{"id":"fee04142-a7f9-4e92-9300-a4ff2ddec41e","arxiv_id":"2508.04671","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ERC20 transfers on Ethereum split by account type reveal stable scaling exponents for human-to-human activity and unstable, sublinear, heavy-tailed scaling for smart-contract-to-smart-contract activity.","lead":"An analysis of 44 million Ethereum token transfers finds that human-controlled accounts (EOAs) show stable statistical patterns, including consistent power-law and Taylor's law exponents, while smart-contract-to-smart-contract activity shows unstable, sublinear, and heavy-tailed behavior. The study offers a statistical fingerprint for distinguishing human versus automated trading in blockchain systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'stable γ≈2.3' claim for EOA-driven transactions is contradicted by the paper's own Table 3: EOA–EOA sender γ drops to 1.76 in Period 3 and EOA–SC receiver γ falls below 2, so the headline stability criterion is not met.","rationale":"The reader's verdict is CONDITIONAL, and my read does not move it. The core EOA-vs-SC contrast may survive, but the abstract overstates temporal stability. The reader's rationale already notes that the abstract overstates stability against Table 3, though the reader's weakest_assumption centers on address-label validation and zero-inflated TL fits. My most load-bearing concern is the internal inconsistency in the stability claim: the paper's own Table 3 shows an EOA-side power-law exponent falling to 1.76 in Period 3, which is precisely the heavy-tailed, unstable regime the paper assigns to SCs. Because the power-law fits lack bootstrap goodness-of-fit p-values and confidence intervals, this discrepancy cannot currently be adjudicated. A focused reanalysis with the Clauset bootstrap would settle whether γ≈2.3 is stable or whether the headline claim must be qualified. Therefore the appropriate verdict remains CONDITIONAL: the manuscript needs revision/reanalysis before the abstract's universal-stability claim can be accepted.","tokens_in":17416,"tokens_out":9395,"duration_ms":118302,"concrete_test":"Recompute all 24 power-law fits in Table 3 from the XBlock-ETH data using the Clauset et al. protocol: for each cell, estimate xmin by KS minimization, MLE γ, then run ≥1000 semi-parametric bootstrap samples to obtain a goodness-of-fit p-value and a 95% CI for γ. If the Period 3 EOA–EOA sender cell (γ=1.76, xmin=105) has bootstrap p<0.05 or a CI excluding the Period 1–2 γ values, the 'stable γ≈2.3' claim fails as stated; if it passes and the CI overlaps 2.3, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim's first pillar—EOA-driven transactions show stable power-law exponents γ≈2.3 across periods—is not supported by the paper's reported results. In Table 3, the purest EOA-side role, EOA–EOA sender, has γ=2.32, 2.26, 1.76 (with xmin jumping to 105 in Period 3), and EOA–SC receiver fits fall from 2.07 to 1.89. Thus an EOA-side exponent drops into the same heavy-tailed regime (γ<2) that the paper uses to characterize SC-driven activity. The text attributes this to the 2018 crypto crash, but the abstract claims time-persistent universal behavior; invoking a one-off market condition is an ad hoc rescue unless the stability claim is explicitly qualified. Moreover, no goodness-of-fit bootstrap p-values (Clauset et al. 2009) or confidence intervals are reported; the KS acceptance thresholds are stated ad hoc, and xmin is estimated from the same data, so we cannot tell whether γ=1.76 vs 2.3 is a genuine regime shift or an artifact of threshold selection and small tail samples.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes 44.86 million ERC20 token transfers on Ethereum (July 2017–March 2018), splits them into three three-month periods, and classifies each transfer into EOA–EOA, EOA–SC, SC–EOA, or SC–SC using the fromIsContract/toIsContract flags of the XBlock-ETH dataset. For each class and period it estimates three scaling quantities: the volume–partner exponent α, the power-law tail exponent γ with KS-selected xmin, and the temporal Taylor's law exponent β from hourly mean–variance regressions. The paper's central claim is that EOA-driven transactions show stable, near-universal scaling (γ ≈ 2.3, β ≈ 2.3), while SC-involved transactions, especially SC–SC, show sublinear scaling, unstable exponents, and larger Taylor-exponent fluctuations (ΔβSC–SC = 0.51). The authors interpret this as a statistical