The Governance of Decentralized Autonomous Organizations: A Study of Contributors’ Influence, Networks, and Shifts in Voting Power Stefan Kitzler1,2, Stefano Balietti3, Pietro Saggese2,1, Bernhard Haslhofer1 and Markus Strohmaier3,4,1 1 Complexity Science Hub Vienna 2 AIT Austrian Institute of Technology 3 University of Mannheim 4 GESIS - Leibniz Institute for the Social Sciences Abstract. We present a study analyzing the voting behavior of con- tributors, or vested users, in Decentralized Autonomous Organizations (DAOs). We evaluate their involvement in decision-making processes, discovering that in at least 7.54% of all DAOs, contributors, on aver- age, held the necessary majority to control governance decisions. Fur- thermore, contributors have singularly decided at least one proposal in 20.41% of DAOs. Notably, contributors tend to be centrally positioned withintheDAOgovernanceecosystem,suggestingthepresenceofinner power circles. Additionally, we observed a tendency for shifts in gov- ernance token ownership shortly before governance polls take place in 1202 (14.81%) of 8116 evaluated proposals. Our findings highlight the centralroleofcontributorsacrossaspectrumofDAOs,includingDecen- tralized Finance protocols. Our research also offers important empirical insights pertinent to ongoing regulatory activities aimed at increasing transparency to DAO governance frameworks. Keywords: DAO · Governance · Ethereum · Networks · Blockchain · Voting 1 Introduction DAOsrepresentorganizationalstructuresdesignedtoofferanalternative,decen- tralized form of governance for decentralized applications (dApps) operating on DistributedLedgerTechnologies(DLTs).TheintentionofDAOsistocircumvent central authorities and hierarchical structures that are prevalent in traditional organizations,anddemocratizethedecision-makingprocessbydistributingvot- ing rights through so-called governance tokens to community members [9]. Anecdotalevidencesuggeststhattheseintentionsarenotalwaysmetinprac- tice. For instance, there are signs of a centralized power circle that has emerged within the Decentralized Exchange (DEX) service Sushiswap [29]. Similarly, the governanceofArbitrumDAOproposedchannelingtokensvaluedat1billionUS dollars into their own treasury [38]. The lending protocol Solend confiscated the 3202 peS 82 ]IS.sc[ 2v23241.9032:viXra 2 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier funds of a prominent user who posed a risk to its financial stability [10]. In an- other instance, major cryptoasset exchanges, significant entities in this context, reportedly colluded and leveraged investors’ tokens to vote on the Steem plat- form [15,27]. Attempts at bribery have been noted among community members in governance forums [40]. Lastly, developers from the mixing service Tornado Casharereportedlyunderinvestigationforfinancialcrimes;it’sallegedtheyma- nipulated its governance to circumvent the introduction of rigorous anti-money laundering controls [18,13]. It is well known that governance tokens are distributed primarily to team members, early investors, or protocol treasuries [6], and decision-making power can be concentrated in the hands of a few [5]. Earlier research has provided preliminary evidence on the involvement of DAO team members and developers in DAOs decision-making processes [34,23,20]. However, there is a surprising gap in studies that systematically investigate the role of vested users in the governance of DAOs and how they determine their trajectories. In this study, we focus on DAO contributors, encompassing project owners, administrators, and developers. These contributors are involved in the technical realization of the dApp overseen by a DAO and thus can be viewed as vested users. Our aim is to empirically examine their influence in decision-making pro- cesses, the structure of their co-voting network, and any sudden shifts in ma- jorities just before voting takes place. Our contributions and findings can be summarized as follows: 1. Wecompiledadatasetcomprising986557votersacross872DAOswith7478 recognized contributions. Additionally, we cross-verified a subset of 438668 votes from 8116 proposals against their on-chain records, determining that 97.48% of these were consistent. 2. We introduce a metric to measure the involvement of contributors in DAO voting: in 66 (7.54%) DAOs contributors held, on average, the necessary majority to steer governance. We also measured contributor self-decisions, discovering that their votes were decisive in 178 (20.41%) DAOs. 3. We analyze the co-voting structures of users through a network approach. Our findings indicate that contributors are more likely to be found towards thecenteroftheDAOgovernanceecosystem.Furthermore,contributorsare highly concentrated in a few communities formed by co-voting patterns. 4. Weobservedmajorityshifts ingovernancetokenownershipin1202(14.81%) outof8116proposalsinthedaysprecedingthevotes.Thenumberofmajor- ity shifts increases sharply prior to governance polls, indicating last-minute token acquisitions. To the best of our knowledge, our study is the first to systematically inves- tigate the role of contributors in the governance of DAOs. It underscores their pivotalroleacrossvariousDAOs,includingleadingDecentralizedFinance(DeFi) protocols. Beyond shedding light on centralization tendencies within DAO gov- ernancestructures,ourfindingsdemonstratethatcontributorspossessthecapa- bility to effectively steer the direction of DAOs. These insights have significant The Governance of Decentralized Autonomous Organizations 3 implicationsregardingaccountability.Moreover,theyarerelevantinthecontext of current regulatory initiatives aimed at pinpointing the individuals who either control or exert notable influence over DeFi operations or structures [31]. We will release our dataset and the implementation of methodologies to en- sure the reproducibility of our findings. 2 Background, Definitions and Related Work 2.1 Voting in Decentralized Autonomous Organizations Decentralized Autonomous Organizations (DAOs) are a novel form of gover- nance model that has become popular in the crypto ecosystem since 2020. They can govern decentralized applications (dApps) and their associated smart con- tracts[45].SeveralDecentralizedFinance(DeFi)protocolsimplementDAOgov- ernance models [2], e.g., MakerDAO [26,36], Uniswap [1], Sushiswap [37], and Compound [25]. DAOs can also operate without an underlying dApp [12]. DAO voting mechanisms can be divided into two primary categories: on- chain andoff-chain voting.TheformeroccursdirectlyonaDLT,throughsmart contracts implementing the voting logic. To vote, token holders delegate an ad- dress that can be controlled by another entity. This approach offers security and transparency, but transaction costs make it economically inefficient [17,16]. The latter takes place on centralized platforms like Snapshot [33], and only the voting outcome is stored on the DLT. This method is more scalable, accessible and efficient, at the cost of higher centralization (e.g., concerns that the DAOs might not enforce the decisions, concurrent voting on different platforms, or non-tamper-proof databases). Our study focuses on the Snapshot platform, the largest off-chain governance platform with a market share of over 90% [43]. Decision-making in DAOs is executed through voting on so-called improve- ment proposals that can determine the evolution of the technical infrastruc- ture[41],modifyparametersaffectingtheeconomicincentivesanddesign[14,11], orreallocatefundsmanagedbyaDAO[39,42].Governanceuserscanparticipate in the voting by possessing specific tokens, known as governance tokens, which representtheirDAOmembershipandtheirproportionaldecision-makingpower. 2.2 Definitions WenowpresentaconceptualmodelofDAOvotingandintroducethekeytermi- nologyandnotationusedthroughoutthispaper,referringtoDAOsasspaces [33]. Figure1illustratestheentities:spaces,proposalsandusers;anditdescribestheir relations of contribution and vote. – Let U be the set of all users exercising voting rights and S be the set of all spaces. Users can also be denoted as voters in this context. – A contribution is a relation C ⊆ U ×S ×P(T), where (u,s,T) ∈ C, if a user u contributes to a space s in one or more role types T ⊆ T = {owner,administrator,developer}.Userscantakemultipleroles.Acontrib- utor is a user that has at least one contribution association to one space. 