{"id":"baf00df3-f533-4944-aceb-476991f8d715","arxiv_id":"2608.08752","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"At the critical temperature, the rescaled two-replica overlap of the Sherrington-Kirkpatrick model converges in quenched law to an explicit measure built from the reflected Airy_1 point process.","lead":"This paper finds the exact random distribution of the overlap (similarity) between two copies of a spin glass at the critical temperature. It answers a question of Talagrand and shows the spherical and Ising versions converge to the same Airy-based law.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ising transfer and part of the spherical argument rest on unverified estimates from the unpublished companion [DH26]; the central claim is not independently checkable until those are supplied.","rationale":"The reader's CONDITIONAL verdict is appropriate. The spherical part of the proof is built on standard random matrix edge theory and appears internally consistent; the formal construction of the limit measure via pinned Gaussian laws is coherent, including the root of the limiting saddle equation once the OCR ambiguity in (1.3) is read as 1/x + Xi(chi) + sum(...). The single most load-bearing weakness is the dependence of the Ising transfer on multiple estimates from the unpublished companion [DH26], and to a lesser extent the use of the same exponential-tail estimate inside the spherical proof. The paper's disclosure that most formal arguments were generated by GPT-5.6 Pro does not by itself invalidate the mathematics, but it raises the standard of verification and makes the missing companion proofs more consequential. The proposed check, re-deriving (1.9) and (3.9) from first principles, would directly settle whether the cited estimates hold and whether the sphere-to-cube comparison is sound. No internal inconsistency was found that would justify a stronger verdict change.","tokens_in":37715,"tokens_out":25966,"duration_ms":253676,"concrete_test":"Independently re-derive the two key imported estimates from [DH26] starting from the explicit formulas in Section 1.3: the ratio bound (1.9) and the local density comparison (3.9). If (3.9) fails to be uniform with error O(N^{-1/3}) on the scale |q| <= M N^{-1/3}, or if (1.9) cannot be closed with the stated O(N^{-1/3}) bound, then Theorem 1.9 and hence Theorem 1.1(a) are unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 1.1(a) is deduced from Theorem 1.9 plus Theorem 1.1(b). Theorem 1.9 (sphere-to-cube comparison) and Proposition 3.1 depend on several imported ingredients from the unpublished companion [DH26]: (i) the ratio bound E[(X_N-1)^2] = O(N^{-1/3}) from [DH26, Theorem 1.6], used through (1.9) to get X_N -> 1 in probability in Proposition 3.1; (ii) the exponential tail estimates quoted as Proposition 3.2, used in Lemma 3.3, Lemma 3.4, Proposition 1.8, and the proof of Theorem 1.9; (iii) the local density comparison [DH26, Lemma 3.9], quoted as (3.9); and (iv) [DH26, Propositions 3.1-3.2], which enter the uniform bound (3.10). Even the spherical part, Theorem 1.1(b), uses Proposition 3.2 for relative compactness in Lemma 3.3 and for moment control in Proposition 1.8. These results are cited but not proved here, and the outline in Section 1.3 does not supply the missing arguments. If any of these estimates fails, the sphere-to-cube comparison breaks, and with it Theorem 1.1(a); if the exponential tail fails, the Wasserstein-2 convergence and the finiteness of V_a also fail. Because [DH26] is not available to the reader, the verification burden for the paper's central claims is presently unresolved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the quenched law of the two-replica overlap R_{1,2} in the Ising and spherical Sherrington-Kirkpatrick models at the critical inverse temperature β=1. The main results, Theorem 1.1 and Corollary 1.2, assert that N^{1/3}R_{1,2} converges in a Wasserstein sense to an explicit random probability measure built from the reflected Airy_1 point process, and that N^{2/3}E< R_{1,2}^2 > converges to an