{"id":"7879d585-851a-4ff9-a9e2-d70ed4f3903b","arxiv_id":"2507.20494","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A two-score system for Uniswap wallets is built from rule-based blueprints refined by a deep residual network, with validation limited to reproducing the rules.","lead":"Dhanashekar and colleagues give Uniswap wallets two scores, one for those who provide liquidity and one for those who swap tokens, by combining hand-made rules with a neural network. The result could inform DeFi risk and incentives, but the paper's own tests only show the network learning to copy the hand-made rules.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central empirical claim is self-referential: the model is trained to predict blueprint scores and then evaluated against those same scores and against features built into the blueprint, so no independent evidence supports the claimed behavioral validity.","rationale":"The reader identified the blueprint assumption as weakest, and I agree with the core direction. The load-bearing defect is slightly sharper: even if we grant the blueprint a reasonable definition, the paper's reported evidence cannot establish the claim because the evaluation target and the score-generation target are the same. In detail: Section 3.2 says the neural net is trained on noise-augmented blueprint scores, Section 4.1-4.2 then report 'predicted vs. blueprint' agreement, and Tables 1-2 show correlations between features and predicted score where the features are the ones used to calculate the blueprint. This is not a hidden implementation flaw; it is a logical property of the method. The only way to break the circularity is to validate the blueprint against something not used in its construction. The external validation report linked in Section 4 is not included in the manuscript and, per the text, only repeats 'behavioral analysis across score bins'—the same internal metrics—so it does not supply the missing evidence. Given that the contribution's stated value is downstream reputation/risk use, the absence of any external outcome test makes the central claim unsupported. I would keep the reader's REJECT verdict, while noting the engineering pipeline itself may be reusable if such validation is added.","tokens_in":8711,"tokens_out":3701,"duration_ms":49937,"concrete_test":"Extend the dataset with a time-based holdout: score wallets using data from a first window (e.g., months 1-6 of the Uniswap v3 sample), then measure an external outcome in a later window (e.g., realized LP returns net of impermanent loss, survival/continued liquidity provision after 6 months, or success on a lending protocol such as avoiding liquidation). Compare zScore quintiles against raw volume/frequency baselines in a logistic or regression model. If high-zScore wallets show no significant incremental separation on such outcomes, the central behavioral-validity claim fails; if they do, the rubric has external support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 and the conclusion assert that 'empirical evaluations confirm that zScores align closely with strategic behavior patterns' and that high-scoring users 'exhibit strong liquidity retention, disciplined trading activity, and broad engagement across pools.' But the only training targets in the supervised regression are the blueprint scores from Section 3.1; the noise-injected labels in Section 3.2 are Gaussian perturbations of those same scores. Therefore the 91.79%/90.83% agreement in Figures 2-3 measures how well a flexible network reproduces its own training targets, not whether the scores capture behavior. The bin-wise tables (Tables 1-2) also use metrics that are components of the blueprint—holding duration, withdrawal volume, swap volume, token diversity—so monotonic trends across bins are entailed by the hand-set weights and caps, not discovered. The blueprint's weights are never justified against an external outcome; without such validation, 'strategic behavior' is just the authors' rubric restated. The caveat that dusk wallets are removed is reasonable, but it does not address target validity. Thus the central claim is an unvalidated assumption, not an empirical result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a dual-role behavioral scoring framework for Uniswap v3 wallets, producing a Liquidity Provision Score and a Swap Behavior Score. Scores are first generated by a hand-designed rule-based blueprint (Section 3.1), then Gaussian noise is added to these scores to create training targets, and a deep residual network (DeepMLP_ResNet) is trained to regress toward those noisy targets (Section 3.2). The empirical evaluation in Section 4 