Written answers to the second review, September 29, 2026

Disposition of Astra's referee review of the Gil note

Review: reviews/astra/REVIEW.md (OpenAI Codex, gpt-6-astra, launched by Claude through the Codex CLI on 2026-09-29 14:25 EDT with the prompt in reviews/astra/PROMPT.md; committed at 24c79d0, 14:37 EDT; the CLI's own final message is LAST-MESSAGE.md, identical to the review up to the trailing newline). Reviewed manuscript: the f37590d text (sha256 42673800...). Applied by Claude (Fable 5.1) on 2026-09-29, 14:42 to 14:50 EDT. Every item was checked before it was applied; the one mathematical addition (Horn's derivation) was verified line by line.

1. MAJOR: Theorem 4.1 is an immediate consequence of Horn's 1950 multiplicative inequality

APPLIED. The derivation was checked independently: with D = diag(sqrt a) and T = DMD skew, its singular values are t_1, t_1, ..., t_m, t_m (+0 for odd n); Horn's product inequality applied twice at each even index 2l gives prod_{k<=l} t_k^2 <= prod_{k<=l} s_k^2 a_{2k-1} a_{2k}; both vectors are nonincreasing, so this is weak log-majorization, which implies weak majorization (exp is increasing and convex), hence sum t_k^2 <= sum s_k^2 a_{2k-1} a_{2k}, and sum_{i<j} a_i a_j |M_ij|^2 = (1/2)||T||_F^2 = sum t_k^2. It is now the "Second proof, from a classical singular-value inequality" right after the layer-cake proof, Horn 1950 is cited (PNAS 36(7) 374-375), and the text after the theorem, the abstract, Remark 4.4 and the README say plainly that as a matrix inequality Theorem 4.1 is a sharp consequence of classical majorization, and that what the note claims is the application to Gil's questions, the polyhedral viewpoint of Lemma 2.1, and the nested-set family of Proposition 4.2. The title is unchanged (the layer-cake proof stands and yields the family). The earlier assessment that "Theorem G" of the cross-review was new is corrected in Claude's memory as well.

2. MAJOR: Corollary 5.1(1) false without the support restriction

APPLIED. Item (1) now says: with free Youla values, the maximizing aligned state's support is a maximum-weight matching for b on the support {i : a_i > 0}, every inactive entry being zero (so for positive populations the matching is on all of [n]). Item (3) reads s_1 <= 1 on the support.

3. MINOR: trace normalization

APPLIED. sum a_i = 1 is imposed in eq. (1), with Gil's definition of a density matrix quoted, and in the hypothesis of Theorem 3.1; the matrix inequality of Theorem 4.1 keeps unrestricted nonnegative a.

4. MINOR: the Tam / Leite-Richa-Tomei comparison

APPLIED. "Do not imply" became "do not directly state"; the n = 3 comparison (which ignored the fixed Frobenius norm) is deleted; Sanyal-Sottile-Sturmfels Prop. 3.10 is described as the linear skew-diagonal projection of an orbitope; Mathias 1992 (LMA 31, 57-70) is cited as the closest subject, DMD = (dd^T) o M; the remark states that Tam, LRT and Mathias were read through their abstracts only and makes no claim about their later sections; the Horn antecedent leads the remark.

5. MINOR: verification-code wording

APPLIED. pga_free's docstring states the coordinate convention (factor 4 in upper-triangular coordinates, 2 under the Frobenius inner product; the direction is normalized); snap_isometry's docstring no longer calls every partial isometry an extreme point; worst_shortfall is clamped at zero for both searches (the prescribed-spectrum value is now 0.0 rather than -1.08e-18); Section 6 says "not exceeded beyond numerical error" for the prescribed-spectrum search. The script was rerun after the edits; no result changed beyond the clamp and the source hash.

6. CORRECT: definitions agree with Gil, with the normalization correction

APPLIED the fix: Section 1 now states the fixed-real-structure reading of the IRB (the populations are the eigenvalues of the real part, not in general of rho) and the zero-extension convention for M on inactive axes, beside the first definition of M.

7. CORRECT: Theorem 3.1's proof

APPLIED the optional addition: the explicit maximizing state (2 x 2 blocks with eigenvalues a_{2k-1} + a_{2k} and 0, trace one, real part A) is written out in the proof.

