{"id":"a7104473-304b-4a17-ad8d-0fe80c0c1f40","arxiv_id":"2607.26217","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"In an agent-based epidemic model, swapping only the mobility assumptions changes peak incidence from about 350k to 650k infected and shifts the peak by up to 5 days.","lead":"This paper runs one-million-person computer simulations of a flu-like outbreak in a synthetic Fairfax County, changing only how virtual people decide where to go. It finds that those mobility choices—not the disease itself—can roughly double the predicted peak of infections and shift it by days.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported contact-network metrics appear inconsistent with the log2(n) contact cap in §3.2; if the graphs are built from full co-location rather than transmission contacts, the paper's mechanism is unverified.","rationale":"The reader's weakest assumption flagged the log2(n) saturation as arbitrary and potentially ranking-altering; my concern extends this to an apparent internal inconsistency: the reported network degrees cannot be produced by the stated contact rule. This is more specific and goes to the heart of the mechanism. I still view the paper as conditionally acceptable because the epidemic curves themselves are genuine simulation outputs and the qualitative sensitivity to mobility assumptions may survive, but the network-level explanation is currently unverified. The concrete test—code inspection plus a contact-function sensitivity analysis—would settle whether the mechanism holds or whether the results are artifacts of the contact construction. I did not find grounds to reject outright, as the code is public and the issue is checkable, not necessarily fatal.","tokens_in":14118,"tokens_out":11273,"duration_ms":106932,"concrete_test":"Inspect the GitHub code to determine how the Table 3 contact graphs were constructed (full co-location vs log2(n)-capped transmission contacts). Then re-run the five mobility models under two conditions: (a) building the network only from the capped contacts as described in §3.2, and (b) using an alternative contact function such as sqrt(n) or no cap while keeping mobility fixed. If the Random Advan average degree drops below ~3,200 under (a), or if the ranking of peak incidence/R0 across models changes under (b), the central attribution to network structure is not robust and the paper's conclusions need revision.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim attributes epidemic differences to contact-network structure rather than movement volume, citing Table 3 metrics. But the connection between those metrics and the transmission model is unclear. §3.2 caps each agent's contacts at log2(n) per co-located group, with a fixed subset that remains constant while co-located. Over 50 days, Random Advan agents make ~160 visits (160M visits / 1M agents). Even if every visit were to a maximally crowded location (n = 1M, log2(n) ≈ 20), each agent could accumulate at most ~160 × 20 = 3,200 unique contacts. Yet Table 3 reports an average degree of 9,037 for Random Advan. This is arithmetically impossible under the stated cap, implying either the contact cap was not implemented as described, or Table 3 was built from full co-location pairs rather than the transmission-relevant capped contacts. If the latter, the network metrics used to explain epidemic outcomes are not the contacts that drove transmission, so the causal mechanism is unsupported. Moreover, the log2(n) saturation is an arbitrary functional form that suppresses per-capita contacts sublinearly; a different contact function (e.g., sqrt(n) or linear) could differentially alter crowded-POI models and change the ranking of epidemic outcomes across mobility models. The paper reports no sensitivity analysis to this choice.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a controlled agent-based modeling experiment with one million synthetic agents in a Fairfax County urban environment. Five mobility models vary two dimensions—activity patterns (empirical NHTS trip chains vs. randomized schedules) and destination choice (Advan popularity weights, distance decay, or random)—while disease parameters are held fixed. The authors find that mobility assumptions alone produce substantially different epidemic curves: NHTS-based models yield lower, later peaks; Random Advan yields the highest R0 and fastest growth; Random Places yields a high peak despite the lowest estimated R0. They interpret these differences as caused by contact network structure—modularity, path length, and degree—rather than by aggregate movement volume, which they argue is broadly comparable across models. The paper concludes that mobility assumptions deserve calibration and sensitivity testing in policy-facing ABMs.","tokens_in":14398,"tokens_out":5308,"duration_ms":58455,"significance":"If the causal attribution holds, the study makes a useful methodological contribution by isolating mobility as a first-class source of uncertainty in ABMs and by demonstrating that network structure, not just contact volume, drives epidemic outcomes. The experiment is large-scale, uses 30 stochastic runs with confidence intervals for the main epidemic metrics, and the code is publicly available. The systematic variation across five models and the discussion of limitations—including the explicit statement that the study