4 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier u 1 v 1 p 1 s v 2 1 p 2 v 3 c u 1 p s 2 3 2 Fig.1. Conceptualization of DAO voting. A proposal p introduces potential changes to a DAO space s, and users u can exert their decision-making power on them with their vote v ( ) and voting power w indicated by the arrow thickness. Governance users can be vested by their contribution c as owner, administrator or developer toaspace( ).WedenotetheirvestedvoteasVP whentheyarecontrib- SS utorsofthesame-space ( )theyarevotingon,andVP whentheyarecontributors OS of an other-space ( ). – AproposalisarelationP ⊆S×P(O)×P(F)×N+,where(s,O,F,h)∈P, ifthereisaproposedchangetoaspacesprovidingasetofchoicesoroptions O to vote on, and a set of strategies F to be applied for determining the outcomeofavoteatagivenblockheighth.Thesetsofoptionsandstrategies are defined as follows: • Op ⊆ P(O) denotes the set of possible options (choices) that can be selectedduringthevotingphaseontheimprovementproposalp.Inmost cases,thealternativesaresimplyayes/noanswer(i.e.,Op ={Yes,No}). • Fp ⊆ P(F) denotes the set of strategy functions that are applied to compute the voting power for the governance user issuing a vote. – A vote is a relation V ⊆U×P×O×R+, where (u,p,o,m)∈V, if a user u votes on a proposal p by selecting an option o ∈ Op, where Op denotes the set of options published as part of a specific proposal p. In rare cases, o can become a vector, e.g., the associated voting strategies allow one to express multiple choices. Then, the magnitude vector m characterizes the weighted (cid:80) preference of each option, and m = 1; if the vote expresses one single i i choice, m is a scalar equal to 1. We further denote as Vp ⊆V the set of all votes related to proposal p. We distinguish between two types of votes: 1. Ausercancontributeandvoteonanimprovementproposalofthesame space.We,therefore,denoteVP ⊆Vp asthesetof same-space votes, SS where, for all tuples (u ,p ,o ,m ) ∈ VP , a tuple (u ,s ,T ) ∈ C such i i i i SS j j j that u = u and s = s for psn exists, i.e., the users u equals u and i j n j i i j s of proposal p equals the space s . n i j 2. Ausercanalsocontributetoonespaceandvoteonanimprovementpro- posalforanotherspace.WedenoteVP ⊆VP asthesetof other-space OS votes, where for all tuples (u ,p ,o ,m )∈VP , a tuple (u ,s ,T )∈C i i i i OS j j j suchthatu =u exists,buttheredoesnotexistonewhereadditionally i j s =s is fulfilled for psn. Note that: VP ∩VP =∅. n j i SS OS The Governance of Decentralized Autonomous Organizations 5 3. Finally, we denote VP =VP ∪VP as the set of contributor votes. C SS OS – Thevoting poweristheweightwassignedtoanoptionoandcharacterizes the influence of a vote v. It is determined by the strategy function f : V × N+ → R+ of the vote v at block height h. For proposals with multiple functions, the weight is defined by their sum Fp(v,h):= (cid:80) f(v,h). f∈Fp – Finally, the options Op can be ranked by aggregated voting power w. We denote as the outcome the options Oˆp = [oˆp, oˆp, ...] ranked in descending 1 2 order by voting power, and denote oˆp as the decision, i.e., the option having 1 the highest accumulated voting power for the proposal p. 2.3 Related work Prior research has extensively documented that the ownership of governance tokens is highly concentrated [34,23,28,6,16], as a result of intentional design decisions and market dynamics (governance tokens carry a market price and can be traded). Furthermore, their total supply, the monetary policy, and the initial token allocation affects their distribution; finally, mechanisms such as airdrops [21] further favor early participants and DAO members. Studiesfocusingspecificallyonon-chainDAOvotingconfirmthatgovernance tokensarehighlyconcentrated.Furthermore,theyshowthatusersrarelyexercise voting rights [5], and that individuals who possess the potential power to alter outcomes rarely exercise it [20,17]. Two related works identify the existence of voters’ coalitions in MakerDAO [36,35]. A preliminary study reports examples of voters who held governance tokens for the duration of a single proposal life- cycle [16]. Ourworkiscloselyrelatedtostudiesonoff-chainvoting.Wangetal.[43]de- liveredacomprehensiveoverviewofthevotingplatformSnapshot.Laturnus[24] utilized data from Snapshot and DeepDAO to investigate the economic perfor- manceofDAOsinrelationtoownershipconcentrationandvotingparticipation. While earlier research has provided preliminary evidence on the involvement ofDAOteammembersanddevelopersinDAOsdecision-makingprocesses,none of these studies has systematically investigated the role of vested users in DAOs and their influence in determining their trajectories. This knowledge gap serves as the motivation for our study. 3 Data ToanalyzetheinvolvementofcontributorsinDAOdecision-makingthroughvot- ing, we gather data from the following sources: Snapshot, Ethereum blockchain, Ethereum Name Service (ENS) and The Graph. We combine them to identify contributions, as defined in Section 2.2. Then, we clean, verify, and validate our dataset, as summarized in Table 1. Additional details on the entire data preparation process and contribution identification are reported in Annex A. 6 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier Raw Cleaned Validated (Sections 4 & 5) (Section 6) Spaces S 12294 872 357 Voters U 1603994 986557 119413 Contributions C 11949 7478 3927 Proposals P 76851 35124 8116 Votes V 8365707 5240622 438668 Contributor votes V 316900 191507 22878 C Table 1. Dataset summary. The raw dataset combines Snapshot data on voters U with additional sources to identify the contributions C and quantify their voting activityV .Usersvote(V)onimprovementproposalsP toDAOspacesS.Tofocuson C DAOswithmaturegovernancestructures,wecleaned andvalidated thedatasetusing selected proposals with Ethereum on-chain data. Raw dataset. Weobtained1603994DAO voters andtheirwalletaddressesfrom theSnapshotdataset.Thesehavecast8365707votesfromNov-2020toDec-2022 on 76851 proposals, using 208 distinct voting strategies in 12294 DAO spaces. Next, we identify voters’ contributions to DAOs by joining their addresses with additional data. We extract their respective roles T by retrieving the domain owner address from ENS references, the administrators’ addresses from Snap- shot,andthecreators,ordevelopers,ofcodeaccounts(CA)fromtheblockchain transaction for all space-related CA from Snapshot. Cleaned dataset. We found that 42.67% of DAO spaces have one proposal only, 54.05% have less than five followers and 50.85% have at most two voters. We consider these as indicators for immature governance structures. Therefore, we apply a cleaning procedure by incorporating related benchmarks that assess minimumrequirementstoincludematureDAOsinthedataset.Wealsoremove non-final proposals and restrict to proposals using the single-choice voting. Validated dataset. Previous studies have shown inconsistencies between the re- ported and actual on-chain data, including instances of flawed data records within a Blockchain explorer [22], emphasizing the need for a validation frame- work. Therefore, we validate the consistency between the voting power values computed by Snapshot and the ground truth reflected in on-chain data. We focus on the Ethereum Blockchain, the most relevant one for Snap- shot[43]intermsofexpressedvotingpower,andonlyconsiderproposalsthatare almost entirely5 covered by strategies bound to F′ ⊆ P({ferc20,ferc721,feth}). With this approach, we could verify that 461402 (97.48%) of 473306 Ethereum Snapshot weights are correct. 5 The selected strategies cover more than 99% of the voting power in the proposals. The Governance of Decentralized Autonomous Organizations 7 100% 10% instadapp−gov.eth [19.2%] 1% 0.1% uniswap [29.9%] frax.eth [15.1%] 0.01% aave.eth [28.1%] lido−snapshot.eth [ 1.2%] 0.001% 0 250 500 750 DAOs ranked in decreasing order ]%[ tnemevlovni rotubirtnoC Fig.2. Contributor involvement across DAO spaces. The DAOs are ranked by contributor involvement w¯s ( ) from highest (left) to lowest (right). Some high-TVL C dApps( )areannotatedforillustrativepurposesandcontributorinvolvementofmore than 50% is colored ( ). 