distinction between human-controlled and algorithmic behavior in decentralized finance.","tokens_in":17689,"tokens_out":4213,"duration_ms":49505,"significance":"If the quantitative claim were fully supported, the paper would provide a useful, interpretable signature separating human and automated agents in blockchain transaction networks, and it would connect ERC20 activity to scaling laws observed in other social and economic systems. The study has clear strengths: it uses a public dataset, it reports detailed per-period tables (Tables 2, 3, and 5), and it combines two complementary scaling frameworks, power-law analysis and Taylor's law. The qualitative contrast — EOA-side exponents are generally higher and more stable, SC–SC exponents are lower and more variable — is visible in the paper's own tables. However, the specific headline claim of time-persistent 'stable γ ≈ 2.3' is contradicted by Table 3, and the statistical validation of the power-law and Taylor-law fits is incomplete in ways that affect the load-bearing conclusions.","major_comments":[{"comment":"The abstract's first pillar — 'stable power law exponents (γ ≈ 2.3)' for EOA-driven transactions — is not supported by the paper's own Table 3. The purest EOA-side role, EOA–EOA sender, has γ = 2.32, 2.26, 1.76 across periods, with xmin jumping to 105 in Period 3; EOA–SC receiver falls from 2.07 to 1.89. These values put EOA-side activity into the γ < 2 regime that the paper elsewhere associates with SC-driven activity. The text's explanation that the 2018 crypto crash caused this is a one-off post hoc qualification; it does not rescue an abstract claim of universal, time-persistent behavior. The authors should either restrict the stability claim to Periods 1–2, model the crash period explicitly, or provide a statistical test showing that γ = 1.76 is consistent with γ ≈ 2.3 once xmin and sample-size effects are accounted for.","section":"Abstract; Table 3"},{"comment":"The power-law validation is incomplete. The paper reports KS distances and LLR values but no bootstrap goodness-of-fit p-values of the type recommended by Clauset et al. (2009), and it gives no confidence intervals for γ or xmin. The KS acceptance thresholds stated in 'Determining Threshold and Model Evaluation' (D < 0.05 good, D > 0.1 bad) are ad hoc, and xmin is estimated from the same data used to evaluate the fit. Consequently, the reader cannot tell whether the Period 3 EOA–EOA sender value γ = 1.76 is a genuine regime shift or an artifact of threshold selection in a thin tail. Please report full Clauset-style p-values, confidence intervals on γ, and a sensitivity analysis over xmin choices.","section":"Power Law Distribution; Table 3"},{"comment":"The >90% KPSS stationarity percentages in Table 4 are not persuasive for hourly count series that are likely zero-inflated. A trader with many zero-activity hours and occasional nonzero hours will pass KPSS almost trivially, and regressing log-variance on log-mean for such sparse series can produce apparent β ≈ 2–2.5 without reflecting a stable fluctuation law. The paper should report the distribution of nonzero hours per trader, the number of traders contributing to each TL fit, and the raw mean–variance scatter with point counts. Without this, the claim that EOA-driven transactions 'adhere to TL with β ≈ 2.3' is not independently grounded.","section":"Stationarity Test Results; Temporal TL Exponent"},{"comment":"The entire four-way classification relies on the fromIsContract and toIsContract flags in the XBlock-ETH dataset, yet the paper never validates this labeling. Misclassified early contracts, proxy contracts, or multi-sig wallets labeled as EOAs could contaminate the exact buckets on which the human-versus-algorithm contrast depends. Please validate the flags against an independent source (e.g., code existence at the address, contract-creation block, or Etherscan labels) or at least quantify the expected error rate and discuss how it would affect the contrasts in Tables 2, 3, and 5.","section":"Section 1.1; Table 1"}],"minor_comments":[{"comment":"The last paragraph of the introduction says 'Section 1.1 presents and discusses the results' and 'Section 1.1 offers concluding remarks,' but the Results and Conclusion sections are not numbered subsections of Section 1.1. Please fix the section cross-references.","section":"Introduction / organization"},{"comment":"The text says SC–SC sender β drops to 1.96 'during Period 2' and recovers