{"id":"ad7e5989-4410-44a5-8d60-c9bd918b28e3","arxiv_id":"2601.17763","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"In edge-based SIR epidemics on static networks, the mean degree, effective degree, and secondary-case count of infected nodes all peak (or monotonically decline) before half the time to peak prevalence.","lead":"This paper derives exact formulas for how three measures of superspreading—an infected person's total contacts, susceptible contacts, and expected onward transmissions—change over the course of a network SIR epidemic. It proves these measures peak or fall well before infection prevalence peaks, so contact-based interventions are most valuable at the very start of an outbreak.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5's strict inequality appears unsupported: the derived asymptotics give lim t_m/t_I = λ1/(2λ1+γ), equal to the stated upper bound, not less.","rationale":"The reader identified the main scope limitation (static configuration-model networks, random initial infections) as the weakest assumption. That is a legitimate caveat but not an internal flaw. My stress-test pass found an internal analytical issue in the proof of Theorem 5: the strict upper bound on the limiting ratio t_m/t_I appears to be equality in the θ(0)→1 limit. The proof in the Supplement derives t_m ~ L/(2λ1+γ) and t_J ~ L/λ1, which yields a limit exactly equal to the stated RHS. Thus the theorem as written is false, though the paper's core applied message (superspreading peaks well before the prevalence peak) remains true because the RHS is < 1/2. This warrants a correction to the theorem statement or a proof of strictness; hence the verdict should be CONDITIONAL rather than unconditional ACCEPT. No other load-bearing flaw surfaced: the derivations of the three metrics, the peak-time comparisons for m_E and m_Z, and the simulation support are consistent with the stated model assumptions.","tokens_in":28021,"tokens_out":13661,"duration_ms":129846,"concrete_test":"Independently re-derive the asymptotic slopes: from θ(t) ≈ 1 − (1−θ0)e^{λ1 t} for t up to t_m, show t_m ~ L/(2λ1+γ) with L=log(1/(1−θ0)), and from the θ dynamics near θ=1 show t_I ~ L/λ1 + O(1). If both slopes are confirmed, the limit equals λ1/(2λ1+γ), so Theorem 5 must be weakened from '<' to '≤'. A numerical check on a Poisson network with μ=5, γ=1, R0=3 (β=1.5), integrating Eqs. (1)–(4) and (20) for θ(0)=1−10^{-k}, k=8,...,14, should show t_m/t_I approaching 5/11≈0.454545 rather than a strictly smaller value.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Theorem 5 claims 0 < lim_{θ(0)→1} t_m/t_I < R where R = 1/2 − γ/(4(β+γ)(R0−1)+2γ) = λ1/(2λ1+γ) with λ1=(β+γ)(R0−1). The proof's own asymptotics contradict the strict '<'. Supplement Eq. (S38) gives t_m ≈ [log(1/(1−θ0)) + c_m]/(2λ1+γ). For t_I, the lower bound in Eq. (26) gives t_J ≥ [log(1/(1−θ0)) + log(1−θ_J)]/λ1, but because θ(t) spends a time ~log(1/(1−θ0))/λ1 near θ=1 before approaching θ_J, t_J (and hence t_I, since t_I = t_J + O(1)) has asymptotic slope 1/λ1. Therefore lim t_m/t_I = λ1/(2λ1+γ) exactly. The strict inequality in Theorem 5 is not justified; the correct statement would be '≤' (or an explicit limit-superior interpretation). The qualitative conclusion 'peak times are less than half the prevalence peak time' survives because R < 1/2, but the precise theorem as stated is false. This is the most load-bearing concern because the reader's strongest claim rests directly on this theorem.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies SIR epidemics on static configuration-model networks using the edge-based model of Miller and Volz. It defines three time-dependent measures of superspreading potential: the mean degree of infected nodes m(t), the mean number of susceptible neighbors of infected nodes m_E(t), and the mean number of secondary cases generated by nodes newly infected at time t, m_Z(t). For each measure, the authors derive closed-form expressions or differential equations, characterize peak times and peak values, and give limiting behavior. The central quantitative claim is that the superspreading peak time t_m of m(t) is less than half the prevalence peak time t_I, with a tighter bound stated in Theorem 5, and that m_E and m_Z peak no later than t_m. The analytic results are supported by configuration-model simulations on Poisson and negative-binomial networks, and code is provided.","tokens_in":28354,"tokens_out":11171,"duration_ms":108613,"significance":"If the technical claims are correct, this is a useful contribution to network epidemic theory. The paper gives explicit, non-empirical formulas for how the composition of the infected population changes over time, and