{"id":"29104446-bd89-4427-8e91-8be254ad79a6","arxiv_id":"2607.23427","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"partial","parameter_count":0,"one_line_summary":"Overlap concentration and O(1/N) free-energy convergence to the RS formula hold in the SK model whenever β² q/arctanh(q) < 1, covering all β<1 and some β>1 with nonzero field.","lead":"The authors extend a classical spin-glass proof so the free energy of the SK model matches the replica-symmetric formula with 1/N error on a larger parameter region, including some temperatures past the old β=1/2 barrier. Specialists tracking rigorous mean-field spin glass theory will care because the argument is short, explicit, and Lean-checked.","discovery_kind":"extension","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"No significant objection identified","rationale":"The reader’s ACCEPT with high confidence is consistent with my stress test. I agree that the weakest point is the imported uniqueness of the relevant RS solution, but I do not regard it as a reason to move the verdict: the note names the dependence in footnote 1, and the new contribution is the interpolation/coupled-free-energy argument, whose constants and inequalities check out. The claimed region is modest and correctly distinguished from the AT line; the quantitative O(N^{-1}) statements follow once ρ<1 gives λ*>0 and the endpoint mgf condition holds. The strongest residual action is verification hygiene rather than a mathematical objection: make sure the Lean artifact and bibliography pin down the uniqueness theorem and remove the λ* notation ambiguity. Those are confirmation steps, not evidence against the theorem.","tokens_in":8493,"tokens_out":6990,"duration_ms":58153,"concrete_test":"Compile the cited Lean 4 formalization and inspect the theorem corresponding to Theorem 1.1: confirm it has no `sorry`/extra axioms beyond mathlib, states λ* as (κ(q)^{-1}−β²)/4, and makes the uniqueness/choice of q an explicit hypothesis or imports a precisely cited Guerra–Talagrand lemma. If uniqueness is only an axiom or the citation does not cover the claimed β>1,h≠0 interval, downgrade; otherwise the ACCEPT verdict stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I did not find a load-bearing internal gap in the central argument. The endpoint Kearns–Saul/Schlemm bound is derived with the correct ±1 constant κ(q)=q/arctanh q (limits q→0 and q→1 behave properly), the Gaussian transform to (9) has the right condition κ(q)Λ<1, the covariance derivative in (13) matches the smart-path Hamiltonian including the harmless −β²/2 diagonal term, and the coupled-energy estimate uses convexity in the correct direction: ∂ΛF(s,0+)=ν_s[Q12²]/4 and Λ_s≥2λ* give the Gronwall rate β²/(2λ*)=2ρ/(1−ρ). The final free-energy identity then yields the one-sided O(1/N) bound exactly as stated. The only real dependence is the explicitly flagged uniqueness of the RS fixed point q, imported via Guerra/Talagrand; that is an external citation rather than an inconsistency in the note. The scope is also honest: ρ<1 is not the AT line, and no full AT resolution is claimed. Minor presentational issue: the display for λ* reads typographically ambiguous, but the equality λ*=(1−ρ)/(4κ) fixes the intended meaning (κ(q)^{-1}−β²)/4.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"This note studies the Sherrington–Kirkpatrick model with a deterministic external field. Let q solve the replica-symmetric fixed-point equation q = E tanh²(h + β√q Z), and set κ(q) = q/arctanh q (κ(0)=1) and ρ(β,h) = β²κ(q). The main result (Theorem 1.1) is that whenever ρ < 1, the overlap concentrates at rate E⟨(R₁₂−q)²⟩ ≤ C/N and the free energy satisfies 0 ≤ ϕ_RS − ϕ_N ≤ C/N. The proof refines Latała's interpolation argument: (i) at the independent endpoint of Guerra's smart path, the Kearns–Saul inequality (proved in §2.1 via concavity of t ↦ log cosh √t) gives a subgaussian bound for the overlap with optimal constant κ(q), and a Gaussian transform yields an exponential moment bound (9); (ii) a coupled free