{"id":"4ea8865a-92ad-4930-b314-950d14453616","arxiv_id":"2608.01934","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Block time distributions in HotStuff-based blockchains can be decomposed via quorum-completion time mixture models, with mainnet fits distinguishing homogeneous (Hyperliquid) from heterogeneous (Aptos) validator deployments.","lead":"This paper models block times in HotStuff-based blockchains as the time needed to collect a quorum of votes, with block time distributions described as mixtures of simple exponential components. Fitting this model to mainnet data from Hyperliquid and Aptos shows that validator homogeneity or heterogeneity leaves a detectable signature in block time statistics.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fitted mixture weights are per-round probabilities, not validator fractions; the paper's reading of w_k as the share of slow validators is an unsupported semantic step on which the diagnostic central claim depends.","rationale":"The reader's weakest assumption is the independence and phase-composition approximation underlying Eq. (14). That is a fair concern, but the more direct threat to the paper's central diagnostic claim is an identification step downstream of the math. Even if Eq. (14) and the mixture in Eq. (8) are correct, the fitted weights w_k are probabilities over rounds under global rate conditions, not probabilities over validators. The text oscillates between those interpretations: Section VI.A correctly says w_k is the fraction of rounds under a condition, while Section V.B and the optimization discussion in Section VI.A treat w_k as the fraction of validators in a slow group. The aggregate expected block time formula in Eq. (29) follows from the round-level mixture, but it does not establish that removing slow validators would reduce the weight of the slow component. Thus the central claim that mixture parameters diagnose validator heterogeneity is conditional on an unstated identification assumption. My proposed simulation would settle the question directly by checking whether the estimator recovers known validator-group parameters under permanent heterogeneity. Because the theoretical derivation is sound and the concern is addressable with a synthetic identifiability check, the appropriate verdict remains conditional acceptance rather than rejection; the condition should be an explicit validation of the validator-level interpretation.","tokens_in":11712,"tokens_out":10778,"duration_ms":106444,"concrete_test":"Simulate a 148-validator HotStuff committee with M_0=99. In each of 200,000 rounds, draw N-1 receiver delays as i.i.d. exponentials, with 10% of validators permanently assigned rate 600 s^-1 and 90% assigned rate 1200 s^-1; record the time of the (M-1)-th arrival. Fit the paper's K=2 EM mixture from Eq. (8) to the simulated block times and compare the recovered w_k and lambda_k to the input 0.10/0.90 and 600/1200. As a control, repeat with the slow group resampled every round, which is the model's own global-rate mixture. If the permanent-heterogeneity case recovers the input validator fractions and rates, the diagnostic mapping is supported; if it yields different weights or rates, or if the fitted mixture is not identifiable from the permanent case, then fitted mixture parameters cannot be interpreted as validator-group shares.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (8) is a mixture over block-time observations: each interval is drawn from one of K order-statistic densities f_{T_M}(t|N, M_0, lambda_k), each with a single global rate lambda_k. The fitted weight w_k is therefore the fraction of rounds that occur under network condition k. Yet Section V.B interprets the same weights as partitioning validators into performance groups ('a fast minority', 'a majority group', 'a slow component'), and Section VI.A turns them into optimization levers by recommending jailing or reputation-weighting slow validators. These are different generative models. A system with permanently slow validators produces a quorum completion time that is the (M-1)-th order statistic of a sample of exponentials with non-identical rates; its distribution is not equal to a finite mixture of identical-rate order-statistic densities. Hence the fitted (w_k, lambda_k) need not correspond to the fraction or the rate of slow validators. The paper itself contains the more careful statement that w_k quantifies 'the fraction of rounds under that condition' (Section VI.A), so the validator-group reading is not derived from the model; it is an additional, unvalidated identification assumption. This is load-bearing because the claimed diagnostic value, localizing heterogeneity and targeting slow validators, depends on that step, not on the correctness of Eq. (14). The limitations section acknowledges correlations but does not flag this semantic gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a