4 Influence of contributors on DAO governance 4.1 Contributor involvement Contributorsarevestedusers,havingintuitivelyhigherincentivestobeinvolved in the decision-making in DAOs. We analyze their involvement by measuring theirvotingpowerexercisedinproposals.AsdiscussedinSection2.2,thevoting powerw istheweightassignedtoanoptiono andcharacterizestheinfluenceof i i a vote v . In most cases, it is equal to the amount of governance tokens held by i the user. Recall that the weight w is the result of a proposal strategy function i Fp(v ,h) applied on a vote v at a specific block height h. i i Wecomputetheinvolvementofcontributorsinagivenspacebyaveragingthe share of voting power they have in proposals associated with that space. Since mintedamountsofgovernancetokens,theirprices,andtheirdistributionsacross user vary, we normalize voting power by total voting power across proposals as follows: w˜ =w ×( (cid:80) w )−1. Next, we consider the set of contributors Vp, thatinclud i essa i me-spa v c l∈ es Vp vot l ersVp aswellasother-space Vp .Then,wecom C - SS OS putethefractionofweightscontrolledbycontributorsw˜p (1)andfinallyobtain C the contributor involvement w¯s (2) as the average weights of contributors’ C votes for all proposals P in a DAO space s. w˜p = (cid:88) w˜, (1) w¯s = |P|−1 (cid:88) w˜p. (2) C i C C vi∈V C p p∈P To give an example, let’s assume proposal p has four votes {v ,v ,v ,v } 1 2 3 4 (cid:80) with normalized voting powers {.1,.4,.3,.2}, where w˜ = 1. Supposing that i the first two voters are contributors, ie Vp ={v ,v }, then w˜p =.1+.4=.5. C 1 2 C 8 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier We determine w¯s for all spaces on the cleaned data set, and show the re- C sults in Figure 2. The DAO spaces are ranked by contributor involvement in descendingorder.Thus,theinvolvementofcontributorsishighfortheDAOson the left-hand side and low for the DAOs on the right-hand side. For illustration purposes, we highlight the data points representing top DeFi protocols in terms of Total Value Locked (TVL), such as Aave, Uniswap or Instadapp. Our results show that the involvement of contributors in terms of average voting power is relatively low for most DAOs. The median value is 4.26% and the standard deviation is 21.22. However, for 297 spaces, the relative voting power of contributors is higher than 10% and for 66 DAOs it is higher than 50%. In these spaces, the contributors have, on average, a majority of voting powerandcandeterminesingle-handedlytheoutcomeofproposals.In9spaces, the contributors were the only voters with 100% voting power. 4.2 Contributor self-decisions Knowing that DAO contributors are involved in decision-making, we now in- vestigate to what extent they decide on proposals related to their own spaces. Thus, we concentrate our analysis on the votes cast by users that contributed on improvement proposals of the same spaces (Vp , see Section 2.2), which we SS herein denote as “self-votes”. Furthermore, we also consider the choices they made with their votes and their influence on the outcome of a proposal. Recall that the decision of a proposal poll is determined by the option with the highest voting power oˆp within the ranked outcome Oˆp = [oˆp, oˆp, ...]. We 1 1 2 denote the set of decisive self-votes as Vp, where the option o of Vp is the D D winning choice oˆp. 1 We are specifically interested in the self-votes where the decision-making was dominated by contributors, which we denote as contributor self-decisions. Intuitively, for a given space s, we determine the share of selected proposals based on two joint conditions. First, we consider the weight of contributor votes within a decision and select those proposals where contributors have a relative majority(≥50%).Second,weconsideralsothesecond-rankedoptionandselect those proposals where the weight of contributors in the decision is higher than the weight of the second ranked option. The underlying intuition is that in a head-to-head race between options, contributors might want to outweigh and overrule a leading option. More formally, we define the set of decisive self-votes Vp = Vp ∩Vp and D oˆ1 SS also the complement set Vp = Vp \Vp; we identify the fractions of relative CV oˆ1 D voting power for decisive self-votes (3), for the complement set (4) and for the second choice oˆp (5) as 2 (cid:88) (cid:88) (cid:88) w˜p = w˜, (3) w˜p = w˜, (4) w˜p = w˜. (5) D i CV i oˆ2 i vi∈V D p vi∈V C p V vi∈V oˆ p 2 Foragivenspaces,wecandefinethecontributor self-decisions δs asfollows: The Governance of Decentralized Autonomous Organizations 9 100.0% 10.0% instadapp−gov.eth [20.0%] 1.0% aave.eth [13.6%] 0.1% 0 50 100 150 DAOs ranked in descending order ]%[ snoisiced−fles rotubirtnoC Fig.3.Contributorself-decisionsacrossDAOspaces.The178DAOsareranked bycontributorself-decisionsδs( )indescendingorder,withathresholdof0.1%.They- axisrepresentsthefractionofproposalsinwhichDAOcontributorsvotedanddecided theiroutcomewithdominantvotingpower.Somehigh-TVLdApps( )areannotated for illustrative purposes. (cid:88) δs :=|P|−1 [(w˜p >w˜p )∧(w˜p >w˜p )] . (6) D CV D oˆ2 p∈P Figure 3 shows the results with DAOs ranked by self-decisions in descend- ing order. Note that we introduced thresholds and only show spaces with self- decisionsabove0.1%.Thisgivesus178(20.41%)differentspaceswherecontrib- utors of the same DAO decided on at least one proposal on their own. In total 2100 out of 35124 proposals were decided by governance users who contributed and voted on the same DAO. Annex B provides more details and analyses on involvement and self-decisions. 5 Co-voting networks In Section 4, we measured the involvement of contributors in decision-making for each space separately; however, users can contribute to multiple proposals and DAO spaces. Herein, we conduct a network-based analysis of users’ co- voting patterns across DAOs. We analyze topological features such as centrality measures and community structures that may indicate whether contributors occupy a central role in the DAO voting ecosystem. Networks construction. The basis for the investigation is the bipartite network G that links users U to the proposals P they voted on, having options O as PU edge features. We derive co-voting networks as a monopartite projection of the G onvoters,bycreatinganetworkofuserswithweightedlinksthatrepresent PU 10 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier Network G G G G AA AW TA TW Daos All All Top-100 Top-100 Votes All Winning All Winning Num Nodes 104863 75879 20401 14494 Num Edges 739813062 107374710 19917792 6045065 Avg. Degree 14110.09 2830.16 1952.63 834.15 Contr. Nodes 1.29% 1.45% 3.25% 4.5% Contr. Edges 1.61% 1.76% 3.4% 8.0% Table 2. Network statistics of four co-voting networks. The top of the table defines the four networks as a unique combination of two features: DAOs and Votes. Top-100 DAOs are ranked by total value locked (TVL); Winning votes are votes for the choice that ultimately won the majority of voting power. the number of proposals they voted together. We introduce a global threshold T on links to focus on users that systematically voted together on the same proposals and for computational reasons. Ultimately, we build four co-voting networks crossing DAOs and votes as shown at the top of Table 2, namely: – G istheentireco-votingnetwork,containingallvotes,regardlessofusers’ AA choices, and all spaces; – G istheco-votingnetworkofdecision-makers.Itonlytakesintoaccount AW votes v for the winning decision oˆp (we hypothesize that co-voting patterns i 1 maybeespeciallyrelevantamonguserswhovotedforthewinningoutcome); – G is the co-voting network of all votes of the top-100 DAOs by TVL. TA – G is the co-voting network of decision-makers in the top-100 DAOs by TW TVL, was constructed in the same fashion of G . AW Networkdescriptivestatistics. Weutilizethecleaneddatasetof5240622voting relations on 872 DAOs and 35124 proposals. We create the networks using the threshold T = 10 on the links among voters; despite this, given the inherent computational challenge on computing metrics on the large network G , we AA focus mainly on the remaining three networks. In all networks, we identify small-world features, i.e., they are characterized by the presence of several hubs conveying information rapidly across connected communities [44]. Interestingly, the share of contributor nodes and edges in- creases in the Top-100 networks (see bottom rows in Table 2), suggesting both that they are more active and that they tend to have a larger weight in shaping theoutcomesofproposalsofthemostimportantDAOs,ratherthantheperiph- eral ones. The rest of the section tries to confirm whether this intuition is true. Appendix C contains additional network statistics and analyses. The Governance of Decentralized Autonomous Organizations 11 Pagerank K-Core 3e+03 9e-05 2e+03 6e-05 Non-Contributor Contributor 1e+03 3e-05 0e+00 0e+00 All All Top-100Top-100 All All Top-100Top-100 All Win All Win All Win All Win Fig.4.Pagerankandk-corestatisticsinthefourco-votingnetworks.Contrib- utorstendtohavehigherpagerankandk-coreacrossnetworks.Allcentralitymeasures makeuseofedgeweightsandareappliedtothegiantcomponent;k-corestatisticsuse geometric mean to limit the effect of outliers. Error bars are 95% confidence intervals of the means. 5.1 Centrality of contributors To understand the influence of contributors on governance voting, we computed several network centrality measures for contributor and non-contributor nodes, namely pagerank, closeness, eigenvector, and betweenness centrality, and a k- core analysis. We present pagerank in Figure 4 as well as the k-core. Across all four networks, contributor nodes score higher in centrality in all measures but eigenvector for G . These differences are generally highly significant (t-test AW p<0.001), only the betweenness centrality shows more variability (p<0.05 for G , and p<0.1 for G ). TA TW We also computed the k-coreness of contributors. A high k-core indicates di- rect connections with other nodes with high k-core nodes, that is, nodes with at least degree k. Across all networks, contributors have, on average, significantly higher k-core (t-test p < 0.001). For these statistics we chose to use the geo- metricmeansbecausetheyarelesssensitivetooutliers.Infact,contributorsare generally less frequent in the portion of the distribution with the lowest k-core, however, there exists a few clusters of mainly non-contributors with very high k-core, which would skew the results. 