to 2.09 in Period 3, which matches Table 5; but the surrounding sentence for EOA–SC says the sender decreases in Period 3 while receiver increases. Please make the period labels consistent and avoid a possible reading that the SC–SC sender drop occurs in Period 3.","section":"Table 5 and text near Figure 9"},{"comment":"The caption says 'Consistent β values (∼ 2) for senders across three periods,' but Table 5 lists EOA–EOA sender β ≈ 2.35. Please write β ≈ 2.3 or 'near-quadratic' rather than '∼2' if the caption is meant to be precise.","section":"Figure 8 caption"},{"comment":"Reference [57] cites 'Is Newey–West optimal among first-order kernels?' by Kolokotrones, Stock, and Walker, but the Newey–West automatic bandwidth estimator is due to Newey and West. Please verify the citation and cite the original Newey–West papers.","section":"References"},{"comment":"Several table captions begin with 'T able' (Tables 1–5). Please remove the stray space throughout.","section":"Typos"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this paper if you are interested in empirical scaling laws in blockchain networks. The genuinely new bit is the four-way EOA/SC split and the application of temporal Taylor's law to ERC20 transfers, and the qualitative result—EOA activity tends to scale more linearly and stably than SC-SC activity—is visible in the tables and plausible. That part is worth taking seriously.\n\nThe soft spots are in the quantitative claims. The abstract says EOA-driven transactions have stable power-law exponents γ≈2.3 across periods, but Table 3 shows EOA–EOA sender γ dropping to 1.76 in Period 3 (with xmin jumping to 105) and EOA–SC receiver γ at 1.89. The paper flags the crypto crash as a possible cause, but the abstract does not qualify the stability claim. There are no bootstrap p-values (Clauset et al. 2009) and no confidence intervals for the γ fits; the KS thresholds are stated ad hoc, and xmin is estimated from the same data, so the Period 3 shift could be partly an artifact of threshold selection. Taylor's law is fitted to hourly counts that are heavily zero-inflated; the KPSS stationarity result is near-trivially easy to pass for sparse series, so β≈2.3 is not independently grounded. They also do not release code or data, so the fits cannot be recomputed. The account-type classification relies entirely on fromIsContract/toIsContract flags without validation; if those flags are noisy, the human-vs-algorithm contrast blurs.\n\nThe paper's internal consistency is decent—the qualitative contrast runs through all three analyses—and the literature is engaged (Somin et al. 2018 for power laws, Eisler et al. for TL). The issue is more overclaiming than a load-bearing flaw. With a rewritten abstract and some robustness checks (bootstrap goodness-of-fit, handling zero-variance accounts in TL, reporting uncertainties), it would be a solid empirical contribution.\n\nRecommendation: send to peer review. The split is novel and the question is timely. A competent referee can push for the needed robustness checks and qualifications.","headline":"Useful empirical split between human- and contract-driven ERC20 activity, but the abstract overstates the stability of EOA power-law exponents; worth peer review with revisions.","tokens_in":18247,"tokens_out":3507,"would_cite":true,"duration_ms":39202,"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 human and automated Ethereum accounts leave statistically distinguishable, time-persistent scaling signatures in ERC20 token transfers: human activity is near-universal, contract activity is contextual and volatile.","keywords":["ERC20 tokens","Ethereum","scaling laws","power-law distributions","Taylor's law","smart contracts","externally owned accounts","blockchain transaction networks"],"falsifier":"Take a random sample of addresses the dataset flags as externally owned, check the Ethereum chain state at those times to see whether each address actually has code, then re-fit $\\alpha$, $\\gamma$, and $\\beta$ after correcting mislabeled cases. If corrected EOA-EOA exponents become unstable or corrected contract-contract exponents become near-linear, the reported contrast is a labeling artifact.","tokens_in":17252,"feed_emoji":"📊","tokens_out":9448,"duration_ms":89655,"temperature":0.7,"pith_summary":"The paper examines 44.8 million ERC20 token transfers on Ethereum over nine months, split into three equal periods, and asks whether human-controlled accounts and