diagnostic framework for HotStuff-based BFT blockchains in which the observed inter-block interval is modeled as the time for a proposal to reach a quorum of validators (quorum multicast completion time). The authors derive a closed-form distribution for this completion time as an order statistic of independent exponential receiver delays (Eqs. 10-16), extend it to finite mixtures representing distinct effective transfer-rate regimes, and fit the resulting body-and-tail models to mainnet block-time data from Hyperliquid and Aptos. They report that Hyperliquid is well described by a unimodal model with exponential tail decay, while Aptos is persistently multimodal and shows a shift after the Baby Raptr upgrade, and they interpret fitted mixture weights and rates as actionable diagnostics for validator heterogeneity.","tokens_in":11989,"tokens_out":6020,"duration_ms":55975,"significance":"The theoretical part of the paper is a genuine contribution: the exact closed-form quorum-completion distribution, its tail asymptotics, moment formulas, and the finite-N bound on the parallel phase are clearly derived and appear mathematically correct. If the identification were valid, the framework would offer a simple, on-chain-observable diagnostic for comparing BFT deployments and monitoring protocol upgrades. The paper gains credibility from using real mainnet data at scale (1.2M-10M blocks). However, the empirical validation has an unaddressed internal inconsistency, and the central diagnostic claim depends on an unvalidated semantic step from per-round mixture weights to validator-group shares. Because those issues are load-bearing, the paper's current significance is conditional on resolving them.","major_comments":[{"comment":"The Hyperliquid fit reports lambda = 267.6 s^-1 with N = 21 and M = 14. Using the paper's own Eq. (16), this implies a theoretical tail rate Lambda = (N - M + 1) lambda / N = (8 * 267.6) / 21 ≈ 101.9 s^-1. The empirical tail fit in the same section reports Lambda_tail = 43.9 s^-1, a factor of about 2.3 smaller. This discrepancy is not mentioned or reconciled. If the 90th-percentile tail fit is not intended to estimate the asymptotic rate, then comparing it to Eq. (16) is misleading; if it is intended as an empirical check, it contradicts the fitted body parameter. The claim that the agreement supports the homogeneous-network interpretation needs either a corrected comparison or an explicit explanation of how the two quantities are related.","section":"Section V-A, Eq. (16)"},{"comment":"Equation (8) is a mixture over block-time observations: each round is drawn from one of K order-statistic densities, each with a single global rate lambda_k, and w_k is the fraction of rounds generated under condition k. The paper itself states this in Section VI-A. However, Section V-B and the optimization discussion in Section VI-A interpret the same w_k as partitioning validators into performance groups ('a fast minority,' 'a majority group,' 'a slow component') and recommend jailing or reputation-weighting slow validators. These are different generative models: a fixed subset of permanently slow validators produces an (M-1)-th order statistic of exponentials with non-identical rates, which is not equal to a finite mixture of identical-rate order-statistic densities. The validator-group reading is therefore an additional identification assumption that is neither derived nor validated. This is load-bearing because the proposed diagnostic action—targeting slow validators—depends on that interpretation.","section":"Section V-B and Section VI-A, Eq. (8)"},{"comment":"The model defines block time as T_h ≈ T_M + T_parallel, and Section IV-D argues that T_parallel contributes only O(log N) asymptotically. Yet all empirical fits use the density f_{T_M} in Eq. (14) alone, with no attempt to quantify T_parallel for the actual committee sizes (N = 21 and N ≈ 148). If T_parallel is not negligible at these finite sizes, the fitted lambda_k values are not clean 'effective transfer rates' but lumped parameters that absorb the omitted phase. To support the physical interpretation used throughout Section V and VI, the authors should either bound T_parallel numerically from their data or explicitly reframe lambda_k as effective aggregate parameters.","section":"Sections III-A, IV-D, and V"},{"comment":"The empirical validation is entirely in-sample: lambda_k and w_k are estimated from the same block times they are later used to 'diagnose,' and no held-out data or predictive check is reported. Tail classification relies on comparing R^2 values of exponential and power-law fits without standard errors, confidence intervals, or formal goodness-of-fit tests. For a framework whose central promise is diagnostic monitoring, at least one falsifiable check—such as fitting on one time window and predicting the next, or a likelihood-ratio