5.2 Communities of contributors To understand the presence of hidden co-voting formations, we performed the Louvaincommunitydetectionmethodonthreeco-votingnetworks.Thismethod optimizes the modularity of the graph so that the connections within each com- munity are dense, while the connections across communities are sparse. As a result, each node in the graph is uniquely assigned to a community. We then tested if contributors can be found with equal probability in all communities or whether they are more likely to cluster together in a few of them. To answer 12 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier All DAOs ; Win Votes Top-100 DAOs ; All Votes Top-100 DAOs ; Win Votes 10000 5000 High concentration High concentration High concentration 0 Contr NonContr Contr NonContr Contr NonContr xedni namhcsriH-lhadnifreH Communities with at least one contributor 21 50 35 % % % Fig.5. Concentration of contributors across network communities. The bar plots show the Herfindahl-Hirschman concentration index for the distribution of con- tributors ( ) and non-contributors ( ) to communities assigned by the Louvain com- munitydetectionalgorithm.Theinsetdonutplotsshowtheshareofcommunitieswith atleastonecontributor;inallnetworks,contributorsareconcentratedinafewofthem. this question, we computed the Herfindahl-Hirschman index of market concen- tration on the distribution of contributors to communities. A higher value of this index indicates a more concentrated market, that is a distribution with fewer groups or communities dominating in size. Figure 5 indeed shows a very high concentration level for contributors in all networks with a peak above 7000 (below 1500 is considered well-mixed, between 1500 to 2500 is moderately con- centrated, and above 2500 is highly concentrated). Counting the communities with at least one contributor (the donut plot inside each panel of Figure. 5), contributors are to be found only in about 21-50% of all detected communities. A Pearson’s Chi-squared test indicates a significant deviation from chance in all networks (p < 0.001, with 100000 bootstrapped replicates). Figure 6 visually confirmsthisresultfortheG network:contributors(darkred)tendtocluster TA in a few central communities. 6 Pre-voting power shifts Governance tokens are cryptoassets and, consequently, can be purchased and sold. Furthermore, previous works provide preliminary evidence that users may hold their voting rights only for the duration of single proposals [16]. Therefore, we hypothesize that changes in the ownership distribution shortly before the voting power is determined could indicate attempts to acquire additional power to influence a proposal’s decision. We investigate to what extent users, and es- pecially contributors, acquire voting rights shortly before the poll execution. For proposals that rely on on-chain data, it is possible to access the current and historical token balances of voters. Thus, we determine their voting power atearlierpointsintimebyre-implementingtheproposalstrategiesonhistorical data and comparing it to their actual voting power. Note that we recompute it

12 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier All DAOs ; Win Votes Top-100 DAOs ; All Votes Top-100 DAOs ; Win Votes 10000 5000 High concentration High concentration High concentration 0 Contr NonContr Contr NonContr Contr NonContr xedni namhcsriH-lhadnifreH Communities with at least one contributor 21 50 35 % % % Fig.5. Concentration of contributors across network communities. The bar plots show the Herfindahl-Hirschman concentration index for the distribution of con- tributors ( ) and non-contributors ( ) to communities assigned by the Louvain com- munitydetectionalgorithm.Theinsetdonutplotsshowtheshareofcommunitieswith atleastonecontributor;inallnetworks,contributorsareconcentratedinafewofthem. this question, we computed the Herfindahl-Hirschman index of market concen- tration on the distribution of contributors to communities. A higher value of this index indicates a more concentrated market, that is a distribution with fewer groups or communities dominating in size. Figure 5 indeed shows a very high concentration level for contributors in all networks with a peak above 7000 (below 1500 is considered well-mixed, between 1500 to 2500 is moderately con- centrated, and above 2500 is highly concentrated). Counting the communities with at least one contributor (the donut plot inside each panel of Figure. 5), contributors are to be found only in about 21-50% of all detected communities. A Pearson’s Chi-squared test indicates a significant deviation from chance in all networks (p < 0.001, with 100000 bootstrapped replicates). Figure 6 visually confirmsthisresultfortheG network:contributors(darkred)tendtocluster TA in a few central communities. 6 Pre-voting power shifts Governance tokens are cryptoassets and, consequently, can be purchased and sold. Furthermore, previous works provide preliminary evidence that users may hold their voting rights only for the duration of single proposals [16]. Therefore, we hypothesize that changes in the ownership distribution shortly before the voting power is determined could indicate attempts to acquire additional power to influence a proposal’s decision. We investigate to what extent users, and es- pecially contributors, acquire voting rights shortly before the poll execution. For proposals that rely on on-chain data, it is possible to access the current and historical token balances of voters. Thus, we determine their voting power atearlierpointsintimebyre-implementingtheproposalstrategiesonhistorical data and comparing it to their actual voting power. Note that we recompute it The Governance of Decentralized Autonomous Organizations 13 Fig.6.The co-voting network of the Top-100 DAOs by TVL (winning votes only).Thecolorsidentifythelargestcommunitiesobtainedbyoptimizingmodularity, while smaller communities are in light gray; contributor nodes are colored in dark red ( ). Contributors are not uniformly distributed across communities, but they tend to cluster in a few of them in the middle of the graph, suggesting a higher network centrality. Network plotted using the OpenOrd layout algorithm in Gephi [7], after removing redundant edges following [30]. consideringthatthevoterv stillselectsthesameoptiono .Assumingthatusers’ i i holdings do not fluctuate with high frequency, we sample their token balances on a daily basis6. More formally, given a proposal p, for each voter u we denote i the actual voting power w (h ) as the voting power at the block h of the vote i τ τ execution, and Oˆp(h ) = [oˆp(h ), ...] as the actual ranked outcome. Next, for τ 1 τ each of the 100 days preceding the vote on the proposal p, we recompute the users’ historical voting power w (h ) and the resulting hypothetical ranked i τ−t outcome Oˆp(h ), where h is a block representative of the tth day before τ−t τ−t thepoll.WethuscompareOˆp(h )toOˆp(h )anddeterminewhetherthere τ−t τ−t−1 wasamajorityshift ifoˆp(h )̸=oˆp(h ).Finally,wemeasurethenumberof 1 τ−t 1 τ−t−1 majority shifts across proposals. Since we are correlating against on-chain data, we utilize the validated dataset described in Section 3, covering 8116 (23.11%) 6 Daysareapproximatedassumingthatblocksareminedonaverageevery15seconds, . i.e., 1d=(86400/15)blocks. 14 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier 60 40 20 h 0 0 −100 −75 −50 −25 0 Time delta before poll [days] tfihs ytirojam htiw slasoporP # Fig.7. Majority shifts occur in temporal proximity of polls. Majority shifts occur when the voters, shortly before a proposal, trade enough governance tokens to swing the final outcome of the poll. We focus on the validated dataset and identify majorityshiftsupto100daysbeforethevotes.Intemporalproximitytotheproposals, the number of shifts increases, indicating last-minute voting power acquisition. proposals. We therefore emphasize that the findings reported in this Section are a lower boundary estimation. In total, we found majority shifts for 1202 (14.81%) proposals in 229 DAOs in the 100 days before the poll. The median number of shifts per proposal is 1, withastandarddeviationof2.64,andthemaximumnumberofshiftsforasingle proposalis30.Toinvestigatewhetherthemajorityshiftsaremorefrequentinthe proximity of vote executions, Figure 7 reports the aggregated count of majority shiftsacrossproposalsasafunctionofthetimedistancefromthevoteexecution. We observe a constant or slightly increasing trend in farther dates from −100d to −50d, and a clearly increasing trend the closer time gets to the vote date 0d.Thisindicatesthatthetradingofgovernancetokensincreasesshortlybefore polls and that users might trade voting power to decide the outcome of the proposal in their preferred way. We acknowledge, however that we only identify apatternandfurtherresearchisrequiredtobetterinvestigatethisphenomenon. Finally, we examine the participation of contributors in the proposals with majority shifts. Out of 1202 proposals with majority shifts, 1362 contributors associated with 1457 different DAO spaces voted in 728 (60.57%) proposals. 