automated smart contracts obey different statistical laws. Transfers are classified by whether the sender and receiver are externally owned accounts (EOAs) or smart contracts, giving four interaction types. For EOA-driven traffic, the authors find near-linear growth of trade volume with partner count ($\\alpha \\approx 1$), stable power-law exponents around $\\gamma \\approx 2.3$, and consistent Taylor scaling $\\beta \\approx 2.3$ across all three periods. For contract-to-contract traffic, scaling is sublinear ($\\alpha$ falls from 0.93 to 0.67), power-law fits are unstable with heavier tails ($\\gamma < 2$), and Taylor exponents fluctuate widely ($\\Delta\\beta = 0.51$). If correct, these signatures give a principled way to separate human from algorithmic behavior in decentralized finance, with human activity resembling universal scaling seen in social and economic systems.","feed_headline":"Human and bot Ethereum trades scale differently","feed_subtitle":"Stable power laws mark human wallets; smart-contract flows are volatile and context-dependent.","key_machinery":"The argument is carried by a four-way transaction classification (EOA-to-EOA, EOA-to-contract, contract-to-EOA, contract-to-contract) built from the dataset's sender-contract and receiver-contract flags, combined with two canonical scaling tools: the volume-partner relation $V \\sim N^\\alpha$ and temporal Taylor's law $\\sigma^2 = a\\mu^\\beta$, with power-law distributions $p(x) \\propto x^{-\\gamma}$ fitted by maximum likelihood and validated by Kolmogorov-Smirnov distances and log-likelihood ratios against an exponential. The four-way classification is the load-bearing object; the contrast between stable human-side exponents and drifting contract-side exponents is the evidence.","core_discovery":"The central discovery is that ERC20 transaction statistics split cleanly along the EOA/smart-contract boundary. EOA-to-EOA and EOA-to-contract transactions show stable, near-universal exponents: volume grows nearly linearly with partner diversity, transaction counts follow power laws with $\\gamma \\approx 2.3$ whose thresholds are small and whose fits are good, and hourly variance obeys Taylor's law with $\\beta \\approx 2.3$ over every period. Smart-contract-to-contract transactions behave differently: volume growth is increasingly sublinear, power-law fits require very high thresholds and show larger errors and KS distances, and Taylor exponents wander, with the sender exponent dropping to $1","pith_inferences":["The paper never validates the contract flags it relies on; checking a sample of labeled addresses against on-chain bytecode would either confirm the human-versus-algorithm split or reveal it as a labeling artifact.","The near-quadratic Taylor exponents ($\\beta \\approx 2.3$) may partly reflect generic overdispersion in zero-inflated hourly counts rather than a specific behavioral mechanism; a zero-inflated or negative-binomial null model would show how much of the scaling is informative.","If the contrast replicates on other chains, the stable EOA fingerprint could serve as a baseline for anomaly detection in DeFi, flagging addresses whose scaling suddenly becomes contract-like as possible bots or wash traders.","The EOA-to-contract channel behaves like the human channel rather than the contract channel, suggesting that when a human initiates an interaction, the statistical signature remains human-like even though the counterparty is automated."],"forward_implications":["EOA-to-EOA and EOA-to-contract activity can be modeled as approximately universal: $\\alpha \\approx 1$, $\\gamma \\approx 2.3$ and $\\beta \\approx 2.3$ hold across nine months including a major market crash.","Contract-to-contract activity is not universal: its scaling exponents drift over time and its fits degrade, so decentralized-finance models need protocol-specific rather than one-size-fits-all assumptions.","The Taylor-exponent instability gap ($\\Delta\\beta = 0.51$ for contract-contract versus $0.17$ for EOA-EOA) provides a quantitative measure of how much more volatile automated flows are than human flows.","Heavy-tailed contract activity ($\\gamma < 2$) means a small set of contracts dominates outgoing or incoming volume, so tail-risk monitoring in ERC20 networks should concentrate on contract-to-contract channels."],"supporting_citations":[{"why":"Supplies the nine-month ERC20 transfer records and the contract flags that define the four transaction