back to table · edit · history · where entries came from · files · download
8181 bytes, as of the version from 2026-11-18 14:24. Recorded here, not run.
# T753 build evidence
Quantity and selection: c_n=[z^n]F(z), C_n=n!*c_n, n=0,...,20, where
F(z)=Z(1+z)-1/(2*pi*z^2)+log(2*pi)/(2*pi*z). Z is the unconditional
secondary zeta over alpha=(rho-1/2)/i, upper-half zeros with multiplicity and
principal powers, with meromorphic continuation. There are 42 stored entries,
21 orders, including equal forms at orders 0 and 1. argument-selection.md
justifies both parameters independently of numerical precision.
## Generating enclosure
The four-source generator uses the pinned unmodified FLINT evaluator and header,
laurent_bridge.c and generate.py. The precise upstream source hashes and commit
are in upstream-provenance.json; runtime versions are in dependency-versions.json.
FLINT 3.4.0 supplies Arb and the certified finite zero prefix. The evaluator's
tail bounds are unconditional, as stated in FLINT's secondary-zeta documentation.
The prefix does not assume RH for the infinite divisor.
F is analytic on |z|<2, by the inspected pole list after subtracting the double
pole. On |z|=R=7/8 the adapter propagates an entire angular interval through
sin/cos and the complex FLINT evaluation. The 64 intervals cover the boundary;
33 are evaluated explicitly and conjugation symmetry covers the lower half.
Each returned rectangular enclosure has |Re F|+|Im F|<4096. Thus the maximum
modulus principle supplies the analytic bound M=4096 on the disk.
For the N-point trapezoidal coefficient sum on |z|=r=1/2, N=512>20,
discrete Fourier orthogonality gives d_n=c_n+sum_{j>=1}c_(n+jN)*r^(jN).
Cauchy's estimate |c_m|<=M/R^m bounds this alias sum by
M*R^(-n)*(r/R)^N/(1-(r/R)^N). The generator constructs this error as an exact
Sage rational and adds its symmetric interval to the numerical Arb ball.
This bounds aliasing, not just arithmetic error. Conjugation symmetry reduces
the required evaluations to 257. The symmetry follows from pairing the alpha
divisor, not from assuming each individual alpha real.
All nodes were computed once at 493 bits, with every original sample retained
in sample-checkpoint.json. cauchy-evidence.json records the measured output
radii: the largest relative radius is 6.8665243e-112, before decimal conversion.
All values are finite nonzero-width RealBall objects, not Python floats or
manufactured error wrappers. The derivative form uses an exact integer factorial.
The 100-significant-digit decimals imply one last-place unit of uncertainty.
The largest radius leaves more than ten guard digits beyond that claim.
The first reduction reached all native samples and coefficients but stopped at
an unavailable convenience serializer on numberdb.sage. No numerical sample
was lost or repeated: compute.py resumed the source-hash-bound balls, used the
existing numberdb._write serializer, and ran the deterministic checks. The
archived generate-sampling.py differs from the attached generate.py only by
correcting its measured-precision sentence from 480 to 493 bits; the sample
checkpoint preserves its original source hashes. Reproduction verification
uses the final generator afresh, without this checkpoint shortcut.
Existing agents/table-build/check.py and dry_run.py pass: exactness/type,
nonfinite enclosure, suspicious decimal, import and prose checks; 42 entries,
longest native representation 139 characters and 7.9 KB block in the dry run.
See deterministic-results.json and dry-run-results.txt. No API upload was used
to experiment with values.
## Independent build calculation
Let D(t)=d/dt log xi(1/2+t), with the xi normalization defined in the table.
The central Hadamard product gives D(t)=sum 2*t/(t^2+alpha_k^2).
For 1<Re(s)<2, integrating each term by the beta integral yields
Z(s)=sin(pi*s/2)/pi * integral_0^infty D(t)*t^(-s)dt.
The sector containing alpha is compatible with the principal branch. The
continued formula follows after subtracting D(t)'s large-t logarithm.