compares model predictions to the blueprint scores and reports 91.79% / 90.83% agreement within a ±50 tolerance, and presents bin-wise behavioral summaries in Tables 1-2 as evidence that higher scores correspond to more strategic behavior. The conclusion claims that zScores align closely with strategic behavior patterns and that high-scoring users exhibit strong liquidity retention, disciplined trading, and broad engagement.","tokens_in":9001,"tokens_out":3494,"duration_ms":44288,"significance":"If the claimed behavioral validity were established, an interpretable, context-aware, dual-role scoring system for DeFi wallets could be a useful building block for risk models, incentive design, and reputation systems in decentralized finance. The paper has some strengths: the blueprint is explicit and interpretable, the architecture is described in detail, and the authors provide a link to a supplementary validation report and code repository. However, the central empirical claim is not supported by the evidence presented: the 'ground truth' used for evaluation is the same blueprint score that generated the training labels, so the reported accuracy is a measure of how well the model fits its own training targets. The bin-wise behavioral trends in Tables 1-2 are also largely entailed by the features and weights built into the blueprint rather than independently discovered. No external outcome, baseline model, or statistical uncertainty is provided. Therefore, the paper's main contribution, as stated in the conclusion, is an unvalidated assumption rather than an empirically confirmed result.","major_comments":[{"comment":"The evaluation is circular. The 'ground truth' scores in Figures 2-3 are the blueprint scores from Section 3.1, which were then corrupted with Gaussian noise in Section 3.2 to form the network's training targets. The reported 91.79% and 90.83% agreement within ±50 therefore measures how well the network reproduces its own training signal, not how well the scores predict any externally meaningful behavior or outcome. This directly undermines the conclusion that 'empirical evaluations confirm that zScores align closely with strategic behavior patterns.'","section":"Section 4.1 and 4.2 (Figures 2 and 3)"},{"comment":"The bin-wise behavioral summaries do not provide independent validation. The metrics shown in Tables 1-2 (liquidity retention, holding duration, withdrawal volume, swap count, token diversity) are the same sub-category features used in the blueprint to construct the scores, with hand-set weights and caps. The increasing trends across score bins are therefore mechanical consequences of the blueprint's construction, not discoveries about wallet behavior. The paper would need to validate the scores against behaviors or outcomes that were not used to define the score.","section":"Tables 1-2 and Section 4.3"},{"comment":"There is no external ground truth, no baseline, and no uncertainty quantification. The blueprint weights, caps, noise standard deviation, dusk thresholds, and tolerance window are all free parameters, and none is tested against an external outcome such as future returns, protocol revenue, loan repayment, or an independent reputation label. Without such validation, 'strategic behavior' is simply a restatement of the authors' rubric. Additionally, the paper does not compare the deep network against linear regression or against the raw blueprint scores, so the claimed benefit of the neural refinement phase is not established.","section":"Sections 3.1-3.2 and Section 4"},{"comment":"The post-hoc exclusion of 'dusk' wallets, defined as wallets with one transaction and negligible volume, is applied without reporting results with those wallets included. Because this exclusion is made after examining the data and without a sensitivity analysis, it is unclear whether the reported accuracy and bin trends hold for the full population. The exclusion may be reasonable, but it should be presented as a robustness check with both full-sample and filtered results, and the threshold definitions should be precise.","section":"Section 4.3 (dusk wallet exclusion)"}],"minor_comments":[{"comment":"The text 'often below 10f orLP sor50 for swaps' appears garbled and should be corrected to clearly state the dusk thresholds (e.g., dollar amounts) for LP deposits and swap volumes.","section":"Section 4.3.2"},{"comment":"The term 'T_scores' is used without definition; the reader must infer that it refers to the blueprint scores. Please define this notation.","section":"Section 4.1"},{"comment":"Section 3.1 states that 'no rigid cutoffs are enforced,' while Section 3.2 states that each sub-category is 'capped by a