test against a nested alternative—should be provided.","section":"Sections III-D and V"}],"minor_comments":[{"comment":"The fitted parameters are reported without confidence intervals or standard errors; with 1.2M-10M samples even small departures will be statistically significant, so quantitative comparisons such as the '+63%' rate increase should include uncertainty estimates.","section":"Table II and Fig. 4"},{"comment":"The Hyperliquid analysis cutoff of t <= 0.4 s is described as 'inferred from histogram separation' without a concrete rule; this choice can materially affect the fitted rate and the tail analysis, so the criterion should be specified.","section":"Section III-C"},{"comment":"The displayed density is difficult to parse because of the typesetting; please write the prefactor explicitly as \\binom{N-1}{M-1}\\frac{M-1}{N}\\lambda and keep the exponential factors on one line.","section":"Eq. (14)"},{"comment":"The statement that upgrading or excluding the slow components 'yields approximately 10% latency reduction' is not derived in the text; please show the calculation using Eq. (29) or remove the specific percentage.","section":"Section VI-A"},{"comment":"The use of 'fraction of validators' in the main text conflicts with the correct statement in Section VI-A that w_k is the fraction of rounds under a condition; the wording should be made consistent throughout.","section":"Section V-B"}],"recommendation":"major_revision","confidential_remarks":"The mathematical derivation is sound and the dataset is substantial, but the central diagnostic claim currently rests on an unvalidated identification step and an unreconciled tail-rate inconsistency. I would ask the authors to either reframe the mixture weights strictly as per-round condition shares (removing the validator-group optimization language) or provide an independent validation that the mixture components correspond to actual validator subsets. The tail-rate discrepancy for Hyperliquid should be resolved before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the theory is the strong part. The exact quorum multicast distribution in Eq. (14) is a real, correctly derived contribution—it replaces the full-broadcast Gumbel asymptotics of prior work with the order statistic that actually matches certificate-driven BFT. Second thing: the empirical story overreaches. The stress-test note is right. In the mixture, w_k is the fraction of rounds governed by component k, not the fraction of validators in group k. A system with a permanently slow minority gives a quorum time based on order statistics with non-identical exponential rates, which is not the same as a finite mixture of identical-rate distributions. The paper even says w_k quantifies 'the fraction of rounds under that condition' in Section VI.A, but Section V.B turns components into 'a fast minority' and 'a slow component,' and Section VI.A recommends jailing slow validators. That semantic jump is load-bearing and it is not derived from the model.\n\nThe math deserves credit. The CTMC construction, the partial-fraction expansion, the beta CDF form, the moments, and the tail-asymptotic argument all check out. Fitting on mainnet data from Hyperliquid and Aptos is a real step beyond simulation, and the paper is honest about its independence assumptions and about the power-law tails in Aptos as a signal of correlation.\n\nThe empirical side has gaps you should know about. For Hyperliquid, the body-fit λ=267.6 s^-1 predicts a tail rate Λ≈101.9 s^-1 via Eq. (16), but the independently fitted empirical tail rate is 43.9 s^-1. That is a large discrepancy and the paper does not address it. The 0.4s truncation cap is justified by histogram separation, which is post-hoc. K is chosen by peak detection, which is reasonable but easily confounded by sample size, and there is no held-out validation, no confidence intervals, and no released code. These are fixable, not fatal.\n\nWho should read this: anyone modeling BFT latency or monitoring production consensus chains. The theory is worth a serious referee, and the empirical section needs major revision, especially the identification of mixture weights with validator fractions. I would accept it for peer review, with the expectation that the authors either defend that identification or soften the diagnostic claims to what the model can actually support.","headline":"Correct quorum-completion theory, but the validator-group reading of mixture weights doesn't follow from the model.","tokens_in":12535,"tokens_out":3210,"would_cite":true,"duration_ms":27948,"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":"Block times expose validator speed gaps in BFT chains","keywords":["Byzantine fault tolerance","HotStuff","block time distribution","quorum multicast","mixture model","hypoexponential distribution","validator heterogeneity","latency diagnosis"],"falsifier":"Instrument a validator