8. CORRECT: Theorem 4.1's layer-cake proof

APPLIED: "one-half the squared Frobenius norm" in the prose after eq. (11); the degree-constraint sentence of Proposition 4.2 reads |B| >= 2 with the singleton case noted; the density-matrix equality case (off-diagonal entries +- i s_k sqrt(a_{2k-1} a_{2k}), determinant a_{2k-1} a_{2k} (1 - s_k^2) >= 0) added at the end of the proof.

9. CORRECT: the theorems answer the questions as posed

APPLIED the three wording fixes: the abstract lists the odd-dimensional zero singular value and puts the positivity statement on the support; "the question would be empty" became "fixes the objective; compatibility with an aligned representative would be a separate feasibility question, which we do not address".

10. CORRECT: Remark 4.5's numbers, with exact certificates

APPLIED. The example now fixes distinct positive populations a = (0.4, 0.3, 0.2, 0.1) so that alignment means matching support; the LP witness (x_12 = 0.19, x_13 = x_24 = 0.81) and the dual certificate (0.6 x the C = {1,2} inequality plus 0.2 x each of C = {1,2,3}, {1,2,4}, all with B = [4]) replace the solver statement; the exact real-orbit certificate for 2.196 (the X, Y coordinates, {||X||^2, ||Y||^2} = {0.9025, 0.0025}, the bound 2.172 - 0.4 (X_1 - 5 Y_1)^2 + 9.6 Y_1^2) replaces the random-search sentence. Both certificates were rederived here before use: the dual combination's coefficients are exactly (2, 1.2, 1.2) on x_12, x_13, x_24 with nonnegative surplus on the other three edges and right side 2.324; the orbit bound follows from X_2^2 <= ||X||^2 - X_1^2, Y_2^2 <= ||Y||^2 - Y_1^2 and completing the square.

11. CORRECT: Section 6 reports the searches faithfully

No change beyond item 5; the candid shortfall reporting stays.

12. CORRECT: US English, no em dashes

No change.

Verdict

Accepted in full: Theorem 3.1 proved with the normalization made explicit; Theorem 4.1 proved as a real and complex skew-matrix inequality, now with two proofs, and reassessed as a consequence of classical majorization; the density-matrix consequence read on the active support; Corollary 5.1(1) corrected. The manuscript was also set in the house format of the Hypnos Math manuscripts (paper/common/) in the same pass, with the attribution block recording both reviews.

Left open, stated in the note

Full-text reading of Tam 1998, Leite-Richa-Tomei 1999 and Mathias 1992 before any stronger novelty statement; the README carries the item.

Consistency pass, 2026-10-01 (fresh Claude Opus 5.5 agent, directed by the Claude Fable 5.1 outreach thread)

A bounded consistency pass on main.tex (the abstract and Section 1 against the theorems, quotations against Gil's text, citations against their sources, the attribution block against paper/common/README.md, and the house language rules), not a third review of the mathematics; "fulltext line N" is a line of paper/antiunitary-matching/reviews/claude/03-literature/gil-2026-fulltext.txt, and "PMC HTML" is paper/antiunitary-matching/reviews/openai/lit-gil-2026-pmc.html with its MathML alttext extracted to text.

