Third review, by GPT-6 Astra (OpenAI), the other company's model, October 1, 2026

Unfolded zeros: fresh adversarial review

Reviewer: GPT-6 Astra (OpenAI), through Codex. Review date: October 1, 2026, 23:58 EDT (America/New_York). Reviewed the frozen main.tex and 16-page main.pdf at repository baseline 5202569, the previous reviews, their dispositions, and the October 1 consistency pass. Page numbers below refer to that PDF. The manuscript was not changed. This review checks the proofs independently; earlier numerical reruns are records, not computations performed again here.

A. Statements against proofs.

Statement and location	Verdict	Reason and cases checked
Theorem 2.1, Kadec	CORRECT	The strict supremum bound is <1/4, in the exp(i lambda t) normalization on an interval of length 2 pi. This is an imported theorem, not a proof supplied here.
Theorem 2.2, Avdonin	CORRECT	Separation, bounded displacement, one fixed partition length, and a strict centered average bound are all present. Translating the frequencies removes the constant c; shifting the integer indexing reconciles the partition convention with the cited formulation. A sliding-window condition is stronger, and the paper separately disproves it.
Theorem 2.3, Hruscev-Nikol'skii-Pavlov	CORRECT	All three conditions, including the strict bound on the small bounded term, agree with Kozma-Lev, Theorem 5.1, PDF p. 14. The manuscript uses only the necessity of BMO for its main obstruction and all three conditions for Remark 5.4.
Theorem 3.1, main.tex:101	CORRECT	At an ordinate with total jump m, the midpoint convention gives d_n=-S(gamma)+(m+1)/2-j. The correction vanishes for a simple ordinate. The proof counts all zeros at the same height, including different zeros sharing an ordinate.
Proposition 3.2, main.tex:115	CORRECT	The jump of S^2/2 is m S(gamma) and its derivative between ordinates is -theta' S/pi. These give the displayed signs and endpoint terms. The m displacement corrections sum to zero over a complete cluster. Endpoints are chosen away from ordinates.
Lemma 3.3, main.tex:130	CORRECT	The new cutoff passes every reflected positive point originating from a negative unfolded point. It covers arbitrarily negative a, not just the customary choices. Finite modifications produce a bounded correction that is eventually constant.
Theorem 4.1, main.tex:142	CORRECT	Positive and negative excursions give the two signs of unbounded block means. The last/next full partition block lies within 2N indices of the excursion; the adjacent sliding block lies within N indices. Monotonicity suffices even with repeated ordinates. Every fixed N and every fixed centering c are covered. The finite-height explanation following the theorem has a separate defect, UZ-1.
Corollary 4.2, main.tex:162	CORRECT	Integration by parts bounds the weighted integral by O(log^2 b), and the two square terms have the same bound. A partially cut cluster costs O(log^2 b) because both its multiplicity and its displacements are O(log b). Two cut endpoints only change the absolute constant. The sufficient inequality is correctly strict: M>4C log^2 b. The RH improvement uses the stated improvements for both S and its integral.
Theorem 5.1, main.tex:178, pp. 6-8	CORRECT	Repetitions surviving in the modified sequence preclude a Riesz basis immediately. Otherwise the counting argument applies beyond a finite prefix. An alleged basis on an interval of length 2 pi d yields d <= D^- <= 1 <= D^+ <= d, hence d=1; ordinary density alone is not being substituted for uniform density. The Jacobian is theta'/pi; its ratio on [T,2T] tends to one. Minimization subtracts the square of the mean. Littlewood bounds that mean, while Selberg makes the centered second moment diverge. The conservative coefficient 1/(8 pi^2) is valid. John-Nirenberg supplies the contradiction. All finite modifications, real a, and frequency/interval translations are covered.
Theorem 5.2, main.tex:212	CORRECT	Replacing Lambda by d Lambda converts density d to density one and the critical interval to length 2 pi. Mean-square oscillation is invariant under this dilation. The final density hypothesis supplies the same Landau argument for other intervals.
Corollary 5.3, main.tex:224	CORRECT	The explicit assumptions suffice: comparable Jacobians, S_F=O(log t) for ordinary density, a bounded dyadic mean from the integral estimate, and a divergent second moment. No assertion that an entire named family satisfies these assumptions is needed. The following paragraph now admits that those applications have not been verified.
Remark 5.4, main.tex:238	CORRECT	For |c|<=1, the change of variables is increasing and bi-Lipschitz. The corrected gap integral is O((1+|n|)^(-2)), so the centered sawtooth has a bounded primitive. Its Poisson extension tends uniformly to zero. The logarithmic term has a bounded harmonic conjugate; the remainder is bounded, Lipschitz, and in L^2, so its Hilbert transform is bounded by local cancellation and a tail Cauchy-Schwarz estimate. This supplies condition (iii), including its strict bound. The case c=0 is included.
Remark 5.5, sine process	CORRECT in its stated, cited scope	Completeness and excess one are imported almost-sure results. The paper correctly declines to turn ensemble number variance alone into an almost-sure BMO argument.
Theorem 6.1, main.tex:259, pp. 10-11	CORRECT	The counting-measure density and support density are now separated. The logarithmic discrepancy gives multiset exterior density one; support density is at most one. The upper long-family argument handles intervals crossing shells and those near the origin. Eventual simplicity of ordinates removes all but finitely many repetitions, including cross-half coincidences, giving the conditional equality. The proof does not establish unconditional equality for the support. UZ-2 corrects the accompanying account of what lower bounds are known.
Proposition 6.2, main.tex:294	CORRECT	Interlacing gives the monotonicities, and trace gives the equality characterization. The closest-pair eigenvalues are 1 +/- |sinc(s_N)|, implying the weaker signed-sinc bounds printed. The quadratic bound holds for every nonnegative gap, with sinc(0)=1. The N>=2 restriction is present.

