Second review, by Claude Fable 5.1 (Anthropic), the other company's model, September 29, 2026

Adversarial review and follow-on investigation: "Approximate Antiunitary Symmetry as a Matching Problem"

Reviewer: Claude (Fable 5.1). Baseline: commit 9820d1ed1e61381cb22185740377f7d2ee99a6f1 (HEAD at review start; paper directory clean; the rendered PDF in Documents/Codex/2026-09-28/che/outputs/ is byte-identical to main.pdf). Date: September 29, 2026 (EDT). All files referenced below live in paper/antiunitary-matching/reviews/claude/.

Verdict

What survives. Everything the manuscript proves is correct. The main theorem (exact reduction of the penalized antiunitary problem to a weighted matching), the lemma that squared entries of a skew-symmetric contraction lie in Edmonds' matching polytope and fill it in convex hull, and every corollary and proposition (commuting tuples, rigidity and stability, parity, hard constraint, perturbation bound, dual certificates, the sorted-spectrum recurrence) were reconstructed independently before the manuscript's own audit was read, and no error was found. The supplied numerical and exact checks reproduce. New exact checks over complex unitaries (Gaussian-rational Cayley transforms), adversarial structured unitaries, and an independent multistart optimizer found nothing below the theorem's value.

What fails. No mathematical statement fails. Two hypotheses are stronger than needed (the rigidity proposition does not need positive costs, only a unique optimal matching; the hard-constraint corollary needs no sign condition). One typo (max outside math mode). The novelty framing needs one correction: the entire single-observable case (one Hermitian matrix, hard or soft penalty) is a routine consequence of Wigner's normal form, Kramers' theorem and Mirsky's inequality, which the manuscript does not say; the substantive content is the general-cost theorem, hence the tuple case with two or more observables, carried by the hull lemma.

What appears new. The hull lemma and the general weighted identity were not found in the literature after 77 targeted searches (reviewer plus a background search agent), framed between Birkhoff-von Neumann (unitaries) and O'Meara-Pereira 2013 (Hermitian unitaries, where the hull is strictly smaller than the degree polytope). Priority is not established; the lemma is elementary enough to be folklore somewhere.

What matters. The lemma turns out to answer two open questions that J. J. Gil posed independently in Entropy 28:877 (published 2026-08-04) about the maximal coherence of density matrices with prescribed intrinsic populations: his third question (global optimality of the adjacent pairing beyond the "aligned class") follows directly from the lemma, and his first question (fixed populations and fixed Youla values) is settled here by a new inequality (Theorem G) whose proof is independent of the manuscript. That is the strongest external evidence that the manuscript's core object is worth having. The other new results are a sharp characterization of when diagonal (vertex) costs keep the problem combinatorial, a quantitative Kramers inequality in every unitarily invariant norm, and the observation that the standard SDP relaxation of the unitary constraint cannot see the parity obstruction.

1. Correctness and completeness

Claim-by-claim table: 06-claim-audit.md (33 items). Independent proof reconstruction, written before reading PROOF-AUDIT.md: 01-proof-reconstruction.md. Summary of the substantive inferences:

• Transpose versus conjugate transpose: S = (U + U^T)/2 and K = (U - U^T)/2 are complex symmetric and complex skew, contractions but not normal; no step uses Hermiticity. The identity U conj(U) + I = 2 S conj(U) rests on U^T conj(U) = (U^* U)^T = I. Correct.
• Congruence versus similarity: the substitution U = V W V^T is right; the key fact is V^T conj(V) = I. Correct.
• Contraction and rank: an odd-order principal skew block has zero determinant over the complex field; its squared Frobenius norm is at most its rank because its singular values are at most one. Correct.
• Constants: two per matched pair, 4 tau per unmatched vertex; verified exactly (rational arithmetic) for real orthogonal and complex unitaries.
• Convex hull versus realizability: correctly separated. Extra fact established here: strict mixtures of two matchings can be realized by one matrix (explicit 3 x 3 real orthogonal example with x = (1/2, 1/2, 0)), so uniqueness of the optimal matching is genuinely needed for rigidity.
• Attainment: real signed swaps, exact for every stated hypothesis.
• Degeneracies: n = 1, tau = 0, zero costs, repeated eigenvalues, ties, odd n all consistent.
• Duality, the sorted recurrence (exchange identities 2(d-a)(c-b) >= 0 and 2(c-a)(d-b) >= 0), the hard constraint, the perturbation bound: correct.

Where each hypothesis is used, and what happens without it: nonnegativity of c is used once (to drop the symmetric-part term). With diagonal costs the exact identity becomes Theorem A(a) in 04-new-results.md, the matching formula holds under 2 c_ij >= c_ii + c_jj, and outside that condition it can fail: for c = I_3 the true minimum is 4 tau + 2 tau/(1 + 6 tau) (a continuous rotation about (1,1,1)), below the formula's 1 + 4 tau for every tau. Negative costs also break it.

