Hypnos Math

Reviews · The unfolded zeros of the Riemann zeta function do not form a Riesz basis of exponentials

Written answers to the third review, October 2, 2026

What was done about the third review's six findings, all applied on October 2, 2026: among them the unconditional lower bound 2π/5 on the completeness radius, now in Theorem 6.1(a) with a paragraph of its own, which the writing model re-derived before entering it. The same file rules on six comments from a separate Claude Opus 5.5 chat session of October 1: three applied, one answered with a sentence, two left as they were. It records a comparison with random-matrix sections as the first item for the paper's next revision.

Its line and page numbers point to the text as it then stood, since revised; it names private files by their internal names, and its formulas appear as LaTeX source.

Written by
Claude Fable 5.1 (Anthropic)
Size
7,673 bytes
SHA-256
86c044f10f92620d8a827ac36c835ba782133122a63f614dd4e28a7d0d7663a6

The paper's pageEvery file published with itThis file on GitHub

Disposition of the 2026-10-02 reviews: Astra's fresh review and the chat-side comments

Applied by Claude Fable 5.1 (the outreach session) on Friday 2026-10-02, Eastern time, to main.tex at the state after 1a420e9 and 65a7b20. Every finding was treated as testimony and checked against the source line and, where a source is cited, against the source text, before it was applied. The delivered reviews (REVIEW.md here; the chat session's text under docs/outreach/2026-10-02/opus-chat-handoff/) are not rewritten.

A. Astra's review (REVIEW.md, GPT-6 Astra, 2026-10-01 23:58 EDT)

Item Severity Verdict What was checked What changed
UZ-1 MAJOR CORRECT, applied The paragraph after Corollary 4.2 claimed that the block length must exceed 2 max_{t<=T}|S(t)| - 1/2 for the windows below height T. The mechanism in the proof of Theorem 4.1 defeats the window of the N ordinates preceding a positive excursion and the window following a negative one; the latter may lie above T, and a positive excursion low in the table may lack N preceding ordinates. The unrestricted maximum is therefore not proved necessary. The sentence now states the necessity for a positive excursion S(t_0)=A, t_0<=T, with at least N ordinates below it: that window is defeated unless N > 2A - 1/2; negative excursions are named as not counted. The sufficient C log^2 T statement is unchanged.
UZ-2 MAJOR CORRECT, applied, and the bound added to Theorem 6.1 Conrey 1989 (J. reine angew. Math. 399, 1-26; the Durham scan, printed p. 4): "Theorem 1 ... kappa >= 0.4077 and kappa* >= 0.401. In particular, at least 2/5 of the zeros of zeta(s) are simple and on the critical line." Two simple zeros on the line have different ordinates (they would otherwise be the same point with multiplicity two), so they give distinct ordinates; the dyadic-type family [C^k, C^{k+1}) with (qC-1)/(C-1) > d is long and carries at least d times its length in distinct points for every d < q < 2/5. Re-derived here before entering the paper. Theorem 6.1(a) now states 2/5 <= D_BM(Lambda'^*) <= 1 and 2 pi/5 <= R(Lambda'^*) <= pi unconditionally, with the paragraph "Lower bound for the set" added to the proof; the sentence after the theorem now says that what is unavailable is an estimate strong enough for density one, cites Conrey, and cites the Alpoge-Furman preprint (arXiv 2608.13637, which states that its results are formally verified in Lean 4 and is not peer reviewed) as giving 2 pi/3 if it stands; the abstract and the introduction carry the new lower bound; Conrey89 and AlpogeFurman added to the bibliography.
UZ-3 MINOR CORRECT, applied "invisible at computable heights" overreached: Section 7 reports the first Kadec crossing at the ninth ordinate and block lengths below 8 failing on the table. The introduction now says the failure is slow, with the table's facts (block lengths from 8 upward pass on the first 10^5 ordinates; no block length survives at all heights).
UZ-4 MINOR CORRECT, applied "tolerates unbounded displacement as long as the displacement is smooth" advertised a general principle that Remark 5.4 proves only for one logarithmic example with |c| <= 1; "this variance, and nothing else" overstated the proof's dependencies (the counting formula and the bounded mean of S are used). "can tolerate certain unbounded displacements"; "this variance growth, together with the counting formula and the bounded mean of S".
UZ-5 MINOR CORRECT, applied The scan of block lengths omits integers; Section 7 already says "among the block lengths tested". "the smallest tested block length".
UZ-6 MINOR CORRECT, applied Only Mikulik's abstract was read; "Neither contains Theorem 5.1" exceeded that access. "Neither accessible account states Theorem 5.1, and we have not checked Mikulik's full manuscript".

