RepStake: A Blockchain-Based Trust System with Reputation Staking

RepStake: A Blockchain-Based Trust System with Reputation Staking Dimitris Mantzonis Department of Digital Industry Technologies National and Kapodistrian University of Athens (NKUA) Evripos Complex, Psachna, 34400, Greece [email protected] Thanasis G. Papaioannou Department of Digital Industry Technologies National and Kapodistrian University of Athens (NKUA) Evripos Complex, Psachna, 34400, Greece [email protected] Abstract Reputation systems are fundamental to fostering trust and cooperation in digital environments, yet existing solutions often struggle with centralization, vulnerability to manipulation, and limited portability. Centralized reputation platforms can be opaque, censored and susceptible to become single points of failure, while decentralized ones face challenges, such as Sybil attacks, malicious strategies (e.g., ballot stuffing, bad-mouthing) exercised by entities of high influence, i.e. “whales”, and privacy concerns. This paper addresses these persistent issues by proposing a blockchain-based reputation framework that integrates robust identity verification, square root voting constraints, and dynamic stake-based incentives. Rating power is linked to the reputation of the rater that puts its reputation at stake. The model aims to ensure that reputation is earned and maintained through verifiable, community-aligned actions, while simultaneously limiting the potential for abuse by malicious actors or disproportionately influential participants. By leveraging decentralized identifiers, zero-knowledge proofs, and transparent incentive mechanisms, the proposed system seeks to balance transparency, fairness, and privacy. Extensive simulation experiments prove that the approach is effective to reveal the true quality of entities, even in presence of 49% colluding voters. The approach is designed to be adaptable across diverse domains, ranging from marketplaces and collaborative platforms to decentralized finance and governance. CCS Concepts •Information systems → Reputation systems;•Computer systems organization →Peer-to-peer architectures. Keywords trust, hidden quality, blockchain, collusion, Sybil attack ACM Reference Format: Dimitris Mantzonis and Thanasis G. Papaioannou. 2026. RepStake: A Blockchain-Based Trust System with Reputation Staking. In The 41st ACM/SIGAPP Symposium on Applied Computing (SAC ’26), March 23–27, 2026, Thessaloniki, Greece. ACM, New York, NY, USA, 8 pages. https://doi.org/10.1145/3748522. 3779898 This work is licensed under a Creative Commons Attribution 4.0 International License. SAC ’26, Thessaloniki, Greece ©2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2294-3/2026/03 https://doi.org/10.1145/3748522.3779898 1 Introduction Trust is a cornerstone of successful digital interactions, enabling collaboration, commerce, and governance among parties who may have no prior relationship. In online marketplaces, peer-to-peer platforms, and decentralized autonomous organizations (DAOs), reputation systems serve as critical infrastructure for signaling trustworthiness, deterring malicious behavior, and guiding user decisions. The effectiveness of these systems, however, is increasingly challenged by the scale, complexity, and adversarial dynamics of modern digital ecosystems. Centralized reputation systems, such as those employed by major e-commerce platforms and service marketplaces, have historically provided convenient means of aggregating and presenting user feedback [ 18 ]. While these systems have contributed to the growth of digital commerce, they are not without significant drawbacks. Centralized control introduces risks of censorship, arbitrary moderation, and single points of failure [ 2 ]. Moreover, users are often unable to transfer their reputation across platforms, limiting the portability and utility of their digital trust capital [ 6 ]. High-profile incidents of review manipulation and biased moderation have further eroded user confidence in centralized reputation providers. Privacy and regulatory compliance add further complexity. Reputation systems must strike a delicate balance between transparency and user privacy. Users increasingly demand control over their personal data and the ability to interact pseudonymously, without sacrificing the verifiability and accountability necessary for robust reputation mechanisms. Decentralized reputation systems, often built on blockchain or distributed ledger technologies, promise greater transparency, immutability and user autonomy. By recording reputation-related events on a public ledger, these systems can provide verifiable and tamper-resistant histories of user behavior [ 5 ]. However, decentralization introduces new challenges. The most serious one is the threat of Sybil attacks, in which adversaries create multiple fake identities to unfairly influence outcomes or extract disproportionate rewards [ 12 ]. Additionally, many decentralized systems struggle with the problem of “whale dominance”, where users with large stakes or extensive participation histories can exert outsized influence, undermining the egalitarian ideals of decentralized governance [ 14 ]. Static or poorly designed incentive structures may also fail to encourage ongoing, honest participation, leading to reputation inflation, stagnation, or even the emergence of echo chambers [ 11 ]. In a PoS blockchain system, derived from stake delegation, staking as a means to assign trust to an entity