predefined maximum'; this apparent inconsistency should be clarified.","section":"Sections 3.1 and 3.2"},{"comment":"The supplementary validation report is a GitHub link rather than an archival, peer-reviewed artifact. If this report is intended to support the paper's claims, it should be made available in a permanent form with reproducibility instructions, and its contents should be summarized in the manuscript.","section":"External Validation Report"}],"recommendation":"reject","confidential_remarks":"The manuscript's main empirical claim is self-referential because the evaluation compares the model's outputs to the same blueprint scores used to construct its training labels. This is not a minor flaw; it means the central thesis—that zScores align with strategic behavior—is not demonstrated by the experiments reported. The paper may be salvageable if the authors can add an external validation study (e.g., predicting future liquidity provision, protocol revenue contribution, or credit-like outcomes) and compare against simple baselines. As it stands, I cannot recommend publication in a serious journal. I also note that the paper's fit to a q-fin.GN audience is plausible, but the empirical rigor is insufficient for the claims made."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"If you need a practical wallet-ranking system for Uniswap v3, this paper gives you a clear, working recipe: two rule-based blueprint scores (LP and swap), noise-injected labels, and a residual MLP with pool context. The engineering is described well, the dataset is real on-chain data, and they point to an external validation report on GitHub. That is worth credit. The U-Net framing is cosmetic—it is a plain MLP with skip connections—but the architecture is sensible for the task.\n\nThe soft spot is the one the stress-test flags, and it is load-bearing. Every accuracy claim in Section 4 compares the network's output to the blueprint scores that were used to generate the training targets. Getting 91% agreement there shows the model fits its own labels, not that the scores capture anything externally meaningful. The bin-wise behavioral tables in Section 4.3 are also entailed by construction: they use features like holding duration, withdrawal volume, and token diversity that are literally components of the blueprint. So the monotonic trends are not discovered, they are baked in.\n\nThere are no baselines, no error bars, and the blueprint weights are hand-set without any external outcome (default, creditworthiness, future returns) to justify them. The paper is honest that the blueprint is the target, but the conclusion overreaches when it says the scores align with 'strategic behavior patterns.' What we actually know is that the scores align with the authors' rubric. That may be useful for internal segmentation, but it is not validation.\n\nOne more thing: the paper asserts independence from Udupi et al. [16], but the author lists overlap heavily (Sahoo, Paul, and others appear on both). That is not a scientific flaw, but it is misleading and a referee should ask for a clearer statement of what 'independently developed' means here.\n\nSo: the paper is a decent technical report for practitioners building DeFi risk or airdrop tools, not a demonstrated scientific result about wallet reputation. I would send it to peer review only if the venue is willing to demand external validation or reframing; as it stands, the central claim is unsupported. A serious referee could get something useful out of it, but the paper needs major revision before it should be published as anything more than a systems note.","headline":"A real engineering contribution with a self-referential evaluation: the dual-score system trains and tests on its own blueprint, so the behavioral-validity claims do not yet hold.","tokens_in":684,"tokens_out":2237,"would_cite":false,"duration_ms":35762,"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 a wallet's Uniswap v3 behavior can be summarized by two blueprint-guided scores—liquidity provision and swap behavior—refined by a deep residual network, and that high-scoring wallets exhibit disciplined…","keywords":["DeFi","zScore","wallet reputation","Uniswap v3","liquidity provision","behavioral scoring","deep residual network","on-chain behavior"],"falsifier":"Take a training window of Uniswap v3 activity, score wallets with zScore, then follow the same wallets in a later window on an outcome the blueprint does not use—say, whether they still hold liquidity after 90 days, whether they repay a DeFi loan, or whether their realized LP returns after gas and impermanent loss are positive. If top-zScore wallets do not outperform