categories.","marker":"[26]"},{"why":"Showed ERC20 transaction volumes follow power-law distributions, the pattern this paper extends to account-type categories.","marker":"[19]"},{"why":"Documented power-law behavior in cryptocurrency market capitalizations, supporting the universality premise for crypto systems.","marker":"[20]"},{"why":"Introduced the variance-mean scaling law that the paper applies to hourly transaction counts.","marker":"[23]"},{"why":"Provides the fluctuation-scaling framework and the interpretation of Taylor exponents in complex systems.","marker":"[24]"},{"why":"Supplies the maximum-likelihood and Kolmogorov-Smirnov methodology used to fit and validate power-law exponents.","marker":"[33]"},{"why":"Maps the ERC20 token ecosystem and motivates the EOA-versus-contract account distinction in Ethereum.","marker":"[12]"},{"why":"Characterized ERC20 token network structure, giving the network baseline that the four-way split refines.","marker":"[13]"}],"fun_headline_variants":["Ethereum trades: humans scale, bots don't","Smart contracts break scaling laws on Ethereum","Blockchain scaling: humans stable, bots chaotic","ERC20 flows: human wallets scale, smart contracts drift","Why Ethereum bots defy power laws"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The whole human-versus-algorithm contrast depends on the dataset's contract flags being correct; if addresses are mislabeled, the two statistical signatures blur together and the claimed universal split disappears.","fun_headline_variants_meta":{"raw":{"variants":["Ethereum trades: humans scale, bots don't","Smart contracts break scaling laws on Ethereum","Blockchain scaling: humans stable, bots chaotic","ERC20 flows: human wallets scale, smart contracts drift","Why Ethereum bots defy power laws"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000412,"raw_usage":{"total_tokens":2008,"prompt_tokens":825,"completion_tokens":1183,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":1113}},"tokens_in":569,"tokens_out":1183,"duration_ms":8449,"temperature":1.0,"reasoning_tokens":1113,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T23:50:02.295471+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of addresses the dataset flags as externally owned, check the Ethereum chain state at those times to see whether each address actually has code, then re-fit $\\alpha$, $\\gamma$, and $\\beta$ after correcting mislabeled cases. If corrected EOA-EOA exponents become unstable or corrected contract-contract exponents become near-linear, the reported contrast is a labeling artifact.","supporting_citations":[{"cited_title":"Xblock-eth: Extracting and exploring blockchain data from ethereum","cited_arxiv_id":null,"evidence_quote":"Supplies the nine-month ERC20 transfer records and the contract flags that define the four transaction categories."},{"cited_title":"Network analysis of erc20 tokens trading on ethereum blockchain","cited_arxiv_id":null,"evidence_quote":"Showed ERC20 transaction volumes follow power-law distributions, the pattern this paper extends to account-type categories."},{"cited_title":"Classification of cryptocurrency coins and tokens by the dynamics of their market capitalizations","cited_arxiv_id":null,"evidence_quote":"Documented power-law behavior in cryptocurrency market capitalizations, supporting the universality premise for crypto systems."},{"cited_title":"Aggregation, Variance and the Mean","cited_arxiv_id":null,"evidence_quote":"Introduced the variance-mean scaling law that the paper applies to hourly transaction counts."},{"cited_title":"Fluctuation scaling in complex systems: Taylor’s law and beyond","cited_arxiv_id":null,"evidence_quote":"Provides the fluctuation-scaling framework and the interpretation of Taylor exponents in complex systems."},{"cited_title":"Power-law distri- butions in empirical data","cited_arxiv_id":null,"evidence_quote":"Supplies the maximum-likelihood and Kolmogorov-Smirnov methodology used to fit and validate power-law exponents."},{"cited_title":"Traveling the token world: A graph analysis of ethereum erc20 token ecosystem","cited_arxiv_id":null,"evidence_quote":"Maps the ERC20 token ecosystem and motivates the EOA-versus-contract account distinction in Ethereum."},{"cited_title":"Measuring ethereum-based erc20 token networks","cited_arxiv_id":null,"evidence_quote":"Characterized ERC20 token network structure, giving the network baseline that the four-way split refines."}],"review_version":1}