UZ-1, MAJOR: the finite-height necessary bound uses excursions whose obstructing block may lie outside the finite sample. Location: main.tex:174, paragraph after Corollary 4.2, p. 6. The claimed lower bracket uses max_{t<=T}|S(t)| for a condition imposed only on windows below T. A negative excursion produces a following block, and that block may extend above T. At the lower end, a positive excursion need not have N preceding positive ordinates. The proof of Theorem 4.1 has no problem because it is an all-heights statement. It does not prove this finite-sample necessity. Smallest fix: delete the necessary lower bracket and retain the sufficient C log^2 T statement. Alternatively, restrict the maximum to positive excursions with a complete preceding window and negative excursions with a complete following window below T, and state the resulting one-sided inequalities with their endpoint convention. Do not claim the unrestricted maximum is necessary for the observed finite table.

UZ-2, MAJOR: the sentence denying an unconditional density estimate for distinct ordinates is too broad. Location: main.tex:267, immediately after Theorem 6.1, p. 10. Conrey, Theorem 1, printed p. 4 proves a positive proportion of simple zeros on the critical line, unconditionally. Each supplies a distinct ordinate. Thus an unconditional positive density lower bound is available, even though density one is not obtained this way. Smallest fix: replace the last clause by “we do not have an unconditional estimate strong enough to establish exterior Beurling-Malliavin density one for the distinct ordinates,” and cite the known positive-proportion results if discussing lower bounds.

For clarity, a conservative consequence of that source is R(Lambda'^*) >= (2/5) pi; this is my deduction, not a theorem stated in Conrey's paper. Let A(x) count distinct positive unfolded ordinates. For any fixed q<2/5, the simple critical-line zeros give A(x)>=qx eventually, while the full counting formula gives A(x)<=x+O(log x). Given d<q, choose C>1 with (qC-1)/(C-1)>d. The disjoint intervals [C^k,C^(k+1)) form a long family and eventually contain at least d times their length in distinct points. Taking suprema gives the stated lower bound, unchanged by finite modifications. Adding this observation is optional; correcting the denial of any estimate is necessary. This does not determine the exact unconditional radius or settle the critical interval.

B. Abstract and introduction against the statements.

The abstract correctly retains: simple ordinate for the scalar identity; midpoint values of S; every centering and fixed block length; the finite-height qualification on growing blocks; every real unfolding constant and every finite modification in the non-Riesz result; the bounded-mean hypothesis in the generalization; strict b>pi for incompleteness and strict b<pi for conditional completeness; eventual simplicity for that latter assertion; and separation of the positive half. Neither endpoint b=pi is claimed settled. The introduction correctly distinguishes varying a from translating the entire sequence. The mean-square coefficient is correct. The core abstract is consistent with the numbered results.

UZ-3, MINOR: “invisible at computable heights” is broader than both the theorem and the data. Location: main.tex:39, p. 2. Failure for small block lengths is already visible in the paper's table; Section 7 even reports the first Kadec crossing at the ninth ordinate. The O(log^2 T) sufficient bound supplies no general lower limit on the first height of failure for a fixed larger block. Fix: say that selected larger block lengths pass the reported finite table despite failing eventually, with the particular length and table size if desired. Delete the universal computability claim.