#	check (1-5)	item checked	source and page	command or method	finding	verdict	what was done
1	1	Abstract: what Gil asked ("whether this remains true over all orientations of the Youla frame, with free and with prescribed Youla values")	Gil Section 3.4 (fulltext line 231) and Section 8.2 (line 701); the note's Section 1	grep of the full text; read against Section 1's mapping of (i) and (iii)	Gil's problem (i) asks to determine the global maximum for fixed Youla values, not whether adjacent pairing stays optimal; Section 1 maps prescribed values to (i) and free values to Section 3.4 and (iii)	CORRECT, applied	The abstract now says Gil asked whether this remains true over all orientations of the Youla frame with free Youla values, and what the maximum over all orientations is when the Youla values are prescribed
2	1	Abstract: "every pairwise coherence functional with nonnegative weights is maximized in the aligned class"	Theorem 3.1, eq. (8) and the zero-population clause; Corollary 5.1(1)	hypotheses read side by side	Carries "with free Youla values"; aligned attainment holds for every population vector by Corollary 5.1(1) (a maximum-weight matching on the support), so this claim needs no positivity hypothesis; the positivity-dependent statements (value equal to the unrestricted matching weight, hull equal to the full polytope) are not made in the abstract	NO FINDING	none
3	1	Abstract: "Gil's adjacent-pairing value is the global maximum"	Theorem 3.1, eq. (7)	read	Matches: populations summing to 1, every orientation, every admissible Youla spectrum	NO FINDING	none
4	1	Abstract: mechanism sentence (odd-order principal blocks singular, positivity equal to the contraction condition on the support, squared cohesion a nonnegative linear functional)	Lemma 2.1 and its proof; eqs. (3) to (5)	read	Matches; the abstract's "real antisymmetric contraction" is a special case of the lemma's real-or-complex statement	NO FINDING	none
5	1	Abstract: the entries in the prescribed-values inequality	Theorem 4.1, eq. (9)	read; numpy spot check over 1,200 random complex unitary orientations, n = 2 to 7 (largest excess 4.4e-16 with moduli squared)	The abstract wrote the plain square of M_ij "for every skew-symmetric M"; the theorem covers real and complex M with the modulus squared, and the plain square is not real for complex M	CORRECT, applied	The abstract's M_ij^2 became the modulus squared, as in eq. (9)
6	1	Abstract: hypotheses on the singular values of M and on s and a	Theorem 4.1, statement	read	Matches case for case: s_1, s_1, ..., s_m, s_m and 0 when n is odd; s sorted; a sorted and nonnegative; no s_1 <= 1 in the matrix inequality (the density-matrix reading adds it); m is fixed by the count of n singular values	NO FINDING	none
7	1	Abstract: "the sharp inequality"	Theorem 4.1 ("with equality for Q = I"); the proof's last paragraph and its density-matrix equality case	read	Equality is proved for every admissible a and s at the Youla block form, and is realized by a state when s_1 <= 1	NO FINDING	none
8	1	Abstract: labels of the two proofs of Theorem 4.1	the "Proof" (layer cake) and the "Second proof, from a classical singular-value inequality" (Horn)	read	The abstract called the layer-cake proof "a second proof", while the body heads the Horn derivation "Second proof"	CORRECT, applied	"a second proof, by a layer-cake decomposition" became "another proof, by a layer-cake decomposition"
9	1	Abstract: the family "contains the degree and odd-set constraints as its two extreme members"	Proposition 4.2	read; R(1, q) = s_1^2 for q >= 2 and R(q, q) = twice the sum of s_k^2 over k <= floor(q/2) rederived from eq. (12)	The two extreme members are the spectral degree constraint (bound s_1^2) and the set constraint, under the body's own names, and they reduce to Edmonds' constraints at s = 1	NO FINDING	none
10	1	Abstract: the n = 4 example of a non-aligned maximum	Remark 4.5	recomputed in numpy	Singular values 1, 1, 0.9, 0.9; value 2.196; best aligned value 2.172 (matching {13, 24}); the abstract's claim is supported	NO FINDING	none
11	1	Section 1: "iM is Hermitian with eigenvalues plus or minus s_k"	Theorem 4.1 and the abstract, which list the zero singular value for odd n	read	The zero eigenvalue for odd n was omitted (the positivity conclusion is unaffected)	CORRECT, applied	"(and 0 when n is odd)" added