has been employed in [ 15 ]. The trustworthiness of an entity for a particular context equals to its relative stake as compared to that of other entities for the particular context. However, in [ 15 ], trust SAC ’26, March 23–27, 2026, Thessaloniki, Greece D. Mantzonis and T. G. Papaioannou manipulation is feasible, yet costly. In [ 19 ], citizens may confirm or refute information posted by others in a smart city context by betting on it their individual reputation and get some reputation back when they are considered to be right. A similar approach, referred to as Intrinsic Integrity-driven Rating Model (IIRM), was proposed in [ 21 ] where reputation is staked for providing ratings. Despite their merit, both approaches in [ 19 , 21 ] have not dealt with collusion resistance. Comparatively, established approaches like the EigenTrust [ 13 ] model trust as a transitive flow, which, while mathematically elegant, often suffers from “costless voting” and relies on pre-trusted peers that reintroduce centralization risks. In the realm of identity, systems like Human Passport [ 10 ] (formerly Gitcoin Passport) and BrightID [ 4 ] rely on signal aggregation and social graph analysis, respectively. While accessible, these models remain vulnerable to “Sybil farming” and privacy leakage through graph analysis. Furthermore, governance models leveraging Soulbound Tokens (SBTs) [ 17 ] or EigenLayer-style restaking [ 20 ] often result in static credentials or rely solely on financial capital for security. RepStake diverges by introducing “Social Restaking”—making reputation a dynamic, stakable asset protected by ZK-biometric assurance rather than financial barriers or static graphs. Also, Bless et al. in [ 3 ] propose a reputation system for scientific contributions where they employ tokens as rewards for high-quality contributions and collateral to ensure high-quality submissions. However, token starvation or inflation are still possible in [ 3 ]. In [ 17 ], they propose a Soulbound Token (SBT)-based reputation system for sustainable supply chains. SBTs, i.e., non-transferable blockchain tokens, track and verify entities’ sustainability practices immutably. The approach in [ 17 ] has privacy and scalability issues. Despite significant advances in the field, there remains a need for reputation frameworks that are not only secure and manipulationresistant but also fair, privacy-conscious, and adaptable to a wide range of digital environments. This paper addresses these challenges by proposing a blockchain-based reputation system that integrates robust identity verification, square root voting constraints, and dynamic, stake-based incentives. By leveraging decentralized identifiers, zero-knowledge proofs, and transparent smart contract logic, the proposed model aims to create a reputation system that is resistant to manipulation based on honest majorities, equitable in influence distribution, and sustainable in its incentive structure. The framework is designed to be interoperable and portable, supporting applications across e-commerce, decentralized finance, collaborative platforms, and beyond. The remainder of this work is organized to systematically address the design, validation, and implications of the proposed decentralized reputation system. Section two delineates the architecture of the decentralized reputation system, with emphasis on its quadratic endorsement protocol, dynamic stake adjustment mechanics, and ERC20-compatible reputation tokenization. Section three presents a rigorous evaluation of scalability under high transaction loads and collusion resistance against coordinated attacks. To quantitatively validate the framework, section four demonstrates the system’s applicability through real-world case studies such as e-commerce feedback aggregation, cloud provision feedback, and decentralized autonomous organization peer assessment. Section five discusses operational aspects such as scalability and transaction costs. Finally, section six synthesizes key findings, discusses limitations, and outlines future research directions for enhancing privacy guarantees and interoperability in decentralized reputation ecosystems. 2 The Decentralized Reputation System This section presents a decentralized reputation system leveraging ERC20 tokens, designed to evaluate trustworthiness in online platforms such as e-commerce and cloud services. The system incorporates a novel dynamic staking mechanism, where voters risk reputation to cast votes, with stakes adjusting based on community alignment. Built on blockchain technology, it ensures transparency, Sybil resistance, and fairness through square root budgeting and vote weight decay. 