back to table · edit · history · where entries came from · files · download
4804 bytes, as of the version from 2026-11-18 14:47 (current). Recorded here, not run.
# T752 numerical evidence
Quantity: Z(s)=sum alpha^(-s) for Re(s)>1, with alpha=(rho-1/2)/i over every upper-half-plane nontrivial zeta zero with multiplicity and principal logarithm; meromorphic continuation supplies the nonpositive even entries. No global RH or simplicity assumption. Both positive parities and all eleven exact continuation entries are present.
Generator: generate.py plus the unmodified official FLINT sources flint-secondary_zeta.c and secondary_zeta.h, and the separate secondary_bridge.c adapter. Upstream commit and file hashes: upstream-provenance.json. FLINT 3.4.0 supplies underlying Arb operations. Actual dependency versions: dependency-versions.json. No production service computes values.
All 60 values are checkpointed in computed-values.json and are byte-identical, in the same parameter order, to stored-values.json after API normalization. The 49 positive values are finite, nonzero-width Arb enclosures serialized at 100 significant decimal digits. Native relative accuracy is at least 358 bits; working precision was 461 or 589 bits. The eleven nonpositive even values are exact Sage rationals. Native types, individual enclosures, precision and timing are recorded per row. FLINT's explicit tail bounds support proven rigour; mpmath's heuristic error estimates are not used as certificates.
Existing deterministic checks: agents/table-build/dry_run.py and check.py, invoked through agents/sage.sh by run_dry.py. Every type, import, finite-enclosure, suspicious-decimal-exactness and prose check passed. Sixty entries, 9.8 KB block in this check, longest native representation 182 characters. Exact output is not inferred from decimal patterns.
Independent mathematical checks:
- independent_check.py and independent_xi.c: all 25 positive even moments, using a fresh Arb power-series expansion of xi(1/2+z) from its gamma/zeta/product definition and a formal logarithm. Computed at 1600 bits; the enclosures are contained in the stored decimal intervals. This shares Arb arithmetic with the generator but uses no secondary-zeta routine, zero list or decomposition.
- Every exact row: Euler numbers reconstructed from cosh(x)*sech(x)=1, independently of Sage's Euler-number routine; all eleven exact rational equalities pass.
- All 49 positive orders: separate mpmath 1.4.1 secondzeta implementation at 170 decimal digits, splitting parameter a=1/200. This shares the Arias decomposition theorem, so its agreement is disclosed as an implementation comparison.
- odd_resolvent_check.py: all 24 positive odd moments from the regularized resolvent integral of the central xi logarithmic derivative, with fresh Arb even moments through order 200 at 2400 bits and 230-decimal-digit quadrature. It uses neither secondary-zeta evaluation nor the decomposition. Taylor depths 90 and 100 agree well within the stored intervals. Results and limitations: odd-resolvent-results.json. This is numerical agreement, not a quadrature or Taylor-tail certificate. The source identity is derived by subtracting the first m terms of the Hadamard logarithmic derivative, then splitting at t=1 and using u=1/t on the upper interval. No finite list establishes a universal assertion or RH.
Controls: Z(0)=7/8, Z(-2)=-9/32 and Z(-4)=3/128; the exact mock divisor 1+i/4,1-i/4 has inverse-square sum 480/289, rejecting both ordinate and magnitude substitution. Wrong sign, factor two, heat scale, a T36 value, NaN and vacuous enclosures are rejected. The resolvent toy divisor alpha=2 gives Z(3)=1/8 and rejects a factor-two error. A failed large-t subtraction sign in the developing checker was diagnosed by this toy control; failed logs are preserved, and generator values were not changed to match it.
Read-back checks: stored_identity_check.py reruns 109 comparisons on all 60 stored identities, reusing expensive independent computations only after confirming identical entry strings. It compares complete certified enclosures for the even rows and exact rationals for the continuation branch; approximate comparisons remain labelled as such. It also reruns known-answer and negative controls. stored-identity-results.json records no failures. verify_stored.py separately reports 60/60 generator matches, no differences, missing or extra entries; this is reproducibility, not independent correctness.
Skill evidence: skill-manifest.json is copied from agents/runs/20261118T130349Z-build-skills/manifest.json. skill-verification.json records its exact SHA-256 and successful verification of all archived resource hashes and matching local/live bundles. Argument selection is justified separately in argument-selection.md. This build report is not an independent scientific verdict; the controller must commission a fresh critic and argument-selection review before offering T752.