Third review, by GPT-6 Astra (OpenAI), the other company's model, October 2, 2026

Gil note: fresh adversarial review

Reviewer: GPT-6 Astra (OpenAI), through Codex. Review completed October 2, 2026, 00:02 EDT (America/New_York). Baseline: 5202569, including the a19cd37 consistency pass. Locations refer to its main.tex and 9-page main.pdf. I read both, both previous reports and dispositions, and the consistency-pass table. I compared Gil's local full text with the stored 30-page publisher PDF, rather than assuming the two renderings had identical numbering. No manuscript or existing review was changed.

A. Statements against proofs.

The six numbered theorem/lemma/proposition/corollary statements are CORRECT. One displayed identity in the proof of the example's exact optimum is WRONG and has a local repair.

Statement and location	Verdict	Independent check
Theorem 1.1, main.tex:99, p. 2	CORRECT	It faithfully restates Gil's Theorem 2, including saturation only on positive-product pairs. Gil's PDF pp. 9-10 gives the value and the exchange argument. The standing definition of populations includes trace normalization.
Lemma 2.1, main.tex:142, p. 3	CORRECT	Row norms give degree inequalities. Every odd principal skew compression is singular and remains a contraction, including in the complex case, so its Frobenius norm supplies the odd-set inequality. Every matching vertex is realized by signed-swap blocks. Convex-hull equality follows, without asserting that the squared-entry image itself is convex. Empty matchings and odd dimensions cause no exception.
Theorem 3.1, main.tex:173, p. 4	CORRECT	Positivity on the active support is exactly ||M||_op<=1. Linear optimization over the hull has an actual matching witness, not merely a relaxed bound. Nonnegative weights permit completion of a partial matching; in odd dimension the unpaired vertex can be moved to the smallest population. Both exchanges decrease total edge length, ensuring termination at adjacent pairs. The explicit density blocks are positive semidefinite and trace one. Zero populations, odd support, repeated populations, zero weights, and rank-one populations are covered. The general-weight and hull assertions carry the necessary support restriction.
Theorem 4.1, main.tex:240, pp. 5-6	CORRECT, both proofs	In the layer-cake proof, 2 Psi=||M_(C,B)||_F^2, and adjacent-pair counting gives 2h=R(p,q). Empty prefixes contribute zero. The three parity cases exhaust 0<=p<=q, including the odd square compression's terminal zero. Ky Fan and singular-value compression give t_k<=s_k, in the real and complex cases. Equal populations merely remove level sets of positive measure. The block matrix attains the bound. In the second proof, Horn's inequality at each even index gives weak log-majorization of the paired squared singular values; exponential convexity gives the required sum bound. Approximation with positive populations and positive Youla values justifies the zero cases. No normalization of a or upper bound on s is needed for the algebraic inequality. The density-matrix consequence separately supplies trace-normalized populations, s_1<=1, and the support-rank condition.
Proposition 4.2, main.tex:360, p. 6	CORRECT	The block-norm argument works for arbitrary nested sets, not only prefixes. The singleton case is explicitly trivial; the degree bound requires |B|>=2. Taking C=B gives the stated spectral set bound. For odd |B| and s=1 it recovers Edmonds' odd-set bound; even sets give the corresponding redundant even-set bound.
Corollary 5.1, main.tex:458, pp. 7-8	CORRECT	Items (1) and (2) retain the active-support and positive-population restrictions. In item (3), every pair of strict population gaps gives a positive-area region of thresholds with constant nonnegative deficit. Equality of the integrated objective therefore forces each listed prefix inequality to be tight. The statement asserts necessity only and makes no unsupported uniqueness or alignment claim at ties.

Remark 4.3's prefix-LP observation follows from the layer-cake proof, while its coarse-LP comparisons are explicitly finite records. The generic Hermitian comparison in Remark 4.4 has the correct factor when compared in full squared-cohesion units. Its explanation should be read with the paired-singular-value argument, not as a theorem that arbitrary zero-diagonal matrices obey the same bound. The nonaligned witness, best aligned value, and LP certificate in Remark 4.5 are correct. The last, exact real-orbit certificate needs the following correction.

GL-1, MAJOR: missing factor of two and a Pfaffian sign qualification in the exact O(4) certificate. Location: main.tex:447, Remark 4.5, PDF p. 7. With the displayed half-sum definitions of X and Y, the correct identities are

(1/2)||M||_F^2 = 2(||X||^2 + ||Y||^2) = 1.81,
|Pf M| = abs(||X||^2 - ||Y||^2) = 0.9.

The manuscript omits the factor 2 in the first line and fixes the positive sign in the second while claiming the whole orthogonal orbit. Its printed equations would give squared norms 1.355 and 0.455, not 0.9025 and 0.0025. The latter pair is the correct one: their sum is 0.905, half of 1.81. An orientation-reversing orthogonal congruence can change the Pfaffian's sign, and exchanges the two norm roles. Fix: replace the two constraints by the lines above, or use (1/4)||M||_F^2=||X||^2+||Y||^2=0.905 and the absolute-Pfaffian identity. Keep the existing symmetry argument and the completion of the square. They then establish the claimed exact maximum 2.196 without any further change. This is a broken displayed step, not a counterexample to the maximum or either main theorem.