2.1 Identity Verification via ZKorum’s Racine Protocol RepStake integrates ZKorum’s Racine Protocol [ 9 ] as the foundation for user identity verification, replacing traditional KYC-verified credentials with a privacy-centric, cryptographically secure method. Developed by ZKorum for their Agora Citizen Network, this protocol enables users to prove their humanity and uniqueness through zero-knowledge proofs without disclosing personally identifiable information. The Racine Protocol serves as the root of trust in digital interactions, aligning with RepStake’s goals of fairness, transparency, and user autonomy in a decentralized reputation system. The core components adapted from ZKorum’s system include the Rarimo Protocol integration, which converts biometric passports into anonymous digital identities stored locally on users’ devices, and Zero-Knowledge Proof Verification, which provides cryptographic assurance of a unique human identity. Additionally, user-controlled authorization networks (UCAN)-based authentication ensures decentralized, secure authentication, while privacypreserving credentials support standards such as W3C verifiable credentials and biometric passport verification, maintaining user anonymity while preventing Sybil attacks. By adopting ZKorum’s methodology, RepStake ensures GDPR compliance by design, as no personal data is stored or processed centrally, and enhances censorship resistance through optional decentralized data broadcasting capabilities inherent in the Racine Protocol. This integration not only strengthens the system’s Sybil resistance but also aligns with the broader vision of creating a human-centric, privacy-preserving digital trust framework. 2.2 Model Specification 2.2.1 Notation. The key variables of the system are described below. •𝑅𝑖: Reputation balance of entity 𝑖. •𝐸𝑖: Endorsement budget per period. •𝑉𝑖,𝑗 : Vote by 𝑖on 𝑗(positive or negative integer). •𝑆𝑖,𝑗 : Stake for any vote 𝑉𝑖,𝑗 , dynamically updated. •𝑇𝑖,𝑗 : Cumulative votes by 𝑖on 𝑗(𝑇𝑖,𝑗 =Í𝑉𝑖,𝑗 ). •𝑊𝑖,𝑗 : Weighted vote for a single action. 2.2.2 Non-Transferability of Reputation. To prevent plutocracy, where wealthy actors buy influence, the RepStake implements the ERC20 interface for compatibility but overrides the transfer RepStake SAC ’26, March 23–27, 2026, Thessaloniki, Greece and transferFrom functions. Reputation tokens are strictly nontransferable (SoulBound). They can only be minted by the protocol (upon earning reputation) or burned by the protocol (upon slashing or decay). This ensures reputation is a meritocratic signal of history, not a financial asset. Furthermore, this property is critical for mitigating 𝑃+𝜖 attacks [ 8 ], where attackers bribe voters. Since reputation is illiquid and essential for future influence, a rational actor will not trade permanent reputation loss for a one-time financial bribe, unlike in systems where the stake is fungible financial capital. 2.2.3 Administrator-Configurable Parameters. The following parameters are set by the system administrator to tune the reputation system’s behavior. Each parameter can be adjusted to balance collusion resistance, fairness, incentive strength, and system responsiveness. • Vote Weight Decay Rate ( 𝑑 ): Controls how quickly the influence of repeated votes from the same entity on the same target diminishes. A higher 𝑑 increases collusion resistance by reducing the impact of repeated or coordinated endorsements makes it harder for groups to manipulate reputation. Conversely, a lower threshold 𝑑 allows more influence from frequent voters, which may be suitable for highly trusted or smaller communities. • System Epoch ( 𝜏 ): Defines the operational time cycle of the system. This parameter sets the frequency for endorsement budget resets, the maximum lock-up duration for staked tokens, and the interval for applying reputation decay. • Endorsement Budget Scaling Factor ( 𝜆 ) and Whale Resistance: The endorsement budget for each entity is defined as: 𝐸𝑖=𝜆·√︁𝑅𝑖(1) This square root scaling ensures that as an entity’s reputation 𝑅𝑖 increases, their endorsement budget grows more slowly than linearly. This mechanism is inspired by quadratic voting [ 14 ] and is a well-established approach to limiting the disproportionate influence of “whales” (entities with very high reputation). By curbing the endorsement power of high-reputation entities, the system prevents any single participant or small group from dominating collective outcomes, promoting fairness and broad participation. • Reward Multiplier ( 𝑘 ): Sets the maximum reward an entity can earn for a vote that aligns with community consensus, typically capping the reward at 𝑘> 1times the original stake. A higher 𝑘 increases incentives for consensus-aligned voting, on the other hand, may also risk reputation volatility. Lower 𝑘 makes rewards more modest, reducing risk but potentially weakening incentives. • Decay Factor ( 𝛾 ): Governs the rate at which all entities’ reputations decay over time, via 𝑅𝑖=𝑅𝑖· ( 1 −𝛾) each period 𝜏 . A higher 𝛾 means reputations decay quickly, ensuring only recent activity is reflected and discouraging entities from relying on past achievements. Lower 𝛾 allows reputations to persist, rewarding long-term contributions but possibly allowing outdated reputations to linger. • Minimum Threshold for Reward/Penalty Distribution ( 𝑞 ): Sets the minimum value for which rewards or penalties are processed. If the calculated adjustment is less than 𝑞 , it is ignored. Higher 𝑞 reduces blockchain transaction overhead and prevents micro-adjustments, while lower 𝑞 ensures that even minor actions are recognized, at the cost of potential computational inefficiency. 2.2.4 Endorsement Budget. The endorsement budget (eq. (1)) limits voting power to prevent dominance by high-reputation players with square root scaling (e.g., for 𝜆= 2and 𝑅𝑖= 100 =⇒𝐸𝑖= 20). The budget resets every period 𝜏, and votes are constrained by: ∑︁ 𝑗∈ J 𝑖 |𝑉𝑖,𝑗 | ≤ 𝐸𝑖(2) where J 𝑖 represents the set of all target entities that entity 𝑖 has rated in the current period. 2.2.5 Vote Weighting. To ensure that repeated votes from the same entity on the same target have diminishing influence, the vote weight coefficients are defined as a linearly decaying function of the administrator -set parameter 𝑑: Weight Decay Function. 