explicit Airy-functional, resolving Talagrand's Conjecture 11.7.5. For the spherical model, the proof proceeds by representing the Gibbs measure as a norm-constrained Gaussian in an eigenbasis of the GOE matrix and passing to the Airy_1 edge scaling; for the Ising model, the proof invokes a sphere-to-cube comparison principle that transfers the spherical limit to the Ising overlap law. The paper is self-contained in its construction of the limiting object (Section 2) and in the spherical edge convergence (Section 4), but several load-bearing estimates are cited from the authors' unpublished companion paper [DH26].","tokens_in":37877,"tokens_out":4328,"duration_ms":50872,"significance":"If the result is correct, it gives a fully explicit non-Gaussian, quenched limit for the critical two-replica overlap and resolves a conjecture stated by Talagrand. The construction of 𝜓_a(χ) as the law of a bilinear form under an infinite-dimensional pinned Gaussian distribution is elegant, and the paper gives a satisfying random-matrix interpretation of the spherical SK overlap in terms of GOE edge eigenvalues and inverse gaps. The proof is genuinely constructive: the limiting law has no fitted parameters, and the finite-N identification for the spherical model, Lemma 4.5, is a clean exact identity. The main limitation is verifiability: the transfer to the Ising model, and even parts of the spherical argument, rest on estimates from the unpublished companion [DH26]. Subject to those estimates being supplied or made available, the paper would be a substantial contribution to the critical spin-glass literature.","major_comments":[{"comment":"Theorem 1.9 and hence Theorem 1.1(a) depend on four imported estimates from the unpublished companion [DH26]: the partition-function ratio bound E[(X_N-1)^2]=O(N^{-1/3}) from [DH26, Theorem 1.6], the exponential tail estimates quoted as Proposition 3.2, the local density comparison (3.9) quoted from [DH26, Lemma 3.9], and the uniform annealed bound (3.10) attributed to [DH26, Propositions 3.1-3.2]. These are precisely the estimates that transfer the spherical overlap law to the Ising model: X_N→1 in probability and the tails are used to pass from the annealed comparison (1.12) to the quenched comparison (1.11) and then to Wasserstein-2 convergence. Since [DH26] is not available to the reader and the present paper does not prove or even fully restate these results, the central claim (a) is not independently checkable from the manuscript.","section":"§1.3, §3"},{"comment":"The spherical result Theorem 1.1(b) also relies on the companion estimates. In Lemma 3.3, relative compactness of the laws of the quenched overlap measures uses the exponential tail bound from Proposition 3.2, which is cited to [DH26]. In Proposition 1.8, finiteness of Eχ V_a(χ) uses the spherical exponential moment bound from Proposition 3.2 through the estimate on E V_{a_N} in (4.21). Thus even the spherical convergence is conditional on [DH26]. The paper should either include proofs of these estimates, state them with full hypotheses and dependencies, or provide a public version of [DH26] so that the verification burden can be discharged.","section":"§3, Lemma 3.3; §4, Proposition 1.8"}],"minor_comments":[{"comment":"The assertion that a closed ball in the fourth-moment Wasserstein metric is compact in the 2-Wasserstein metric is stated without proof; a short justification via Prokhorov's theorem plus uniform integrability of second moments would improve readability.","section":"§3, Lemma 3.3"},{"comment":"The sentence crediting an 'idea due to GPT-5.6 Pro' is informal for a mathematical proof section; such acknowledgments are better placed in the acknowledgments or in the AI-use disclosure rather than in the proof outline.","section":"§1.3"},{"comment":"The paper would benefit from an explicit table or list of which results are proved here and which are imported from [DH26], especially because the reader may otherwise have difficulty tracking the verification status of assumptions used later.","section":"§1.5, organization"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is