The Schwarzian Bridge in a Single Sigmoid Neuron: Oracle Gram Resolution, Joint Spectral Lexicography, and Spherical Saturation Phases

Strengths

• Theorem 2.1 gives a clean closed determinant identity and correctly separates its explicit logistic specialization from the classical order-two Schwarzian criterion and serial-composition law.
• Theorems 4.1 and 5.1-5.3 expose the transition-layer mechanism with exact constants, including the inverse-radius phase boundary and a concrete covariance-isotropy counterexample.
• Theorem 6.1 links the adjacent-point Loewner defect to radial-tangential anisotropy through an exact fourth-moment multiplier, and Theorem 6.2 turns the limit identity into a finite-scale error bound.
• The fixed-teacher empirical analysis distinguishes bilateral relative Loewner loss from weaker eigenvalue and eigenspace targets, preventing the common but false inference that an r^-3 population eigenvalue automatically requires r^3 samples.
• The spectral-lexicography section cleanly separates the deterministic exterior-power flag mechanism, the exact Gaussian-inverse angle distribution, and the conservative finite-saturation certificate.
• The manuscript repeatedly states what is not claimed: no deep-network theorem, no label-based estimator result, no global optimization result, and no crossover-sharp joint projector rate.

Weaknesses

• Theorem 8.11 relies on a concise sequential dominated-convergence argument for uniform angular overlap and then invokes Sudakov, Borell, and a fixed-grid central limit transfer; these steps are plausible but leave less audit margin than the elementary earlier proofs.
• The claimed literature search is broader than the bibliography evidence independently checked here. The newest exact angle-law and marked angular-process priority claims remain bounded novelty candidates rather than verified firsts.
• The README declares deterministic hashes for five PDFs and five JSON certificates, but the archive omits the five PDFs and four supplied JSON files have different hashes from the README values despite matching PUBLIC-CODE-MANIFEST.json.
• The exact dependency set could not be installed from the available cache. Runs used NumPy 2.3.5, SciPy 1.17.0, and Matplotlib 3.10.8 rather than the pinned versions, and python-flint 0.9.0 was unavailable.
• The Cramer-Rao display on page 27 omits the regular-unbiased-estimator qualification at the point of use, although page 32 later supplies that limitation.

Potential errors

• README.md expects finite_sample_resolution.json hash 4bd3e130d11670c3a51d51ff0a4d962078ae415e95f4e90260e23da9293f9def, while the supplied manifest-verified file hashes to 07cc7821ebf74a82a59f3ccf5324fa7e8f53c0de7947d56b3b866b40131fc9a6.
• README.md expects radial_phase_transition.json hash 26ed3f0460847f79296301d5ecf087694358de819d976c671415762cc76c1d9a, while the supplied manifest-verified file hashes to 6e61b4b37980fb2964d045339a46f7bea5f90174f493391a483abe6752f95e1e.
• README.md expects spectral_lexicography.json hash 93d226200a59e3772b11ead7563eeed9e69a3a1d77951deacab97f4a5f7813c2, while the supplied manifest-verified file hashes to f740cac2ec06627edf012fc9eeeb1281109a62c50ac522087147550bc16540bb.
• README.md expects joint_spectral_resolution.json hash a961a7b695707d24315473aa228e2690d2a84621cff286238cc3319276e4a493, while the supplied manifest-verified file hashes to f09133a9b739a0008f5dd3589b18c75fbd0cee7d0be2ce7d57ae51adb5240d52.
• The five README-listed deterministic PDF outputs are not present in the supplied public-code tree. Newly generated PDFs under the unpinned environment do not match the README hashes, which is not evidence against a correctly pinned run but leaves the advertised byte replay unverified.
• The page-27 Cramer-Rao inequalities require regular unbiased estimation assumptions; without the qualification later supplied on page 32, the displayed statement is too broad.

Strong novelty candidates

• The coefficient identity in Theorem 6.1 equating the local Loewner determinant defect with spherical radial-tangential anisotropy, including q_X, appears both precise and conceptually useful.
• The exact three-phase spherical radial law in Theorem 5.3, including the critical r^-3 log r constant and the subcritical beta-function constant, is a strong extension candidate.
• Theorem 8.8's nearest-transition-hyperplane projector limit and exact tan(theta)=||zeta||/|G| law, combined with matching stated-range complexity, is a distinctive contribution candidate.
• The iterated angular-process lower bound in Theorem 8.11 and its separation from the non-sharp full-shell net upper bound is a credible new obstruction candidate.

Unresolved material objections

None reported.