2. Independent falsification (record in 02-checks/)

Check	What it does	Result
supplied-rerun/	Reruns verify.py, verify_exact.py from baseline	0 failures; counts identical; BFGS success flag 82 vs 80 (environment)
exp01	Exact Gaussian-rational complex unitaries (Cayley), decomposition, odd sets, lower bound; exact complex skew unitaries for the hard constraint	44 unitaries, 232 odd-set checks, 176 bounds, 20 skew unitaries; 0 failures
exp02	Phased permutations, DFT and circulants, odd-cycle rotations, Haar near a 3-cycle, tie mixtures; Cayley-chart L-BFGS multistart (60 starts)	worst energy minus bound -7e-15; 0 alarms; non-swap minimizers only at ties
exp03, exp03b	Gil objective: free and prescribed Youla values, random + local search; LPs	bounds never exceeded; coarse polytope LP not tight (0.045), full-family LP tight to 1e-17
exp04	Vertex-penalty identity on Haar unitaries; formula under w >= 0; c = I_3 curve; random violators	identity exact to 4e-14; formula never violated under w >= 0 (540 cases); c = I_3 minima equal the closed form at 10 values of tau; 28 of 108 violators break the formula
exp05	Kramers bound, all Schatten norms and Ky Fan partial sums, random skew unitaries; minimization over T	1,200 checks, min slack -5e-15; bound attained to 1e-15
exp06	Shor SDP relaxation (cvxpy/SCS) and explicit certificate	value 0 for odd n with c = 0 (true 4 tau); certificate feasible
exp07	Noncommuting pairs: multistart truth vs sandwich bound from a commuting approximant vs projection bound	see Section 4
exp08	Synthetic Zeeman ring: physical T commutator vs spectral bound	bound holds; operator-norm bound within 0.05 percent at weak field

Failed or ambiguous runs are recorded, not hidden: the first version of exp07 used a Jacobi joint-diagonalizer that did not converge (it reported a nonzero commuting-approximant error even for an exactly commuting tuple); it was replaced by an eigenbasis-of-random-combination start with L-BFGS refinement, and the run was repeated. Multistart minima are estimates, not certificates; they are only used as upper bounds on the truth or as agreement with closed forms.

3. Novelty and significance (07-literature-comparison.md, agent report in 03-literature/)

• Correctly labeled classical: Edmonds, Wigner, Kramers, odd-dimension facts (citable to Wigner 1960 or Garcia-Tener 2009, Lemma 4.2).
• Routine consequences of classical theorems, presented without saying so: the single-observable case (Prop D(a) gives the derivation). The one-pair Frobenius distance is in Pinter et al., Quantum 10, 2047 (2026).
• Apparently new: Lemma 2.2, Theorem 2.1 for general nonnegative costs, hence the tuple corollary, the dual certificates and stability estimates.
• Gil 2026 (full text read from PubMed Central; formulas verified from the MathJax source): no overlap of results; his Theorem 2 is the aligned-class special case; his open questions are answered here (Theorems C and G). The paper has no citations yet on Semantic Scholar.
• Skew Schur-Horn results (Sanyal-Sottile-Sturmfels 2011 Prop. 3.10; Tam 1998) treat one linear entry per block, not squared moduli of all entries; Theorem G is outside them.

4. Applications (05-application.md)

1. Gil's fixed-population coherence problem (worked in full): closed-form global maxima with and without prescribed Youla values, a new family of valid inequalities, and numerical confirmation. Value: converts a subclass statement into a global one for a quantity that enters his purity and entropy identities.
2. Time-reversal breaking bounds (synthetic demonstration): the free-T theorem yields, for a prescribed physical T, a certified spectral lower bound; the useful form is the operator-norm version of Theorem B, nearly tight at weak field in a spin-orbit ring with Zeeman splitting.
3. Noncommuting tuples (exp07, corrected run): for nearly commuting pairs the theorem gives a practical near-optimal construction (the matching optimizer of a joint-diagonalization approximant, evaluated on the true tuple, is within 0.1 percent of the multistart optimum at perturbation 0.1 and within 6 percent at perturbation 1.0), while the manuscript's sandwich lower bound is loose (factor 2 to 4 off already at perturbation 0.6) and the reviewer's projection lower bound sup_{|t|=1} Phi_tau(t . H) sits about one third below the truth. No exact formula for noncommuting tuples; the gap was not closed.

5. Extensions and discoveries (04-new-results.md, LaTeX: extension-note.tex/.pdf)

Proved: Theorem A (vertex penalties, sharp condition, explicit continuous minimizer); Theorem B and Corollary B' (quantitative Kramers in every unitarily invariant norm; distance to even-multiplicity matrices); Prop D (classical route for one observable, tuple obstruction); Prop E (Shor gap certificate); Prop F (rigidity under uniqueness); Theorem C and Theorem G (Gil's questions); Prop H (normal matrices, Floquet relation). Abandoned or open: a combinatorial description when w is not nonnegative; the joint hull of (|S_ij|^2, |K_ij|^2); Theorem G for general (non-product) weights; exactness for noncommuting tuples.

6. Separation statement

The directory reviews/openai was never listed, read, or searched; no conversation-search, memory or agent tool was used to retrieve the other reviewer's work; the background literature agent was instructed to use web tools only and reported no local file access. reviews/ did not exist when this review started. Two incidental observations are disclosed: at the end of the session a listing of the parent directory reviews/ showed that reviews/openai/ now exists (its contents were never listed or read), and the git push output showed that the shared branch had advanced by four commits that are not mine before my first commit was made (their messages, diffs and files were not inspected; only their count was read; my commits stage only reviews/claude/). No content of the other review reached this reviewer.

7. Strongest defensible result

Lemma 2.2 of the manuscript, in the form: for every complex skew-symmetric contraction K, the vector of squared moduli (|K_ij|^2)_{i<j} lies in Edmonds' matching polytope, and the convex hull of these vectors is that polytope. It is short, it carries the manuscript's main theorem for general costs and tuples, and it resolves an explicitly posed open question in an unrelated literature (Gil 2026). Its significance is that it exposes a parity constraint that no degree-based or first-level semidefinite relaxation of the unitary group can see, and that it converts a Grassmannian-orbit optimization into linear optimization over a polytope whose vertices are matchings. The reviewer's Theorem G shows the same mechanism survives when the singular values of K are prescribed, which the lemma alone does not cover.