{"id":"12186cd3-2c6a-4cd1-b327-d29bc5ea016d","arxiv_id":"2608.05096","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A GNN-based projection for WBF di-Higgs at 3 TeV CLIC claims kappa_lambda in [0.76,1.31] and Z about 20, but the reported yields and efficiencies give only about 9.","lead":"This paper predicts that the 3 TeV CLIC collider, paired with a graph neural network, can probe the Higgs self-coupling via W-boson fusion di-Higgs production, claiming a 20 sigma signal and kappa_lambda sensitivity of [0.76,1.31] at 95% CL. The paper's own cross sections and efficiencies imply roughly 9 sigma at 5 ab^-1, so the headline numbers do not match the stated inputs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported Z≈20 at 5 ab^-1 is inconsistent with the paper's own cross-section table and efficiencies; Eq. (3) with those inputs gives Z≈9, undercutting the headline and all derived coupling intervals.","rationale":"The central claim is discovery-level sensitivity Z≈20 at 5 ab^-1. The paper supplies all ingredients needed to check this claim: cross sections (Table I), post-classification efficiencies (main text), and the significance formula (Eq. (3)). Plugging them in yields Z≈8.8—a factor of about 2.3 lower than claimed. This is not an external systematic assumption or a detector-transfer risk; it is an internal inconsistency in the paper's own stated numbers. Because the likelihood analysis for κ_λ and κ_2V uses the same event counts and efficiencies, the coupling intervals are built on the same unsupported significance. Even if the GNN performs exactly as simulated, the quoted yields cannot produce the stated result. The missing C_ij coefficients in Eq. (14) further prevent independent reproduction of the sensitivity contours, but the arithmetic check alone is decisive. The reader's weakest-assumption (fast-simulation transfer) is plausible but secondary; once the headline significance is corrected to about 9, the projected intervals would widen substantially regardless of systematics. I therefore recommend keeping the REJECT verdict and requesting a corrected version with consistent event counts or a detailed yield table.","tokens_in":12331,"tokens_out":5523,"duration_ms":56632,"concrete_test":"Recompute the significance from the paper's own numbers: take Table I cross sections (σ_hh = 0.845 fb, σ_hbb = 1.509 fb, σ_ttX = 21.19 fb) and the quoted post-classification efficiencies (ε_S = 0.02, ε_B = 6×10^-4) at 5 ab^-1; evaluate Eq. (3) with S = 0.845×0.02×5000 = 84.5 and B = (1.509+21.19)×6×10^-4×5000 = 68.1. The result is Z ≈ 8.8, not 20. If the authors intend the efficiencies to apply after the Table I selection, recompute as S = 135.2×(0.02/0.032) and B = (15.2+317.9)×(6×10^-4/0.002), which gives a similar Z. If the claimed Z≈20 still does not emerge, the headline significance and all derived intervals are unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline significance, Z≈20 at 5 ab^-1, is not reproducible from the paper's own quoted yields and efficiencies. Using the cross sections in Table I (σ_hh = 0.845 fb; backgrounds: hb\\bar{b}/E 1.509 fb + t\\bar{t}X 21.19 fb, total 22.70 fb) and the post-classification efficiencies stated in the main text (ε_S ≈ 2×10^-2, ε_B ≈ 6×10^-4), Eq. (3) gives S ≈ 84.5, B ≈ 68.1 at 5 ab^-1, hence Z ≈ 8.8. If the efficiencies are instead meant to apply after the Table I preselection, the product of selection and classification efficiencies yields a similar Z (≈9). A significance of 20 would require S/B ≈ 1.9 with S ≈ 90, i.e., a background of ≈47, roughly 30% lower than quoted, or correspondingly larger signal rates. The 1D coupling intervals in Eqs. (8)–(10) are derived from the same binned likelihood, so they inherit this inconsistency. Additionally, Eq. (14) presents the cross-section parameterization without the C_ij coefficients, making the κ_λ–κ_2V sensitivity contours impossible to verify. The internal arithmetic failure is decisive; the detector-transfer concern is secondary.