it connects those formulas to practical questions about when contact-based interventions are likely to be effective and how dispersion estimates depend on the data source. Strengths of the manuscript include the derivation of exact moment equations from a low-dimensional edge-based model, the explicit analytic formulas for peak values, the provided simulation code, and the careful discussion of limitations (uniform random initial infections, static configuration-model networks, negligible clustering). The main mathematical results are stated as theorems with proof sketches in the main text and fuller arguments in the supplement, and the simulation curves agree with the analytic trajectories. However, one of the headline theorems, Theorem 5, contains a strict inequality that is not supported by the authors' own asymptotic analysis; this must be corrected.","major_comments":[{"comment":"The strict inequality in Theorem 5 appears false as stated. The supplement's Eq. (S38) gives t_m ~ [log(1/(1-θ0)) + O(1)]/(2λ1+γ), where λ1=(β+γ)(R0-1). The paper's own bounds for t_J, Eq. (26), together with the argument that t_I = t_J + O(1) (Eq. (S33) and surrounding text), imply t_J ~ t_I ~ log(1/(1-θ0))/λ1. Combining these gives lim t_m/t_I = λ1/(2λ1+γ), which is exactly equal to the right-hand side of Eq. (25), not strictly less than it. The statement should be corrected to a non-strict inequality, or better, to the equality of the limit with the displayed bound. The qualitative conclusion that t_m < t_I/2 survives, since the bound λ1/(2λ1+γ) is itself strictly less than 1/2. Because Theorem 5 is highlighted as a main result and used in Table 2 and the discussion, the theorem and all dependent statements need adjustment.","section":"Theorem 5, Eq. (25); Supplement Eq. (S38), Eq. (26), Eq. (S33)"}],"minor_comments":[{"comment":"The peak-time entry for m_E(t) is reversed relative to Theorem 8: it reads '0 if ν/µ >2, else >0 and < t_m', but Theorem 8 states the opposite (peak at 0 if ν/µ ≤2, and later peak if ν/µ >2).","section":"Table 2, m_E(t) row"},{"comment":"Multiple occurrences of the limit 'θ(0)→∞' should read 'θ(0)→1' (main text after Eq. (27) and Supplement proof of Theorem 5).","section":"Theorem 5 proof and Supplement"},{"comment":"The sentence 'then the moving target will always to µ' appears to have a missing phrase; it should read 'will always be ≤ µ' or similar.","section":"Theorem 8 proof, Eq. (44)"},{"comment":"The caption says 'effective degree distribution’s variance v_Z(t)', but v_Z is the variance of the secondary case distribution Z(t), not the effective degree distribution.","section":"Supplement, Fig. S3 caption"},{"comment":"The statement 'm_Z(0)<R0' is derived under log-convexity of ψ′; the wording in Table 2, 'm_Z(t)<R0 (if ψ′ log-convex)', is fine because m_Z is decreasing, but the main text should make the distinction between the general formula and the inequality under the log-convexity assumption explicit.","section":"Theorem 10, Eq. (49)–(50)"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically strong overall and the error in Theorem 5 is localized; I do not see grounds for rejection. However, the theorem as stated is false, and because it is one of the headline results, the authors must correct the strict inequality (or replace it with the asymptotic equality) and update Table 2 and the discussion accordingly. After that, the paper should be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the reader's verdict is about right on the merits, with one real correction — the stress-test note lands. This is a careful paper; the three metrics (mean degree of infected nodes, mean susceptible-neighbor count, mean secondary cases) are genuinely new within the edge-based framework, and the moment equations (Theorem 1), the ODE for m_E (Theorem 7), and the integral formula for m_Z (Theorem 10) are real analytic contributions. Simulations match the analytic curves, the code is public, and no parameters are tuned. Earlier work had the qualitative fact of declining superspreading; this paper pins it down exactly.\n\nBut Theorem 5 overstates its case. As claimed, 0 < lim t_m/t_I < λ1/(2λ1+γ), where λ1 = (β+γ)(R0−1). The proof's own Eq. (S38) gives t_m ≈ log(1/(1−θ0))/(2λ1+γ), and t_I is asymptotic to t_J ≈ log(1/(1−θ0))/λ1, since the epidemic spends a time of order log(1/(1−θ0))/λ1 near θ≈1 where the growth rate is λ1. The ratio therefore converges to λ1/(2λ1+γ) exactly — the stated upper bound, not something strictly below it. The correct statement is '≤' or a limsup formulation. This matters because Theorem 5 is the paper's headline result. The good news: the qualitative conclusion survives, because λ1/(2λ1+γ) < 1/2, so the 'peak before half the prevalence peak' claim still holds as a strict inequality against 1/2. This is a theorem-statement fix, not a fatal flaw.