UZ-4, MINOR: smoothness alone is not a sufficient Riesz-basis criterion. Location: main.tex:41, p. 2, first sentence of the final introduction paragraph. The smooth-displacement wording advertises a general principle, whereas Remark 5.4 proves a specific logarithmic example with a coefficient restriction. Smooth drifts can change density or destroy separation. Fix: “A Riesz basis of exponentials can tolerate certain unbounded displacements; for example, ...”, preserving |c|<=1. In the last sentence of that paragraph, replace “this variance, and nothing else about the zeros” by “this variance growth, together with the counting formula and bounded mean”; that is the actual proof dependency.

UZ-5, MINOR: restore “tested” in the earlier minimum-block claim. Location: main.tex:174, p. 6, “the smallest block length”. Section 7, p. 13, correctly says “among the block lengths tested” and lists a scan that omits some integer lengths. Fix: “the smallest tested block length”. The values D(6)=0.330 and D(8)=0.216 establish the reported scan outcome, not an exhaustive minimum over all lengths.

The remaining summary comparisons are sound. “Avdonin's condition” in finite-height summaries should continue to mean its averaged inequality alone: the paper correctly explains that these observations do not establish the global hypotheses of Avdonin's theorem.

C. Sources.

There are no attributed prose quotations requiring a verbatim transcription in this manuscript. Quotation marks around short labels such as the mean condition and the earlier centering convention are descriptive. Mathematical statements, constants, and cited locations were checked as follows. Source-access limits are explicit; this is not a certification that every original historical article was retrieved.

Bibliography entries	Source check and outcome
[1] Alemany-Nitzan; [2] Avdonin; [14] Kadec	Alemany-Nitzan PDF, introduction pp. 1-2 and bibliography, confirms the modern formulations and historical references. The original Russian articles and English translations were not independently retrieved. No discrepancy found in the manuscript's conditions or bibliographic data.
[10] HNP; [16] Kozma-Lev; [21] Pavlov	Kozma-Lev PDF, Theorem 5.1, p. 14, explicitly identifies HNP p. 240 and supplies the formulation used. Its bibliography supports the historical references. The original HNP chapter and Pavlov article were not read in full. The manuscript correctly states the attribution chain.
[3] Beurling-Malliavin; [15] Koosis; [22] Seip-Ulanovskii	BM metadata agrees with DOI 10.1007/BF02392477. Koosis's year is confirmed by Cambridge's front matter. The author-uploaded Seip-Ulanovskii text, p. 1745 explicitly uses exponential polynomials for multiplicities and gives the stated journal, volume, year, and pages. The AMS PDF endpoint returned HTML in this review. The counting/support distinction is now correct.
[4] Bober-Hiary	PDF, Table 2, PDF p. 7: 3.3455 occurs just after the indicated zero near 7.7573 x 10^27. The manuscript treats it as the published value known to the authors, not a proved present-day record. The final journal citation agrees with the source record.
[5] Bufetov; [8] Ghosh	Bufetov PDF, Theorem 1.1, and Ghosh's author abstract support excess one and critical completeness, respectively. The former is correctly cited as a 2019 preprint; Ghosh's journal reference is consistent with its publication record. These sources do not prove the corresponding assertions for zeta ordinates.
[6] Burnol; [20] Mikulik	Burnol PDF, printed pp. 65-67, concerns complete/minimal systems in Sonine spaces. Its journal metadata is correct. Mikulik's April 2026 record supports the Gaussian-Gram comparison; a complete downloadable manuscript was not obtained. See UZ-6.
[7] Dyson-Mehta	Title, year, volume 4, and pp. 701-712 agree with publisher DOI 10.1063/1.1704008. The density-one sine-kernel variance coefficient is also obtained directly by integrating its covariance kernel. Original full-text verification remains incomplete.
[9] Goldston-Gonek	Publisher record confirms volume 39 (2007), pp. 482-486. No quoted page or equation is attributed to this item.
[11] Hypnos methods manuscript	The title, August date, author-model attribution, and set affiliation agree with the local methods manuscript and house chronology.
[12] Inoue	PDF, Theorem 1.1, pp. 1-2: the conditional bound 0.50895 implies the manuscript's weaker 0.509. RH is retained. This supplies a small-gap bound, not gaps tending to zero.
[13] John-Nirenberg	Publisher record confirms 1961, volume 14, pp. 415-426. The L^2 mean-oscillation consequence is the correct consequence of exponential integrability. Original full-text access was not completed.
[17] Landau	The normalization and inequality directions are checked through Kozma-Lev, introduction. The original Acta article was not independently read; its bibliographic data agree with that reference chain.
[18] Littlewood; [26] Titchmarsh	Titchmarsh, second edition PDF: Theorem 9.4, printed p. 214, proves S=O(log T); Theorem 9.9(A), printed p. 222, proves S_1=O(log T); Theorem 14.13, printed p. 350, gives the RH improvements. The title page confirms the 1986 revised edition. Littlewood's original 1924 article was not fetched in full.
[19] Lyubarskii-Seip	Author preprint matches the Paley-Wiener/Muckenhoupt topic. The manuscript now treats it only as an uncarried-out alternative route. The final journal pagination was not independently verified in this pass.
[23] Selberg 1946, zeta	Goldston's exposition, Theorem 8 and equation (10.6), pp. 21-22, records the unconditional moment formula and its error term. Setting k=1 and subtracting the estimates at 2T and T gives exactly (S1). This checks the theorem through an identified secondary formulation; the original 1946 full text was not retrieved.
[24] Selberg 1946, Dirichlet; [25] Selberg 1992	Neither original was retrieved in a form permitting verification of a fixed-function height-aspect theorem here. The manuscript explicitly states that limitation and does not use those entries to assert a verified family-wide corollary. Their full bibliographic verification remains incomplete; no correction is asserted without evidence.
[27] Tsang	Scanned PDF, inspected visually: printed p. 369, equation (1.2), supplies the Selberg two-sided Omega bound with exponents 1/3 and -7/3; p. 370, Theorem 3, supplies both endpoints of the range for h. The journal header and publisher record agree with the citation.