or testnet to record actual proposal-receipt times and compare the empirical $M$-of-$N$ order statistic with Eq. (14); if known correlated delays produce a power-law tail when the true rates are exponential, the independence assumption fails. Alternatively, measure block times on a chain where validator groups are deliberately split by latency and check whether the fitted mixture components track those groups.","tokens_in":11478,"feed_emoji":"⛓️","tokens_out":8109,"duration_ms":66319,"temperature":0.7,"pith_summary":"This paper claims that the interval between consecutive blocks in HotStuff-style BFT blockchains is governed by quorum formation rather than full-committee broadcast, because a block advances once a quorum of validators has received it. The authors derive an exact closed-form distribution for that quorum-completion time and show that observed block-time histograms can be decomposed as finite mixtures of this distribution, one component per distinct effective transfer rate. On mainnet data, the fit is unimodal for Hyperliquid and persistently multimodal for Aptos, with a visible shift after Aptos's consensus upgrade. If the model holds, mixture weights and tail decay become practical signals for spotting validator heterogeneity and monitoring protocol changes.","feed_headline":"Block times expose validator speed gaps in BFT chains","feed_subtitle":"A closed-form quorum-completion distribution separates fast and slow groups on Hyperliquid and Aptos mainnets.","key_machinery":"The load-bearing object is the quorum multicast completion time $T_M$: a continuous-time Markov chain on states $i=1,\\dots,M$ where each uninformed validator receives the proposal at normalized rate $\\lambda/N$, so the transition rate from state $i$ is $\\alpha_i=(N-i)\\lambda/N$. Because the sojourn times are independent exponentials, $T_M$ is a sum of exponentials—a hypoexponential distribution—whose Laplace transform factorizes and whose density after partial-fraction expansion is Eq. (14). The same object yields a Beta-form CDF (Eq. 15), harmonic-number mean and variance (Eqs. 17–18), an $O(\\log N)$ correction for the parallel vote-collection phase (Eq. 25), and the mixture likelihood used for fitting (Eq. 8). The tail rate $\\Lambda=(N-M+1)\\lambda/N$ is what separates homogeneous (exponential) from heterogeneous (power-law) behavior in the empirical analysis.","core_discovery":"The central discovery is a distributional identity: the quorum multicast completion time $T_M$, defined as the time for a proposal to reach $M$ of $N$ validators under independent exponential per-validator delays with rate $\\lambda/N$, has the closed-form density $$f_{T_M}(t)=\\binom{N-1}{M-1}\\frac{M-1}{N}\\$\\lambda$\\,$e^{{-(N-M+1)\\lambda t/N}}$\\left(1-$e^{{-\\lambda t/N}}$\\right)^{M-2}$$ and its tail decays as a single exponential with rate $\\Lambda=(N-M+1)\\lambda/N$. The paper extends this to finite mixtures $p(t)=\\sum_k w_k f_{T_M}(t;N,M_0,\\lambda_k)$, fitted by expectation maximization, where each component represents a network condition with effective rate $\\lambda_k$ and weight $w_k$. Applied to production data, Hyperliquid's block times are unimodal ($\\hat\\lambda=267.6$ s$^{-1}$, exponential tail), while Aptos before the BR upgrade needs three components and after it needs four, with all rates higher and a power-law tail; the authors read this as homogeneous versus heterogeneous validator deployments.","pith_inferences":["The authors leave implicit that sliding the fitting window over time and tracking $\\{w_k,\\lambda_k\\}$ would act as a change-point detector, flagging protocol upgrades or infrastructure shifts before the histogram shape alone makes them obvious.","A testable extension is that jailing and reputation mechanisms act on weights before rates: removing slow validators should lower the slow-component weight while leaving the remaining $\\lambda_k$ roughly unchanged, a separation the model predicts and on-chain data could check.","If correlated cloud-region delays violate the independence assumption, the Beta-form CDF (Eq. 15) overestimates quorum-completion probability; simulating Eq. (14) under explicit correlated delays would show how far the fitted diagnostic parameters drift before the homogeneity reading flips."],"forward_implications":["Mixture weights $w_k$ and rates $\\lambda_k$ give an operational partition of rounds by network condition, so a deployment can see what fraction of blocks are produced under each speed regime.","The aggregate expected block time reduces to $E[T_h]=N(H_{N-1}-H_{N-M})\\sum_k w_k/\\lambda_k$, meaning slow components contribute disproportionately through $w_k/\\lambda_k$ and are the natural target for jailing or reputation-based leader selection.","Tail decay rate diagnoses homogeneity: exponential tails match a single effective rate, while power-law tails signal correlated delays or path-dependent