{"id":"7db0d237-b5d7-4011-9fed-64736daafd64","arxiv_id":"2506.03729","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"IntLevPy is a Python library that fits and classifies intermittent and Lévy walk processes by comparing adjusted R-squared values of their moment models.","lead":"This paper presents IntLevPy, a Python library for simulating, fitting, and classifying intermittent search and Lévy walk processes using second and fourth moments. It is relevant to anyone analyzing movement trajectories, such as eye-tracking, foraging, or financial data, who needs to decide which random-walk model generated the data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed classification performance rests entirely on unverified moment expressions from companion preprint Ref [14]; paper shows no quantitative error analysis or independent validation.","rationale":"The reader's weakest_assumption is that the classification and fitting accuracy depend on the analytical moments from the companion preprint Ref [14] that are not verified in this paper; my independent read of the full text confirms this is the single most load-bearing point. The workflow, Figure 1, uses moments m2 and m4 computed from the Ref [14] expressions as the basis for d2/d4 optimization and the Gamma classifier; therefore every displayed result inherits correctness from Ref [14]. The paper's own claims of synthetic parameter recovery are only stated, never quantified, and no data or scripts are provided to reproduce them (Table 1 gives a repository link but the code metadata entry 'Current code version v0.0.4' and PyPI link are given without a specific commit or validation script). I do not see an internal inconsistency; the external validation gap is a correctness risk, not a circularity. My concrete test would settle whether the concern lands: if a labeled synthetic benchmark with both classes shows high classification accuracy and small parameter recovery errors across the parameter space, then the central claim is supported, and the concern is resolved. If the benchmark exposes regions where one process is misclassified or parameters are biased, the paper needs restriction or correction. Thus the verdict should remain CONDITIONAL: conditional on release of the benchmark, exact definitions, and validation results. I agree with the reader rather than merely partially because the reader's weakest_assumption is precisely the assumption I identified as load-bearing, even though I would add that the absence of a quantitative benchmark is the immediate way to test it.","tokens_in":6023,"tokens_out":1715,"duration_ms":15447,"concrete_test":"Run a labeled synthetic benchmark: generate, say, 500 trajectories, half IS with parameters drawn uniformly over each parameter's stated physical range and half LW with gamma in (1,2), tau0 and v varied; feed the moment computation, optimization, and Gamma classifier without human tuning for both classes; report parameter recovery errors and confusion matrix. Then re-run the benchmark with the Ref [14] moment expressions replaced by direct Monte Carlo estimates of the moments from long simulations; if classification accuracy and parameter recovery change materially, the Ref [14] expressions are biased and the paper must disclose the failure range. Also report the precise adjusted-R^2 formula used, with a stated effective number of fitted parameters and sample points.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that IntLevPy can accurately classify and model IS and LW processes. This claim depends on the analytical moments of IS and approximate moments of LW taken from Ref [14] (arXiv:2505.00864), a companion preprint that this paper neither reproduces nor verifies. The optimization minimizes d2 and d4 in Eq. (5) and the classifier uses Gamma = R^2_IS - R^2_LW from Eq. (6); if the Ref [14] moment expressions have systematic errors for some parameter ranges, both parameter recovery and classification would be biased. The paper does not report a benchmark on labeled synthetic data: the Section 3 statement that 'the package successfully recovered the known parameters of the simulated processes' is not supported by quantitative results, error bars, or convergence statistics. The only real-data illustration shows fitted moment curves that 'align closely' with the IS model, but no adjusted-R^2 values, no Gamma values, no uncertainty quantification, and no statistical test of classification confidence are reported. For a software paper whose purpose is trustworthy classification of experimental trajectories, the absence of any measured classification accuracy or parameter-recovery error leaves the headline claim unsubstantiated. The reader's weak-assumption identification is exactly right: the moment expressions are load-bearing and unverified within this manuscript. Additionally, the adjusted-R^2 measure is not defined precisely (degrees-of-freedom formula), and the eye-tracking dataset is not released, so the real-data example cannot be independently