𝑤𝑚=1−𝑑·𝑚, 𝑚 =0,1,2, . . . ,𝑛 −1(3) where 0 <𝑑< 1controls the rate of decay. For example, with 𝑑= 0 . 2, the weights for the first five votes are 𝑤0= 1 . 0, 𝑤1= 0 . 8, 𝑤2= 0 . 6, 𝑤3= 0 . 4, 𝑤4= 0 . 2. Beyond the point where 𝑤𝑘≤ 0, additional votes have no effect. Note that, although this work primarily utilizes a linear function, other vote weight decay methods, such as exponential decay, were also examined. Cumulative Weight Sum. The cumulative weighted sum for 𝑛 votes is given by the following function: 𝜓(𝑛)= min(𝑛−1,𝑁max ) ∑︁ 𝑚=0 𝑤𝑚= min(𝑛−1,𝑁max ) ∑︁ 𝑚=0 (1−𝑑·𝑚)(4) where 𝑁max =1 𝑑 is the maximum number of nonzero-weighted votes. Initial Weight. Let 𝑇𝑖,𝑗 be the current vote count. The initial weight is: 𝑊initial =(𝜓(𝑇𝑖,𝑗 )if 𝑇𝑖,𝑗 ≥0, −𝜓(−𝑇𝑖,𝑗 )if 𝑇𝑖,𝑗 <0.(5) Final Vote Count. After adding the new votes 𝑉𝑖,𝑗 , the updated vote count is: 𝑇′ 𝑖,𝑗 =𝑇𝑖,𝑗 +𝑉𝑖,𝑗 .(6) Final Weight. The final weight is computed similarly: 𝑊final =(𝜓(𝑇′ 𝑖,𝑗 )if 𝑇′ 𝑖,𝑗 ≥0, −𝜓(−𝑇′ 𝑖,𝑗 )if 𝑇′ 𝑖,𝑗 <0.(7) Final Vote Weight. The final vote weight change is then: 𝑊𝑖,𝑗 =𝑊final −𝑊initial.(8) 2.2.6 Reputation and Stake Dynamics. When an entity 𝑖 casts a vote 𝑉𝑖,𝑗 on target 𝑗 , they stake 𝑆𝑖=|𝑉𝑖,𝑗 | reputation tokens, which are escrowed by reducing their available reputation: 𝑅𝑖←𝑅𝑖− |𝑉𝑖,𝑗 | The target’s reputation is immediately updated by the weighted vote: 𝑅𝑗←max(0, 𝑅𝑗+𝑊𝑖,𝑗 )(9) SAC ’26, March 23–27, 2026, Thessaloniki, Greece D. Mantzonis and T. G. Papaioannou where 𝑊𝑖,𝑗 is the signed, decayed weight as defined in the Vote Weighting section. Collective Stake Adjustment via Consensus. As subsequent entities cast votes 𝑉𝑚,𝑗 on the same target 𝑗 , the stakes of all previous voters are collectively adjusted in accordance with the emerging consensus. Specifically, all prior voters whose vote direction aligns with 𝑉𝑚,𝑗 receive a share of a reward equal to 𝑘· |𝑊𝑚,𝑗 | , distributed proportionally to their original weighted contributions. Conversely, all previous voters whose votes oppose 𝑉𝑚,𝑗 collectively incur a penalty. Formally, for each prior voter 𝑙: •If sign(𝑉𝑙,𝑗 )=sign(𝑉𝑚,𝑗 )(aligned), 𝑆𝑙,𝑗 ←𝑆𝑙,𝑗 + |𝑊𝑚,𝑗 | · 𝑘·|𝑊𝑙,𝑗 | Í𝑎∈𝐴|𝑊𝑎,𝑗 |(10) •If sign(𝑉𝑙,𝑗 )≠sign(𝑉𝑚,𝑗 )(opposing), 𝑆𝑙,𝑗 ←𝑆𝑙,𝑗 − |𝑊𝑚,𝑗 | · 𝑘·|𝑊𝑙,𝑗 | Í𝑜∈𝑂|𝑊𝑜,𝑗 |(11) where 𝐴 is the set of all aligned voters, and 𝑂 is the set of all opposing voters with respect to 𝑉𝑚,𝑗 . This ensures that every new vote redistributes value among all previous stakers, rewarding those who align with the emerging consensus and penalizing those who do not. This mechanism ensures that early voters who take the risk of assessing an entity, early before consensus is established, are disproportionately rewarded if their assessment is vindicated by later participants. This creates a “race to vote”, directly countering the prevalent issue of lazy voting [ 7 ], where token holders passively wait for a quorum to form and then simply copy the leading vote to farm governance rewards. Under this model, the first accurate voter stands to gain rewards from every subsequent accurate voter. This incentivizes prompt and honest participation. Staking Finalization. The staking process for each voter 𝑖 is resolved when one of the following occurs: • Reward: If 𝑆𝑖,𝑗 ≥𝑘· |𝑉𝑖,𝑗 | , the voter receives their original stake plus a reward: 𝑅𝑖←𝑅𝑖+𝑘· |𝑉𝑖,𝑗 |(12) •Loss: If 𝑆𝑖,𝑗 ≤0, the entire staked amount is burned. • Timeout: To prevent indefinite locking of reputation tokens and ensure system liquidity, RepStake enforces a strict timeout mechanism: all stakes are automatically finalized after a duration 𝜏 , regardless of the current state of community consensus. At the timeout, any remaining positive stake is returned to the voter: 𝑅𝑖=𝑅𝑖+max(0, 𝑆𝑖)(13) This timeout mechanism serves two essential purposes. First, it prevents indefinite lockup of stakes by ensuring that participants regain access to their reputation within a fixed period. Second, for the calculation of rewards and penalties, only stakes and voting activity within the most recent window of length 𝜏 are considered. By excluding older data, this rolling window approach limits computational complexity and ensures that only active participants influence ongoing reward and penalty distributions, thereby maintaining both efficiency and fairness as the system scales. Reward/Penalty threshold. For scalability, rewards or penalties below a minimum threshold 𝑞 are ignored and not processed by the system. This prevents micro-transactions and trivial updates from overwhelming the blockchain computation while maintaining meaningful incentive structures. Reputation Decay. To ensure that reputation scores reflect recent behavior and to prevent the indefinite accumulation of reputation, the system applies a periodic decay to all reputation balances. At the end of each period 𝜏, every entity’s reputation is updated as: 𝑅𝑖←𝑅𝑖· (1−𝛾)(14) This linear aggregate decay mechanism, inspired by dynamic weighting strategies for reputation calculation [ 22 ], ensures that outdated reputations gradually lose influence, addressing the problem of stale or inflated scores and incentivizing ongoing, positive contributions. 