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript studies W-boson fusion di-Higgs production at a future 3 TeV e+e− collider (CLIC) within the κ-framework, focusing on the e+e−→hhνν̄ final state with four b-jets. The authors simulate signal and backgrounds with MG5_aMC@NLO, Pythia8, and Delphes3, apply a set of kinematic preselections, and then use a graph neural network (GATv2) classifier to separate signal from backgrounds. They report a signal significance of Z≈20 at 5 ab−1 and derive projected 95% CL constraints on κλ and κ2V for two benchmark scenarios (κV=1 and κV=κ2V). The central claim is that this sensitivity substantially exceeds projected HL-LHC reach and discriminates between SMEFT-like and HEFT-like realizations of electroweak symmetry breaking.","tokens_in":12607,"tokens_out":10909,"duration_ms":120179,"significance":"If the reported sensitivity were reproducible from the numbers given, the study would provide a valuable projection for multi-TeV lepton colliders, demonstrating both a physics opportunity and a methodological application of graph-based machine learning to Higgs pair production. The paper includes a detailed description of the GNN architecture, a two-benchmark treatment of the κ-parameter space, and a binned likelihood analysis, and it correctly recognizes the qualitative importance of interference between κλ, κV, and κ2V amplitudes. However, the central quantitative claim is not supported by the paper's own inputs: the quoted cross sections, luminosities, and efficiencies do not yield Z≈20, and the derived coupling intervals inherit this problem. As a result, the significance of the work, even in the best case, is substantially diminished unless the analysis is reworked and the numbers corrected.","major_comments":[{"comment":"The claimed significances Z≈8, 15, 20 at integrated luminosities of 1, 3, and 5 ab−1 are inconsistent with the cross sections and efficiencies stated in the manuscript. Using the total signal cross section σ(hhνν̄)=0.845 fb and the total background cross section σ(hb b̄νν̄)+σ(tt̄X)=22.70 fb from Table I, together with the post-classification efficiencies ε_S≈2.0×10−2 and ε_B≈6×10−4 quoted in the text, Eq. (3) yields S=84.5 events and B=68.1 events at 5 ab−1, giving Z≈8.8, not 20. At 1 ab−1 the same calculation gives Z≈3.9, not 8. If the efficiencies are instead meant to be applied to the post-selection yields N in Table I, the significance is even smaller (approximately 1.4 at 1 ab−1). The significance-versus-threshold curve in Supplement Fig. 4 reaches Z≈20 only at thresholds near P≈0.3–0.4, but that is not what the text says: the text explicitly ties the quoted efficiencies at P(hhνν̄)≥0.5 to the Z values. The manuscript must present a consistent set of efficiencies, yields, and significances; as written, the headline result is not reproducible from the paper's own numbers.","section":"Event Classification, Eq. (3), Table I, Supplement D"},{"comment":"The cross-section parameterization σ(κλ,κ2V,κV) is presented in terms of coefficients C_ij, but the numerical values (or even closed-form expressions) of these coefficients are never provided. The reader therefore cannot reproduce the dependence of the signal rate on the coupling modifiers, cannot obtain the cross-section contours in Supplement Fig. 3, and cannot verify the binned likelihood results that lead to the intervals in Eqs. (8)–(10). The plots are not a substitute for the coefficients themselves. This is a reproducibility gap in a central part of the analysis.","section":"Supplement B, Eq. (14)"},{"comment":"The detector simulation is performed with Delphes3 without specifying a CLIC-specific detector card, and the systematic uncertainty treatment is limited to a single 10% normalization nuisance parameter (Eq. 24 in the supplement). Because the projected significance and the coupling constraints depend directly on the post-classification background rejection (ε_B≈6×10−4), the absence of an explicit validation of the Delphes model against CLIC performance, and the omission of shape uncertainties in the GNN score distribution, b-tagging efficiency, and jet energy scale, leave the results vulnerable to sizable corrections. The authors should either justify the adopted detector response and include a more realistic systematic model, or significantly temper the sensitivity claims.","section":"Event Classification and Supplement E"}],"minor_comments":[{"comment":"The title contains a typographical error: 'viaW' should be 'via W' with a space.","section":"Title"},{"comment":"The resolution parameters σ_h and σ_Δ in the