energy ϕ_N(s,Λ) for two replicas with quadratic interaction ΛNQ²₁₂/2 is introduced, its s-derivative is computed by Gaussian integration by parts along the smart path, and convexity in Λ converts the bound into a Gronwall inequality (Proposition 2.1) with rate 2ρ/(1−ρ); (iii) overlap concentration follows by a further convexity argument, and integration of the interpolation identity (21) yields the one-sided O(1/N) free-energy bound. The region ρ<1 strictly contains β<1 and, for h≠0, a nontrivial interval with β>1; the authors do not claim the full Almeida–Thouless region. A Lean 4 formalization is provided.","tokens_in":8722,"tokens_out":7636,"duration_ms":254317,"significance":"The result extends Latała's simple interpolation proof of replica symmetry from β<1/2 to the strictly larger region ρ<1, which contains all of β<1 and, for every h≠0, a nontrivial interval of β>1; the region is not the full AT region, and the authors are explicit about this. Two features add weight beyond the (correct) analysis: the quantitative rate O(N^{-1}) for both overlap concentration and the free-energy gap is optimal in order and is obtained with an explicit constant; and the authors ship a Lean 4 formalization of the argument (reference [9]), which materially raises confidence in the computational steps of §2.2. The conceptual contribution — that the restriction to β<1/2 in Latała's argument is an artifact of the method rather than of the interpolation scheme itself, removable via a coupled free energy — is a genuine, if modest, advance. Novelty is limited by the stronger condition already obtained in [4] and the announced full-AT result of Lopatto [8], but as a short, self-contained note with a machine-checked proof this is a useful addition to the literature.","major_comments":[{"comment":"Footnote 1 states that in the parameter region considered, 'the relevant solution is unique by Guerra's result; see [13]' — but [13] is Talagrand's book, and no precise theorem or section is given. This matters for the reading of Theorem 1.1: the condition ρ(β,h)<1 in (3) is strictly larger than Talagrand's classical high-temperature region (it covers β>1 with h≠0), so it is not obvious that the cited uniqueness results cover the whole region ρ<1. Two concrete requests: (a) give a precise citation (volume, theorem) establishing uniqueness of the fixed point of (2) throughout {ρ<1}, or restrict the statement accordingly; (b) note that the proof itself does not appear to need a priori uniqueness: applied to each solution q_j of (2) with β²κ(q_j)<1, the argument yields ϕ_RS(q_j) = ϕ_N + O(1/N), so all such replica-symmetric values coincide in the thermodynamic limit. Making this explicit wo","section":"§1, footnote 1 and Theorem 1.1"}],"minor_comments":[{"comment":"The display λ∗ := κ(q)−1 −β 2/4 is typographically ambiguous (it reads as κ(q) − 1 − β²/4). The subsequent equality λ∗ = (1−ρ)/(4κ(q)) fixes the intended meaning (κ(q)^{-1} − β²)/4; please add braces or parentheses.","section":"Eq. (10)"},{"comment":"The author name 'Latała' is repeatedly rendered as 'Lata la' (title, abstract, §1, references), presumably an encoding issue.","section":"Throughout"},{"comment":"The Almeida–Thouless condition is displayed without citing [2]; conversely, references [5], [11], [14] appear in the bibliography but are never cited in the text. Please reconcile the citation list.","section":"§1 (after Remark 1.2) and References"},{"comment":"The one-sided inequality 0 ≤ ϕ_RS − ϕ_N is obtained here directly from (21) and ν_t[Q²₁₂] ≥ 0; it would be helpful to remark that the same bound is Guerra's interpolation upper bound [6], for orientation.","section":"§2.4"},{"comment":"The constant C(β,h) ∝ e^{2ρ/(1−ρ)} log(2/(1−ρ)) / λ∗ diverges as ρ ↑ 1. A one-sentence remark that the O(1/N) rate is not uniform up to the boundary of the region would help the reader interpret Theorem 1.1.","section":"§2.3, Eq. (20)"},{"comment":"In §2.1 the claim that X_i = σ¹_iσ²_i are i.i.d. under the annealed law E⟨·⟩_0 uses independence of the