All numbered internal references checked in the PDF resolve. I inspected the rendered table/figure page as well as extracting all 16 pages; the formulas and the corrected numerical qualifiers are present in the PDF, not merely in the TeX.

UZ-6, MINOR: the negative content claim exceeds the stated access to Mikulik. Location: main.tex:356, Section 8, p. 14. The paper says it has seen only the abstract and then says neither adjacent work contains Theorem 5.1. The first fact does not license the second conclusion about the whole unavailable manuscript. Fix: “Neither accessible account states Theorem 5.1; we have not checked Mikulik's full manuscript.” The Gaussian/local-cluster distinction is supported by the accessible abstract and should remain.

D. Priority and positioning.

Fresh bounded search receipt, October 1, 2026, approximately 23:54-23:58 EDT. Queries included "zeta zeros" "Riesz basis", "unfolded" "zeros" "BMO", "zeta" "Avdonin" "Selberg", "zeta zeros" "complete interpolating", the exact Mikulik title, and "simple zeros" "Conrey" "two fifths". The exact subject combinations produced mostly unrelated Riesz-function, spectral-model, and general zeta pages. The useful adjacent returns were Mikulik's author record, Burnol's journal article, the existing BMO/exponential-basis literature, and Conrey's positive-proportion theorem. Search results announcing newer zero-proportion work were not needed for UZ-2 and are not treated here as verified theorems.

I found no source stating or proving the exact unconditional obstruction for the theta-unfolded multiset, every real a, and finite modifications. This is a bounded search outcome, not a proof of priority. The substantive novelty wording is appropriately narrow once UZ-6 is fixed. The classical character of the ingredients is disclosed. The completeness discussion needs UZ-2 so that an open exact radius is not confused with the absence of any lower estimate.

E. The attribution block and the house rules.

The block immediately after the abstract names Claude Fable 5.1/Anthropic as writer, David Ross's direction, the harness origin, the Opus 5.5 self-review with no blocking error and five major items, and the GPT-6 Astra/OpenAI cross-vendor review with one blocking error and five major items. Those counts match the original reports, and the changed completeness statement is identified. It includes the required human-review sentence and gives the manuscript's chronological place among the three September 28/29 papers and its independence. The main title and page design follow the house format. No new claim of human mathematical verification is made. I found no US-English or em-dash defect in the manuscript text. The overbroad prose claims are the specific items above, not defects in the attribution block.

F. Verdict.

• BLOCKING: 0 findings. No invalid numbered theorem or unresolved proof gap found in the revised manuscript.
• MAJOR: 2 findings. UZ-1: repair the finite-height necessary bound. UZ-2: correct the denial of unconditional distinct-ordinate density estimates.
• MINOR: 4 findings. UZ-3: narrow the computable-height claim. UZ-4: narrow the smooth-drift and proof-dependency wording. UZ-5: say smallest tested length. UZ-6: limit the Mikulik content claim to the text actually accessed.

First-reader decision: REVISE BEFORE SENDING. The smallest required set is the two local prose corrections UZ-1 and UZ-2; applying the four minor wording fixes in the same pass would make the account consistent throughout. The main non-Riesz-basis proof needs no change. Historical full-text and some bibliographic checks remain incomplete as identified in Section C; this review does not certify those unavailable originals or establish priority.