7 Discussion and Conclusions Our study augments the existing body of knowledge on decision-making in DAOs. It substantiates the results of previous studies, highlighting that the distribution of governance tokens is highly concentrated [6,20], and the exercise of voting rights is very low [5]. Going beyond these findings, we discovered evi- dence that contributors, who are essentially users vested in DAOs, are involved in decision-making and, in some cases, have the power to effectively influence The Governance of Decentralized Autonomous Organizations 15 the trajectories of DAOs. Furthermore, we provide evidence that contributors are more likely found at the center of the DAO governance ecosystem and that majority shifts happen, especially shortly before the votes. Thesefindingshaveseveralimplications.First,theysuggestthatcontributors are overrepresented in the decision-making process of certain DAOs compared to other governance users. This is in line with known concerns that contribu- tors may have differing interests and that users with smaller stakes might be discouraged from voting [8]. Second, we found only limited evidence for DAO contributors influencing other, possibly competing, spaces. This is relevant be- causearationalgovernancetokenholdervestedinaspacemightbeinterestedin votingagainstproposalsbenefitingtheevolutionoradoptionofother,competing DAOs (see [19]). Third, we found evidence of co-voting patterns among contrib- utors, which is an indicator of the existence of inner circles of power in DAOs. ThesefindingsrefutetheconventionalwisdomthatDAOsaredecentralizedand run autonomously without being under anyone’s control. This is relevant for re- solving questions of accountability, as vested users and large governance token holders may be considered members of a legally recognized entity and therefore responsible for the underlying dApp. The ongoing Tornado Cash investigation claimingthatthedevelopersinfluenceditsgovernanceisaprimeexampleofthis line of argumentation. Our work clearly faces some limitations and opens directions for further re- search.Currently,itfocusesonoff-chainvotingandononeplatformalone(Snap- shot), whilst voting is executed also on-chain on multiple DLTs, as well as on other off-chain platforms. Extending the study to other governance platforms and to on-chain DLT voting would be a straightforward improvement. Further- more, our results provide preliminary evidence that majority shifts take place before voting. It would be important to investigate and explain more formally the factors influencing governance participation and voting, also combining on- chaindatawithtraditionalmethodssurveyingcryptousers[4,3].Lastly,weview the contributors in our dataset as a lower boundary. Incorporating more data sources, such as Github, would likely elevate this baseline. WhogovernsDAOs?Thisquestionhasbeenthedrivingforcebehindourre- searchandisalsoasignificantconcernforregulatorscurrentlyformulatingpolicy recommendations for Decentralized Finance (DeFi) systems [31]. Although our studydoesnotaimtounveiltheidentitiesoftheresponsibleindividuals,asreg- ulatoryeffortssuggest,itoffersasystematicinvestigationintotheroleofvested users in the governance of DAOs. This, in turn, can provide valuable insights and inform ongoing regulatory debates on that topic. 16 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier References 1. HaydenAdams,NoahZinsmeister,MoodySalem,RiverKeefer,andDanRobinson. Uniswap v3 core. Technical report, Uniswap, 2021. Available at https://uniswa p.org/whitepaper-v3.pdf. 2. Raphael Auer,Bernhard Haslhofer,StefanKitzler, PietroSaggese,and Friedhelm Victor. The technology of decentralized finance (DeFi). Digital Finance, August 2023. doi:10.1007/s42521-023-00088-8. 3. Raphael Auer and David Tercero-Lucas. Distrust or speculation? the socioeco- nomic drivers of us cryptocurrency investments. Journal of Financial Stability, 62:101066, 2022. 4. Stefano Balietti, David Tercero-Lucas, Can Celebi, Myriam Kassoul, and Anca Balietti. State of crypto 22 report, 2023. Available at https://stateofcrypto. net. 5. Tom Barbereau, Reilly Smethurst, Orestis Papageorgiou, Johannes Sedlmeir, and Gilbert Fridgen. Decentralised finance’s timocratic governance: The distribution and exercise of tokenised voting rights. Technology in Society, page 102251, April 2023. URL: https://www.sciencedirect.com/science/article/pii/S0160791 X23000568, doi:10.1016/j.techsoc.2023.102251. 6. TomJosuaBarbereau,ReillySmethurst,OrestisPapageorgiou,AlexanderRieger, and Gilbert Fridgen. Defi, not so decentralized: The measured distribution of voting rights. In Proceedings of the Hawaii International Conference on System Sciences 2022, page 10, 2022. 7. MathieuBastian,SebastienHeymann,andMathieuJacomy.Gephi:anopensource software for exploring and manipulating networks. In Proceedings of the interna- tional AAAI conference on web and social media, volume 3, pages 361–362, 2009. 8. VitalikButerin. Movingbeyondcoinvotinggovernance,2021. Availableathttps: //vitalik.ca/general/2021/08/16/voting3.html. 9. Usman W. Chohan. The Decentralized Autonomous Organization and Gover- nance Issues. Technical Report 3082055, Centre for Aerospace & Security Studies (CASS);CriticalBlockchainResearchInitiative(CBRI);InternationalAssociation ofHyperpolyglots(HYPIA);UniversityofNewSouthWales(UNSW),Rochester, NY, December 2017. URL: https://papers.ssrn.com/abstract=3082055, doi:10.2139/ssrn.3082055. 10. Cointelegraph. Whatdecentralization?solanalendersolendapproveswhalewallet takeover to avoid defi implosion, 2022. Available at: https://cointelegraph.co m/news/what-decentralization-solend-approves-whale-wallet-takeover-t o-avoid-defi-implosion. 11. Compound. Improvement proposal: Compound v2 to v3 migration phase 1, 2023. Available at: https://compound.finance/governance/proposals/152. 12. ConstitutionDAO. Documentation. Technical report, ConstitutionDAO, 2021. Available at https://www.constitutiondao.com/. 13. UnitedStatesDistrictCourt. plaintiffs’replyinsupportofmotionforpartialsum- mary judgment and response to defendants’ cross-motion for summary judgment, 2022. Available at: https://assets.ctfassets.net/c5bd0wqjc7v0/6UF4IrKU07 F1tWhnBT678a/fea1f8fecc839b9f625519247e8e938e/Plaintiffs-_Summary_Ju dgment_Reply___Response.pdf. 14. Curve. Improvement proposal: reduction to crv token inflation, 2020. Available at: https://gov.curve.fi/t/discussion-reduction-to-crv-inflation/851. The Governance of Decentralized Autonomous Organizations 17 15. Decrypt. Steem vs tron: The rebellion against a cryptocurrency empire, 2020. Available at: https://decrypt.co/38050/steem-steemit-tron-justin-sun-c ryptocurrency-war. 16. Maya Dotan, Aviv Yaish, Hsin-Chu Yin, Eytan Tsytkin, and Aviv Zohar. The Vulnerable Nature of Decentralized Governance in DeFi. Technical report, The Hebrew University, August 2023. arXiv:2308.04267 [cs] type: article. URL: http: //arxiv.org/abs/2308.04267, doi:10.48550/arXiv.2308.04267. 17. RainerFeichtinger,RobinFritsch,YannVonlanthen,andRogerWattenhofer. The hidden shortcomings of (d) aos–an empirical study of on-chain governance. arXiv preprint arXiv:2302.12125, 2023. 18. FIOD. Arrest of suspected developer of tornado cash, 2022. Available at: https: //www.fiod.nl/arrest-of-suspected-developer-of-tornado-cash/. 19. Bryan Ford and Rainer Bo¨hme. Rationality is self-defeating in permissionless systems. arXiv preprint arXiv:1910.08820, 2019. 20. RobinFritsch,MarinoMu¨ller,andRogerWattenhofer. Analyzingvotingpowerin decentralized governance: Who controls daos?, 2022. arXiv:2204.01176. 21. Michael Fro¨wis and Rainer Bo¨hme. The operational cost of ethereum airdrops. In Data Privacy Management, Cryptocurrencies and Blockchain Technology: ES- ORICS 2019 International Workshops, DPM 2019 and CBT 2019, Luxembourg, September 26–27, 2019, Proceedings 14, pages 255–270. Springer, 2019. 22. ZhiguoHe,JiasunLi,andZhengxunWu. Don’ttrust,verify:Thecaseofslashing from a popular ethereum explorer. In Companion Proceedings of the ACM Web Conference 2023, WWW ’23 Companion, page 1078–1084, New York, NY, USA, 2023. Association for Computing Machinery. doi:10.1145/3543873.3587555. 