I checked the example independently in exact rational arithmetic where possible. The displayed matrix gives ||X||^2=361/400, ||Y||^2=1/400, and Pf M=9/10. The objective is 549/250=2.196; the best aligned objective is 543/250=2.172. The LP witness satisfies all 81 nested-set constraints with largest excess zero. The dual combination has right side 581/250=2.324 and adds nonnegative surplus 1.2 x_14+1.2 x_23+0.4 x_34 to the objective. These are dimensionless objectives. No random optimization rerun is claimed.

B. Abstract and introduction against the statements.

The abstract now correctly separates the free-spectrum and prescribed-spectrum questions. It restricts the arbitrary nonnegative-weight alignment assertion to free Youla values, uses modulus squares in the complex inequality, includes the odd-dimensional zero, and describes Horn's result as a classical implication. Its product-weight restriction is substantiated by the explicit counterexample independently of the damaged exact-optimum step in GL-1. The singular-value wording already implies that the s_k are nonnegative; spelling out s_m>=0 would be optional clarity, not a substantive omission.

The introduction uses intrinsic populations as eigenvalues of the real part in a fixed real structure, not as eigenvalues of the full density matrix. It defines the zero extension of M, so invisible inactive-axis variables do not enter either optimization. Positivity, the factor of two in squared cohesion, existential alignment under residual rotations, and the interpretation of the third question all agree with Gil's definitions. The note explicitly limits its reading of spectral constraints to the Youla data and leaves arbitrary fixed-density-spectrum feasibility unaddressed. Fixing the full spectrum and the populations indeed fixes tr rho^2-sum a_i^2, the objective.

The statement after Theorem 3.1 also retains free Youla values through its immediate context. The later support qualification in Corollary 5.1 is present in the rendered PDF. I found no remaining false abstract inequality or quantifier.

C. Sources.

The two long quotations and the short phrase about spectral constraints match the words in the stored publisher PDF after joining PDF line wraps. The first quotation changes its terminal sentence punctuation to fit the surrounding sentence; the substantive quoted text is unchanged. The second appears on p. 27, Section 8.2, and the three problems are indeed unnumbered there. They are different from the directions in Section 8.3. The first quotation has a version-dependent section locator:

GL-2, MINOR: the first quotation's section number does not resolve in the stored publisher PDF. Location: main.tex:105, PDF p. 3. The note cites Gil Section 3.4. The text in paper/antiunitary-matching/reviews/claude/03-literature/gil-2026-fulltext.txt does label the relevant section 3.4, but paper/antiunitary-matching/reviews/openai/lit-gil-2026.pdf labels it 3.5, with Theorem 2 on p. 9 and the quoted paragraph on p. 10. In that PDF, Section 3.4 is the scalar-complementarity recovery. Fix: cite “the paragraph following Theorem 2, p. 10 of the publisher PDF,” preferably with “Section 3.5” and an explicit PDF version/access date. If the authors intend the HTML numbering instead, label it as HTML. The quotation itself is accurate; the two stored renderings must not be treated as identical pagination or numbering.

Gil equation/source map, checked against that PDF: equation (3), p. 4, gives the IRB; equations (7) and (8), p. 5, give the support-defined metaspin and factorization; equation (9), p. 6, gives the paired normal form; Definition 1, p. 7, includes residual IRB freedom; equation (20), p. 8, gives the aligned norm formula; Theorem 2 and equations (23)-(26), p. 9, give the maximum and first exchange; equation (27), p. 10, gives the other exchange; equation (33), p. 11, gives the general norm expansion. Sections 2.1 and 2.3 resolve correctly. No factor from Gil's separately normalized correlation-asymmetry quantity has been imported into cohesion.