Strengths

• Theorems 2.1 and 6.1, pp. 4 and 11: the closed determinant has the correct sign, and the anisotropy coefficient is -q_X p Sg(0). The fourth-moment multiplier and sixth-moment remainder are accounted for; the paper credits the underlying Schwarzian criterion.
• Theorems 4.1 and 5.1, pp. 6-8: the change of variables gives the r^-1/r^-3 laws with the correct moments. The d=2 exponential-localization hypothesis controls the angular endpoint singularity; evenness is correctly unnecessary for the Gaussian reduction.
• Lemma 5.2 and Theorem 5.3, pp. 8-10: Tonelli gives the beta-function coefficient, the kernel asymptotic supplies the critical logarithm, and the R>=eta contribution is controlled without extra density assumptions away from zero. Independent Fubini quadrature agrees with the phase constants.
• Theorem 7.2, pp. 13-14: strict likelihood-ratio ordering establishes monotonicity for every fixed p>0. Referee exact polynomial arithmetic independently recovers 1+p*r^2/2-p*r^4/8+p*(p+1)*r^6/16.
• Theorems 8.1 and 8.4, pp. 15-18: conditional Gaussian blocks support the relative Loewner upper bound and the d-2 off-diagonal lower obstruction. Lemma 8.3 supplies an explicit confidence lower tail rather than an unnamed moderate-deviation step.
• Lemmas 8.5-8.7 and Theorems 8.8-8.9, pp. 19-23: the exterior-power/Plucker argument, independence of selected tangential rows, exact Gaussian inverse law and Beta(1,n) spacing bound fit together. The finite perturbation lemma also survived 126 additional numerically stable referee stress cases.
• Theorem 8.11, pp. 25-26: the overlap estimate separates order-R many marked directions; Sudakov and Gaussian concentration precede a finite-grid CLT with n tending to infinity first. The claim is appropriately weaker than a nonasymptotic shell lower bound.
• Sections 9-13, pp. 27-32: the manuscript explicitly limits stationary-step optimization, regular unbiased information bounds, whitening, oracle access and the order of limits. These restrictions prevent several tempting but unsupported learning-theory interpretations.

Weaknesses

• Abstract, pp. 1-2: the sharp Theta_p(r[d+log(1/delta)]/epsilon^2) statement should explicitly inherit Theorem 8.4's d>=3, large-r and small-delta restrictions. Its displayed high-confidence cutoffs are about 1.47e-34 and 6.00e-29 for p=1,2; the theorem does not directly provide that necessity at ordinary 95% confidence.
• Section 12 and repro/README.md: the supplied finite_sample_resolution, radial_phase_transition, spectral_lexicography and joint_spectral_resolution JSON hashes differ from the README expectations before any execution. PUBLIC-CODE-MANIFEST.json does match. Update the replay documentation to identify the exact certified artifacts.
• The advertised lualatex main.tex build from src cannot run as documented: src/main.tex and the referenced source-layout figure set are absent. paper.tex is supplied at a different path and was not substituted. This limits artifact reconstruction, not the validity of the reviewed PDF proofs.
• Theorem 8.9 resolves existence of a joint regime with a very conservative r threshold; it does not locate the r comparable to n crossover. Theorem 8.11 supplies only an iterated lower obstruction, so neither result closes the sharper finite-sample problems identified in Section 13.
• The v4 and v5 verifiers partly inspect supplied certificates. Their successful exit is not an independent rerun of everything named in their PASS messages. Separate fresh generators and independent referee calculations provide additional, explicitly bounded evidence.
• Sections 2, 3 and 11 contain substantial classical material. The bridge is a useful coefficient identification but not a broad equivalence between matrix order and learning geometry; its centered-inflection and symmetry hypotheses matter. Significance and priority should remain tied to the specific new refinements.

Potential errors

• Reproducibility documentation has a confirmed internal inconsistency: four expected JSON hashes in README.md disagree with the manifest-verified supplied certificates. The final verification record lists both values. This is not evidence that the mathematical constants are wrong.
• Section 9.2, Eq. (93), p. 28 should place regular-unbiased-estimator assumptions beside the Cramer-Rao display. They appear only in the limitations on p. 32. Without that qualification an unrestricted variance lower bound is false, for example for a constant biased estimator.
• The 80-digit referee calculation puts the witness eigenvalues and determinant inside Proposition 2.2's displayed intervals, but it is not outward-rounded Arb certification. The original main replay fails at import, so the manuscript's directed-rounding execution claim was not independently reproduced here.

Strong novelty candidates

• Theorem 6.1, Eq. (35): the spherical fourth-moment coefficient identifies the same local Schwarzian in an adjacent-point Loewner defect and radial/tangential onset; Theorem 6.2 adds explicit finite-scale errors. Candidate novelty is this precise identification, not the Schwarzian criterion itself.
• Theorems 8.8-8.9: nearest-transition-hyperplane samples determine the extreme-saturation projector, with tan(theta)=||zeta||/|G| and an explicit finite-saturation bridge. Classical scaled-SVD and Gaussian invariance are ingredients rather than new claims.
• Theorem 8.11: a marked angular Gram process yields a sqrt(log R) penalty in the stated high-precision iterated regime. The elementary proof is credible, but absence of an earlier exact formulation has not been established by an exhaustive literature search.

Unresolved material objections

None reported.