heavily dependent on the authors' own unpublished companion [DH26]. This is not a matter of style: several of the most important transitions in the Ising transfer and even the spherical Wasserstein-2 convergence explicitly invoke [DH26] results without proof. I would recommend requiring the authors either to include the needed estimates (or their proofs) in the present paper or to post a complete, checked version of [DH26] before acceptance. The AI-generation disclosure is not itself a mathematical concern, but it increases the importance of having the cited companion results independently examinable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is the first explicit characterization of the critical overlap distribution in both the Ising and spherical SK models at beta=1, with scale N^{-1/3} and a limit built from the reflected Airy_1 point process. It also gives the constant in the N^{2/3} second-moment asymptotics, resolving Talagrand's Conjecture 11.7.5. That is a real result. Second, the load-bearing wall is not all in this paper: several key estimates—the ratio bound E[(X_N-1)^2]=O(N^{-1/3}), the exponential tails in Proposition 3.2, and the local density comparison in (3.9)—are imported from the companion [DH26], which is cited but not proved here. If those estimates are solid, the sphere-to-cube transfer likely works; if [DH26] is not available at review time, no referee can verify Theorem 1.1(a). The stress-test note is right, and I'd add that even the spherical part borrows Proposition 3.2 for relative compactness, so this is not just an Ising-transfer concern.\n\nThe paper deserves credit where it is earned. The spherical derivation is coherent: GOE edge spectrum, the saddle equation, the pinned Gaussian measure, and the Airy_1 limit form a natural chain. Section 2's construction of the conditioned infinite-dimensional Gaussian measure is genuine work and looks correct. The Hilbert-space embedding in Section 3 is a clean way to upgrade annealed convergence to quenched convergence. The citation pattern is not abusive: the authors cite their own companion because that is where the quoted estimates live, and the overlap limit is constructed from the Airy process, not fitted to overlap data.\n\nThe soft spots are proportionate. The paper openly states that most formal arguments were generated by GPT-5.6 Pro. That alone is not a mathematical defect, and the specific attribution of the Hilbert-space idea to the model is honest. But it does raise the verification burden: the proof chain is long, there is no machine-checking, and the companion is unpublished. I would not desk-reject on AI use, but I would want the companion posted in a form referees can check, and I would want the authors to explicitly confirm which steps in this paper were human-verified.\n\nWho is this for? Spin glass and random matrix people, definitely. It answers a named conjecture and gives a concrete limiting object that will be cited. A serious referee gets real value from engaging with this. My recommendation: send it to peer review, but make acceptance conditional on the companion being available and the AI-generated steps receiving human verification. The scientific claim is likely correct, but the current verification burden is too high for an unconditional acceptance.","headline":"A likely-correct answer to Talagrand's critical-overlap question, but the main theorem leans on estimates living in an unpublished companion; referee it seriously and require that companion to be verifiable.","tokens_in":38539,"tokens_out":1729,"would_cite":true,"duration_ms":19938,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["82B44","60K35","60B20","82B27"],"pacs":[],"model":"deepseek-v4-flash","headline":"At criticality, both Ising and spherical SK overlaps have scale $N^{-1/3}$ and converge to an explicit Airy$_1$-point-process random measure.","keywords":["SK model","spherical SK model","overlap distribution","critical temperature","Airy_1 point process","GOE spectral edge","2-Wasserstein