does not validate against real outbreak data—are commendable. However, the central mechanism claim relies on contact-network metrics that must be generated by the same contacts used for transmission; the current manuscript contains an apparent inconsistency between the stated log2(n) contact cap and the reported mean degree for Random Advan, and this must be resolved before the mechanism can be accepted.","major_comments":[{"comment":"The reported mean degree for Random Advan (9037.76) appears arithmetically impossible under the stated contact cap. Random Advan generates ~160 visits per agent over 50 days (Table 2: 159,988,454 visits / 1,000,000 agents). With the cap of at most log2(n) contacts per co-located group, the absolute upper bound is 160 × 20 = 3,200 unique contacts per agent if every visit were to a maximally crowded location (n=1M) and contacts did not overlap. The reported degree is nearly three times this bound. Either the contact cap was not applied as described to the network construction, or Table 3 is built from full co-location pairs rather than from the capped contacts that drive transmission. In the latter case, the network metrics do not describe the actual transmission process, and the causal narrative linking modularity/degree to epidemic outcomes is unsupported. Please clarify precisely which","section":"§3.2, Table 3"},{"comment":"The R0 estimation method is not described. The text defines R0 as the average number of secondary cases generated in a fully susceptible population by initially infectious individuals, but the algorithm used to compute it from the simulation is absent. This matters because the paper draws a notable conclusion from the Random Places model having the lowest R0 (14.69) but the highest peak incidence. Without knowing whether R0 is measured from the first generation, from an exponential fit, or by contact tracing, readers cannot assess whether this counterintuitive result is an artifact of the estimator. Please provide the exact estimation procedure and, ideally, a supplementary figure showing the per-run distribution.","section":"Table 1, §3.2"},{"comment":"The claim that aggregate movement volumes are 'broadly comparable' across models is load-bearing for the central conclusion that epidemic differences are not due to agents moving more or less overall. Table 2 reports only point estimates for total visits and dwell time, with no confidence intervals or per-run variation, despite 30 independent runs. The 12% difference in total visits between NHTS models (~180M) and Random models (~160M), and the 0.84-hour difference in mean dwell time, might be statistically significant. Please report confidence intervals or distributions for these mobility metrics, and if the differences are statistically significant, explain why they are still too small to explain the much larger epidemic contrasts.","section":"Table 2"},{"comment":"The log2(n) contact saturation is an arbitrary functional form with no sensitivity analysis. The paper does test sensitivity to the distance-decay exponent (Figure 3), but not to the contact-rate function. Because Random Advan concentrates agents at high-traffic locations, the sublinear cap disproportionately suppresses contacts in that model. A different saturation form (e.g., sqrt(n) or a linear cap) could alter the ranking of epidemic outcomes across models, which would weaken the specific attribution to 'who meets whom.' Please either justify the log2 form with empirical evidence or add a sensitivity test across plausible contact-rate functions.","section":"§3.2"}],"minor_comments":[{"comment":"The wording 'every infectious agent interacts with the subset of co-located agents' is ambiguous: does the cap apply to all agents or only infectious ones? If only infectious agents initiate contacts, how are edges between two susceptible agents ever formed in the contact network? Please clarify the contact-formation rule and how it maps to the undirected graphs analyzed in Table 3.","section":"§3.2"},{"comment":"Figure 1 shows single epidemic curves per model without confidence bands, even though 30 runs were performed. Adding shaded intervals would help readers judge whether the qualitative differences in peak height and timing are consistent across runs.","section":"Figure 1"},{"comment":"The column header 'Unique PoIs Visited' contains a capitalization inconsistency; also, confidence intervals should be added as noted in the major comments.","section":"Table 2"},{"comment":"The text says the synthetic population includes 480,000 workers, 210,000 students, and 310,000 homemakers, which sums to 1,000,000. Please clarify whether these categories are mutually exclusive and exhaustive, since agents could in principle be both worker and student.","section":"§3.3"},{"comment":"Reference 51 has a spelling inconsistency in the URL ('codebook v1.2.pdf' may be a placeholder) and the reference list entry would benefit from a stable DOI or URL.","section":"References"},{"comment":"In the paragraph on environment, 'places 6)' appears to be a truncated reference or typo; please complete it.","section":"§3.3"}],"recommendation":"major_revision","confidential_remarks":"The arithmetic inconsistency between the log2(n) cap and the reported Random Advan degree is the key technical issue. If the authors can show that the network was constructed from capped