bottom-zScore wallets on those held-out outcomes, the claim that the score ranks strategic behavior is falsified.","tokens_in":8532,"feed_emoji":"📊","tokens_out":6579,"duration_ms":70097,"temperature":0.7,"pith_summary":"The paper proposes a way to score DeFi wallets on Uniswap v3 without any identity data, using two complementary numbers: a Liquidity Provision Score and a Swap Behavior Score. The scores start from hand-built 'blueprint' rules that decompose behavior into volume, frequency, holding time, and withdrawal patterns, then a deep residual neural network is trained on noise-augmented blueprint scores to smooth out the rules and capture feature interactions. Pool context—TVL, fee tier, pool size—is folded in so the same behavior counts differently in a deep, low-fee pool versus a shallow, high-fee one. The authors report that on held-out wallets, 91.79% of LP scores and 90.83% of swap scores fall within ±50 points of the blueprint targets, and that higher-scoring wallets show stronger liquidity retention, longer holding, and broader token engagement. The larger claim is that these scores can support risk assessment, airdrop design, and governance weighting in DeFi.","feed_headline":"zScore ranks DeFi wallets by liquidity and trading discipline","feed_subtitle":"Uniswap v3 wallets get two behavior scores, and 91% of predictions land within 50 points of the blueprint target.","key_machinery":"The load-bearing machinery is a blueprint-guided, noise-augmented supervised scoring pipeline: a hand-specified, weighted, per-category-capped scoring schema for LPs and swappers, plus Gaussian noise injection that creates soft training targets, plus a DeepMLP_ResNet (four residual blocks with projection shortcuts, layer normalization, dropout, and a regression head) that learns to reproduce those targets under wallet-level train/validation splits. The blueprint supplies interpretability and structure; the network supplies smooth, context-aware generalization; and pool-level features (TVL, fee tier, relative pool size) are appended so identical behavior is scored differently across pool environments.","core_discovery":"The central discovery is that a two-phase hybrid—interpretable rule-based blueprint followed by a noise-injected deep residual network with U-Net-style dense skip blocks—can produce a continuous 0–1000 behavioral score for Uniswap v3 wallets that differentiates roles and aligns with the authors' definition of strategic behavior. The LP score rewards long holding, low withdrawal, and high retention; the swap score rewards volume, frequency, token diversity, and routing complexity while penalizing micro-swaps and wash-like patterns. On their validation set, high-score bins concentrate exactly those behaviors: retention ratios rise from 0.00 in the lowest LP bin to 1.00 in the highest, and swap volumes climb from $4K to $82M across swap bins. The paper frames this as a bridge between anonymous on-chain interaction and trust-based protocol alignment.","pith_inferences":["The reported 91% accuracies measure agreement with the authors' own rubric, not truth about reputation; a real test would compare zScore against external outcomes such as default rates, realized LP profitability, or continued participation after a market shock.","The deliberate exclusion of one-transaction wallets may inflate the observed score-behavior gradient, since the lowest-activity band is removed before bin analysis; re-including these wallets could change the shape of the tables.","Pool-level context suggests a natural stress test: check whether wallet rankings stay stable when a pool changes fee tier or experiences a sharp TVL change, to see whether the score reflects durable behavior or merely current pool conditions.","The dual-score output could be treated as a two-dimensional behavioral embedding, allowing clustering and anomaly detection to run directly on the score pair rather than on raw transaction features."],"forward_implications":["Liquidity and swap scores give protocols a cheap, on-chain-only way to segment users by role—LP or trader—without any identity verification.","According to the bin tables, high-zScore wallets retain liquidity longer and trade more broadly, so the score could be used to gate airdrops, weight governance votes, or adjust lending terms toward users with stable behavior.","The noise-injected blueprint pipeline turns a fixed rule system into a differentiable scoring function, allowing pool context to be updated as markets shift without re-specifying all thresholds.","The architecture is reusable: swapping the blueprint to another role or protocol, such as lending or borrowing, yields a comparable