23. Johannes Rude Jensen, Victor von Wachter, and Omri Ross. How decentralized is the governance of blockchain-based finance: Empirical evidence from four gov- ernance token distributions, 2021. arXiv:2102.10096. 24. ValerieLaturnus.Theeconomicsofdecentralizedautonomousorganizations.Avail- able at SSRN 4320196, 2023. 25. Robert Leshner and Geoffrey Hayes. Compound: The money market protocol. Technical report, Compound, 2019. Available at https://bit.ly/3ioWOjW. 26. MakerDAO. The maker protocol: Makerdao’s multi-collateral dai (mcd) system. Technical report, MakerDAO, 2020. Available at https://makerdao.com/en/wh itepaper. 27. Messari. Crypto theses for 2022, 2022. Available at: https://messari.io/pdf/m essari-report-crypto-theses-for-2022.pdf. 28. Matthias Nadler and Fabian Scha¨r. Decentralized Finance, Centralized Own- ership? An Iterative Mapping Process to Measure Protocol Token Distribution. arXiv:2012.09306 [cs, econ, q-fin], December 2020. arXiv: 2012.09306. URL: http://arxiv.org/abs/2012.09306. 29. Rektnews. Sushiswapscandal,2021. Availableat:https://rekt.news/sushiswa p-scandal/. 30. Arlind Nocaj, Mark Ortmann, and Ulrik Brandes. Untangling the hairballs of multi-centered, small-world online social media networks. Journal of Graph Algo- rithms and Applications, 19(2):595–618, 2015. 31. TheBoardoftheInternationalOrganizationofSecuritiesCommissions.Policyrec- ommendationsfordecentralizedfinance(defi)consultationreport,2023. Available at: https://www.iosco.org/library/pubdocs/pdf/IOSCOPD744.pdf. 32. EvisSinani,AnnaStafsudd,SteenThomsen,ChristoferEdling,andTrondRandøy. Corporategovernanceinscandinavia:Comparingnetworksandformalinstitutions. European Management Review, 5(1):27–40, 2008. 18 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier 33. Snapshot. Documentation. Technicalreport,Snapshot,2023. Availableathttps: //docs.snapshot.org/. 34. KaterinaStroponiati,IlyaAbugov,YiannisVarelas,KostasStroponiatis,Modesta Jurgeleviciene,andYashodaSavanth. Decentralizedgovernanceindefi:Examples and pitfalls. Technical report, DappRadar, 2020. 35. Xiaotong Sun, Xi Chen, Charalampos Stasinakis, and Georgios Sermpinis. Voter coalitionsanddemocracyindecentralizedfinance:Evidencefrommakerdao,2023. arXiv:2210.11203. 36. XiaotongSun,CharalamposStasinakis,andGeorigiosSermpinis. Decentralization illusionindecentralizedfinance:Evidencefromtokenizedvotinginmakerdaopolls, 2023. arXiv:2203.16612. 37. Sushiswap. Documentation. Tech. rep., Sushiswap, 2022. Available at https: //docs.sushi.com/. 38. UnchainedCrypto. Arbitrumbacktracksonaip-1aftercommunitybacklash,2023. Available at: https://unchainedcrypto.com/arbitrum-backtracks-on-aip-1 -after-community-backlash/. 39. Uniswap. Improvementproposal:Makingprotocolfeesoperational,202. Available at: https://gov.uniswap.org/t/making-protocol-fees-operational/21198. 40. Uniswap. A user posts advances to try and buy votes, 2021. Available at: https: //gov.uniswap.org/t/consensus-check-abolish-delegates-and-change-the -uni-governance-voting-system/13458. 41. Uniswap. Improvementproposal:deployuniswapv3onbnbchain,2022. Available at: https://gov.uniswap.org/t/rfc-update-deploy-uniswap-v3-1-0-3-0-0 5-0-01-on-bnb-chain-binance/19734. 42. Uniswap. Improvementproposal:Uniswapdaocanhelp80kunitoturkiyetorelief after the big earthquake disaster, 2023. Available at: https://gov.uniswap.org/ t/governance-proposal-uniswap-dao-can-help-80k-uni-to-turkiye-to-rel ief-after-the-big-earthquake-disaster/20758. 43. Qin Wang, Guangsheng Yu, Yilin Sai, Caijun Sun, Lam Duc Nguyen, Sherry Xu, and Shiping Chen. An empirical study on snapshot daos. ArXiv, abs/2211.15993, 2022. URL: https://api.semanticscholar.org/CorpusID:254069759. 44. Duncan J Watts and Steven H Strogatz. Collective dynamics of ‘small- world’networks. nature, 393(6684):440–442, 1998. 45. WeiqinZou,DavidLo,PavneetSinghKochhar,Xuan-BachDinhLe,XinXia,Yang Feng, Zhenyu Chen, and Baowen Xu. Smart contract development: Challenges andopportunities.IEEETransactionsonSoftwareEngineering,47(10):2084–2106, 2019. The Governance of Decentralized Autonomous Organizations 19 A Data This Annex provides additional material to Section 3. We start by listing more detailed information on the data sources (A.1) and how we extracted the con- tributor roles (A.2). Additionally, we include how we gathered information on Total Value Locked (TVL) of DAOs and describe the scrape and join procedure in(A.3).Wedescribethecleaningprocedureoftheraw datain(A.4),addmore informationonthevotingstrategies(A.5)andfinallyoutlinethedatavalidation procedure (A.6). A.1 Sources We gathered data from the following sources: – Snapshot:WereceivedacomprehensivedatabasedumpfromtheSnapshot off-chain voting platform, provided by the platform’s developers. It encom- passesallinformationabouttheDAOspaces,associatedproposals,andvotes on each proposal, spanning from the platform’s inception in November 2020 up until the 7th of December 2022. – Ethereum blockchain:WerunafullErigonEthereumarchivenodetoac- quirefurtherinformationoncodeaccountcreatorsandcryptoassetbalances. – Ethereum Name Service (ENS): Taking advantage of Snapshot’s data structure, all DAO spaces are linked to their corresponding ENS domains7, which reveal information about addresses associated with a DAO and the roles of users in charge. – The Graph: This platform makes DLT data accessible through GraphQL and supplies additional information about ENS domains and their sub- domains8, which we are able to utilize and interlink. A.2 Contribution roles We extract contributions with different roles using the following sources: – Owner: we retrieve the domain owner address from the ENS service of the Ethereum node and The Graph. – Administrator: we retrieve the addresses of administrators from Snapshot. – Developer: from Snapshot we obtain all contract account (CA) addresses associated with a certain space. If a CA represents a governance token, we keep it only if it can unambiguously be associated with that space9. Then we retrieve the addresses of contract creators from executed on-chain cre- ate transactions. Adopting a conservative approach, we only consider direct creationsbyusersinitiatedviatheirexternallyownedaccounts(EOAs).We discern indirect creations via contract accounts (CAs). 7 Refer to https://docs.snapshot.org/user-guides/spaces/space-roles 8 https://thegraph.com/hosted-service/subgraph/ensdomains/ens 9 Some spaces might utilize external governance tokens or prominent tokens, like the stablecoin USDT, for voting. 20 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier A.3 Total Value Locked We use the Total Value Locked (TVL) as the benchmark of the financial value of a DAO, and therefore as an indicator of its relevance. We collect the TVL by scraping available dApps from the DeFiLlama API10. We focus on those with TVLavailableonEthereum,andexcludecentralizedentitiessuchascentralized exchanges (CEX). To link the TVL data to snapshot DAO spaces, we directly included all scraped protocols that had governance identifiers in Snapshot. For those with- out such identifiers, we employed a name-matching approach based on Jaccard similarity. We accomplished this by utilizing an algorithm that compares names andidentifierstringsbetweenSnapshotandDeFiLlamaentries.Weintroduceda thresholdof0.5fortheJaccardsimilarityofthenamesand0.6fortheidentifiers to link similar entries. To ensure consistency, we conducted a manual check of the entries and removed those with uncertain matches. For the remaining matches that passed both the automatic and manual checks, we collected the TVL on Ethereum at the latest available time point in 2022. In cases where DeFi protocols had multiple versions, resulting in a 1:n relationshipbetweenSnapshotandDeFiLlama,weaggregatedallversionsofthe same snapshot identifier to establish a 1:1 relationship. A.4 Data cleaning procedure We now provide additional details on the data cleaning procedure described in Section 3. To identify DAO spaces with mature governance structures, we computedfourfeaturesforeachspace:thetotalnumberofproposals;thenumber of proposals with more than ten votes; the number of users following a space; and the TVL obtained from DeFiLama. We then applied the following cleaning steps: first, we removed DAOs with TVL smaller than 100k USD and selected the set of DAOs in the first 500 positions for at least one of the aforementioned measures. Next, we removed 1806 non-final proposals with pending or invalid status and only took into account those with positive final status. We also restricted our dataset to proposals using the single-choice voting type, which reflects the majority of proposals (86.97%). In a single-choice vote, users can choose only oneoption,whichsimplifiestheevaluationofvotingoutcomes.Finally,weverify the consistency of the scoring information of each proposal, i.e., we measure if thesumofindividualvotesandaggregateSnapshotdatacorrespondandremove 17 inconsistent proposals. A.5 Voting strategies There are several ways to enable voting rights on proposals. We denote the methodstodeterminethepowerofavoteasstrategies,followingtheterminology of Snapshot. 