contacts (e.g., the cap applies per tick and the graph accumulates temporal edges differently than I assumed), then the degree might be plausible; otherwise the mechanism claim needs substantial rework. The paper fits the journal's scope and the qualitative finding that mobility assumptions matter is likely robust, so I would not reject at this stage, but the revision must be substantive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core experiment is worth taking seriously. Five mobility models, one million agents, fixed disease parameters — that is a clean, controlled way to show that mobility assumptions alone can roughly double peak incidence and shift timing by days. Figure 1 is stark, and the authors deserve credit for a systematic comparison, for holding the disease model fixed, and for putting the code on GitHub. The qualitative finding — mobility assumptions matter, a lot — does not depend on any sophisticated network metric. That is the paper's solid core.\n\nThe soft spots are real, and one is load-bearing. Section 3.2 caps each agent's contacts at log2(n) per co-located group, with a fixed subset that stays constant while co-located. Over 50 days a Random Advan agent makes roughly 160 visits; even in a maximally crowded venue the cap is about 20, so an agent cannot accumulate more than about 3,200 unique contacts. Yet Table 3 reports an average degree of 9,038 for Random Advan. That is arithmetically impossible under the stated cap. Either the cap was not implemented as described, or the graphs in Table 3 were built from full co-location pairs rather than the transmission-relevant contacts. Either way, the paper's central mechanism claim — that epidemic differences arise from who-meets-whom as measured by modularity, path length, and degree — is not supported by the reported metrics. This is not a minor typo; it is the explanatory bridge between the mobility models and the epidemic curves.\n\nThe reader's other concerns are secondary but legitimate: Table 2 has no uncertainty bounds for the 'broadly comparable movement volumes' claim; the R0 estimation method is unspecified; and the log2 saturation is an arbitrary functional form with no sensitivity analysis. A different contact function, say sqrt(n) or linear, could change the ranking of crowded-POI models. These are fixable in revision.\n\nFor all that, the paper is not a waste of time. The demonstrated range of outcomes is the takeaway, and the limitation section is refreshingly honest. The mechanism story needs to be repaired, not discarded. I would send it to peer review and ask the authors to clarify exactly what edges go into the contact networks, recompute the metrics consistently with the transmission model, and add sensitivity analysis on the contact cap. With that, the paper could be a useful contribution. For this reading group, it is actually a nice case study in how a mismatch between model description and output metrics can undermine a causal narrative.","headline":"Worth a referee's time, but the network mechanism claim has an internal inconsistency that needs resolution before the causal story holds.","tokens_in":14910,"tokens_out":3066,"would_cite":false,"duration_ms":29846,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"In a one-million-agent epidemic simulation, changing only how agents move changes outbreak curves.","keywords":["agent-based modeling","human mobility","contact networks","epidemic dynamics","mobility assumptions","network modularity","disease transmission","sensitivity analysis"],"falsifier":"Track total distance traveled and time spent traveling in each of the five models; if the empirical-schedule models travel far shorter distances than the random-schedule models, the claim that movement volumes are comparable fails and the attribution to contact structure is undercut. Alternatively, rerun the comparison with the contact cap changed from log2(n) to a linear or different saturating function; if the ranking of models reverses, the epidemic differences are partly an artifact of the contact rule rather than mobility alone.","tokens_in":13993,"feed_emoji":"🦠","tokens_out":4521,"duration_ms":46470,"temperature":0.7,"pith_summary":"This paper asks whether the simplifying assumptions modelers make about human mobility—daily schedules and destination choice—are harmless background choices or load-bearing parts of an epidemic simulation. Simulating one million people in an urban county with a fixed disease model, the authors varied only mobility: empirical travel-survey schedules versus random schedules, and destinations chosen by real foot-traffic popularity, distance decay, or at random. They found that these mobility choices alone changed the contact network structure and shifted peak incidence, time to peak, growth rate, and the basic reproduction number. Because total visits and dwell times were similar across models, the differences are attributed not to how much people moved but to who met whom, where, and how often. The conclusion: simulated epidemics and intervention evaluations may be driven as much by unvalidated mobility assumptions as by disease biology.","feed_headline":"Who-meets-whom, not how much people move, drives simulated epidemics","feed_subtitle":"A one-million-agent study finds mobility assumptions alone change outbreak size and timing, even when movement volumes look