checked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes IntLevPy, a Python library for simulating, analyzing, and classifying intermittent search (IS) and Lévy walk (LW) processes. The workflow computes empirical second and fourth moments from a trajectory, fits the parameters of both model classes with an iterative optimizer, evaluates each fit with an adjusted-R2-type measure, and classifies the process through a score Gamma = R^2_IS - R^2_LW. The manuscript includes a flowchart, code metadata, module descriptions, and an illustrative application to a real eye-tracking trajectory that is classified as intermittent. For synthetic data, the paper states that known parameters were successfully recovered, but provides no quantitative results, tables, or error metrics.","tokens_in":6325,"tokens_out":2828,"duration_ms":26509,"significance":"If properly validated, a user-friendly library that correctly fits and distinguishes IS and LW processes would be a useful resource for anomaly-diffusion research in ecology, neuroscience, and related fields. The paper's strengths are its open-source availability, modular architecture, unit tests, documentation link, and a clear conceptual flowchart of the classification procedure. However, the central claim of reliable parameter estimation and classification is currently unsupported: no benchmark on labeled synthetic data is reported, no uncertainty quantification is given for the real-data classification, and the theoretical moment expressions that drive the entire pipeline are imported from a companion preprint (Ref. [14]) without independent verification in this manuscript. The contribution as presented is therefore more of a user guide than a validated software paper.","major_comments":[{"comment":"The statement in Section 4 that \"the package successfully recovered the known parameters of the simulated processes\" is not supported by any reported results. Section 3 shows only a visual comparison of synthetic and model moments in Figure 2, with no parameter values, errors, or convergence statistics. Please add a quantitative benchmark on labeled synthetic data, including tables of true versus estimated parameters, mean errors or relative errors, and, where relevant, standard deviations across independent realizations.","section":"Section 3 and Section 4"},{"comment":"The classification of the real eye-tracking trajectory rests on the visual claim that the IS model \"aligns closely\" with the empirical moments. The paper never reports the adjusted-R2 values for the two models, the value of Gamma, confidence intervals, or any statistical test of classification confidence. Please report these quantities, and ideally add a bootstrap or synthetic-label validation that quantifies classification accuracy.","section":"Section 3, Eq. (6) and Figure 2"},{"comment":"The parameter fitting and the classification score Gamma depend entirely on the analytical moments of IS and the approximate moments of LW, which are taken from Ref. [14] (a companion preprint by overlapping authors). The present paper neither reproduces nor independently verifies those moment expressions. If they have systematic errors in any parameter regime, both parameter recovery and Gamma would be biased. Please state explicitly whether Gamma is defined here or in Ref. [14], provide the functional form of the moment expressions used, and include a Monte Carlo or other numerical check of the moments over the relevant parameter ranges.","section":"Section 2.2 and Section 3, Eqs. (5) and (6)"},{"comment":"The text refers to \"adjusted-R2-type measures\" but does not give an exact definition. Since the classification decision is based on a difference of such values, the formula matters: how are the residuals d2 and d4 combined, and how many degrees of freedom are counted? Please define the exact adjusted-R2 calculation used by the library, including how the number of fitted parameters enters.","section":"Section 3, Eq. (5) and Eq. (6)"}],"minor_comments":[{"comment":"The text says \"an LW with nu=1 (Cauchy walk)\" but the exponent in Eq. (2) is denoted by the Greek letter nu; please use consistent notation and clarify whether this is nu (ν) or gamma (γ), since elsewhere gamma is the Lévy exponent.","section":"Section 1"},{"comment":"The phrase \"a noble performance measure\" appears to be a typo for \"a novel performance measure\"; please correct it.","section":"Abstract and Section 2"},{"comment":"References [12] and [15] are the same paper (Bartumeus et al., Optimizing the encounter rate in biological interactions). Please remove the duplicate.","section":"References"},{"comment":"Reference [14] is listed as \"Submitted (2025)\" with an arXiv identifier; please clarify its current status (under review, accepted, or preprint) and cite the arXiv version consistently.","section":"References"},{"comment":"The code metadata lists pomegranate as a dependency, but the paper does not describe its role in the package; please either mention its use or remove it from the dependency list.","section":"Table 