10 https://api.llama.fi/protocols The Governance of Decentralized Autonomous Organizations 21 Name Number of proposals Percentage erc20-balance-of 24359 53.85 delegation 5815 12.86 multichain 3842 8.49 erc721 3806 8.41 ticket 3543 7.83 contract-call 3514 7.77 erc721-with-multiplier 2549 5.64 decentraland-estate-size 2260 5.0 pancake 2231 4.93 cake 1350 2.98 erc20-balance-of-delegation 1111 2.46 erc20-with-balance 700 1.55 uniswap 699 1.55 balance-of-with-min 660 1.46 pagination 637 1.41 eth-balance 625 1.38 whitelist 566 1.25 sushiswap 460 1.02 erc1155-balance-of 447 0.99 masterchef-pool-balance 377 0.83 Table A.1. Top 20 Strategies by number of proposals. Strategies are used to determinethevotingpowerofvotes.Welistmostfrequentlyusedonesbynumberand percentage of proposals. TheSnapshotraw datasetintroducedinSection3consistsof8365707votes on 76851 proposals. In these proposals, we identified 208 distinct voting strate- gies, some of which are specific to individual DAOs. The most prominent one is erc20-balance-of (ferc20), which defines the voting power to be proportional to the balance of a specific token, typically the DAO governance token. It is used in53.85%ofproposals.Similarly,thestrategyerc721 (ferc721)computesvoting powerproportionallytotheamountofsomespecificNFTsheldbyeachuserand eth-balance from the ETH balance. Also, the ticket strategy is relevant, imple- menting the “one person, one vote” approach. We note that strategy functions can be implemented arbitrarily and be more complex and also exploit holdings of other cryptoassets such as stablecoins (USDT) or governance tokens of other DAOs. In Table A.1, we show the top 20 strategies by number of proposals that use them. Note that strategies can vary significantly, e.g., from using tokens of multiple chains to being tailored to a specific DAO. The retrieval of this infor- mation necessitates a substantial investment in blockchain infrastructure and ongoing maintenance, concurrently making more complicated for outsiders to validate the correctness. 22 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier A.6 Data validation procedure DAO space Token Error Count bgansv2.eth 0x31385d3520bced94f77aae104b406994d8f2168c [1] 3003 theopendao.eth 0x3b484b82567a09e2588a13d54d032153f0c0aee0[1] 2312 purrnelopes- 0x9759226b2f8ddeff81583e244ef3bd13aaa7e4a1 [1] 2168 countryclub.eth goopsnapshot.eth 0x15a2d6c2b4b9903c27f50cb8b32160ab17f186e2 [1] 1677 9x9x9.eth 0x5219c2f6f8ed1e76c937ed1269eda2658ba3c721 [1] 1296 cryptohoots.eth 0x5754f44bc96f9f0fe1a568253452a3f40f5e9f59 [1] 1168 pandaparadise.eth 0x24998f0a028d197413ef57c7810f7a5ef8b9fa55 [1] 202 cryptohoots.eth 0x196f7e9c5769fc777909a3f1b9bd65959f3f64fb [1] 102 gfanchan11.eth 0xa0b86991c6218b36c1d19d4a2e9eb0ce3606eb48[4] 14 shibvinci.eth 0xe9615071341c6f0392a5dfde1645ad01b810cb43 [3] 9 gfanchan11.eth 0x6b175474e89094c44da98b954eedeac495271d0f [1] 7 ilvgov.eth 0xaebd9bd588f044cbdec8f3cf1e80277a7a52dc69 [2] 4 gfanchan11.eth 0xdac17f958d2ee523a2206206994597c13d831ec7 [4] 3 Table A.2: Report of DAO spaces and patterns of mis- matches. This table contains the DAO spaces with inconsistent valuesbetweenSnapshotandon-chaindata.Patternsoferrorscan be found and categorized in classes. We aim to validate the voting power w of the cleaned data set with on-chain datawhereverpossible.Therefore,wefocusontheon-chainEthereumblockchain andonproposalscoveredbytheselectedstrategiesF′ ⊆P({ferc20,ferc721,feth})11. ThereasoningfortherestrictiontoF′ isthatstrategiesmustbere-implemented manually,andF′containsthethreemostwidelyadoptedstrategies(whichdefine the voting power to be proportional to the balance of a specific token, typically the DAO governance token or to the ether balance). We start by coding the aforementioned strategies, i.e., we gather all asset holdingsofthevotersandrecomputetheirvotingpowerfromon-chainEthereum data. We apply the strategy information given in the Snapshot data. For gov- ernance tokens, this includes the token address as well as the decimal, i.e., the decimal floating point value necessary to convert on-chain data to a floating point number. That is, a token balance is reported on-chain as an integer and must be converted to its actual value. For instance, an illustrative token TKN canbereportedon-chainwithvalue10,000,000butitsactualvalueis10because its decimal is 6. If no decimal is given, we assume it to be zero. When we compare the queried voting power to the Snapshot data, discrep- ancies arise. A closer inspection revealed that some Snapshot proposals report incorrect values for the decimal value, e.g., there is no correspondence with the 11 Note,however,thatinprinciplethesameapproachcanbeappliedtomultiplechains to retrieve additional voting balances. The Governance of Decentralized Autonomous Organizations 23 value reported in the smart contract of the token, or there is no ”BalanceOf” functionandthereforeitisnotpossibletoconverttheon-chainvaluetothefloat- ing point. To avoid considering rounding errors as mismatches, we introduce a tolerance interval ∆ϵ = 10−3 and consider the on-chain and off-chain values as equal if the difference is smaller than the threshold. We compare the voting power computed with that reported by Snapshot, and found high consistency: with this approach, we could verify that 461402 (97.48%) of 473306 Ethereum Snapshot weights are correct. We documented, however, larger deviations for some DAO spaces. The list in Table A.2 reports the spaces and the number of mismatches for each space. We find recurring patterns of errors, which we can categorize and also provide solutions to get also consistent values: – Solution[1]:InsteadofthedecimalvaluegivenbytheSnapshotstrategy,we apply the decimal contained in the token smart contract. If none is given, we consider the decimal equal to 0. – Solution[2]:WerepeattheapproachofSolution[1]butconsiderthedecimal equalto18.18istheconversionfactorfromweitoether,andsomecurrencies adopted it as well. – Solution [3]: We apply the log -function for both Snapshot and on-chain 10 data before comparing them because the on-chain data reported as integers are very large, and therefore the deviation exceeds ∆ϵ, even though the actual values after the conversion would be small. – Solution[4]:Wesubstitutethedecimalwith18toconvertthefloatingvalue. These DAO spaces leave doubts about the reliability of the reported values. Thus, we selected only 357 out of 367 DAOs where all votes have been fully validated.Theresultingvalidateddataset,usedtoconducttheanalysesinSection 6, comprises 119413 users, who voted 438668 times on 8116 proposals of 357 spaces. These include 22878 contributor votes out of 3927 contributions. A.7 Snapshot activity The voting platform Snapshot gained popularity in recent years, and especially after2020.Togiveanoverview,FigureA.1showstheevolutionovertimeofthe monthlynumberofproposalsandtheaveragevotesfromJuly2020toDecember 2022. We observe a trend of increasing proposals and average votes, with peaks in November 2021 for the former and in November 2022 for the latter.

The Governance of Decentralized Autonomous Organizations 23 value reported in the smart contract of the token, or there is no ”BalanceOf” functionandthereforeitisnotpossibletoconverttheon-chainvaluetothefloat- ing point. To avoid considering rounding errors as mismatches, we introduce a tolerance interval ∆ϵ = 10−3 and consider the on-chain and off-chain values as equal if the difference is smaller than the threshold. We compare the voting power computed with that reported by Snapshot, and found high consistency: with this approach, we could verify that 461402 (97.48%) of 473306 Ethereum Snapshot weights are correct. We documented, however, larger deviations for some DAO spaces. The list in Table A.2 reports the spaces and the number of mismatches for each space. We find recurring patterns of errors, which we can categorize and also provide solutions to get also consistent values: – Solution[1]:InsteadofthedecimalvaluegivenbytheSnapshotstrategy,we apply the decimal contained in the token smart contract. If none is given, we consider the decimal equal to 0. – Solution[2]:WerepeattheapproachofSolution[1]butconsiderthedecimal equalto18.18istheconversionfactorfromweitoether,andsomecurrencies adopted it as well. – Solution [3]: We apply the log -function for both Snapshot and on-chain 10 data before comparing them because the on-chain data reported as integers are very large, and therefore the deviation exceeds ∆ϵ, even though the actual values after the conversion would be small. – Solution[4]:Wesubstitutethedecimalwith18toconvertthefloatingvalue. These DAO spaces leave doubts about the reliability of the reported values. Thus, we selected only 357 out of 367 DAOs where all votes have been fully validated.Theresultingvalidateddataset,usedtoconducttheanalysesinSection 6, comprises 119413 users, who voted 438668 times on 8116 proposals of 357 spaces. These include 22878 contributor votes out of 3927 contributions. A.7 Snapshot activity The voting platform Snapshot gained popularity in recent years, and especially after2020.Togiveanoverview,FigureA.1showstheevolutionovertimeofthe monthlynumberofproposalsandtheaveragevotesfromJuly2020toDecember 2022. We observe a trend of increasing proposals and average votes, with peaks in November 2021 for the former and in November 2022 for the latter. 