2.3 Properties The RepStake system exhibits several key properties that ensure its effectiveness as a decentralized reputation framework. Sybil resistance is achieved through the adoption of ZKorum’s Racine Protocol for entity authentication, which uses zero-knowledge proof-verified Decentralized Identifiers (DIDs) to prevent the creation of multiple identities by a single entity, maintaining the integrity of the system while preserving privacy. Anti-whale mechanisms are implemented through quadratic endorsement budgeting, which curbs dominance by high-reputation players by scaling their voting power non-linearly. Dynamic incentives are incorporated via a staking mechanism that adjusts based on community alignment, with rewards capped at a multiplier of the initial vote to encourage fair participation. Lastly, simplicity is ensured by designing ERC20compatible operations, facilitating seamless integration with existing blockchain infrastructures. These properties collectively make the model a robust and accountable solution for trust evaluation in decentralized systems, as elaborated in subsequent sections. 3 Evaluation To rigorously test the RepStake framework, we adopted an AgentBased Model (ABM) [ 16 ]. This choice was deliberate, as the core mechanics of our system—involving heterogeneous strategies, nonlinear dynamics, and history-dependent interactions make it challenging for traditional analytical models to capture system dynamics. An ABM is uniquely capable of simulating the complex, emergent behavior of a decentralized trust ecosystem populated by autonomous actors with competing interests. The RepStake design includes several features that violate the assumptions of simpler models. Furthermore, the system’s resilience hinges on the interplay between two distinct populations: honest players, who vote to correct discrepancies between reputation and true quality, and colluding players, who execute strategies to artificially boost their group’s reputation. Modeling these divergent, strategic behaviors within a single analytical framework would be intractable. An ABM, however, is perfectly suited to capture these dynamics. It allows us to observe if a globally trusted and accurate reputation system can emerge from a set of simple, local interaction rules. We RepStake SAC ’26, March 23–27, 2026, Thessaloniki, Greece can directly implement and test the efficacy of nonlinear mechanisms like the square root endorsement budget, which is critical for preventing high-reputation “whales” from dominating the system and ensuring fairness. Most importantly, an ABM allows us to simulate the collective consensus process of the staking mechanism, where the reward or penalty for a past vote is determined by the unpredictable sequence of future community votes. This dynamic social computation is the heart of RepStake’s incentive structure and can only be meaningfully studied through simulation. 3.1 Experimental Setup We model a population of 1000 digital entities, referred to as players, divided into honest and colluding ones, to analyze how reputation evolves over time under different collusion scenarios. Each player has a true ability or quality expressed in range [ 0 , 100 ] , with 100 denoting the highest capacity or expertise. To model the diversity of ability and reputation as observed in real-world systems, we initialize honest and colluding players according to a Pareto-like distribution [1], as follows: • The true quality of the 20% of each group (honest and colluders) is assumed to be very high and follow the uniform distribution in [ 90 , 100 ] , while their initial reputation is equal to their true quality plus some noise that follows the normal distribution with 𝜇=0, 𝜎=5. • The remaining 80% of each group (honest and colluders) are assigned true quality values from a normal distribution ( 𝜇 =50, 𝜎 =15), truncated between 10 and 90, while their initial reputation equals their true quality plus some noise derived from a normal distribution with 𝜇=0, 𝜎=10. This hybrid approach ensures that each population contains a small elite with high ability and reputation, while the majority represents a broader, more average spectrum. Such a distribution reflects the “80/20 rule” (Pareto principle) commonly observed in social and economic systems. Note that we also experimentally assessed the case where all players are assigned an initial low reputation with similar results. Honest players vote to correct discrepancies between a target’s true quality and reputation. Colluders coordinate to maximize their group’s reputation via ballot stuffing (voting positively for colluding members with reputations below the group average) and damage competitors via badmouthing (i.e., vote negatively) non-colluders. The simulation results presented in the following section were obtained by empirically using the following system parameters: Vote Weight Decay Rate 𝑑= 0 . 2, System Epoch 𝜏= 30 days, Reward Multiplier 𝑘= 2, Endorsement Budget Scaling Factor 𝜆= 2, Decay Factor 𝛾=0.05, and Minimum Threshold 𝑞=0.05. 3.2 Results Initially, in order to show the convergence of reputation values of the different players to their appropriate values, we plot reputation with respect to true quality for all honest and colluding players, as shown in Fig. 1. As depicted therein, the Pearson correlation between reputation and true quality for all players is close to 1, meaning that the reputation value of each entity converges to its true quality. This is true for either 20% (Fig. 1a) or 49% (Fig. 1b) of colluders present in the system. Moreover, the distribution of reputation values follows the quality distribution of our experimental setup, despite collusive raters, as shown in Fig. 2 irrespectively of the fraction of colluders present in the system. Clearly, as long as a majority of honest players exist in the system, our mechanism reveals the true quality of