convergence","sphere-to-cube comparison"],"falsifier":"Simulate the spherical model at large $N$: diagonalize a GOE matrix, sample the anisotropic Gaussian conditioned on $\\|\\boldsymbol{\\xi}\\|^2=N$, and measure the 2-Wasserstein distance between the empirical law of $N^{1/3}R_{1,2}$ and the predicted $\\mathfrak{P}_{a(\\chi)}$; if the distance does not tend to zero, or $N^{2/3}\\mathbb{E}\\langle R_{1,2}^2\\rangle^{\\mathrm{sph}}$ drifts from $\\mathbb{E}_\\chi V_a(\\chi)$, the central claim is false. A standalone simulation of the reflected Airy$_1$ point process gives a concrete numerical value for the limit to compare against.","tokens_in":37379,"feed_emoji":"🎲","tokens_out":11329,"duration_ms":105208,"temperature":0.7,"pith_summary":"This paper proves that at the critical inverse temperature $\\beta=1$, the two-replica overlap $R_{1,2}$ in both the Ising and spherical Sherrington–Kirkpatrick models fluctuates on scale $N^{-1/3}$. The quenched distribution of $N^{1/3}R_{1,2}$ — the random probability measure one sees after averaging over the disorder — converges, in the 2-Wasserstein metric, to an explicit random probability measure built from a realization of the reflected Airy$_1$ point process. The same object gives the sharp second-moment asymptotics $N^{2/3}\\mathbb{E}\\langle R_{1,2}^2\\rangle\\to\\mathbb{E}_\\chi V_a(\\chi)>0$, answering Talagrand's Conjecture 11.7.5. A reader should care because this is the first exact description of the critical overlap law for the original SK model, tying spin-glass criticality to the GOE spectral edge.","feed_headline":"Critical spin-glass overlaps settle on Airy_1 law","feed_subtitle":"Two-replica overlap in Ising and spherical SK models gains a universal N^{-1/3} law; Talagrand's conjecture now follows.","key_machinery":"The load-bearing object is the random probability measure $\\mathfrak{P}_{a(\\chi)}$. For a realization $\\chi=(\\chi_k)_{k\\ge1}$ of the reflected Airy$_1$ point process, set $d_k=\\chi_k-\\chi_1$ and define $\\Delta(\\chi)$ as the unique positive zero of $\\Psi(x;\\chi)=1/(x+\\Xi(\\chi))+\\sum_{k\\ge2}(1/(x+d_k)-1/d_k)$, with $a_k=(\\Delta(\\chi)+d_k)^{-1}$. The measure $\\nu_a$ is the law of independent standard Gaussians $g_k$ conditioned on the constraint $\\sum_k a_k(g_k^2-1)=0$, constructed rigorously via projectively consistent finite-dimensional pins; two independent samples give $Q_a=\\sum_k a_k g_k^{(1)}g_k^{(2)}$, and $\\mathfrak{P}_a=\\mathrm{Law}(Q_a)$. This object carries the argument because the spherical Gibbs measure is exactly such a pin with coefficients $a_{N,k}=N^{-2/3}(\\gamma-\\lambda_k)^{-1}$; the reflected Airy$_1$ edge scaling of GOE eigenvalues (1.2) turns the finite pins into $\\mathfrak{P}_{a(\\chi)}$.","core_discovery":"The paper's central claim is Theorem 1.1: for both the Ising and the spherical SK model at $\\beta=1$, as $N\\to\\infty$, the law of the quenched random probability measure $\\langle\\delta(N^{1/3}R_{1,2})\\rangle$ converges in the 2-Wasserstein metric $\\mathcal{W}_2$ to the law of $\\mathfrak{P}_{a(\\chi)}$, an explicit random probability measure measurable with respect to a realization $\\chi$ of the reflected Airy$_1$ point process. The companion Corollary 1.2 identifies the limit of $N^{2/3}\\mathbb{E}\\langle R_{1,2}^2\\rangle$ with $\\mathbb{E}_\\chi V_a(\\chi)$, settling Talagrand's Conjecture 11.7.5. For the spherical model the proof is a direct random-matrix calculation: the Gibbs measure is an anisotropic Gaussian conditioned on $\\|\\boldsymbol{x}\\|^2=N$, and the overlap becomes a pinning of GOE edge eigenvectors. For the Ising model the proof is a sphere-to-cube transfer: the quenched overlap distributions of the two models asymptotically coincide, so the explicit spherical limit carries over.","pith_inferences":["The paper's own limitation is that the sphere-to-cube transfer is proved only at $\\beta=1$; for $\\beta_N=1+bN^{-1/3}$ with $b>0$, the authors state the proof breaks down, so the Ising analogue of the critical-window interpolation remains an open prediction.","The same conditioning