χ² pairing statistic are not numerically specified; provide the values used in the analysis.","section":"Eq. (4)"},{"comment":"The text states that the background efficiency ε_B depends on κV (ε_B(κV=2)=0.00077, ε_B(κV=3)=0.00086), but it is unclear why a SM background rate should depend on an anomalous coupling modifier; clarify whether the backgrounds are reweighted to non-SM κV values and how this is implemented.","section":"Supplement D"},{"comment":"The test statistic Q(κ) is defined but the connection between Q(κ) and the quoted confidence intervals is not explicit; state the threshold used for 1D and 2D intervals (e.g., Δχ²=3.84 for 95% CL on one parameter).","section":"Section 'Sensitivity of Couplings'"}],"recommendation":"reject","confidential_remarks":"The internal arithmetic failure is decisive: the paper's headline significance Z≈20 at 5 ab−1 is not derivable from the cross sections, luminosities, and efficiencies reported in the manuscript. The same inconsistency affects the derived coupling intervals, which are the main physics output. The missing C_ij coefficients in Eq. (14) further prevent verification of the coupling sensitivity. While the study could in principle be reworked with corrected numbers, the current version does not meet the standard for publication in a serious journal; the central claim is unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the stress-test note is right and it lands. This is a useful study with a real amount of work in it, but the headline significance cannot be reproduced from the paper's own numbers, and the coupling intervals inherit that problem.\n\nWhat is genuinely good: the paper treats κ_V and κ_2V under two physically motivated scenarios rather than one default relation; the GATv2 architecture is described in unusual detail (features, edge definition, pooling, training recipe); and the point that hb\\bar{b}/E is the hard background while t\\bar{t}X separates cleanly is well made. That is a solid basis for a revised study.\n\nThe soft spots are load-bearing. Using the cross sections from Table I (0.845 fb signal, 22.70 fb background before selection) with the quoted post-classification efficiencies ε_S≈2e-2 and ε_B≈6e-4, Eq. (3) gives S≈84, B≈68, Z≈8.8 at 5 ab^-1, not ≈20. If the efficiencies are instead meant to be applied after the Table I selection, the result is far worse. No reading of the numbers I tried gives 20. The 1D and 2D coupling intervals come from the same likelihood, so they inherit the error. The withheld C_ij coefficients in Eq. (14) make independent checks impossible; this is a minor issue if the authors release them, but as submitted it is a real transparency gap.\n\nDetector transfer is a secondary concern, not the main one. I would not hang the rejection on Delphes without a CLIC card, though the paper should say which card it used and should at least discuss shape systematics on the GNN score. A single 10% normalization nuisance is not enough.\n\nNet: this is not a desk-reject-quality manuscript. It is a serious phenomenological study with an internal arithmetic failure in its central claim. A serious referee should see it, and the right editorial outcome is major revision with consistent numbers, released coefficients and samples, and a corrected significance. If the corrected Z is ~9 rather than 20, that still may be a useful result, but the paper has to say so. I would not cite the current version's projections.","headline":"The GNN design and the HEFT/SMEFT split are worth reading, but the headline Z≈20 does not survive arithmetic on the paper's own table; the current version needs major correction before any coupling projection can be trusted.","tokens_in":13252,"tokens_out":2934,"would_cite":false,"duration_ms":33055,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that W-boson fusion di-Higgs production at a 3 TeV lepton collider, analyzed with a graph neural network, can reach 20σ significance and constrain the Higgs self-coupling modifier to roughly 30% precision at 95%…","keywords":["Higgs self-coupling","di-Higgs production","W boson fusion","CLIC","graph neural network","kappa framework","electroweak symmetry breaking","future lepton collider"],"falsifier":"Run the same analysis through a full CLIC detector simulation with per-jet