similar.","key_machinery":"The central object is the pairing of activity-pattern and destination-choice assumptions inside a single agent-based model, with disease parameters locked. Activity patterns come from empirical trip chains (structured schedules) or uniform random schedules; destination choice is weighted by empirical origin-destination foot traffic, a power-law distance-decay from home, or uniform random. Contacts emerge from co-location at hundreds of thousands of points of interest, with each agent interacting with at most log2(n) co-located others—a saturating contact rule meant to reflect that people in crowded places do not contact everyone present. Network metrics—Newman modularity, average path length","core_discovery":"The paper's central claim is that mobility assumptions alone—independent of disease parameters—shape emergent contact networks and epidemic trajectories in an agent-based model. Holding population, disease probabilities, and initial infections fixed, five mobility models (empirical daily schedules paired with popularity-, distance-, or random-destination selection; and randomized schedules paired with popularity- or random-destination selection) generated contact networks with different modularity, average degree, and path length. These network differences produced a wide spread of outcomes: empirical-schedule models gave modular networks and delayed, flattened peaks; randomized schedules wi","pith_inferences":["If real urban mobility resembles structured schedules with recurring routines, models that assume random schedules plus popularity-weighted destinations may overestimate outbreak speed; conversely, fully random destination choice may overestimate the breadth of the peak.","A direct next step would be to compare each model's emergent contact network—degree distribution, mixing patterns, and repeated contacts—against empirical contact data from surveys or wearable sensors to see which mobility representation best reproduces real interaction patterns.","The saturating log2(n) contact rule is an arbitrary functional form; changing it to a linear contact rule or a different saturation rate could reorder the models, so the ranking's robustness to the contact function deserves testing.","Tracking distance traveled and time spent traveling, rather than only visits and dwell time, would sharpen the claim that movement volumes are truly comparable across models."],"forward_implications":["Mobility assumptions should be reported and sensitivity-tested in agent-based epidemic models, since they can change conclusions as much as disease parameters.","Structured daily schedules produce sparser, more modular contact networks that delay and flatten epidemic peaks, while removing schedule structure with popularity-weighted destinations generates dense networks and rapid spread.","Aggregate movement-volume metrics such as total visits and dwell time are not enough to judge whether a mobility model is adequate; structurally different models can have similar volumes but different epidemic outcomes.","Destination popularity amplifies transmission most strongly when movement is not constrained by structured daily schedules.","Standard model evaluation that focuses on disease parameters alone may miss a hidden source of uncertainty in many existing agent-based models."],"fun_headline_variants":["Contact structure, not travel volume, steers outbreak curves","How agents meet shapes outbreak size more than how far they go","Mobility assumptions alone can shift epidemic peaks, timing","Simulated outbreaks hinge on who meets whom, not movement totals","Mobility model choice changes outbreak outcomes without disease changes"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The conclusion that epidemics differ because of who-meets-whom rather than how much people move rests on measuring 'movement' only as total visits and dwell time; if the models differ in distance traveled or travel-time budget, the movements may not actually be comparable.","fun_headline_variants_meta":{"raw":{"variants":["Contact structure, not travel volume, steers outbreak curves","How agents meet shapes outbreak size more than how far they go","Mobility assumptions alone can shift epidemic peaks, timing","Simulated outbreaks hinge on who meets whom, not movement totals","Mobility model choice changes outbreak outcomes without disease changes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000257,"raw_usage":{"total_tokens":1411,"prompt_tokens":737,"completion_tokens":674,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":607}},"tokens_in":481,"tokens_out":674,"duration_ms":6667,"temperature":1.0,"reasoning_tokens":607,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T00:26:41.173759+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Track total distance traveled and time spent traveling in each of the five models; if the empirical-schedule models travel far shorter distances than the random-schedule models, the claim that movement volumes are comparable fails and the attribution to contact structure is undercut. Alternatively, rerun the comparison with the contact cap changed from log2(n) to a linear or different saturating function; if the ranking of models reverses, the epidemic differences are partly an artifact of the contact rule rather than mobility alone.","supporting_citations":[],"review_version":1}