Specifically, with z=s-1,
I(z)=integral_0^1 D(t)*t^(-1-z)dt
+integral_1^infty [D(t)-(log(t)-log(2*pi))/2]*t^(-1-z)dt,
so Z(1+z)=cos(pi*z/2)/pi*[1/(2*z^2)-log(2*pi)/(2*z)+I(z)].
I is analytic for |Re(z)|<1. This suffices for all derivatives at zero,
although the full regular part extends farther.
independent_check.py computes every logarithmic moment through order 20.
Below 1 it integrates a freshly computed central-xi power series, from the
gamma/zeta/product definition, using independent_xi.c at 2400 bits. It never
reads T752 values. Above 1, u=log(t) gives a smooth integral, evaluated with
mpmath zeta derivatives and digamma through t=1024. Inverse-power digamma
terms beyond 1024 are integrated analytically. The two complete passes use
170/210 decimal digits, 80/112 Gauss-Legendre nodes per each of 16 panels,
and central/asymptotic depths 80/100. The final 42 comparisons pass within
each intended stored decimal interval. The largest high-order derivative
two-pass difference is below 7e-152. Checkpoints, actual outputs, source and
limitations are retained in independent-results.json and mellin-checkpoints.json.
This is an independently derived resolvent-Mellin route, replacing the family
dossier's suggested heat-Mellin checker. It shares xi/Hadamard mathematics and
Arb arithmetic for the small-t series, but no secondary-zeta evaluator,
A-P+E-S implementation or finite zero list. Its quadrature and asymptotic-depth
agreement are heuristic checks, not a second rigorous error certificate. The
generating enclosure has its own explicit tail and alias bound.
Controls: the mock divisor alpha=2 has Z(s)=2^(-s); numerical logarithmic
moments and the same cosine convolution recover its first five coefficients.
Deliberately wrong factor-two, factorial and single-pole signs are rejected.
The exact mock pair alpha=1+i/4,1-i/4 has inverse-square sum 480/289,
distinct from both its ordinate-only and modulus-only replacements. The stored
checker also compares Kawalec Table 9 at n=0,...,10,20; these published values
are lower-precision controls, not the generating source or a 100-digit proof.
## Stored-value checks and limitations
stored_identity_check.py uses exact rational endpoints of each decimal to
check all 42 against the independent integral results, all 21 factorial
relations and all 12 published controls. Its developing parser initially used
Decimal.scaleb under the default 28-digit context, causing false factorial
disagreements. The preserved failed program/report diagnose this checker
rounding; generator values were not changed. The corrected parser extracts
sign, integer digits and exponent without Decimal arithmetic, and has a
1e-100 difference control. A named real_mpfr import initializes the modular
Sage coercion machinery before the exact parser control.
The API read-back, all-row stored check, generator verify(), attachment hashes,
history, authenticated audit and rendered preview checks are recorded in the
handoff and their named artifacts. verify() establishes reproducibility only.
This builder report is not an independent scientific or argument-selection
verdict; only the controller's fresh critic can issue those.
Skill evidence: skill-manifest.json is the matching archived run manifest at
../20261118T135803Z-build-skills/manifest.json. skill-verification.json validates
all local/live resources and the common bundle SHA-256
d6e85d49873fac5bc748d068b0288cf242b0f9aebf3a9a0fff70295019227420.
evidence-manifest.json indexes retained files and actual SHA-256 hashes.
Final read-back result: all 42 stored identities and strings match the offline checkpoint in the intended n-major order; stored-identity-results.json records 75 successful comparisons. The fresh verify() returns 42/42 matched, 0 differing, missing or extra. A subsequent reporting-only vars() call failed because VerifyReport is slotted; its exact successful report and exception are retained, and the logging adapter is corrected. verify-stored-results.json records that limitation rather than hiding it. All ten renderer cases are successful; every intended companion relation appears. History retains the exact proposal #342 and wanted #275 message with gpt-6.1-sol provenance. The authenticated audit has only the documented invertible-convention advisory.