12	1	Section 1: Theorem 1.1, the restatement of Gil's Theorem 2	Gil Section 3.4, Theorem 2, eqs. (23) and (24) (PMC HTML)	MathML alttext extraction	Matches, including "s_k = 1 on every pair of positive population product"	NO FINDING	none
13	1	Section 1: "We determine the maximum asked for in (i) (Theorem 4.1) and answer (iii) affirmatively (Theorem 3.1)", and the trace remark on fixed spectra	Theorems 3.1 and 4.1; eq. (1)	read; tr(rho^2) = sum of a_i^2 plus the squared Frobenius norm of N rederived (N real antisymmetric with zero diagonal)	Matches	NO FINDING	none
14	1	Lemma 2.1 "taken from the companion manuscript"	paper/antiunitary-matching/main.tex line 161, Lemma "Squared-entry convex hull"	grep	Present in the companion	NO FINDING	none
15	2	Quotation "Whether the same maximum is globally optimal over arbitrary Youla orientations is a separate optimization problem and is not assumed here"	Gil Section 3.4, fulltext line 231	grep -n "globally optimal"	Verbatim; section correct	NO FINDING	none
16	2	Quotation "deliberately left open because they require optimization over Grassmannian orbits rather than the elementary matching argument used here"	Gil Section 8.2, fulltext line 701	grep -n "deliberately"	Verbatim; section correct	NO FINDING	none
17	2	Quotation "suitable spectral constraints"	Gil Section 8.2, fulltext line 701	grep -n "suitable spectral"	Verbatim	NO FINDING	none
18	2	Unquoted paraphrases of problems (i), (ii) and (iii)	Gil Section 8.2 (PMC HTML)	MathML alttext extraction	(i) matches word for word; (ii) is Gil's "image of the map (a, Q_M) to G(a, Q_M) subject to Plücker constraints", fairly paraphrased; (iii) differs only in "decide" for "identify"	NO FINDING	none
19	2	Gil's terms used without quotation marks: intrinsic reference basis, metaspin tensor, cohesion, density matrix, aligned class	Gil Section 2.1 (fulltext line 59) and eq. (3); eq. (7) and Section 2.3 (line 91); Section 2.4 (line 129: the cohesion of the antisymmetric block is the Frobenius norm of N) and Theorem 3 ("cohesion squared"); Definition 1	grep for each term over the extracted text	All match the note's usage, including the zero extension of M on inactive axes	NO FINDING	none
20	2	Tam, Leite-Richa-Tomei and Mathias: anything quoted from their abstracts	OpenAlex abstracts (Tam in full, Mathias truncated, Leite-Richa-Tomei absent and publisher-elided on Semantic Scholar too)	OpenAlex and Semantic Scholar APIs	Nothing is quoted; the Tam paraphrase (entries on one fixed matching under unitary congruence, weakly majorized by the singular values, one extra inequality in the even full-matching case) matches his abstract; the Mathias paraphrase matches his title; the attribution of the skew-symmetric Schur-Horn theorem to Leite-Richa-Tomei is confirmed by Sanyal-Sottile-Sturmfels citing it as their [22]. Semantic Scholar lists the Leite-Richa-Tomei DOI as bronze open access; not opened, per the instruction	NO FINDING	none; the abstracts-only disclosure is unchanged
21	3	Gil citations: eqs. (3), (7), (8), (20), (26), (27), (33); Definition 1; Theorem 2; Sections 2.1, 2.3, 3.4, 8.2	Gil full text (fulltext lines 59, 91, 223, 227, 231, 701) and PMC HTML	grep for each equation tag and heading	Every one resolves to the content the note attributes to it	NO FINDING	none
22	3	Gil's problem numbering (i) to (iii)	Gil Section 8.2 ("One is ... Another is ... A third is ...") and Section 8.3	read	Unnumbered in Gil, as the note says; Section 8.3 ("Open Questions") lists three different, also unnumbered, directions; the note numbers the Section 8.2 problems itself and says so	NO FINDING	none
23	3	Sanyal-Sottile-Sturmfels Prop. 3.10	arXiv:0911.5436v4 (8 Mar 2011), p. 14	pdftotext; the first-page stamp shows v4	Prop. 3.10 is "The skew-symmetric Schur-Horn theorem [22]", stated for the linear skew-diagonal map SD; their [22] is Leite-Richa-Tomei, LAA 286 (1999) 149-173	NO FINDING	none
24	3	Horn 1950, the product inequality for singular values	PNAS 36(7) (1950) 374-375	curl: PNAS answers 403, PMC and Europe PMC answer with bot challenges (not bypassed); Crossref and the Europe PMC REST API for the metadata	Bibliographic data confirmed; the text could not be retrieved; Astra's review (its item 1) records checking both original pages	NOT CHECKABLE	none
25	3	Fan 1949, the maximum principle	PNAS 35(11) (1949) 652-655	as in row 24	Bibliographic data confirmed; the text could not be retrieved	NOT CHECKABLE	none
