{"id":"5ca4581b-23b3-47f9-b5c3-649a3418de6e","arxiv_id":"2607.29546","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A formalized agent-based model of criminal propensity yields persistent oscillations and predicts that a universally neutral perception of the environment drives polarisation to extremes.","lead":"This paper turns a verbal criminological theory—retribution and reciprocity—into a mathematical model of how criminal propensity changes through social interactions and observation. In simulations the model produces persistent oscillations and predicts that a population with a shared neutral perception of its environment splits into extremes, while a shared non-neutral perception produces consensus.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Neutral-perception polarisation is a degeneracy of the 'shock' threshold rather than a robust RRM consequence: at P_i=0, eqs. (3)/(5) make every non-neutral encounter shocking; a c>0 threshold perturbation should dissolve the polarisation band.","rationale":"I read the paper in good faith as a proof-of-concept formalisation of RRM. The model is specified clearly; the boundedness proof (Thm 4.1) and convergence theorems (Thm 4.3–4.4) appear sound; and the authors are honest about lack of empirical calibration and about the threshold dependence in §6. The strongest advertised result is the neutral-PoE polarisation, because it is presented as a novel prediction that RRM's verbally stated mechanisms could not generate. That result is where the argument is least secure: the threshold in (3)/(5) has a codimension-one degeneracy at P_i=0. At that point the inequality's RHS is zero, so any deviation of another agent from zero is 'shocking'; the natural-drift branch is suppressed and the extreme branch dominates. A small positive constant on the RHS removes this degeneracy and, plausibly, the polarisation band. The authors' own statement in §6 confirms that the result is a direct consequence of the threshold, so the issue is not hidden. The reader's conditional verdict already reflects this concern and the lack of reproducible simulation code; I find no reason to make it harsher or milder. The paper remains a valuable formal exercise, but the central behavioural claims should be read as conditional on the singular threshold and on independent reproduction.","tokens_in":15983,"tokens_out":8865,"duration_ms":89511,"concrete_test":"Re-run the Section 5.2 grid (N=10 and N=100, same fixed initial C_i, pairings and interaction types) with the threshold regularised to r|C_j−P_i| > |P_i| + c for c=0.01, 0.05, 0.1 (and, for contrast, threshold r|C_j−P_i| > c). Classify consensus/polarisation at P=0 and measure the width of the polarisation band in P. If polarisation at P=0 disappears or the band width collapses as c goes from 0 to 0.1, the neutral-polarisation prediction is an artifact of the c=0 degeneracy and should be downgraded from a model prediction to a threshold property.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's most original advertised prediction—uniform neutral PoE → polarisation, non-neutral → consensus (abstract; §5.2, Fig. 6)—is not a stable consequence of the RRM mechanisms. It is produced by the threshold inequality in (3) and its witness analogue (5), r_i^±|C_j−P_i|>|P_i|. When P_i=0 the RHS is zero, so any encounter with C_j≠0 counts as 'shocking' and activates the extreme-pulling top line; the drift-to-PoE branch is effectively disabled for almost every interaction. Thus the neutral case is a singular limit of the update rule. The authors concede in §6 that the result is 'a direct consequence of the threshold specification rather than an artefact of implementation,' but that is precisely the problem: the headline result is an artifact of an arbitrary threshold with c=0. No argument is given that c=0 is the empirically correct threshold for RRM; the amplitude of the polarisation band in Fig. 6 is therefore not a robust prediction. This does not invalidate the boundedness/convergence theorems, but it makes the central non-convergence/polarisation claim conditional on a singular modelling choice.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formalises the verbal Retribution and Reciprocity Model (RRM) of criminology as an agent-based system of criminal propensity. Agents update their propensity C_i(t) in [-1,1] through pairwise interactions and third-party witnessing, combining a bounded-confidence-style influence term with drift towards a fixed perception-of-environment P_i. The analytical results include boundedness of propensities, sufficient conditions for convergence to the extremes or to P_i, and a necessary condition for oscillatory dynamics. Simulations with 100 agents are used to claim that persistent oscillations are the dominant regime and that a uniform neutral perception of the environment produces polarisation to the extremes while a uniform non-neutral perception produces consensus, with a narrow near-neutral polarisation band whose width decreases with population size and increases with the strength of reciprocal/retributive tendencies. The paper presents these as novel predictions generated by the formalised RRM.","tokens_in":16321,"tokens_out":8700,"duration_ms":93563,"significance":"There is genuine value in this manuscript. It offers, to my knowledge, the first formal dynamical-systems treatment of propensity dynamics from RRM, extends