and each comparison recorded in check-results.json.
- Exact monomial multiplication of H=8h gives CT(H^2)=104 and CT(H^4)=17880,
  all odd moments through 7 vanish, and moments through 8 match b_r for r<=4.
  Kernel corner extrema and strict decrease of stored values were checked.
- Known nonzero controls: zero-kernel quadrature at t=3 gives 1/3,
  2F1(...;0)=1, K(0)=pi/2. Wrong triple-product sign, missing angular measure
  factor, modulus instead of parameter, additive potential shift, and a
  1e-90 perturbation of a stored 100-digit value are all rejected.
- `Generator.verify(sample=None)` reports 66/66 matched, 0 differing,
  0 missing and 0 extra. This tests reproducibility, not scientific independence.
- The displayed Programs example at unlisted 81/8 runs and agrees with a
  separate 50-digit mpmath original-kernel quadrature, difference <3e-53;
  this is heuristic evidence for the example.

Execution: all Sage was run with the unchanged `agents/sage.sh` on the builder.
Mount generate.py, entries.json for offline checks, or stored-table.json for
readback checks, alongside checks.py. Execute `sage -python checks.py` on a
normal Sage installation, or the wrapper here. checks.py prints RESULTS_JSON
and writes the same result to NUMBERDB_CHECK_OUTPUT (otherwise a unique local
temporary directory); this runner preserves it as check-results.json. Versions:
Sage 10.9, numberdb 0.1.10, Python 3.12.3, SciPy 1.17.1. Exact logs, versions,
source hashes and checkpoint entries are retained in the run artifact directory.

Limitations: this build does not issue a fresh independent actor's verdict.
High-precision formula checks share the published mathematics; original-kernel
quadrature checks every argument at lower precision. The source's unprinted
arithmetic identities at J(14),J(322) are not controls. No claim is made about
interior spectral values, complex branches, or finite G(2).