\n\nEverything else holds up. Theorem 3's convergence to the neighbor degree distribution and Corollary 4 are fine. The m_Z monotonicity claim is honestly qualified by log-convexity of ψ'. The assumptions — static configuration network, negligible clustering, uniform initial infections — are stated plainly and are the standard price of exactness. The paper is also honest that the most direct superspreading metric (m_Z) is the least tractable, which is worth noting but not a flaw. Minor typos: 'θ(1)' in Theorem 3's proof sketch, 'ν/µ > µ' where they mean ν/µ > 2, and 'lim θ(0)→∞' in the supplement.\n\nWho should read this: network epidemic modelers and anyone estimating dispersion from incidence or contact-tracing data — the discussion of how the choice of X versus Z changes inference is useful. It deserves referee time. I'd send it out with instructions to fix Theorem 5 and clean up the typos; after that, it is publishable.","headline":"Solid, genuinely new analytic account of time-varying superspreading in network SIR; but Theorem 5's strict inequality is unsupported — their own asymptotics give the limit equal to the stated bound, so '≤' is the defensible statement.","tokens_in":28801,"tokens_out":8917,"would_cite":true,"duration_ms":86346,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92D30","05C82"],"pacs":[],"model":"deepseek-v4-flash","headline":"In SIR epidemics on random networks, all three superspreading metrics peak in less than half the time to peak infection prevalence, so contact-based control is most effective at the very start of an outbreak.","keywords":["superspreading","SIR epidemic","contact networks","edge-based model","infected degree distribution","effective degree distribution","secondary case distribution","peak time"],"falsifier":"Run the model on a highly clustered or assortative network with the same degree distribution and measure the ratio of the superspreading peak time to the prevalence peak time; if the observed ratio equals or exceeds 1/2 - γ/(4(β+γ)(R0-1)+2γ), the configuration-model assumption is violated and the central claim fails. A simpler check is a simulation on a small-world network with the same degree sequence.","tokens_in":27950,"feed_emoji":"🦠","tokens_out":4691,"duration_ms":53591,"temperature":0.7,"pith_summary":"This paper asks whether the role of superspreading changes over the course of an epidemic and provides exact mathematical answers for SIR epidemics on static random networks. It defines three time-varying statistics: the average number of contacts of infected individuals, the average number of susceptible contacts, and the average number of secondary transmissions caused by newly infected individuals. The central finding is that all three statistics peak earlier than the population-level infection peak, with the infected-degree peak occurring in less than half the time to peak prevalence. This implies that superspreading is already waning as an epidemic approaches its worst phase, which has practical consequences for when contact-based interventions are worth deploying and for how dispersion estimates from different data sources should be interpreted.","feed_headline":"Superspreading peaks at less than half the epidemic peak time","feed_subtitle":"The mean contacts, exposures, and transmissions of infected nodes all peak early, so contact-based controls work best at the start.","key_machinery":"The edge-based SIR model for static configuration networks, expressed through probability-generating functions ψ(θ) and φ(y), together with the newly defined infected-degree, effective-degree, and secondary-case distributions X(t), E(t), and Z(t). The central identity is the moment equation ṁ_n = -(J/I)(m_n - φ^(n+1)(log θ)/φ'(log θ)), where J/I is the infection turnover rate; this shows that every moment of the infected-degree distribution is continuously attracted to the corresponding moment of the neighbor-degree distribution, and the rate of that attraction controls the timing of the superspreading peak.","core_discovery":"The paper proves that, in the limit of a small uniformly random initial infection on a large configuration-model network, the mean degree of infected nodes (the superspreading peak m(t)) peaks at a time t_m that satisfies 0 < lim t_m/t_I < 1/2 - γ/(4(β+γ)(R0-1)+2γ), where t_I is the peak time of infection prevalence and R0 the basic reproduction number. The mean effective degree (susceptible neighbors of infected nodes) peaks even earlier, before t_m, and the mean secondary-case distribution peaks at