bounded-confidence opinion dynamics with witness effects and fixed individual attractors, proves boundedness cleanly, identifies a necessary parameter region for oscillations, and is explicit about its limitations (fixed PoE, no empirical calibration, no crime rate). These are concrete strengths: the analytical results are nontrivial and the modelling framework is reusable. However, the two headline claims—persistent oscillations as the dominant behaviour and neutral-perception polarisation—rest respectively on visually assessed single simulations and on a singular threshold specification. The paper's own discussion in Section 6 concedes that the polarisation result is a direct consequence of the threshold choice. As a result, the advertised conclusions are not yet established, although the underlying framework is defensible and the issues appear fixable within the manuscript's scope.","major_comments":[{"comment":"The polarisation-at-neutral prediction is a singular-limit consequence of the threshold inequality. For P_i=0 the condition r_i^±|C_j-P_i|>|P_i| becomes r_i^±|C_j|>0, so every interaction with C_j≠0 triggers the extreme-pulling branch; the witness condition in (5) behaves similarly. The natural-drift branch is effectively disabled. The manuscript's Section 6 states that the result is 'a direct consequence of the threshold specification rather than an artefact of implementation,' but that is precisely the vulnerability: no criminological derivation of a zero threshold is given, and a perturbed threshold r|C_j-P_i|>c with c>0 would remove the degeneracy and may well eliminate or shrink the neutral-polarisation band. Since this band is a headline result, the paper needs either a principled justification of c=0, a robustness analysis over threshold perturbations, or a reframed claim explicit","section":"§3, Eqs. (3)/(5); §5.2; §6"},{"comment":"The central simulation claims—that persistent oscillations are the dominant system behaviour and that the Fig. 6 grid separates consensus from polarisation—are not quantified. There is no operational definition of an oscillating trajectory as opposed to a slowly converging one, no criterion for polarisation (e.g., fraction of agents within tolerance of ±1), and no explanation of how the 'inconclusive' class in Fig. 6 is determined. Figures 2–5 show single runs without replication statistics, error bars, or convergence diagnostics, and no code or seed details are provided. Because these simulations carry the abstract's claims about dominant behaviour and phase regimes, the paper should supply precise convergence/oscillation criteria, averages over multiple realisations, and code/data availability.","section":"§5.1–5.2, Figs. 2–6"},{"comment":"The proof after Eqs. (14)–(15) is incomplete. If j(t_m) does not converge, the statement that 'letting t→∞ ... immediately yields |C_i^∞|=1' does not follow: the update is evaluated at t_m+1, not along the subsequence t_m, and a non-convergent partner sequence can in principle produce vanishing increments if 1-|C_i| tends to zero. A complete proof should use monotonicity of C_i(t) and the δ-lower-bound on interaction products to show that infinitely many positive increments bounded below preclude an interior limit. The theorem is plausible and likely fixable, but as written this is a genuine gap in a stated sufficient condition.","section":"§4.1, proof of Theorem 4.3"},{"comment":"The inference from the necessary condition (19)–(20) to 'the model setup therefore enables propensity oscillations to be the dominant mode of behaviour' overstates the analytical content. Condition (20) is necessary for the shock branch to be reachable; it does not say that trajectories actually oscillate. The further claim that a truncated-normal parameter distribution yields more than half of agents capable of oscillation is an assumption about parameter priors, stated without criminological data. The dominance claim must be carried by simulations, which currently lack the quantitative support discussed above.","section":"§4.2"}],"minor_comments":[{"comment":"The stochastic generative process is underspecified: how n(t) is sampled, whether witness assignment is with replacement, and how the 20×20 grid fixes pairings across runs would need to be stated for reproducibility.","section":"§5, first paragraph"},{"comment":"The terms 'blue circle', 'red square', and 'black triangle' are not defined in the text; the classification criterion should be given explicitly, as should the choice of 10,000 steps as the convergence horizon.","section":"Fig. 6 caption"},{"comment":"There are typographical slips: 'larger e_i' should be 'larger r_i^e', and 'serves as the their' should be 'serves as their'.","section":"§6"},{"comment":"For r_i^±=1, the text says the range is infinite; since P_i is constrained to [-1,1], the effective range is simply [-1,1].","section":"Eq. (20)"},{"comment":"The phrase 'direct consequence of the threshold specification rather than an artefact of implementation' is confusing. The threshold specification is a modelling choice, so this wording does not address the concern that the result may be an artefact of that choice.","section":"§6"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern raised by the reader lands: the neutral-polarisation result is a degeneracy of the threshold with c=0, and the authors' own Section 6 confirms that it is a direct consequence of that