numerical-evidence.md

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

7180 bytes, as of the version from 2026-11-17 11:56 (current). Recorded here, not run.

# T737 numerical evidence, builder 20261117T112154Z

This is build evidence, not an independent scientific verdict. A fresh critic
must judge the live revision and argument selection before the controller offers it.
The archived canonical skill manifest is `skill-manifest.json`, copied from
`../20261117T112154Z-build-skills/manifest.json`; its matching bundle hash is
984965681e2f8925c511a121c118efcde59581c7b0d906cac8b641bf6925e55a.

## Quantity and selection

The controlling source is Dunn--Radziwill arXiv:2109.07463v3 §5.1 (5.6)--(5.8),
with trace additive character and positive imaginary embedding as in §1.2.
David--de Faveri--Dunn--Stucky arXiv:2410.03048v2 §3.3 and (6.8) give the
shift, CRT and integral-frequency checks. Original Patterson 1977 was not
retrieved independently. No conditional asymptotic is used.

The exact product enumeration and an independent full numerator-lattice scan
agree on all 511 identities AND c,d,n,j decompositions. Of 1821 sign
representatives of norm<=1000, exclusions are 758 by unit support, 414 by
ramified valuation and 138 by prime exponents congruent to 2 mod3.
Branch B counts n=0,1,2 are 363,15,1. Branch A counts are 42 per unit phase
at n=1 and 2 per phase at n=2. There are 11 nonunit cube-factor entries,
355 denominators and largest denominator norm997. Every allowed shared prime
is retained. See selection.json and argument-selection.md.

## Arithmetic checks and precision

The production generator uses Sage native ideal residues and residue fields,
then exact cyclotomic polynomial reduction. For each row, an exact witness
specifies root order L, a rational polynomial V, and positive rational r,
meaning tau=V(exp(2*pi*i/L))*sqrt(r). The two exact-witnesses JSON parts retain
these witnesses. Component-square reductions decide exact zeros and rational
components. Arb evaluates all operations at 512 bits. Every approximate
nonzero component is required to have relative radius <10^(-115); the
stored decimal claims have 100 significant digits. There is no truncation tail
for a finite coefficient. Five entire values are exactly 27.

`dry-run.log` is the existing deterministic dry_run.py/check.py outcome:
511 entries, all exact or carrying finite bounds, no suspicious decimal
exactness or comment faults. Its size estimate prints full working-precision
balls (188.7 KB), while the actual 100-digit flat entries JSON is 143190 bytes;
the longest stored number is 212 characters. No limit exception is needed.

`independent.py` uses integer pairs, primary-prime trial division, explicit
modular finite-field arithmetic and character-free cubic exponential sums.
For every split prime, the sum fixes 2*Re(G) and G^3=-N(pi)*pi fixes three
candidates; both wrong roots are rejected. Inert-prime sums are exact.
Nontrivial cubic CRT factors reconstruct every composite, and DR (1.5)
supplies the shifted symbols. At 640 bits, independent enclosures lie within
every stored interval. Known controls include both ninth-root phases,
tau(lambda^-3)=9*sqrt(3), tau(1)=27, tau(lambda^-1)=9*sqrt(3), structural zeros,
and c=d=-2. Conjugation, a 10^-60 perturbation, removal of a ramified factor
and reversal of a ninth-root phase are rejected. The checker independently
simplifies exact rational c before ball evaluation: otherwise an enclosure
of 5*sqrt(729/25) straddles exact27 and cannot be contained in the singleton.
This was a checker representation issue; no stored value was changed.

The companion checks read the real published decimal intervals as intervals:
all 176 T735 enclosures agree, and T142 (q,n)=(7,2) encloses G_3(1+3*omega).
DR's two supplementary laws agree with separate residue-symbol arithmetic at
all 355 denominators. All 22 integral frequencies obey DFDS's n>=2/n>=1
restriction. A deliberately omitted norm91 CRT phase is rejected.
Sibling values are extra controls, not the independent computation itself.

## Fourier interpretation and its tail

`automorphy.py` independently constructs 5014 auxiliary coefficients of
norm<=10000. These are private check data, not additional public coverage.
The saved checkpoint contains 80-digit intervals and exact decompositions;
its input errors are propagated when resumed. Using the constant term and
both signs, the program checks theta(Ew)=theta(w), E=[[0,-1],[1,0]], at
(0,1/2) versus (0,2) and (omega/3,1) versus (-3*conjugate(omega)/10,9/10).
DR §5.1 states E's multiplier is one.

The proved bound in tail-bound.md sums all omitted Eisenstein lattice points,
including zeros. Each point uses radius 100,25,50,56 respectively; the larger
auxiliary coefficient calculation is reused. Tails are <7e-47 and final
difference component radii <1e-38. Both correct pairs contain zero, and missing
branch, doubled constant term and reversed ninth-root phase mutations fail.
This establishes a normalization diagnostic exceeding 30 digits, not a proof
of automorphy everywhere or a separate validation of all 100 stored digits.
The first uniform-radius Bessel attempt was too wide and is preserved as
`automorphy-320bit-wide-bessel.log`; the resumed check uses the same saved
coefficients with point-specific, rigorously controlled cutoffs.

## Bounded proof-facing experiment

`local_experiment.py` treats the ramified and cube recurrences as known source
consequences, not discovered theorems. The development slice has 478 entries;
33 nonunit-cube or higher-level cases were withheld. Independent evaluations
check 22 ramified pairs and 11 cube pairs, including shared primes.
The candidate scalar annihilator Df(mu)=f(lambda^3*mu)-sqrt(3)*f(mu) cancels
the infinity coefficients but fails on 48 DR auxiliary tau_1/tau_2 cases
(§5.1 (5.13),(5.14), Appendix A d_26,d_17): the extra lambda^-4 support has
nonzero base values and vanishes after this shift. Small controls are 9*omega
and 9*omega^2, with the wrong phases rejected. The next lemma needs an operator
on the full cusp vector, its nonunit projections, decay, admissible weights
and reflection/Voronoi compatibility. These are unresolved. No RH or improved
zero-free-region conclusion follows.

## Reproduction and limitations

Use the preserved invocation scripts under agents/sage.sh. The standalone
attachments use SageMath/passagemath and numberdb; all keys and agent metadata
remain session configuration. Generator verification is reproducibility,
while independent.py checks a different arithmetic construction. Both share
the published coefficient theorem, exact arithmetic and Arb. The build does
not independently prove Patterson's theorem. The authenticated readback independently contains all 511 stored entries;
verify(sample=None) reports 511/511 matched, none differing/missing/extra.
Native returns are 26 RealBall, 480 ComplexBall and 5 Rational; minimum
relative supported precision is 148.377 decimal digits at 512 bits.
The final audit has only two advisories (definition length and tag reach,
explained in environment-findings.md). The actual page renders without JS or
math errors; all four companion relations are visible and their links give
HTTP200. The initial and later history messages retain #332 and #272 and the
correct gpt-6.1-sol/codex-cli provenance.