1 and Section 2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is essentially a software description whose scientific validation relies entirely on the companion preprint Ref. [14] by the same group. The editor should verify that this submission is not redundant with Ref. [14] and that a user-guide paper of this type fits the journal's scope without independent validation. The authors should be encouraged to include the missing quantitative benchmarks, as these are crucial for a trustworthy software paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know this is a reasonably useful software paper with a validation hole at the center. The library itself looks well-structured: four clean modules (simulation, moments, optimization, classification), code on GitHub, a PyPI release, and a clear workflow. The Gamma = R2_IS - R2_LW score is a simple, interpretable heuristic for choosing between two process classes, and the package genuinely lowers the barrier for someone wanting to test IS versus LW on their own trajectories.\n\nThe problem is the paper never shows the core claim working. We are told the package \"successfully recovered the known parameters\" for synthetic data, but there is no table, no error bar, no convergence statistic. The real-data example compares fitted curves visually in Figure 2; no adjusted R2 values, no Gamma value, no uncertainty. Since the entire method reduces to minimizing d2 and d4 and then comparing R2s, the absence of any measured classification accuracy is a real gap, not a cosmetic one.\n\nThe load-bearing input is the moment expressions from Ref [14], a companion preprint by largely the same authors. They are not re-derived or verified here, and the paper doesn't even clearly state which parts of Gamma come from Ref [14] and which are new to IntLevPy. That is a circularity burden, though not a fatal one: the authors have shared code, so a referee can in principle check the moments against simulation. But the paper as written does not do that work.\n\nOne smaller thing: \"adjusted R2\" appears without a formula. Given that the number of parameters differs between the two models (4 vs 3), the degrees-of-freedom correction matters for the Gamma comparison. And the eye-tracking dataset is not released, so the illustrative example cannot be independently re-run.\n\nThe citation pattern is fine; the field's standard references are there, and Ref [14] is properly cited as the source of the moments. The prose in Sections 4 and 5 is a bit boosterish, but that's normal for software papers.\n\nWho is this for? Researchers in movement ecology, anomalous diffusion, and eye-tracking who want a ready-made tool and are willing to trust Ref [14]. It's a legitimate niche. I'd send it to peer review, but with the clear expectation that the authors add a synthetic benchmark with error metrics, publish the data, and define adjusted R2. Without those, the classification claim is unsubstantiated.","headline":"IntLevPy is a usable, well-structured software package, but the paper needs a benchmark and error metrics before the classification claim is credible.","tokens_in":6866,"tokens_out":1938,"would_cite":false,"duration_ms":18791,"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":"IntLevPy classifies intermittent and Lévy processes from a single trajectory using moment fitting and a score function.","keywords":["intermittent search","Lévy walk","anomalous diffusion","moment fitting","process classification","parameter estimation","eye-tracking","Python library"],"falsifier":"Generate synthetic Lévy walks and intermittent trajectories with known parameters, then run IntLevPy's fitter and classifier: if Lévy walks with $\\nu \\in (1,2)$ frequently get $\\Gamma > 0$, or if the recovered parameters deviate systematically from the generating values, the central claim fails.","tokens_in":5844,"feed_emoji":"🐍","tokens_out":6191,"duration_ms":57613,"temperature":0.7,"pith_summary":"This paper presents IntLevPy, a Python library for simulating, fitting, and classifying two families of random movement: intermittent search and Lévy walks. Its claim is that the two can be told apart from one trajectory by comparing second and fourth velocity moments with analytical model moments, optimizing parameters, and reading a single score, $\\Gamma = \\bar{R}^2_{\\mathrm{IS}} - \\bar{R}^2_{\\mathrm{LW}}$. On synthetic data the package recovered known simulation parameters, and on a real eye-tracking trajectory it returned $\\Gamma > 0$, classifying the movement as intermittent search. If the claim holds, researchers in ecology, neuroscience, and finance gain a ready-made way to decide which movement strategy produced their data.","feed_headline":"Python library separates intermittent search from Lévy walks","feed_subtitle":"The package fits and classifies movement trajectories using second and fourth velocity moments and a single score.","key_machinery":"The load-bearing mechanism is the moment-distance optimization loop. For time scales $t_s$, the empirical $k$-th order