z_i together with E m_i² = q; one extra line making this conditioning explicit would smooth the reading. Also, in (9) the Gaussian transform being used (completion of the square in u) could be named.","section":"§2.1"}],"recommendation":"minor_revision","confidential_remarks":"Two points for the editor's awareness. (1) Novelty: the condition (3) is slightly weaker than the Brennecke–Yau [4] condition obtained by Bolthausen's Morita method, and the preprint by Lopatto [8] (cited by the authors) claims the full AT region. The contribution here is therefore methodological (a short interpolation with a new coupled-energy ingredient and optimal O(1/N) rate) plus the Lean formalization, rather than a new phase diagram; the authors present it honestly as a note, and I think it clears the bar on those terms. (2) The manuscript includes an unusually detailed AI-assistance disclosure (ChatGPT/Codex for calculations, Codex/Aristotle for the Lean development); the authors take responsibility, and the presence of the machine-checked proof mitigates verification concerns, but the editor may wish to confirm the journal's policy on such disclosures."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This note does exactly what the title says: it pushes Latała’s interpolation for SK with field past the old β<1/2 barrier. The new region is ρ=β² κ(q)<1 with κ(q)=q/arctanh q. That covers all β<1 and, for h≠0, a strip of β>1. They get both overlap concentration E⟨(R12−q)²⟩≤C/N and the one-sided free-energy bound 0≤ϕ_RS−ϕ_N≤C/N.\n\nWhat is actually new is the combination of the coupled free energy along Guerra’s smart path, the specific λ*=(κ(q)^{-1}−β²)/4, and the endpoint MGF bound (derived via concavity of log cosh√t, or Kearns–Saul). The derivatives, Gronwall step, and convexity readout are standard and written out carefully. The authors also ship a Lean 4 formalization; that is real credit, not decoration.\n\nSoft spots are minor and already flagged by the authors. Uniqueness of the relevant fixed point q is taken from Guerra/Talagrand rather than reproved; if that failed the target formula would be ambiguous, but it is an external citation, not an internal hole. The condition is strictly weaker than the AT line, and they say so; a concurrent preprint claims the full AT region by other methods. The λ* display is typographically a bit messy, but the algebra fixes the intended value. None of this breaks the argument.\n\nThis is for people who work on rigorous mean-field spin glasses and care about quantitative rates and interpolation technology. It is short, honest about scope, and checkable. I would send it to referees without hesitation; it is a solid note, not a claim of the full AT resolution.","headline":"Clean extension of Latała past β=1/2 to ρ=β²q/arctanh q<1, with full write-up, O(1/N) rates, and a Lean formalization; not the AT line, and uniqueness of q is imported.","tokens_in":9757,"tokens_out":499,"would_cite":true,"duration_ms":8427,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","82B44","82D30"],"pacs":[],"model":"grok-4.5","headline":"Refining a classic interpolation shows replica symmetry and 1/N overlap concentration in the SK model whenever β² q/arctanh(q) is less than 1.","keywords":["Sherrington–Kirkpatrick model","replica symmetry","overlap concentration","Guerra interpolation","Kearns–Saul inequality","external field","free energy"],"falsifier":"Exhibit parameters (β,h) with β² q/arctanh(q) < 1 for which the annealed second moment of (R₁₂ − q) stays bounded away from zero as N → ∞, or for which ϕ_RS − ϕ_N fails to be O(1/N).","tokens_in":9665,"feed_emoji":"❄️","tokens_out":958,"duration_ms":21188,"temperature":0.7,"pith_summary":"The Sherrington–Kirkpatrick spin glass with an external field is expected to be replica-symmetric in a large region of temperature and field, but elementary proofs have lagged behind that expectation. This note sharpens an interpolation argument of Latała, previously stuck at β < 1/2, by adding a coupled free energy and a sharp