back to table · edit · history · where entries came from · files · download
7169 bytes, as of the version from 2026-11-18 20:02 (current). Recorded here, not run.
# Numerical evidence for T756
This is build evidence, not an independent scientific-review verdict. The
controller must commission a fresh critique and an argument-selection review.
The private draft stores g=-Psi, not Psi or its absolute value. Its sole real
parameter is t>=0; negative arguments follow by evenness. All logarithms are
natural. There are no branch or normalization choices left implicit.
## Sources and exact normalization
Suzuki, arXiv:2206.03682v4, equations (1.1) and (1.8), supplies the prime formula
and g=-Psi. The paragraph following (1.1) gives Phi(1,2,1/4)=pi^2+8G and the
exact origin. Equations (3.9)-(3.11) identify Psi=W(Delta_t), with
Delta_t(x)=(t-|x|)_+/2. Theorem 1.2 and its following paragraph distinguish
the unconditional named function from membership in the global screw class,
which is equivalent to RH. The build makes no such positivity assumption.
Suzuki, arXiv:2606.09096v1, introduction, gives the explicit Weil functional;
its C_0 is the Euler-Mascheroni constant, not the C in the Lerch formula.
Both source HTML files are preserved with hashes in research.md.
## Production enclosure
At 100 significant decimal digits, generate.py uses RealBallField(513), namely
numberdb.bits(100,losing=180). The exact integer q determines the support and
t=log(q) is recomputed as a ball. The m=q summand is identically zero, so no
rounded comparison determines whether it contributes. Primes are sieved and
their positive powers are generated. Python division never produces a stored
float: all relevant quotients involve Sage integers, rationals or balls.
For z=q^(-2), after the terms j=0,...,N-1 of Phi(z,2,1/4), the remaining
positive sum is bounded by
B_N = z^N / ((N+1/4)^2*(1-z)).
The denominator increases with j, and the geometric majorant proves this bound.
Its outward ball upper bound is added as a symmetric error. Truncation stops
when B_N < 2^(-(working_bits-32)). Arb propagates all special-function,
logarithm, prime-sum and arithmetic errors. The 70 nonzero values retained at
least 479 relative-accuracy bits, more than 144 decimal digits. The exact anchor
is returned as a native Sage Integer(0), not a zero-width approximate value.
The package converts these enclosures to 100 significant-digit decimals; a
stored decimal denotes the exact rational centre plus/minus one last-place unit.
production-balls.json retains every pre-serialization ball and native type.
## Separate triangular-Weil evaluation
independent_check.py does not import generate.py. It factors every integer
separately for the von Mangoldt support, uses no Lerch evaluation, no digamma
and no Catalan constant, and computes W(Delta_t) directly from the integral
formula. The pole integral is evaluated by quadrature too:
P(t) = integral_0^t 2*(t-x)*cosh(x/2) dx.
The arithmetic contribution is sum_{m<q} Lambda(m)*(t-log(m))/sqrt(m).
The remaining formula is
W = P - prime_sum - t*(log(4*pi)+EulerGamma)/2 - J,
J = integral_0^t ((t-x)*exp(x/2)-t)/(exp(x)-exp(-x)) dx
- t*atanh(exp(-t)).
For x>t, the last integrand before splitting is -t/(exp(x)-exp(-x)); its
integral to infinity is exactly -t*atanh(exp(-t)). Thus no unbounded
integration tail is discarded or guessed.
For y=exp(x/2), the finite-interval integrand is rewritten exactly as
t*y^2/((y+1)*(y^2+1)) - y/(2*sinhc(x)),
sinhc(x)=sinh(x)/x, sinhc(0)=1.
Its removable value at x=0 is t/4-1/2. When a complex quadrature ball contains
zero, the checker encloses sinhc by sum x^(2k)/(2k+1)! and an explicit tail.
After term k, the absolute next-term ratio is bounded by
r^2/((2k+2)*(2k+3)), where r bounds |x|; once this is <1, a geometric
majorant bounds the tail. Balls containing poles yield a nonfinite enclosure,
which Arb subdivides; nonfinite final values are rejected. The actual
integration path is real with t>0, and all functions are analytic near it.
The origin is handled by its exact definition, avoiding 0*atanh(1).
Quadrature uses ComplexBallField(512), absolute and relative tolerances 2^-420.
Every nonzero result has at least 421 relative-accuracy bits. All 71 values
were checked offline and checked again against the API-read stored document.
For each nonzero row, the independent ball is strictly inside the exact
rational endpoints of the stored decimal interval, is finite, has at least
380 relative-accuracy bits, and has radius below 1/1000 of the stored width.
check-results.json gives all 71 actual comparisons. The checker is stronger
than a midpoint residual or bare overlap test.
The routes share the explicit-formula theorem, elementary support structure
and Sage/Arb/FLINT arithmetic. They differ in analytic representation,
implementation, arithmetic-support enumeration and special-function reduction.
This is a separate numerical method within the build, not independent actor
approval. No spectral zero-sum computation or infinite-prefix certificate is
claimed. T3 and the zero-count paper were inspected as leads for an optional
coarser spectral check, which was not needed to obtain the full quadrature
comparison and was not executed.
## Controls and reproducibility
The checker validates a known nonzero quadrature, integral_0^1 x^2 dx=1/3,
and the exact nonzero pole term P(log(4))=2, together with the exact origin
and the removable sinhc value. Wrong sign and a factor of two at log(4) are
rejected. NaN and a ball enlarged by radius 1 are also rejected. The results
JSON contains the actual control outputs and booleans, not hand-written
passing results.
agents/table-build/dry_run.py and check.py ran on all 71 generator values:
native types, finite enclosures, suspicious decimal exactness, ring imports,
entry prose and size checks all passed. The measured full-ball block was
14.6 KB, with longest value 169 characters. The stored block is smaller.
Generator.verify(sample=None) read the private draft with its key and matched
71/71, with no differing, missing or extra entries. This verifies
reproducibility only. The displayed mpmath example also agreed to 120 decimal
places at 130-digit working precision; it is an illustrative noncertified
calculation, not the basis for the proven rigour label.
Reproduce from downloaded attachments with Sage 10.9, FLINT 3.4.0,
numberdb 0.1.10 and Python 3.12.3:
sage -python generate.py
sage -python independent_check.py stored-document.json --output new-results.json
A private draft read needs authorized credentials in NUMBERDB_API_KEY. No
credential is embedded in these files. On this deployment all Sage invocations
used agents/sage.sh; wrappers and complete stdout logs remain in the run
working directory. versions.json and verify.log record actual versions.
## Skill binding
The coordinated archived skill manifest is
/home/ubuntu/numberdb-campaign-20261116-sol/repo/agents/runs/20261118T173437Z-build-skills/manifest.json.
All core and reference resource hashes of both archived bundles were checked;
the complete local and live bundles match. skill-manifest.json is an exact
copy of that manifest. skill-verification.json records its SHA-256 and the
matching bundle digest. No older remembered skill was substituted.