energy-scale and b-tagging uncertainties, then verify the GNN score distribution in background-enriched control regions; if the observed background efficiency at the signal working point is more than about twice the simulated 6×$10^{-4}$ while signal efficiency stays fixed, the projected Z≈20 at 5 $ab^{-1}$ would fall well below 20 and the 95% CL κλ interval would widen beyond the quoted [0.76,1.31].","tokens_in":12095,"feed_emoji":"⚛️","tokens_out":10446,"duration_ms":112577,"temperature":0.7,"pith_summary":"This paper aims to show that the Higgs self-coupling can be measured precisely at a future high-energy lepton collider, specifically the 3 TeV CLIC, using W-boson fusion di-Higgs production in the four-bottom-quark plus missing-energy final state. It argues that a graph neural network classifier, which treats each event as a graph of jets and uses attention-based message passing, can separate the small signal from much larger backgrounds well enough to reach a signal significance of about 20σ at 5 $ab^{-1}$. If that holds, CLIC would constrain the Higgs self-coupling modifier κλ to about [0.76,1.31] and the double Higgs-gauge coupling modifier κ2V to about [0.95,1.05] at 95% confidence, well beyond projected HL-LHC sensitivity. The same measurement would also distinguish linearly realized electroweak symmetry breaking from non-linearly realized frameworks because the two scenarios predict different correlations between κV and κ2V.","feed_headline":"Machine learning boosts CLIC Higgs probe to 20σ","feed_subtitle":"Simulated W-fusion Higgs pairs at 3 TeV would pin κλ to ±30% and κ2V to ±5% at 95% CL.","key_machinery":"The central object is a heterogeneous graph neural network classifier. Each reconstructed jet is a node with a 14-component feature vector (transverse momentum, angles, mass, particle multiplicities, energy ratios, b-tag and flavour information, time of flight, and jet-shape variables), a missing-energy node carries the MET information, and directed edges connect jets within angular distance ΔR≤1.5. Three GATv2 attention layers perform message passing, then pooled jet and MET embeddings feed a small multilayer perceptron that assigns each event a probability of being signal. This classifier is the machine that pulls the hhννbar signal out of single-Higgs and top-quark backgrounds; the coupling reach is then set by a binned profile likelihood over the reconstructed di-Higgs kinematic observables m_hh, Δη_hh, and Δφ_hh, using the cross-section parameterization σ = $κ_λ^{2}$ $κ_V^{2}$ C_11 + $κ_2V^{2}$ C_22 + $κ_V^{4}$ C_33 plus interference terms.","core_discovery":"The central claim is that WBF di-Higgs production at a 3 TeV e+e− collider, with each Higgs decaying to bbbar, is a discovery-level and precision-level probe of the Higgs potential. For the Standard Model benchmark, the paper's GNN classifier yields a micro-averaged AUC of 0.95 and post-classification signal and background efficiencies of about 2×$10^{-2}$ and 6×$10^{-4}$, translating to Z≈8, 15, and 20 at integrated luminosities of 1, 3, and 5 $ab^{-1}$. A binned profile-likelihood fit over reconstructed di-Higgs mass, rapidity separation, and azimuthal separation then gives, at 5 $ab^{-1}$ and 95% CL, κλ∈[0.76,1.31] and κ2V∈[0.95,1.05] when κV=1, and κ2V∈[0.90,1.10] when the SMEFT-like relation κV=κ2V is imposed. The author's conclusion is that high-energy lepton colliders combined with graph-based machine learning provide a powerful, discriminating probe of new physics in the electroweak sector.","pith_inferences":["Because only one global normalization systematic is included, the quoted intervals would likely widen once shape systematics in the GNN score distribution, b-tagging efficiency, and jet energy scale are propagated in a full detector simulation.","The 4b final state is only one of several hh decay channels; combining bbWW and bbττ in the same graph framework would probably tighten the κλ constraint beyond the reported values.","The strong κλ–κ2V correlation visible in the 2D contours suggests that a global simultaneous fit, rather than separate 1D intervals, is the right way to interpret a future measurement; a shift in one coupling could masquerade as a shift in the other.","If the GNN score calibration is validated in data control regions, the same