players. Also, the colluding players fail to maliciously boost their reputation values, while the distribution of reputation values is fair. Then, we show the evolution of the reputation values of all players in Fig. 3. As depicted therein, the average reputation of honest and colluding players converge to their final values quite fast for the different fractions of colluders in the system. This means that our reputation mechanism can effectively cope with dynamic contexts with high churn rates, i.e., when players either change their quality/behavior over time or leave the system and new players enter frequently. Also, the average reputation of colluding players converges to a lower value than their true average quality, meaning that they suffer some reputation penalty due to their malicious rating behavior. Finally, to rigorously assess the effectiveness of RepStake against related work, we repeat our simulation experiments for the Intrinsic Integrity-driven Rating Model (IIRM) [ 21 ], which we consider as a baseline reputation system that is using dynamic stake mechanisms as well. As shown in Fig. 4, in IIRM the reputation values do not follow the true quality of low-quality colluders: the latter are able to successfully boost their reputation values. In contrast, in RepStake, the combination of square root endorsement budgets and dynamic staking mechanisms significantly limits the ability of colluders to manipulate the system. Even at high collusion rates, the reputation gap between colluders and honest entities remains small, and the correlation between colluder reputation and true quality is near to one. This indicates that colluders cannot systematically boost their reputation beyond what is justified by their actual quality. 3.3 Key Findings 3.3.1 Resistance to Ballot Stuffing and Bad-Mouthing. Square root voting constraints and dynamic staking mechanisms curb collusion. The system effectively neutralizes ballot stuffing attempts, as colluders gain no reputation advantage, and mitigates badmouthing, ensuring honest entities maintain accurate scores despite coordinated negative attacks and even get punished for their voting behavior. 3.3.2 “Whale” Resistance. RepStake’s square root endorsement budget ensures that the voting power of high-reputation entities (whales) increases sub-linearly with reputation. This mechanism prevents any single participant or small group from dominating the outcome, promoting fairness and broad participation. As a result, even entities with very large reputation balances are unable to exert disproportionate influence over reputation outcomes. 3.3.3 Fairness and Accuracy. High correlations between reputation and true quality for honest players indicate accurate trustworthiness reflection. Colluders’ reputations correlate less with true quality due to getting punished by the honest majority, mitigating manipulation. SAC ’26, March 23–27, 2026, Thessaloniki, Greece D. Mantzonis and T. G. Papaioannou (a) (b) Figure 1: RepStake: Reputation vs. True Quality for (a) 20% and (b) 49% of colluders. (a) (b) Figure 2: RepStake: Distribution of Final Reputation by Player Type for (a) 20% and (b) 49% of colluders. (a) (b) Figure 3: RepStake: Reputation evolution for (a) 20% and (b) 49% of colluders. (a) (b) Figure 4: IIRM: Reputation vs. True Quality for (a) 20% and (b) 49% of colluders. 3.4 Sensitivity Analysis of Parameters We systematically examine the impact of key system parameters on reputation dynamics and collusion resistance, using the same simulation setup as in previous experiments. For each parameter, we vary its value across a representative range while keeping others constant. 3.4.1 Vote Weight Decay Rate 𝑑 .At low values, repeated votes from the same entity retain significant weight, which enables colluding groups to gradually amplify each other’s reputations and manipulate the system, delaying convergence to the steady state. In contrast, higher values of 𝑑 reduce the risk of manipulation and allow the system to converge more quickly but also introduce greater short-term reputation fluctuations. As shown is the Fig. 5a increasing the vote weight decay rate 𝑑 increases the reputation gap between honest entities and colluders, while causing colluders RepStake SAC ’26, March 23–27, 2026, Thessaloniki, Greece (a) (b) Figure 5: Impact of parameters (a) vote weight decay rate 𝑑 at mean reputation before convergence and (b) reward multiplier 𝑘 at the convergence speed. to gain less reputation before convergence without negatively affecting honest entities’ ability to reveal the system’s true quality. Repetitive simulation experiments with different population sizes and mixes have shown that the choice of 𝑑=0.4 is optimal. 3.4.2 System Epoch ( 𝜏 ). Higher 𝜏 extends the time required for reputation values to converge to the players’ true quality. This occurs because voters exhaust their endorsement budgets and are obliged to wait longer for the budget reset before they can vote again. 3.4.3 Reward Multiplier 𝑘 .Higher 𝑘 increases rewards and penalties, incentivizing honest participation, but excessive values can lead to reputation volatility and slower convergence speed, as illustrated in Fig. 5b. We repeated the simulation experiments for different population sizes and mixes and found that the choice of 𝑘=3 is the optimal choice. 3.4.4 Endorsement Budget Scaling Factor 𝜆 .Lower values of 𝜆 effectively limit the influence of high-reputation entities, while higher values increase the risk of disproportionate influence. 