t=0 and decreases monotonically under a log-convexity condition. These results formalize and quantify the intuition that infections cascade from high-degree to low-degree nodes, so the potential","pith_inferences":["The peak-time bound is proven for static configuration-model networks; on real clustered or dynamic contact networks the ratio t_m/t_I could be larger or smaller. A natural test is to simulate SIR on empirical networks with the same degree sequence and check whether the superspreading peak still precedes half the prevalence peak.","Because the small-seed limit sends both t_m and t_I to infinity while keeping their ratio bounded away from zero, the bound is an asymptotic statement; finite outbreaks with larger seed fractions may show a different ratio, and simulation studies varying the initial seed size could map this boundary.","The monotone decrease of the secondary-case mean implies that the most overdispersed phase is the earliest one, so case data collected after an outbreak is formally detected may systematically miss the phase where superspreading matters most; early outbreak-size data are the most informative for estimating the dispersion parameter.","The framework's generating-function approach could be extended to time-varying effective reproduction numbers on networks and to SIS or SEIR dynamics, provided the moment attraction structure is preserved; such extensions would need to check whether the peak-time inequality continues to hold."],"forward_implications":["Contact-based control strategies that target high-degree individuals are best deployed as early as possible, before the potential for superspreading has already declined, and may be much less effective once incidence approaches its peak.","Estimates of dispersion from incidence, contact tracing, or mobility data will differ systematically depending on whether they track transmissions, exposures, or contacts, and on the time window over which they are aggregated, because the three metrics peak at different times.","The mean secondary-case distribution decreases from the start (when ψ' is log-convex), so the effective reproduction number measured later in an epidemic will be lower and may understate the early superspreading that drove the outbreak.","The infected-degree distribution at the superspreading peak equals the neighbor-degree distribution, giving concrete predictions: for a Poisson network the infected degree is approximately Poisson shifted by one, and for a negative-binomial network it is approximately negative binomial shifted by one.","Since the superspreading peak occurs before the infection peak, epidemic phases defined by rising or falling incidence are not aligned with the phases of superspreading importance; using the superspreading peak to define phases could improve inference and intervention design."],"fun_headline_variants":["Superspreading peaks before the outbreak does","Contact controls work best early, math proves","Superspreading drops before infection peaks—act fast","High-contact superspreaders matter at outbreak start","Proved: superspreading peaks in first half of outbreak"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The core results assume the contact network is a large random graph with no clustering, no degree correlations, and no short cycles, and that the initial infections are uniformly random; real networks with household or workplace structure, or with seeding biased toward high-degree individuals, may not obey the exact peak-time inequality.","fun_headline_variants_meta":{"raw":{"variants":["Superspreading peaks before the outbreak does","Contact controls work best early, math proves","Superspreading drops before infection peaks—act fast","High-contact superspreaders matter at outbreak start","Proved: superspreading peaks in first half of outbreak"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000835,"raw_usage":{"total_tokens":3506,"prompt_tokens":796,"completion_tokens":2710,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":2635}},"tokens_in":540,"tokens_out":2710,"duration_ms":22572,"temperature":1.0,"reasoning_tokens":2635,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T08:09:49.813254+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the model on a highly clustered or assortative network with the same degree distribution and measure the ratio of the superspreading peak time to the prevalence peak time; if the observed ratio equals or exceeds 1/2 - γ/(4(β+γ)(R0-1)+2γ), the configuration-model assumption is violated and the central claim fails. A simpler check is a simulation on a small-world network with the same degree sequence.","supporting_citations":[],"review_version":1}