moment-generating bound of Kearns–Saul type. The outcome is quantitative: whenever β² times q/arctanh(q) is strictly less than 1, the overlap concentrates about its replica-symmetric value at rate 1/N and the free energy sits within C/N of the replica-symmetric formula. That region contains every inverse temperature below 1 and, for any nonzero field, a nonempty interval of temperatures above 1. A sympathetic reader cares because the argument stays elementary, gives an explicit error rate, and pushes a transparent method past the artificial 1/2 barrier into part of the physically interesting high-temperature side of the Almeida–Thouless line.","feed_headline":"SK overlap concentrates at 1/N past the old β=1/2 barrier","feed_subtitle":"A refined interpolation reaches all β<1 and some β>1 whenever β² q/arctanh(q)<1","key_machinery":"Guerra’s smart-path interpolation together with a two-replica coupled free energy whose Λ-derivative controls the quadratic overlap; the path is closed by a Kearns–Saul-type bound that produces the factor κ(q) at the decoupled endpoint, then Gronwall.","core_discovery":"Under the condition ρ(β,h) = β² κ(q(β,h)) < 1, where q solves the replica-symmetric fixed-point equation and κ(q) = q/arctanh(q) (with κ(0)=1), the SK model with deterministic field satisfies E⟨(R₁₂ − q)²⟩ ≤ C/N and 0 ≤ ϕ_RS(β,h) − ϕ_N ≤ C/N. The same bound covers the whole half-plane β < 1 and, when h ≠ 0, a strip of β > 1 near the critical temperature.","pith_inferences":["Closing the remaining gap to the full Almeida–Thouless line likely needs a tighter endpoint inequality than κ(q), or a different coupling, since the present constant is locked by that factor.","The explicit 1/N rate suggests the method could supply quantitative input to dynamical or algorithmic analyses that only need high-temperature concentration.","Because the proof is short and already formalized, it is a natural test case for machine-checked extensions to diluted or multi-species SK-type models."],"forward_implications":["Overlap concentration at rate 1/N holds for every β < 1, with or without field.","For every nonzero external field the method reaches a nonempty interval of β > 1.","The free-energy error is explicitly O(1/N) with a constant controlled by ρ and κ(q).","The same interpolation-plus-coupling template can be checked against any candidate improvement of the endpoint mgf bound."],"fun_headline_variants":["SK overlap concentrates at 1/N past β=1/2 via refined Latała","1/N SK free-energy error under β²q/arctanh(q)<1","Overlap concentration for all β<1 and some β>1 in SK model","Refined Latała plus Kearns-Saul yields SK RS convergence at 1/N","SK model reaches RS formula at 1/N whenever β²κ(q)<1"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The argument treats the replica-symmetric fixed point q as the unique relevant solution in the region, importing uniqueness rather than proving it inside the note.","fun_headline_variants_meta":{"raw":{"variants":["SK overlap concentrates at 1/N past β=1/2 via refined Latała","1/N SK free-energy error under β²q/arctanh(q)<1","Overlap concentration for all β<1 and some β>1 in SK model","Refined Latała plus Kearns-Saul yields SK RS convergence at 1/N","SK model reaches RS formula at 1/N whenever β²κ(q)<1"]},"model":"grok-4.5","effort":"low","cost_usd":0.00568,"raw_usage":{"total_tokens":1521,"prompt_tokens":807,"num_sources_used":0,"completion_tokens":103,"cost_in_usd_ticks":56804000,"prompt_tokens_details":{"text_tokens":807,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":611,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":807,"tokens_out":103,"duration_ms":10382,"temperature":1.0,"reasoning_tokens":611,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T22:20:55.589976+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Exhibit parameters (β,h) with β² q/arctanh(q) < 1 for which the annealed second moment of (R₁₂ − q) stays bounded away from zero as N → ∞, or for which ϕ_RS − ϕ_N fails to be O(1/N).","supporting_citations":[],"review_version":1}