3.4.5 Reputation Decay Factor 𝛾 .Higher 𝛾 promotes ongoing engagement, ensures that reputations reflect recent activity, and makes the system deflationary. However, it can lead to reputation volatility and discourage long-term engagement. 3.4.6 Minimum Threshold for Reward/Penalty Distribution 𝑞 .While a lower 𝑞 captures more granular reputation adjustments, it does so at the cost of higher transaction volume, confirming that 𝑞 must be carefully tuned to balance incentive precision with blockchain scalability. Simulations confirmed a direct inverse relationship between the threshold 𝑞 and the system’s computational load. As 𝑞 was lowered, the number of messages produced by the system increased significantly. This validates the scalability analysis in Section 5.1, which identifies 𝑞as the primary lever for controlling messaging overhead. 4 Real-World Use Cases The RepStake reputation system has a wide range of practical applications across various domains; specific examples follow. 4.1 E-commerce Implementation Framework In the e-commerce context, RepStake provides a decentralized framework for reputation management that mitigates manipulation and ensures fair influence distribution. Entities are on-boarded through privacy-preserving identity verification and receive initial reputation allocations. Transactional feedback is implemented as a staking process: buyers rate sellers by staking reputation tokens, which are dynamically adjusted as additional ratings accumulate. This mechanism rewards entities whose assessments align with community consensus and penalizes those whose ratings diverge. The result is a robust reputation signal that reflects genuine market behavior and resists common attacks such as Sybil or collusionbased manipulation. 4.2 Federated Cloud Service Provision For federated cloud service provision, RepStake enables transparent, cross-platform trust evaluation among distributed providers and users. Upon registration, users receive reputation tokens based on verified credentials and usage history. After consuming cloud services, users evaluate providers by staking reputation, with the same dynamic adjustment and consensus-driven mechanisms as in the e-commerce scenario. This approach ensures that only consistently accurate evaluators are rewarded, while reputation remains portable and meaningful across federated networks. 4.3 Decentralized Autonomous Organization Peer Assessment Within DAOs, RepStake supports merit-based peer assessment by allocating reputation tokens for verified contributions and enabling members to rate each other’s expertise and output. These ratings are staked and subject to dynamic adjustment according to community consensus, aligning governance influence with demonstrated value rather than simple token holdings. This model promotes accountability, discourages manipulation, and ensures that reputation within the DAO accurately reflects meaningful participation and expertise. 5 Discussion Section While the evaluation demonstrates the robustness of RepStake’s core mechanics, a practical implementation must also consider operational aspects like scalability and transaction costs. SAC ’26, March 23–27, 2026, Thessaloniki, Greece D. Mantzonis and T. G. Papaioannou 5.1 Scalability A key concern in any blockchain-based system is its ability to scale as the number of interactions grows. In RepStake, a single vote can trigger stake adjustments for all previous voters on the same target within the active window 𝜏 . This could potentially lead to a high volume of on-chain transactions. However, the system’s design provides direct control over this messaging overhead. The distribution of rewards and penalties is governed by two administratorconfigurable parameters: the Reward Multiplier 𝑘 and the Minimum Threshold for Reward/Penalty Distribution 𝑞 . For any new vote, the total reward or penalty pool is proportional to 𝑘 , while the smallest adjustment processed is 𝑞 . In the worst-case scenario, where a reward is fragmented into the smallest possible chunks, the maximum number of stake-update messages that can be generated per vote is bounded by 𝑘/𝑞 for both aligned and opposing voters. By setting a reasonable ratio of 𝑘 to 𝑞 , an administrator can effectively cap the computational load and transaction volume generated by any single vote, ensuring the system remains scalable and efficient even with a high number of participants. 5.2 Gas Costs Executing operations on a blockchain typically incurs transaction fees, known as “gas costs”, which could become a barrier to adoption for a high-frequency system like RepStake. This challenge can be effectively mitigated through strategic implementation. The system can be deployed on an economical chain, such as a proprietary gas-less network where the operator covers costs, or a Layer-2 scaling solution that offers significantly reduced fees. Alternatively, the business model of the platform integrating RepStake can be designed to sponsor transactions, absorbing the gas costs as an operational expense. This approach creates a seamless user experience, allowing participants to interact with the reputation system without needing to manage cryptocurrency or pay for individual transactions. 