specification. I do not think this requires rejection, because the formal framework, boundedness proof, and necessary condition are meaningful and the paper could be revised to justify or robustify the threshold. However, the current abstract and conclusion overstate what is established. Given the weight placed on simulations, I would encourage the editor to require code/data release and a quantitative classification of the simulated regimes before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something useful: it takes the verbal Retribution and Reciprocity Model of criminal propensity and writes down a precise dynamical system, extending Deffuant bounded confidence with fixed agent-specific perception-of-environment attractors and third-party witnesses. That is a real novelty, and the analytical core is mostly solid. Theorem 4.1 (boundedness) is clean and correct. The sufficient conditions for convergence to an extreme or to the PoE are believable, and the necessary condition for oscillations in Section 4.2 is a genuine contribution. The authors are also admirably explicit that the model is not empirically calibrated and is a proof-of-concept, not a validated theory of crime.\n\nBut the two strongest advertised results are both softer than they look. First, the claim that a uniform neutral perception of the environment produces polarisation is, as the authors themselves concede in Section 6, 'a direct consequence of the threshold specification rather than an artefact of implementation.' The threshold in (3) and (5) has the form r|C_j - P_i| > |P_i|. When P_i = 0, the RHS is zero, so every non-neutral interaction is 'shocking' and the drift-to-PoE branch is disabled. That is a singular limit. The polarisation band in Figure 6 is therefore not a robust prediction of RRM; it is an artifact of setting the threshold constant to zero, and no criminological justification is given for that choice. A perturbation c > 0 would likely dissolve the band. This is a load-bearing issue for the paper's most novel claim.\n\nSecond, the pervasive oscillations—the other headline result—rest on a visual reading of time-series plots. There are no code, no error bars, no quantitative classification of runs, and no sensitivity analysis. For a paper whose main empirical evidence is simulation, this is a serious gap. The theoretical necessary condition shows oscillations are possible, but not that they are 'dominant' in any measured sense. Theorem 4.3 also has a lax step: when the interaction partner sequence does not converge, the proof simply asserts that taking t→∞ in the update yields |C_i^∞| = 1, without handling the non-convergent factor. That is fixable, but it is a real hole.\n\nThe paper is worth engaging with. It brings a formal toolkit to a verbal theory and produces at least one new mathematical finding (the oscillation condition). I would send it to a competent referee, but with a strong request for code, quantitative oscillation/polarisation measures, and a robustness test of the threshold specification. The core formalisation and the boundedness/convergence theorems are likely to survive; the polarisation prediction may not.","headline":"A genuine formalisation of a verbal criminology theory with some real analysis, but the headline neutral-perception polarisation result is a thin artefact of the shock threshold and the simulation evidence is too visual to carry the 'dominant oscillations' claim.","tokens_in":16800,"tokens_out":2960,"would_cite":false,"duration_ms":30765,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91D10","37N99"],"pacs":[],"model":"deepseek-v4-flash","headline":"Formalising a criminological theory predicts that populations sharing a neutral view of their environment split into rule-abiding and rule-breaking extremes, while a shared non-neutral view leads to consensus.","keywords":["criminal propensity","bounded confidence","agent-based model","opinion dynamics","polarisation","oscillations","retribution and reciprocity","perception of environment"],"falsifier":"Re-run the simulations with the shock condition changed to r|C_j − P_i| > c for a fixed c > 0 (e.g., c = 0.1) and otherwise identical parameters; if a uniform neutral perception P = 0 still produces polarisation to ±1 for c well above 0, the central finding is contradicted. Alternatively, an empirical longitudinal study of a population with genuinely neutral perceptions of their environment should show a bimodal split in offending propensity if the model is right.","tokens_in":15825,"feed_emoji":"⚖️","tokens_out":3868,"duration_ms":38598,"temperature":0.7,"pith_summary":"The paper formalises the Retribution and Reciprocity Model of crime causation as an agent-based dynamical system, called mRRM, and asks what population-level patterns the posited mechanisms can generate. Its central claims are that criminal propensity is dominated by persistent oscillations rather than convergence, and that a population sharing a neutral perception of its environment tends to polarise to the extremes, while a shared non-neutral perception produces consensus. The authors prove individual propensities stay bounded, give sufficient conditions for convergence to extremes or to one's perception of the environment, and derive a necessary condition for oscillations. A sympathetic reader would care because this is a rare case of a verbal criminological theory being made precise enough to