Written answers to the first review, September 29, 2026

Disposition of the self-review of the Gil note

Review: self-review/REPORT.md (fresh Claude Opus agent, 2026-09-29 ~13:45 EDT, independent of the authoring thread's audit files and of the OpenAI review directory). Applied by Claude (Fable 5.1) on 2026-09-29, 14:20 to 14:50 EDT, to main.tex, verify_gil.py and README.md. Theorems 3.1 and 4.1 were found proved; nothing in the mathematics changed. Item numbers are the report's. Remark numbers in "Now" refer to the rebuilt manuscript, in which the new Remark 4.5 (the counterexample) follows the old Remarks 4.3 and 4.4, whose numbers are unchanged.

MAJOR

1. Abstract's aligned-extremality claim false with prescribed Youla values. APPLIED. The sentence now reads "with free Youla values, every pairwise coherence functional with nonnegative weights is maximized in the aligned class" and stands before the prescribed-value sentence; the abstract ends by saying the product form of the weights is essential and that an n = 4 example shows a non-aligned maximum. The example (s = (1, 0.9), b_12 = 2, b_13 = b_24 = 1.2, M with entries 0.3, sqrt(0.84), -sqrt(0.84), 0.2) is Remark 4.5, with its singular values (1, 1, 0.9, 0.9), value 2.196, best aligned value 2.172 (matching {13, 24}), LP value 2.324, and a random search over O(4) finding nothing above 2.196; every number was recomputed independently before it went in. "What is not settled" now says the maximizer need not be aligned and the LP need not be tight for general weights, value open.
2. Tam 1998 claim made without the full text. APPLIED as the report specified: Remark 4.4 now says the skew-symmetric Schur-Horn theorem (Leite, Richa and Tomei 1999; Sanyal, Sottile and Sturmfels Prop. 3.10) and Tam's main theorem describe the entries on one fixed matching (weak majorization of the skew-diagonal moduli by the Youla values, with Tam's parity inequality) and do not imply (9), with the n = 3 comparison, and that no earlier statement of (9) was found; it states that Tam's paper was available only through its abstract. Leite, Richa and Tomei added to the bibliography with DOI. README's "before submission" item kept and sharpened.
3. Section 6 misreported check A. APPLIED. verify_gil.py check A now uses projected gradient ascent over the contraction ball (Euclidean projection = singular values clipped at 1, 20 starts x 4000 normalized steps, a final snap of the singular values to 1) and records the maximum AND the minimum over instances of (best - bound), the worst shortfall, and the counts of instances within 1e-6 and 1e-4; check B records its minimum too. The run of 2026-09-29 14:18 EDT: check A never exceeded the bound (5.6e-16), came within 1e-6 in 19 of 30 instances and within 1e-4 in all 30, worst shortfall 7.2e-5 at n = 7; check B reached its value within 1e-6 in all 30. Section 6 quotes those recorded values and no longer says the bound was attained where it was not.

MINOR

1. "cohesion" vs "squared cohesion": APPLIED throughout (abstract, Section 1, both theorems' proofs, Corollary 5.1(3), Section 6); Section 1 says Gil calls ||N||_F the cohesion and cites his eqs. (20) and (33) for the formula.
2. Aligned class per Gil's Definition 1: APPLIED ("in some admissible IRB representative").
3. Problem numbering and "affirmatively": APPLIED ("unnumbered there, and we number them (i) to (iii)"; "we determine the maximum asked for in (i) and answer (iii) affirmatively"; the reading of "suitable spectral constraints" as Youla-value constraints stated, with the tr rho^2 reason).
4. Eq. (8) and the hull claim with zero populations: APPLIED (Theorem 3.1 states (8) and the hull claim for positive populations, and on the support of A otherwise; the proof carries the a = (1/2, 1/2, 0, 0), b_12 = 1, b_34 = 10 example; Corollary 5.1(2) says "for positive populations").
5. Symbol clash M: APPLIED (matchings are mu, incidence vectors 1_mu; the reservation of M for the metaspin tensor stated).
6. Exchange argument: APPLIED (complete to a matching of maximal size; for odd n move the unpaired vertex to n; each uncrossing lowers sum |i - j| by 2(k - j) > 0, so the process stops at the adjacent pairing, the only perfect matching of {1, ..., 2m} without a crossing or nested pair).
7. Density-matrix reading of Theorem 4.1: APPLIED (s_1 <= 1 in the statement; the support condition, at most floor(r/2) positive s_k when r populations are positive; positive populations realize every orientation with s_1 <= 1).
8. "Theorem 3.1 is the special case s = 1": APPLIED (Theorem 4.1 recovers (7) through monotonicity of the right side in every s_k; it does not recover (8) or the hull statement, and Remark 4.5 shows it cannot).
9. Remark 4.3's units and inference: APPLIED with the script's own numbers rather than the referee's scratch runs. The remark quotes the excess of the C = B linear program in both units (the LP's own and ||N||_F^2), says that the degree constraints added to C = B still fail in the recorded instances, and states the prefix result (the n(n+1)/2 prefix inequalities alone reproduce the value); verify_gil.py check C now computes all four families (full, C = B, C = B plus degree, prefix) and verification-gil.json carries them: full and prefix agree with the theorem to 5.6e-17; C = B alone exceeds by up to 0.0097 in the LP's units (0.019 in ||N||_F^2); C = B plus degree still exceeds in 4 of 30 instances, by up to 2.2e-4 (4.4e-4 in ||N||_F^2). The random instances differ from the first run's because the ascent consumes the seeded stream differently, which is why the C = B excess is no longer the 0.0126 of the original JSON.
10. Corollary 5.1(3) equality condition: APPLIED (||M_{[p],[q]}||_F^2 = R(p, q) for every p <= q with a_p > a_{p+1} and a_q > a_{q+1}, a_{n+1} = 0; whether this forces alignment not examined).
11. "Since s is sorted decreasingly" misplaced: APPLIED (moved to the interlacing step, t_k <= s_k).
12. Script wording, SO(n) vs O(n), theorem numbering in the docstring: APPLIED (docstring uses 3.1 / 4.1 / 4.2 / Lemma 2.1; Section 6 says "floating-point checks" for the odd-set inequalities and that the SO(n) search covers O(n) up to a coordinate sign flip, which preserves squared entries).
13. Bibliography: APPLIED (DOIs and issue numbers for Youla, Fan, Sanyal-Sottile-Sturmfels, Tam, Horn-Johnson Topics; Horn and Johnson, Matrix Analysis, 2nd ed., cited for the real normal form under orthogonal congruence beside Youla; Leite-Richa-Tomei added).
14. Typography: APPLIED (the two exchange inequalities set in align*, no overfull box in the rebuilt log; Pl\"ucker).
15. Title: APPLIED ("... Answered by a Matching Polytope and a Layer-Cake Inequality").

CORRECT items C1 to C5

No change required. C5's byte-identical reproduction refers to the previous script; the script is now a new file with a new source_sha256, and its verification-gil.json and verify_gil.stdout.txt were regenerated by one run on the laptop (Python 3.12.10, numpy 2.2.6, scipy 1.16.3).

Not changed

• Authorship, the research attribution paragraph, the location of the note inside reviews/claude/ (the owner's call), and the date line.