Bibliography entry	Verification and limits
[1] Gil	The local 30-page PDF, the local full text, and PubMed's article record confirm the title, author, August 4, 2026 publication date, volume 28, issue 8, article 877, and DOI. The current publisher endpoint was rate-limited, so the substantive check used the committed source copy. The relevant claims and pages are mapped above.
[2] Edmonds	NIST PDF, printed pp. 125-126, gives the nonnegative, degree, and odd-set constraints and the integral polyhedron result. Title, year, volume, page range, and DOI match.
[3] Youla	Cambridge PDF, printed p. 694, equation (2), and p. 701, Corollary 2, supports the complex unitary-congruence normal form. Metadata agrees with the bibliography. The note correctly supplies a different citation for the real orthogonal form.
[4] Fan; [5] Horn	Direct Crossref records for Fan's DOI and Horn's DOI confirm titles, years, volumes, issues, and pages. PNAS returned HTTP 403 for both PDFs; PMC returned a browser challenge, and the Europe PMC full-text API did not supply Horn's article. The previous Astra report records having read Horn's two original pages, but that is not a fresh retrieval by this review. I independently rederived the used product inequality from submultiplicativity of the exterior-power operator norm, and Ky Fan's bound from maximizing the trace over rank-p projections. These verify the mathematics, not the original-page attribution.
[6] Bhatia	Matrix Analysis, Springer, 1997, Chapter III, is the appropriate variational-principles reference. Full chapter retrieval was not completed; no original page is certified here.
[7] Horn-Johnson, Topics	The publisher DOI confirms the 1991 book. The exact Theorem 3.1.2/Corollary 3.1.3 locators were not verified from the book's full text in this pass. The compression inequality used is correct independently, by multiplying on both sides by coordinate contractions.
[8] Horn-Johnson, Matrix Analysis	The Cambridge edition record is consistent with the cited second edition. The exact Section 2.5 normal-form passage was not independently retrieved. This leaves a source-location check open, not an unresolved real skew normal-form fact.
[9] Leite-Richa-Tomei	Publisher abstract and metadata match the title, volume 286, 1999, pp. 149-173, and DOI. The PDF request returned HTTP 403. The real/complex Schur-Horn comparison is supported by the abstract and the explicit attribution in [11].
[10] Mathias	Publisher record and its Crossref record confirm volume 31 (1992), pp. 57-70, and the stated subject. The PDF request returned HTTP 403. Its unavailable formulas are not evidence for absence of the note's result.
[11] Sanyal-Sottile-Sturmfels	arXiv PDF, Section 3.2, Proposition 3.10, p. 14, gives the linear skew-diagonal projection and weak majorization, and credits Leite-Richa-Tomei. That description is accurate. The journal reference and DOI identify the published article.
[12] Tam	Publisher DOI and Crossref confirm title, volume 19, issue 3, 1998, pp. 737-754. The full PDF request returned HTTP 403. The abstract-level comparison is expressly limited in the manuscript; I do not certify absence of an equivalent result in the unread applications.
[13] Companion	Title, writer model, director, month, set affiliation, and hull lemma agree with the local antiunitary manuscript. The corrected attribution is appropriate.

All internal equation and statement references checked in the nine-page PDF resolve. Some exact historical source locations remain unverified, as listed; a table of contents alone was not treated as verification of a theorem number. The original proofs supplied in the note do not depend on an unverified substantive statement beyond the standard inequalities independently checked above.

The numerical claims in Section 6 agree with the committed JSON: 30 free-spectrum instances, 19 within 10^-6 and all 30 within 10^-4; worst shortfall 7.178675857100192 x 10^-5; prescribed-spectrum excess 1.500466417780899 x 10^-13; full/prefix LP error 5.551115123125783 x 10^-17; 4 coarse-plus-degree failures; and 8,760 odd-set checks with largest residual approximately -0.03216. The conversions between the LP objective and squared-cohesion units are correct. This is a record comparison, not a repeat of the searches.

D. Priority and positioning.

Fresh bounded receipt, October 2, 2026, 00:00-00:02 EDT. Queries included "Gil" "global" "Youla" coherence matching, "skew-symmetric" "nested" "squared entries", and the exact titles of Horn 1950, Tam 1998, Leite-Richa-Tomei 1999, and Mathias 1992 with pdf. Useful returns were Gil's author/publication record, the publisher abstracts for the three specialized matrix papers, the Orbitopes PDF, Youla's original PDF, and the historical product-inequality references. No returned primary source supplied an earlier statement of the exact population formula or the full nested-set squared-entry family. Sparse search returns do not establish novelty.

The decisive antecedent is already acknowledged: the prescribed-spectrum inequality follows directly from classical multiplicative singular-value inequalities. The paper now gives that derivation and does not market it as an independent new spectral inequality. The more modest claims about the application to Gil's questions, the matching-polytope viewpoint, and the nested-set family are appropriately sized for first-reader circulation. Full-text comparison with Tam, Leite-Richa-Tomei, and Mathias remains necessary before a stronger priority statement. The existing abstract-only disclosure correctly preserves that limitation.

E. The attribution block and the house rules.

The writer and vendor, David Ross's direction, and the origin of the questions while reviewing the companion are stated. Both prior reviews are named with their findings. The self-review's three framing defects and the cross-vendor review's two major items match their delivered reports; Horn is correctly called the second proof. The required human-review sentence is present. The note's place as the last of the three September 28/29 mathematics manuscripts and its independence from the unfolded-zeros paper are stated without claiming independence from the antiunitary paper it reuses. The companion bibliography entry now follows the house authorship pattern. No em dash or prohibited English spelling was found in manuscript prose. I inspected the rendered p. 7 and all extracted PDF text; GL-1 is present in the actual PDF.

F. Verdict.

• BLOCKING: 0 findings. Both main theorems, the hull lemma, and the corollary are correct.
• MAJOR: 1 finding. GL-1: repair the factor of two and Pfaffian sign in Remark 4.5's exact-orbit certificate.
• MINOR: 1 finding. GL-2: make the Gil quotation's locator resolve in the publisher PDF despite the HTML/PDF numbering difference.

First-reader decision: REVISE BEFORE SENDING. Correct the two constraints in Remark 4.5 and the source locator, then rebuild the PDF. The stated optimum and both main proofs remain unchanged. Original full-text access for several historical sources remains incomplete as disclosed in Section C; no stronger priority claim is certified.