26	3	Youla 1961, "the complex case under unitary congruence"	Canad. J. Math. 13 (1961) 694-704, Cambridge Core PDF, p. 694 eq. (2) and p. 701 Corollary 2	curl; pdftotext	A complex skew-symmetric C satisfies U'CU = E_1 + ... + E_k + 0 with 2 x 2 blocks of off-diagonal entries a_r and -a_r, a_r > 0; matches	NO FINDING	none
27	3	Edmonds 1965, the polytope of eq. (2)	J. Res. Nat. Bur. Standards 69B (1965), NIST PDF pp. 125-126	curl; pdftotext	Inequalities (1) to (3) and Theorem (P) match eq. (2) (the note's singleton odd sets add only 0 <= 0)	NO FINDING	none
28	3	Horn and Johnson, Topics in Matrix Analysis, Thm. 3.1.2 and Cor. 3.1.3	Cambridge Core table of contents	curl	Chapter 3 is "Singular value inequalities" (pp. 134-238), consistent; theorem numbers are not shown	NOT CHECKABLE	none
29	3	Horn and Johnson, Matrix Analysis, 2nd ed., Section 2.5 (the real normal form)	Cambridge table of contents	curl	Chapter 2 is "Unitary similarity and unitary equivalence" (pp. 83-162), consistent; section numbers are not shown	NOT CHECKABLE	none
30	3	Bhatia, Matrix Analysis, Chapter III, for Ky Fan's maximum principle	Springer table of contents	curl	Chapter III is "Variational Principles for Eigenvalues", consistent at the chapter level the note cites	NO FINDING	none
31	3	Bibliographic data of the ten DOIs	Crossref API	scripted query	Titles, volumes, issues and pages match the bibliography	NO FINDING	none
32	3, 4	The companion entry (cite key ross)	the companion's title block (paper/antiunitary-matching/main.tex lines 12-15); the set's convention in paper/unfolded-zeros/main.tex line 391	grep	It named "D. Ross" as the author of an "AI-assisted research manuscript ... (in preparation)"; the companion was written by GPT-6 Astra at the direction of David Ross and is dated September 28, 2026, and the set cites its own papers as MODEL, at the direction of D. Ross, TITLE, manuscript (MONTH YEAR), Hypnos Math, URL	CORRECT, applied	The entry now reads "GPT-6 Astra, at the direction of D. Ross, Approximate antiunitary symmetry as a matching problem, manuscript (September 2026), Hypnos Math, https://hypnosmath.org"
33	4	Writer, director and the division of labor	paper/common/README.md, pattern items 1 and 2	read	Matches the pattern; the source of the questions (Gil's paper, found while reviewing the companion) is stated	NO FINDING	none
34	4	The self-review, named with its verdict in one clause	self-review/REPORT.md lines 9 and 15-23	read	"Claude Opus 5.5, 2026-09-29 ... found the proofs complete and three framing defects" matches the report (proofs complete, no BLOCKING item, three MAJOR framing items)	NO FINDING	none
35	4	The Astra review, named with its verdict in one clause	reviews/astra/REVIEW.md items 1 and 2 and the verdict (line 278)	read	"no blocking defect and two major items" matches; but "now given as its first proof" was false: the Horn derivation is headed "Second proof" and follows the layer-cake proof, as item 1 of this ledger records	CORRECT, applied	"now given as its first proof" became "now given as its second proof"
36	4	"No human mathematician has reviewed this paper."	README pattern item 4	grep	Present verbatim	NO FINDING	none
37	4	Place in the set; no paper called "the first"	README table; the sibling sentences in paper/antiunitary-matching/main.tex and paper/unfolded-zeros/main.tex	grep	The last of the three of the night of 2026-09-28/29, after Astra's paper and the Riesz-basis paper, written without reading the latter; consistent with both siblings; no paper is called the first	NO FINDING	none
38	5	Em dashes	main.tex	Python count of U+2014 and of "---"	0 and 0; the file has no non-ASCII character	NO FINDING	none
39	5	US English	main.tex	regex scan for -our, -tre, -ise, -yse, doubled -ll- and similar British forms	None (the only hits were pairwise, termwise and pointwise)	NO FINDING	none
40	5	A Clay Mathematics Institute (A)/(B) statement	main.tex	grep for Clay, Millennium, (A), (B)	None	NO FINDING	none
41	5	Claims beyond the theorems	the abstract, Section 1, Remark 4.4	read	Every claim maps to Lemma 2.1, Theorem 3.1, Theorem 4.1, Proposition 4.2 or Remark 4.5; "its use for density matrices is new here" is relative to the companion; Remark 4.4 limits the claim to the application, the polyhedral viewpoint and the family	NO FINDING	none

Build: bash build.sh (tectonic) succeeded after the edits; main.pdf has 9 pages, as before; no new warning (only the two existing package notices, on the relative style path and on inputenc) and no Overfull or Underfull box.