generate novel, testable predictions about when offending tendencies fluctuate and when they split.","feed_headline":"Neutral perceptions of society split crime propensity to extremes","feed_subtitle":"Agent-based formalisation of retribution-reciprocity theory yields oscillation and a narrow polarising band at neutral perceptions.","key_machinery":"The central object is a bounded-confidence opinion dynamics model extended with two features: an agent-specific attractor, the perception of the environment (PoE) P_i, and third-party witnesses whose propensities also update. The decisive mechanism is the threshold condition r|C_j − P_i| > |P_i| (and its witness analogue), which decides whether an interaction 'shocks' an agent into moving toward the interaction partner or lets the agent drift back toward P_i. When P_i = 0, every nonzero interaction is shocking, which makes neutral perceivers maximally exposed to influence and drives convergence to the absorbing extremes ±1; this is the mechanism behind the polarisation claim.","core_discovery":"The core discovery is that the two RRM mechanisms—reciprocal and retributive updating from interactions and witnessing, plus a drift toward each agent's fixed perception of the environment—are sufficient to produce sustained non-convergence and polarisation with no external driver. In simulations, most agents' criminal propensities oscillate persistently around their perception of the environment; when everyone shares the same perception, the population converges to that value unless the shared perception is near zero, in which case the population splits into two factions converging to the extremes. The authors show the neutral-polarisation effect follows directly from the shock threshold in","pith_inferences":["A testable extension: the neutral-polarisation prediction hinges on the threshold's zero baseline; modifying the threshold to include a positive constant would likely erase the effect, so empirical studies of populations with genuinely neutral environmental perceptions are the natural check.","The individual-level oscillations may connect to RRM's population-level free-rider cycles only if perceptions are allowed to evolve; the paper itself flags this as future work, so linking the two timescales is a plausible next step.","The paper's self-averaging explanation for why larger populations polarise less suggests a network-structure experiment: hubs that concentrate shocking interactions could counteract averaging and preserve polarisation in large populations."],"forward_implications":["If correct, persistent oscillations of criminal propensity arise from deterministic rules without adaptive confidence bounds, offering a mechanistic account of fluctuating offending tendencies.","A population sharing a non-neutral view of how it is treated will tend toward consensus in criminal propensity; a population sharing a neutral view will tend to split into rule-abiding (C=1) and rule-breaking (C=−1) factions.","The polarisation is confined to a narrow band of near-neutral perceptions whose width shrinks with population size and grows with retributive and reciprocal strength, so small changes in the shared perception near zero can flip a population between consensus and division.","The model traces propensity dynamics rather than generating crime rates; its predictions are qualitative regime claims, not calibrated quantitative forecasts."],"fun_headline_variants":["Criminal propensity oscillates in agent-based model without external shocks","Neutral shared views trigger extreme split in crime propensity","Narrow neutral-perception band polarises crime propensity","Reciprocity and retaliation drive persistent crime oscillations","Agent-based model reveals persistent crime propensity oscillations"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The prediction that neutral perceivers polarise depends on the modelling choice that an interaction shocks an agent whenever r|C_j − P_i| exceeds |P_i|, so that at P_i = 0 every interaction is shocking; a different threshold with a positive constant would likely destroy the effect.","fun_headline_variants_meta":{"raw":{"variants":["Criminal propensity oscillates in agent-based model without external shocks","Neutral shared views trigger extreme split in crime propensity","Narrow neutral-perception band polarises crime propensity","Reciprocity and retaliation drive persistent crime oscillations","Agent-based model reveals persistent crime propensity oscillations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001153,"raw_usage":{"total_tokens":4595,"prompt_tokens":706,"completion_tokens":3889,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":3813}},"tokens_in":450,"tokens_out":3889,"duration_ms":25977,"temperature":1.0,"reasoning_tokens":3813,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T04:53:19.918805+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the simulations with the shock condition changed to r|C_j − P_i| > c for a fixed c > 0 (e.g., c = 0.1) and otherwise identical parameters; if a uniform neutral perception P = 0 still produces polarisation to ±1 for c well above 0, the central finding is contradicted. Alternatively, an empirical longitudinal study of a population with genuinely neutral perceptions of their environment should show a bimodal split in offending propensity if the model is right.","supporting_citations":[],"review_version":1}