6 Conclusion and Future Work This paper introduced RepStake, a blockchain-based reputation system addressing trust and manipulation in digital environments. By integrating identity verification via zero-knowledge proofs, square root voting constraints, and dynamic staking, RepStake ensures reputation is earned through verifiable, community-aligned actions. Simulation results demonstrate that the system effectively counters collusion and outperforms existing models like IIRM in resisting collusion [ 21 ]. The architecture supports diverse applications, from e-commerce to DAOs, while upholding user privacy. Future work will prioritize the study of a rigorous theoretical framework to prove the system’s equilibrium properties and the development of analytical methods for optimal parameter selection. Additionally, we plan to investigate reputation transferability across chains, enhance privacy guarantees, and explore additional monetary staking mechanisms to further strengthen system resilience. Acknowledgments This work has been funded by the EU project NGI TrustChain (grant no. 101093274). References [1] Barry C. Arnold. 2015. Pareto Distribution. John Wiley and Sons, Ltd, New York, 1–10. [2] Alex Biryukov and Daniel Feher. 2019. ReCon: Sybil-resistant consensus from reputation. Pervasive and Mobile Computing 61 (12 2019), 101109. [3] Christof Bless, Alexander Denzler, Oliver Karras, and Sören Auer. 2024. A Reputation System for Scientific Contributions Based on a Token Economy. In Linking Theory and Practice of Digital Libraries, Apostolos Antonacopoulos, Annika Hinze, Benjamin Piwowarski, Mickaël Coustaty, Giorgio Maria Di Nunzio, Francesco Gelati, and Nicholas Vanderschantz (Eds.). Springer Nature Switzerland, Cham, 41–61. [4] BrightID. 2022. Universal Proof of Uniqueness. White paper 1, 1 (2022), 1–27. https://www.brightid.org/whitepaper [5] Richard Dennis and Gareth Owen. 2015. Rep on the block: A next generation reputation system based on the blockchain. In 2015 10th International Conference for Internet Technology and Secured Transactions. IEEE, ICITST, London, UK, 131–138. [6] Tassos Dimitriou. 2021. Decentralized Reputation. In Proceedings of the Eleventh ACM Conference on Data and Application Security and Privacy (Virtual Event, USA) (CODASPY ’21). Association for Computing Machinery, New York, NY, USA, 119–130. doi:10.1145/3422337.3447839 [7] Edith Elkind, Evangelos Markakis, Svetlana Obraztsova, and Piotr Skowron. 2015. Equilibria of Plurality Voting: Lazy and Truth-Biased Voters. In Algorithmic Game Theory, Martin Hoefer (Ed.). Springer Berlin Heidelberg, Berlin, Heidelberg, 110–122. [8] William George and Clément Lesaege. 2021. An Analysis of p+ 𝜖 Attacks on Various Models of Schelling Game Based Systems. Cryptoeconomic Systems 1, 2 (10 2021). [9] Nicolas Gimenez and Yuting Jiang. 2025. Racine Protocol. Retrieved December 12, 2025 from https://github.com/zkorum/racine [10] Gitcoin. 2025. What is Human Passport? Retrieved December 12, 2025 from https://support.passport.xyz/passport-knowledge-base [11] Audun Jøsang, Roslan Ismail, and Colin Boyd. 2007. A survey of trust and reputation systems for online service provision. Decision Support Systems 43, 2 (2007), 618–644. Emerging Issues in Collaborative Commerce. [12] Nesrine Kaaniche and Maryline Laurent. 2017. A blockchain-based data usage auditing architecture with enhanced privacy and availability. In IEEE 16th International Symposium on Network Computing and Applications (NCA). IEEE, Cambridge, United States, 1–5. [13] Sepandar D. Kamvar, Mario T. Schlosser, and Hector Garcia-Molina. 2003. The Eigentrust algorithm for reputation management in P2P networks. In Proceedings of the 12th International Conference on World Wide Web (Budapest, Hungary) (WWW ’03). Association for Computing Machinery, New York, NY, USA, 640–651. doi:10.1145/775152.775242 [14] Steven P. Lalley and E. Glen Weyl. 2018. Quadratic Voting: How Mechanism Design Can Radicalize Democracy. AEA Papers and Proceedings 108 (May 2018), 33–37. doi:10.1257/pandp.20181002 [15] Mario Larangeira. 2022. Reputation at stake! A trust layer over decentralized ledger for multiparty computation and reputation-fair lottery. In International Conference on Information Security and Cryptology. Springer, Springer Nature Switzerland, Cham, 195–215. [16] Charles M. Macal and Michael J. North. 2009. Agent-based modeling and simulation. In Proceedings of the 2009 Winter Simulation Conference (WSC). IEEE, Austin, TX, USA, 86–98. [17] Massimiliano Pirani, Alessio Cacopardo, Alessandro Cucchiarelli, and Luca Spalazzi. 2023. A Soulbound Token-based Reputation System in Sustainable Supply Chains.. In EWSN. Junction Publishing, Rende, Italy, 363–368. [18] Paul Resnick, Ko Kuwabara, Richard Zeckhauser, and Eric Friedman. 2000. Reputation systems. Commun. ACM 43, 12 (2000), 45–48. [19] Luis Silva, Carlos Senna, and André Zúquete. 2019. Using Reputation as a Coin to Bet on Information Items Distributed in a Smart City. In 2019 Fifth Conference on Mobile and Secure Services (MobiSecServ). IEEE, Miami Beach, FL, USA, 1–2. doi:10.1109/MOBISECSERV.2019.8686570 [20] EigenLayer Team. 2024. Eigenlayer: The restaking collective. White paper 1, 1 (2024), 1–19. https://docs.eigencloud.xyz/assets/files/EigenLayer_WhitePaper88c47923ca0319870c611decd6e562ad.pdf [21] Haibao Wen, Shengmin Sun, Tao Huang, and Dunhui Xiao. 2024. An intrinsic integrity-driven rating model for a sustainable reputation system. Expert Systems with Applications 249 (2024), 123804. [22] Yan Wu, ChunGang Yan, ZhiJun Ding, GuanJun Liu, PengWei Wang, ChangJun Jiang, and MengChu Zhou. 2013. A Novel Method for Calculating Service Reputation. IEEE Transactions on Automation Science and Engineering 10, 3 (2013), 634–642.