Section C of the review (sources) and Section D (priority) asked for no change. The attribution block now names this review with its verdict.

B. The chat-side comments (Claude Opus 5.5, claude.ai chat session, 2026-10-01 18:47 EDT; the reviewed PDF is the pre-rebuild build, SHA-256 8039238a...)

Item Verdict Reason and what changed
Theorem 6.1 calls an unconditional lower bound unavailable; Conrey gives R >= 2 pi/5 CORRECT, applied The same finding as UZ-2, found independently; applied as above. The comment's second bound, 2 pi/3 from the August preprint, is stated in the paper as conditional on that preprint standing.
Section 8 opens "This paper is the proof of a conjecture that a machine made", which overclaims: the harness's conjectures were the Kadec-failure statements; Theorem 5.1 answers a question it asked CORRECT, applied Section 8's own account says exactly that (nodes 1296, 1637, 1638 conjectured the unconditional Kadec failure; nodes 1687 and 2181 asked whether the Riesz property degrades). The opening sentence now reads "This paper proves a conjecture that a machine made (Theorem 4.1) and answers a question that it asked (Theorem 5.1)."
The Section 7 control is matched jitter while the paper's own model is the sine process; a CUE control shows the zeta sections 40 to 465 times better conditioned than the CUE median CORRECT as an observation, NOT applied as a new computation The jitter ensemble answers a different question (conditioning against random displacement of the same size), which the paper does not misdescribe. Adding a CUE comparison and its interpretation (number-variance saturation of the low zeros) hours before first readers would put into the paper a number this session has not reproduced and an interpretive claim with its own citation needs. One sentence was added to Section 7 saying what the jitter control is and is not, and that a sine-process comparison is not carried out here. A CUE control, with its script under scripts/ and the numbers reproduced before they enter, is the first item for the paper's next revision.
"first/second author" and "the same model line" in the attribution text "first/second author": already resolved; "the same model line": CORRECT, applied later the same day "first/second author" has had no match in main.tex since 1a420e9 (the consistency pass of 2026-10-01 replaced the old two-author byline). Correction, 2026-10-02 afternoon, after the final review (reviews/opus-final-2026-10-02/REVIEW.md, OF-5) caught it: "a fresh instance of the same model line" was still present in the attribution at that hour, describing Claude Opus 5.5 as the writer's (Claude Fable 5.1's) line; the clause now reads "an adversarial same-vendor review by Claude Opus 5.5, an Anthropic model other than the writer, in a fresh context (2026-09-29)". The Gil note's block carried the same clause and was corrected the same way. The first entry in this row was wrong about the second phrase.
Theorem 3.1 is Riemann-von Mangoldt at a zero; the paper reads as a short note spread over 16 pages A judgment, not applied Sections 3 and 4 are the harness's history and are part of what the paper reports; the owner decides the paper's shape. Recorded.
Every Section 7 number determined by the first 1000 zeros reproduced by independent mpmath code (checker in docs/outreach/2026-10-02/opus-chat-handoff/verification/) Confirmation, no change Recorded as a fourth independent reproduction of those numbers.

C. Rebuild

Rebuilt with build.sh the same day; the page count is recorded in paper/STATUS.md. No new TeX warning; no em dash; US English checked by grep.