classifier could be reused as a model-independent search for anomalous couplings, not just the two benchmark scenarios."],"forward_implications":["At 5 ab^-1, a 3 TeV CLIC run would make WBF di-Higgs production a discovery channel with Z≈20 for the Standard Model, and already Z≈8 at 1 ab^-1.","The trilinear Higgs coupling would be measured to about ±30% at 95% CL, a direct test of the shape of the Higgs potential.","The double Higgs-gauge coupling κ2V would be pinned to roughly ±5% in the HEFT-like scenario, offering a sharper test of the hhVV vertex than the LHC.","Comparing the two scenarios would give an indirect discriminator between SMEFT and HEFT: the allowed region in the κλ–κ2V plane changes shape and orientation depending on whether κV=κ2V.","The same analysis strategy, applied at a muon collider with similar center-of-mass energy, would likely deliver comparable sensitivity."],"supporting_citations":[{"why":"Provides the current ATLAS 95% CL limits on κλ and κ2V that the projected CLIC sensitivity is compared against.","marker":"[9]"},{"why":"Supplies the projected HL-LHC di-Higgs sensitivity used as the baseline that CLIC's κλ reach is claimed to exceed.","marker":"[11]"},{"why":"Supplies the profile-likelihood significance formula and statistical treatment used to convert simulated event counts into Z and coupling intervals.","marker":"[29]"},{"why":"Generates the leading-order signal and background samples whose cross sections and kinematics enter the expected yields.","marker":"[30]"},{"why":"Hadronizes the generated events, producing the jet multiplicities and particle-level features that feed the classifier.","marker":"[31]"},{"why":"Fast detector simulation that defines reconstructed jets, b-tags, and missing energy; the projected significance inherits its response.","marker":"[32]"},{"why":"Defines the R=0.4 anti-kT jet collection used for event selection and for reconstructing the two Higgs bosons.","marker":"[33]"},{"why":"Supplies the GATv2 graph attention mechanism that forms the core of the signal-background classifier.","marker":"[34]"}],"fun_headline_variants":["GNN classifier pins Higgs self-coupling to ±30% at 3 TeV CLIC","3 TeV W-fusion di-Higgs probe reaches 20σ with graph ML","CLIC 3 TeV W-fusion di-Higgs: 20σ Higgs self-coupling probe","High-energy lepton collider GNN probes Higgs potential to 20σ","Disentangle linear vs nonlinear EW breaking with CLIC Higgs pairs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The projection rests on the assumption that the GNN's simulated separation power—about 2% signal efficiency and 0.06% background efficiency at the chosen working point—will hold in the real CLIC detector, since the analysis uses fast simulation and propagates only a single global 10% normalization systematic, not shape systematics in the classifier score, b-tagging, or jet energy scale.","fun_headline_variants_meta":{"raw":{"variants":["GNN classifier pins Higgs self-coupling to ±30% at 3 TeV CLIC","3 TeV W-fusion di-Higgs probe reaches 20σ with graph ML","CLIC 3 TeV W-fusion di-Higgs: 20σ Higgs self-coupling probe","High-energy lepton collider GNN probes Higgs potential to 20σ","Disentangle linear vs nonlinear EW breaking with CLIC Higgs pairs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001041,"raw_usage":{"total_tokens":4396,"prompt_tokens":983,"completion_tokens":3413,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":3300}},"tokens_in":599,"tokens_out":3413,"duration_ms":26933,"temperature":1.0,"reasoning_tokens":3300,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:24:06.994863+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same analysis through a full CLIC detector simulation with per-jet energy-scale and b-tagging uncertainties, then verify the GNN score distribution in background-enriched control regions; if the observed background efficiency at the signal working point is more than about twice the simulated 6×$10^{-4}$ while signal efficiency stays fixed, the projected Z≈20 at 5 $ab^{-1}$ would fall well below 20 and the 95% CL κλ interval would widen beyond the quoted [0.76,1.31].","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the projected HL-LHC di-Higgs sensitivity used as the baseline that CLIC's κλ reach is claimed to exceed."}],"review_version":1}