{ "source_url": "https://github.com/twinheterodox/beta4/blob/main/beta-intervals/review/independent-audit-2026-11-17.txt", "report": "Independent AI audit. Publication copy with reproduction paths updated.\nThe verdict and calculations below concern the original audited version.\n\nAudit of Two shorter intervals of even Dirichlet beta values\nManuscript dated 17 November 2026\n\nVerdict\n\nI found no mathematical error or unproved step affecting Theorem 1. Each of the\ntwo lists, beta(6), beta(8), ..., beta(66) and beta(8), beta(10), ..., beta(98),\ncontains an irrational value. They have 31 and 46 entries. A single irrational\nvalue in their overlap would satisfy both assertions. The earlier powered-kernel\npaper gives endpoints 74 and 110 for those same starting indices.\n\nMathematical checks\n\n1. Formation of the linear forms. Counting multiplicities gives gamma = 341/16,\nG = 221/16 and degree -Gn-24. Reflection preserves R_n and pairs the order-i\nprincipal parts with sign (-1)^i. Because n is even, this eliminates every\nodd i in the aggregate beta coefficients. The prefactor removes one nominal\norder at the central pole. The derivative of order 2q at every required\ninteger zero vanishes, since that zero has multiplicity 3q. Differentiating\nand shifting the half-integer fractions gives the rising factorial, power\nof two and alternating sign in (2.6)-(2.8). The even exponents range from\n2q+2 through the largest even integer at most 33q.\n\n2. Integer coefficients. For a local series with coefficient valuations at\nleast v-lambda*h in degree h, the qth power has valuations at least\nqv-lambda*h. Applying the rule before extracting degree qB_a-i gives\n(3.5) and (3.6), also at the central pole. At small odd primes the harmonic\ndenominator costs at most (i+2q)*lambda, leaving q times the original bound.\nAt 2, all harmonic denominators are odd. For a repeated large-prime class\nthe bound i-qF_p(a) loses at most i+2q powers at a bad denominator. For an\nisolated class the change of variable ell=a-h rewrites the rational constant\nas (3.7). The zero identity at -ell lets one take the side that omits the\nsole bad address. This also handles primes larger than n. These calculations\ngive the very same D_n raised to q.\n\n3. Growth and decay. I checked the Legendre estimates for the small primes\nin the fifteen-value paper's Appendix A. After subtracting G*log(n!), the\nsmall odd primes contribute o(n) to the upper bound and the prime 2 costs at\nmost 15*n*log(2)+o(n). For large odd primes I checked the real envelope,\nincluding the restricted domain 0 <= y <= x-1 for 1 <= x < 2, the endpoint\nconvention, the prime-number-theorem integration, and the tail estimate.\nMy code generated the partition from affine-offset coincidences. It computed\neach maximum directly at every eligible y-breakpoint and just to its right.\nOn 38,371 open intervals, the sum rounded upwards cell by cell reproduced\n\n body = 73794151590486879261 / 9223372036854775808\n tail upper bound = 34253 / 319200\n total upper bound = 8.10808684435616...\n\nI checked the uniform Stirling estimate including t=rn+1, the fixed-order\nderivative bound, and summability on t>3n. The critical polynomial has\ndegree 33 and one sign change. Exact rational bisection and outward interval\nlogarithms put its maximum point at x = 0.757334986669258889... and the\nphase there at -18.835112366808193600.... For x >= 3 the elementary\nderivative estimate gives x*Phi'(x) <= -383/64. Proposition 3's coefficient\nper q is consequently at most -0.2327922916008203... using its stated\nconstants. Using the verified density and phase bounds directly gives the\nstronger upper bound -0.3298 per q.\n\n4. Prime dilation and the final inference. At n=96p the selected zeros are\nthe odd multiples of p up to 239p in absolute value, with the prefactor\naccounted for separately. The zero address class has total pole count\n2077 and the other classes have count 2046. Accounting for the prefactor\nand 720 selected zero factors gives excess 1356 in the selected class.\nThe two extra endpoints give the unit -1. At an unselected pole the\ncorresponding deficit is at most 1326. These counts give the local\nprincipal-part estimates of Lemma 4. After multiplying by p^(1358q)\nevery beta coefficient is divisible by p. The harmonic denominators cost\nat most i+2q at an ith principal part, leaving at least 30q powers of p in\neach unselected constant contribution. For a selected address the surviving\nterms have h=pv+(p-1)/2, giving the sign chi_4(p). I also tested the actual\nlarge kernel at n=96p for p=347 and p=349, checking the local counts, the\nscalar reduction at all unit residue classes away from the selected poles\nand zeros, and the normalized harmonic sums for all 97 selected addresses\nand exponents through 99. These finite checks supplement the general proof.\n\nIf all the beta values in the indicated list were rational, one fixed T\nwould clear them. For all sufficiently large primes p avoiding the\nnumerators and denominators in the nonvanishing argument, the scaled\nconstant is a p-adic unit while every other surviving term is divisible\nby p. This proves nonvanishing on n=96p. The real numbers T D_n^q L_{q,n}\nare then nonzero integers on that subsequence and tend to zero, giving\nthe contradiction for each q separately.\n\n5. Independent calculation of the finite constants. I constructed the global\nnumerator P and denominator Q of the trimmed seed over each finite field.\nI translated these global polynomials to each of the 97 pole centers with\nbinomial coefficients, divided the resulting series after removing the\nnominal pole, and computed the powered principal parts. I reconstructed\nthe full rational function at q*2077 distinct regular field points for\neach q and each modulus. This establishes the full identity in each case,\nsince multiplying its difference by Q^q gives a polynomial of degree\nstrictly less than q*2077. I also checked reflection and odd-coefficient\ncancellation and then evaluated the finite harmonic sums.\n\n q k_q mod 10007 k_q mod 10009 k_q mod 10037\n 1 3738 5333 938\n 2 8279 5481 5660\n 3 9978 6289 4500\n\nThe first two columns reproduce all six printed residues, covering 24,924\nnominal principal-part coefficients. The last column uses an additional\nprime. I wrote the verification programs separately from the supplied\ncheckers, and used the prime number theorem, Chebyshev's upper bound,\nStirling's formula, Legendre's formula, Wilson's theorem and Descartes'\nrule of signs as standard inputs. I checked the earlier powered-kernel\nresult for the comparison endpoints. The proof here needs only q=2 and 3.\n\nExposition\n\nOn pages 7-8, introduce 1358q as 1356q+2q at its definition. A selected\ncoefficient of pole order i then carries i+2q powers of p, exactly the\namount the harmonic sum can consume. An unselected contribution has a\nsurplus of at least (1356-1326)q=30q. This explains why the normalization\npreserves precisely the part of the constant the finite calculation detects.\n\nOn page 8, replace “Either nonzero residue in either of the last two rows”\nwith “Each of the last two rows has a nonzero entry, proving respectively\nthat k_2 and k_3 are nonzero.” One prime per row is sufficient. In the final\nproof, it would also help to recall that D_n^q clears the form globally,\nwhile the power of p tests nonvanishing on the selected subsequence.\n\nReproduction in this source ZIP\n\nFrom the extracted ZIP root, run\n python3 -B anc/checks/independent-audit/run_checks.py\nThe original audit used a C++ density evaluator. The deposit includes a Python/NumPy translation checked on every partition cell. The four checks require Python 3.9+, NumPy and mpmath.\n" }