back to table · edit · history · where entries came from · files · download
5736 bytes, as of the version from 2026-11-17 06:17. Recorded here, not run.
# T733 build evidence (not an independent scientific verdict)
Skill manifest: ../20261117T060056Z-build-skills/manifest.json. Both archived bundles validated; SHA-256 984965681e2f8925c511a121c118efcde59581c7b0d906cac8b641bf6925e55a. A copy is attached as skill-manifest.json. Inspected source downloads and hashes are in sources/manifest.json. Source passages and normalization decisions are summarized in research.md.
Runtime: Sage 10.9, FLINT 3.4.0, numberdb 0.1.10, Python 3.12.3; exact output in versions.json. Online FLINT docs are 3.7.0-dev, not the installed version. All Sage work used agents/sage.sh in the configured builder. Working precision is 512 bits (about 154 decimal digits); the production route's worst relative-radius bound retains at least 135.717 significant digits at alpha=1. Only 100 significant digits are serialized. Conversion uses numberdb._write.to_text on the computed RealBall, without invented precision.
All 46 default values were computed from the unconditional explicit formula, with complete prime-power enumeration and rigorous archimedean quadrature. The removable integrand at x=0 is evaluated through 1F1(1;2;u). Both finite input errors and cancellation errors are included. All default cutoffs have L=12; M is at least 100000 and is increased by the stated digit/width formula. The maximum omitted-prime bound is below 3.0e-182. The archimedean tail uses a Gaussian majorant after subtracting atanh(exp(-L)). The bounds are derived in generate.py; raw enclosures and components are in components.json. Prefix checkpoints are preserved in build-compute-2.log, one flushed row per width.
The existing dry_run.py/check.py checks passed on all 46 production values: native RealBall type, finite non-vacuous enclosures, no suspicious exact decimals, imports, prose and size. The entry block was approximately 9.2 KB, with no oversized values. Original failed probes/logs are retained; they reflect interface issues resolved before uploading, not hidden mathematical discrepancies.
check_values.py imports no production code. It recomputes the full heat sum independently from 150 consecutively indexed certified zeta zeros, retaining the 138 below T=300; zero 139 is above 300. The library's rigorous isolation and index/count guarantee supplies completeness. Above T, the critical-strip inequality |exp(t*(rho-1/2)^2)|<=exp(t/4-t*Im(rho)^2) and HSW's zero-count estimate give an unconditional tail. For u>=300, log(u/(2*pi*e))<=log u<=u and log log u<=u, so U(u)<=u^2/(2*pi)+0.3611*u+9.3675<=u^2. Stieltjes integration, dropping the negative boundary term, gives tail <=2*exp(t/4-t*T^2)*(T^2+1/t). This includes possible off-line zeros and multiplicities above T. The worst zero-route relative radius is below 3e-152. No global RH or unproved zero simplicity is assumed.
All 46 independent enclosures were contained in the complete intervals represented by the planned decimals. check-results.json records every value, tail and comparison. The Gaussian transform (including both poles) and nonzero mock two-point and complex quartet controls passed. The quartet is negative at t=2*pi and rejects replacing complex frequencies by real parts or magnitudes. Deliberately missing factors of two, an omitted p^2=4 term, the wrong archimedean sign, NaN and vacuous intervals were rejected. Endpoint/midpoint calculations at 640 bits agree with 512-bit enclosures (precision-summary.json). The routes share Arb/FLINT arithmetic and the explicit-formula theorem, but neither zeros nor prime/quadrature components are reused across the two routes.
Stored-value checks and reproducibility were repeated after upload, with authenticated readback: all 46 passed again, and verify(sample=None) matched 46/46 (stored-check-results.json and verify-results.json). verify(sample=None) checks reproduction, while check_values.py checks the separate zero formula. Neither is an independent critic's verdict. Only the controller may advance this draft after a fresh independent scientific and argument-selection review.
## Bounded proof-facing remainder experiment
Connes's exact kernel decomposition yields, for t=alpha/(4*pi),
R(t)=2*integral_0^infinity F_t(e^u)*(1/4-r(u))du-P(alpha),
where F_t(e^u)=exp(-u^2/(4*t))/(2*sqrt(pi*t)) and
r(u)=exp(u/2)/(2*sinh u)-1/(2*u), continuously r(0)=1/4.
The incomplete-gamma terms in Connes (17) cancel exactly, rather than merely being asymptotically negligible. This is a deduction from the inspected formula.
The archimedean part B(t)=R(t)+P(alpha) is positive. For u>0, r(u)<1/4 is equivalent to 2*u*exp(u/2)<(u+2)*sinh(u). In the power series the u and u^2 coefficients agree; every subsequent coefficient difference is positive: for odd u^(2k+1), k>=1, compare 2/(2k+1)! with 2/(2^(2k)*(2k)!); for even u^(2k+2), k>=1, compare 1/(2k+1)! with 1/(2^(2k)*(2k+1)!). The resulting strict positivity has an elementary proof for the kernel alone. It gives R(t)>-P(alpha), not a positive full arithmetic remainder or RH.
The tempting uniform sign R>=0, suggested by the positive archimedean remainder alone, fails already at the training width alpha=1: R is about -0.2360669544. The opposite sign fails at alpha=1/8: R is about +0.0029439281. The exact experiment scope is 22 widths of denominator<=8 (4 positive, 18 negative), followed by all 24 withheld widths of denominators 9-12 (6 positive, 18 negative). No zero or unresolved sign appeared. All B values were rigorously positive. remainder-results.json preserves all intervals; summary_check.py reproduces them. Both uniform-sign candidates are refuted; no fitted inequality, all-width arithmetic theorem, or proof of RH is asserted. Denominator 13 and alpha=1/24 were not added or tested in this experiment.