mechanisms like payload fetch misses.","Protocol upgrades can be tracked by shifts in mixture structure; the Aptos BR upgrade raised all component rates and split one slow component into two, so mixture fits act as a monitoring instrument.","The model implies that as committee size grows, maintaining unimodal, exponentially-tailed block times becomes harder without explicit homogeneity enforcement, since the mean scales as $N$ and the parallel phase adds only $O(\\log N)$."],"supporting_citations":[{"why":"Defines the chained HotStuff protocol whose pipeline phases the quorum multicast model abstracts.","marker":"[11]"},{"why":"Supplies the full-committee broadcast/Gumbel baseline that the quorum-completion model replaces, along with the block-time decomposition.","marker":"[27]"},{"why":"Provides the coupon-collector broadcast model and the $T_{\\text{block}}=T_{\\text{create}}+T_{\\text{broadcast}}+T_{\\text{validate}}$ decomposition.","marker":"[28]"},{"why":"Gives single and parallel broadcast time formulas that the paper extends to the $M$-of-$N$ quorum multicast setting.","marker":"[29]"},{"why":"Defines the Aptos system and committee parameters used in the empirical fits.","marker":"[5]"},{"why":"Defines Hyperliquid and its validator and latency assumptions used as the homogeneous comparison case.","marker":"[7]"},{"why":"Introduces Baby Raptr's optimistic data availability, the upgrade whose post-BR block-time distribution is analyzed.","marker":"[22]"},{"why":"Documents the Order Votes consensus path that motivates treating block time as pipeline cadence rather than finality latency.","marker":"[21]"}],"fun_headline_variants":["Block time mixtures unmask validator heterogeneity in BFT chains","Mixture model reveals multimodal block times on Aptos post-upgrade","Closed-form quorum delay separates validator groups on Aptos and Hyperliquid","Mixture modeling spots distinct block-time modes on BFT mainnets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole diagnostic reading rests on the assumption that each validator's receipt delay is independent and exponential at a common rate, so the quorum-completion time is an exact sum of exponentials, and that the observed block interval is essentially that quorum phase rather than a mix of other pipeline stages.","fun_headline_variants_meta":{"raw":{"variants":["Block time mixtures unmask validator heterogeneity in BFT chains","Mixture model reveals multimodal block times on Aptos post-upgrade","Closed-form quorum delay separates validator groups on Aptos and Hyperliquid","Mixture modeling spots distinct block-time modes on BFT mainnets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001303,"raw_usage":{"total_tokens":5333,"prompt_tokens":984,"completion_tokens":4349,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":4275}},"tokens_in":600,"tokens_out":4349,"duration_ms":26948,"temperature":1.0,"reasoning_tokens":4275,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:03:45.639175+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Instrument a validator or testnet to record actual proposal-receipt times and compare the empirical $M$-of-$N$ order statistic with Eq. (14); if known correlated delays produce a power-law tail when the true rates are exponential, the independence assumption fails. Alternatively, measure block times on a chain where validator groups are deliberately split by latency and check whether the fitted mixture components track those groups.","supporting_citations":[{"cited_title":"HotStuff: BFT consensus in the lens of blockchain,","cited_arxiv_id":null,"evidence_quote":"Defines the chained HotStuff protocol whose pipeline phases the quorum multicast model abstracts."},{"cited_title":"Mitigation and resurgence of the scalability trilemma: Mathematical analysis for optimal sharding in dual–layer Byzantine fault–tolerant blockchains,","cited_arxiv_id":null,"evidence_quote":"Provides the coupon-collector broadcast model and the $T_{\\text{block}}=T_{\\text{create}}+T_{\\text{broadcast}}+T_{\\text{validate}}$ decomposition."},{"cited_title":"Aptos White Paper,","cited_arxiv_id":null,"evidence_quote":"Defines the Aptos system and committee parameters used in the empirical fits."},{"cited_title":"HyperCore Overview,","cited_arxiv_id":null,"evidence_quote":"Defines Hyperliquid and its validator and latency assumptions used as the homogeneous comparison case."},{"cited_title":"Raptr: Prefix Consensus for Robust High-Performance BFT","cited_arxiv_id":"2504.18649","evidence_quote":"Introduces Baby Raptr's optimistic data availability, the upgrade whose post-BR block-time distribution is analyzed."},{"cited_title":"AIP-89: Consensus Latency Reduction using Order V otes","cited_arxiv_id":null,"evidence_quote":"Documents the Order Votes consensus path that motivates treating block time as pipeline cadence rather than finality latency."}],"review_version":2}