behavioral score from the same network structure."],"supporting_citations":[{"why":"Provides foundational analysis of the Uniswap automated market maker structure that motivates the behavioral features the blueprint scores.","marker":"[2]"},{"why":"Quantifies impermanent loss for liquidity providers, supporting the risk-adjusted view of LP strategy that informs the blueprint.","marker":"[5]"},{"why":"Gives empirical evidence that Uniswap v3 LP returns vary strongly with strategy and volatility, which the dual-role scoring is designed to capture.","marker":"[11]"},{"why":"Introduces a prior cross-protocol zScore reputation system that the paper explicitly distinguishes its own metric from as methodologically different.","marker":"[16]"},{"why":"Presents a network-propagation wallet risk model, a related approach the paper differentiates from fine-grained behavioral scoring.","marker":"[6]"},{"why":"Argues for transparency and auditability in machine-learning credit scoring, grounding the paper's interpretability requirement for the architecture.","marker":"[14]"},{"why":"Surveys statistical and machine-learning credit-scoring models and supports the hybrid interpretable-ML approach the paper applies to wallets.","marker":"[20]"}],"fun_headline_variants":["Dual zScore ranks Uniswap wallets for risk models","zScore splits Uniswap users into LP and swap profiles","Deep residual net scores DeFi wallets from on-chain actions","Uniswap reputation: two scores from liquidity and trading","Hybrid zScore: rule-based plus deep learning for DeFi risk"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The hand-set blueprint weights and caps are assumed to be a valid measure of wallet reputation, and nothing outside the blueprint is used to check that assumption.","fun_headline_variants_meta":{"raw":{"variants":["Dual zScore ranks Uniswap wallets for risk models","zScore splits Uniswap users into LP and swap profiles","Deep residual net scores DeFi wallets from on-chain actions","Uniswap reputation: two scores from liquidity and trading","Hybrid zScore: rule-based plus deep learning for DeFi risk"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000238,"raw_usage":{"total_tokens":1514,"prompt_tokens":950,"completion_tokens":564,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":478}},"tokens_in":566,"tokens_out":564,"duration_ms":7266,"temperature":1.0,"reasoning_tokens":478,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:33:02.398623+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a training window of Uniswap v3 activity, score wallets with zScore, then follow the same wallets in a later window on an outcome the blueprint does not use—say, whether they still hold liquidity after 90 days, whether they repay a DeFi loan, or whether their realized LP returns after gas and impermanent loss are positive. If top-zScore wallets do not outperform bottom-zScore wallets on those held-out outcomes, the claim that the score ranks strategic behavior is falsified.","supporting_citations":[{"cited_title":"Uniswap and the emergence of the decentralized exchange.Journal of financial market infrastructures, 10(2):1–25, 2021","cited_arxiv_id":null,"evidence_quote":"Provides foundational analysis of the Uniswap automated market maker structure that motivates the behavioral features the blueprint scores."},{"cited_title":"zScore: A Universal Decentralised Reputation System for the Blockchain Economy","cited_arxiv_id":"2503.05718","evidence_quote":"Introduces a prior cross-protocol zScore reputation system that the paper explicitly distinguishes its own metric from as methodologically different."},{"cited_title":"Riskprop: Account risk rating on ethereum via de-anonymous score and network propagation.IEEE Transactions on Dependable and Secure Computing, 2024","cited_arxiv_id":null,"evidence_quote":"Presents a network-propagation wallet risk model, a related approach the paper differentiates from fine-grained behavioral scoring."},{"cited_title":"Transparency, auditability, and explainability of machine learning models in credit scoring.Journal of the Operational Research Society, 73(1):70–90, 2022","cited_arxiv_id":null,"evidence_quote":"Argues for transparency and auditability in machine-learning credit scoring, grounding the paper's interpretability requirement for the architecture."},{"cited_title":"Statistical and machine learning models in credit scoring: A systematic literature survey.Applied Soft Computing, 91:106263, 2020","cited_arxiv_id":null,"evidence_quote":"Surveys statistical and machine-learning credit-scoring models and supports the hybrid interpretable-ML approach the paper applies to wallets."}],"review_version":1}