construction should apply to other spherical-model observables, such as $k$-replica overlaps or linear statistics of the pinned Gaussian field, yielding further Airy$_1$-type limits beyond $R_{1,2}$.","A direct numerical evaluation of $\\mathbb{E}_\\chi V_a(\\chi)$ from simulated reflected Airy$_1$ point processes would pin down the limiting constant, providing a sharp test of the finite-$N$ convergence and its rate."],"forward_implications":["At $\\beta=1$ the overlap $R_{1,2}$ fluctuates on scale $N^{-1/3}$, and its quenched law is non-Gaussian and random for both the Ising and spherical SK models.","The sharp constant in Talagrand's Conjecture 11.7.5 is identified as $\\mathbb{E}_\\chi V_a(\\chi)$, a positive, explicitly defined number, instead of the previous $N^{-2/3}$ order-of-magnitude bound.","The critical Ising and spherical SK models share the same asymptotic quenched overlap law, so the previously mysterious Ising overlap is now controlled by the GOE spectral edge.","For the spherical model, the same proof adapts throughout the critical window $\\beta_N=1+bN^{-1/3}$, with $\\Delta_b$ the root of $\\Psi(\\Delta_b)=b$, interpolating from Gaussian to bimodal overlap laws."],"supporting_citations":[{"why":"States Conjecture 11.7.5 and supplies the critical-window overlap bounds this paper resolves and sharpens.","marker":"[Tal11]"},{"why":"Provides the Ising-to-sphere transfer tools: the $O(N^{-1/3})$ second-moment bound for $X_N$, exponential overlap tails, and local density comparison used in Section 3.","marker":"[DH26]"},{"why":"Supplies the Airy$_1$ edge fluctuation estimates and asymptotic identities (Propositions 1.3 and the gap estimates in Section 4) that define the limiting object.","marker":"[LS22]"},{"why":"Gives the textbook GOE edge eigenvalue convergence to the reflected Airy$_1$ point process invoked in (1.2).","marker":"[For10]"},{"why":"Provides GOE eigenvalue rigidity estimates used in Lemma 4.3 to control inverse-gap tails.","marker":"[EYY12]"},{"why":"Establishes the Gaussian high-temperature overlap CLT that forms the noncritical baseline the limit must match.","marker":"[GT02a]"}],"fun_headline_variants":["Critical SK overlap follows Airy_1 law","Overlap scaling N^{-1/3} in SK models proven","Talagrand's conjecture on SK overlap resolved","Airy_1 governs spin-glass overlap at criticality","Two-replica overlap in SK gets explicit limit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The Ising result rests on the companion bounds of [DH26]: $\\mathbb{E}[(Z_N/Z_N^{\\mathrm{sph}}-1)^2]=O(N^{-1/3})$, exponential tails for $N^{1/3}R_{1,2}$, and a local density comparison; if any of these fail, the sphere-to-cube comparison of Theorem 1.9 collapses.","fun_headline_variants_meta":{"raw":{"variants":["Critical SK overlap follows Airy_1 law","Overlap scaling N^{-1/3} in SK models proven","Talagrand's conjecture on SK overlap resolved","Airy_1 governs spin-glass overlap at criticality","Two-replica overlap in SK gets explicit limit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000662,"raw_usage":{"total_tokens":3102,"prompt_tokens":1102,"completion_tokens":2000,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":718,"completion_tokens_details":{"reasoning_tokens":1923}},"tokens_in":718,"tokens_out":2000,"duration_ms":18193,"temperature":1.0,"reasoning_tokens":1923,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:25:04.053790+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the spherical model at large $N$: diagonalize a GOE matrix, sample the anisotropic Gaussian conditioned on $\\|\\boldsymbol{\\xi}\\|^2=N$, and measure the 2-Wasserstein distance between the empirical law of $N^{1/3}R_{1,2}$ and the predicted $\\mathfrak{P}_{a(\\chi)}$; if the distance does not tend to zero, or $N^{2/3}\\mathbb{E}\\langle R_{1,2}^2\\rangle^{\\mathrm{sph}}$ drifts from $\\mathbb{E}_\\chi V_a(\\chi)$, the central claim is false. A standalone simulation of the reflected Airy$_1$ point process gives a concrete numerical value for the limit to compare against.","supporting_citations":[],"review_version":1}