velocity moments $m^{\\mathrm{exp}}_k(t_s)$ are computed; for each model family, parameters are adjusted to minimize $d_k(t_s) = [\\log(m^{\\mathrm{exp}}_k / m^{\\mathrm{model}}_k)]^2$, and the quality of the optimized fits is summarized by adjusted $R^2$ for IS and LW. The classification score is $\\Gamma = \\bar{R}^2_{\\mathrm{IS}} - \\bar{R}^2_{\\mathrm{LW}}$. The analytical moment expressions for the intermittent model and the approximate moment expressions for the Lévy walk come from the companion derivation the paper cites as Ref. [14].","core_discovery":"The central discovery is that a movement trajectory can be classified between these two process families from its velocity moments alone. The package estimates the empirical second and fourth moments $m_2(t_s)$ and $m_4(t_s)$, fits each candidate model by minimizing a logarithmic distance between empirical and model moments, and then compares the fit quality through adjusted $R^2$ measures. The score $\\Gamma = \\bar{R}^2_{\\mathrm{IS}} - \\bar{R}^2_{\\mathrm{LW}}$ decides: positive means intermittent search, non-positive means Lévy walk. The authors report successful parameter recovery on synthetic trajectories and a positive $\\Gamma$ for the analyzed eye-tracking data, which they read as evidence that the library can separate the two classes in practice.","pith_inferences":["The paper reports one real-data classification; a natural extension would be to compute confidence intervals for $\\Gamma$ by bootstrap-resampling trajectories, which would show whether the classification is stable.","Since only second and fourth moments are used, the method could confuse two process classes that happen to share these moments; a stress test with deliberately matched moments would map the limits of the score.","The same comparison could be extended to three-dimensional trajectories or to mixed strategies where a searcher switches between intermittent and Lévy phases, neither of which the current package claims to cover."],"forward_implications":["Researchers can classify real trajectories without reimplementing simulation, moment formulas, or optimizer loops.","The same moments that drive classification also yield parameter estimates, so fitting and testing hypotheses about movement parameters becomes one workflow.","Applications that depend on knowing whether movement is intermittent or Lévy, such as eye-tracking, animal foraging, human travel, and financial time series, have a single entry-point tool.","Because the package provides unit tests and example scripts, published analyses can be reproduced and inspected."],"supporting_citations":[{"why":"Supplies the analytical second and fourth moments for intermittent search, the approximate moments for Lévy walks, and the iterative fitting scheme the package implements.","marker":"[14]"},{"why":"Defines the intermittent search model with ballistic and diffusive phases that IntLevPy simulates.","marker":"[4]"},{"why":"Provides the Lévy walk framework and the step-length statistics that IntLevPy uses as the Lévy model.","marker":"[8]"},{"why":"Motivates the classification problem by showing Lévy walks as search strategies, the ecological context the package targets.","marker":"[10]"}],"fun_headline_variants":["Velocity moments reveal intermittent vs Lévy walks","Gamma score classifies movement with Python","Fit trajectory moments to identify walk type","Python library separates search and Lévy patterns"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline stands on the assumption that the built-in moment formulas for intermittent and Lévy processes actually match those processes; if Ref. [14]'s derivations are inaccurate, every fitted parameter and the $\\Gamma$ score inherit that error.","fun_headline_variants_meta":{"raw":{"variants":["Velocity moments reveal intermittent vs Lévy walks","Gamma score classifies movement with Python","Fit trajectory moments to identify walk type","Python library separates search and Lévy patterns"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000231,"raw_usage":{"total_tokens":1421,"prompt_tokens":815,"completion_tokens":606,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":431,"completion_tokens_details":{"reasoning_tokens":553}},"tokens_in":431,"tokens_out":606,"duration_ms":6284,"temperature":1.0,"reasoning_tokens":553,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:55:42.703851+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic Lévy walks and intermittent trajectories with known parameters, then run IntLevPy's fitter and classifier: if Lévy walks with $\\nu \\in (1,2)$ frequently get $\\Gamma > 0$, or if the recovered parameters deviate systematically from the generating values, the central claim fails.","supporting_citations":[{"cited_title":"Viswanathan, et al., Optimizing the success of random searches, Nature 401 (1999) 911–914","cited_arxiv_id":null,"evidence_quote":"Motivates the classification problem by showing Lévy walks as search strategies, the ecological context the package targets."}],"review_version":1}