24 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier slasoporP # setov egarevA 70−0202 80−0202 90−0202 01−0202 11−0202 21−0202 10−1202 20−1202 30−1202 40−1202 50−1202 60−1202 70−1202 80−1202 90−1202 01−1202 11−1202 21−1202 10−2202 20−2202 30−2202 40−2202 50−2202 60−2202 70−2202 80−2202 90−2202 01−2202 11−2202 21−2202 4000 2000 0 600 400 200 0 Time Fig.A.1. Evolution of Snapshot activity in time. Number of proposals created each month on the Snapshot platform (top) and the average votes per proposal (bot- tom).ThepanelontopshowsthatthenumberofproposalscreatedmonthlyonSnap- shothasgrownsignificantlyin2021,withapeakofmorethan5,000proposalsexecuted inNovember2021.ThebottomPanelshowsthattheaveragenumberofvotescastper proposalwasaround100untilthebeginningof2022.Sincethen,thisnumberisfollowed an increasing trend. B Influence of Contributors on DAO Governance We provide more details on the involvement of individual DAOs (Section B.1) and on self-decisions (Section B.2). First, we provide additional statistics on thecontributorinvolvementandsecond,wemeasuretowhatextentother-space contributors overrule the same-space contributors in making decisions. B.1 Contributor involvement We complement the analysis in Section 4.1 by providing additional statistics on the top spaces by TVL and their contributor involvement. Table B.1 lists DAOs with contributor involvement above 50% and additional statistical measures. Figure B.1 shows the distributions across proposals for the top five spaces by TVL. In all these plots, the median is less than the average, but we also find outliers even above 50%, meaning that contributors had majorities in certain proposals. The Governance of Decentralized Autonomous Organizations 25 DAOs #Proposals ContributorInvolvement Mean Max Std Min Median levidao.eth 185 100.0% 100.0% 0.0% 100.0% 100.0% cryptodigger.eth 29 100.0% 100.0% 0.0% 100.0% 100.0% zingan.eth 4 100.0% 100.0% 0.0% 100.0% 100.0% 651792.eth 22 100.0% 100.0% 0.0% 100.0% 100.0% strikeorg.eth 1 100.0% 100.0% -% 100.0% 100.0% alongdomainisacheapdomain.eth 23 100.0% 100.0% 0.0% 100.0% 100.0% oligamy.eth 2 100.0% 100.0% 0.0% 100.0% 100.0% atato.eth 22 100.0% 100.0% 0.0% 100.0% 100.0% loganzc.eth 28 100.0% 100.0% 0.0% 100.0% 100.0% test4.shot.eth 28 50.2% 100.0% 42.0% 0.0% 50.0% longbtc20090103.eth 20 50.23% 100.0% 23.02% 30.77% 41.43% castledao.eth 3 50.26% 56.98% 6.04% 45.28% 48.51% mcv.eth 59 50.48% 95.01% 16.08% 0.0% 47.65% comp-vote.eth 5 51.15% 100.0% 36.55% 0.02% 49.62% coinpirates.eth 2 51.23% 70.08% 26.65% 32.38% 51.23% seedifyfund.eth 22 51.3% 100.0% 42.89% 0.0% 28.57% apeswap-finance.eth 24 52.02% 99.37% 40.76% 0.0% 62.19% 5021314.eth 17 52.94% 100.0% 43.79% 0.0% 66.67% sarcophagus-ambassadors.eth 41 54.81% 100.0% 28.86% 0.13% 59.03% rabbitholes.eth 6 56.95% 100.0% 45.48% 0.0% 71.27% lxdao.eth 37 57.5% 100.0% 32.53% 18.18% 42.86% ffdao.eth 25 57.69% 86.59% 17.32% 19.03% 58.99% usdaogov.eth 21 57.9% 100.0% 43.55% 0.0% 77.78% sdbal.eth 31 58.02% 100.0% 50.12% 0.0% 100.0% layer2finance.eth 2 58.73% 84.87% 36.96% 32.59% 58.73% meebitsdaopool.szns.eth 22 59.02% 100.0% 41.14% 0.0% 74.94% metrox.eth 34 59.54% 100.0% 46.08% 0.0% 99.93% ginoct.eth 64 60.99% 100.0% 47.5% 0.0% 100.0% weenus 44 61.61% 100.0% 47.63% 0.0% 100.0% dorg.eth 161 62.81% 100.0% 27.54% 0.0% 67.56% al409.eth 17 65.11% 100.0% 36.51% 0.0% 66.67% dakshow.eth 25 66.67% 100.0% 43.03% 0.0% 100.0% alexec.eth 3 66.69% 100.0% 57.7% 0.06% 100.0% wintersun.eth 11 68.79% 76.04% 13.53% 28.55% 73.37% cryengine.eth 13 69.01% 85.03% 17.13% 31.0% 75.4% pharo.eth 1 69.37% 69.37% -% 69.37% 69.37% insuretoken.eth 36 69.4% 100.0% 46.69% 0.0% 100.0% wdefi.eth 23 70.94% 100.0% 42.61% 0.0% 99.54% thanku.eth 21 71.43% 100.0% 46.29% 0.0% 100.0% primerating.eth 249 72.06% 100.0% 34.75% 0.0% 100.0% cvx.eth 265 72.44% 99.58% 26.34% 0.0% 77.57% testbsw.eth 17 76.4% 100.0% 43.68% 0.0% 100.0% venus-xvs.eth 17 77.47% 99.88% 36.07% 0.0% 91.47% lemu.dcl.eth 42 78.16% 100.0% 41.4% 0.0% 100.0% huangwenchao.eth 72 78.24% 100.0% 37.7% 0.0% 100.0% melson.eth 14 78.82% 100.0% 19.84% 27.28% 82.33% cabindao.eth 4 79.53% 97.16% 20.28% 56.86% 82.05% grantsdao.eth 36 80.42% 100.0% 20.64% 33.33% 77.5% polywrap.eth 132 80.58% 100.0% 20.03% 0.0% 83.57% cgpool.eth 1 81.04% 81.04% -% 81.04% 81.04% 4.spaceshot.eth 9 81.19% 100.0% 34.68% 0.24% 100.0% xunfa.eth 35 81.86% 100.0% 5.91% 79.83% 79.83% tommyg.eth 14 83.33% 100.0% 36.4% 0.0% 100.0% retokendao.eth 36 84.46% 100.0% 10.09% 54.55% 85.43% frami.eth 206 85.17% 100.0% 34.98% 0.0% 100.0% adaocompany.eth 20 85.85% 100.0% 32.06% 0.0% 100.0% 3.spaceshot.eth 69 86.22% 100.0% 23.12% 0.0% 100.0% mycontext.eth 9 87.04% 100.0% 26.06% 33.33% 100.0% * 15 87.78% 100.0% 28.5% 0.0% 100.0% btc1.eth 2 89.49% 100.0% 14.87% 78.97% 89.49% szns.shean.eth 61 91.15% 100.0% 27.93% 0.0% 100.0% bridgeswap.eth 25 91.73% 98.7% 5.11% 83.9% 94.69% oracles.opiumprotocol.eth 25 94.77% 100.0% 18.74% 8.77% 100.0% ecashxec.eth 1 95.82% 95.82% -% 95.82% 95.82% halodao-kovan.eth 41 97.56% 100.0% 15.62% 0.0% 100.0% alexjyoung.eth 24 98.75% 100.0% 6.11% 70.05% 100.0% TableB.1.Statisticsofcontributorvotingpower.Additionalstatisticsonvoting power for the 66 DAOs with mean contributor involvement above 50%. * DAO name contains character that cannot be displayed. 26 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier aave.eth frax.eth instadapp−gov.eth lido−snapshot.eth uniswap 100% 75% 50% 25% 0% DAOs ]%[ srotubirtnoc fo rewop gnitov evitaleR Fig.B.1.Contributors’involvementintop-TVLDAOs.Box-andviolinplotsof relative voting power of contributors across top TVL DAOs. We find the mean values (X) above the median, for some proposals but outliers even above 50% relative voting power. Thus, also in high TVL DAOs, contributors had majorities in some proposals. B.2 Contributor self-decisions Decisions by other-space contributors To provide a broader picture, we complement the measure of contributor self-decisions, introduced in Section 4, with the contributor other-decisions, that is, we measure to what extent other- spacecontributorsoverrulethesame-spacecontributorsinmakingdecisions.We definethreejointconditionsforproposalstobedecisiveother-votes.Thefirsttwo conditionsareinspiredbydecisive self-votes appliedtovotesofother-spacecon- tributors,whichwedenoteas“other-contributorvotes”.First,other-contributor votes have a relative majority within the voting power for the decision. Second, we consider also the second-ranked option and select those proposals where the weight of other-space contributors in the decision is higher than the weight of the second-ranked option. Third, we take into account only proposals where same-spacecontributorsareinvolvedinthevotingforthesecondrankedoption. Formally, we define the set of decisive other votes Vp =Vp ∩Vp and also DO oˆ1 OS thecomplementsetVp =Vp \Vp ;weidentifythefractionsofrelativevoting CO oˆ1 DO powerfordecisiveothervotes(7),forthecomplementset(8)andthesame-space contributors for the second option Vp =Vp ∩Vp as SS2 oˆ2 SS (cid:88) (cid:88) (cid:88) w˜p = w˜, (7) w˜p = w˜, (8) w˜p = w˜, (9) DO i CO i SS2 i vi∈V D p O vi∈V C p O vi∈V S p S2 We consider proposals to be decided by other-contributors, when all three conditions are fulfilled: [(w˜p >w˜p )∧(w˜p >w˜p )∧(w˜p >0)], (10) DO CO DO SS2 SS2 The Governance of Decentralized Autonomous Organizations 27 In rare cases, we find that the outcome of a proposal has been dominated by contributors of other spaces over the self-contributors of the proposal’s space. WeshowtherelativevotingpowerfortheseproposalsinFigureB.2,byastacked bar plot of votes on different choices. Each proposal is illustrated by three bars, respectively corresponding to the relative voting power of the winning choice, the second choice by voting power and all other choices aggregated. 520crypto.eth 651792.eth 651792.eth adaocompany.eth adaocompany.eth adaocompany.eth bridgeswap.eth bridgeswap.eth 0x7c2... 0x567... 0x761... 0xd2d... 0xde6... 0xe6f... 0xd8b... 0xe69... crisisdao.eth ethip.eth evoms002.eth evoms002.eth fabien.eth frax.eth melson.eth melson.eth 0xc7b... 0xc15... 0x261... 0xff8... 0x806... 0x2fe... 0x4f6... 0x6ed... melson.eth orangedaoxyz.eth qhsky.eth tecommons.eth test4.shot.eth wintersun.eth wintersun.eth wintersun.eth 0xe63... 0x877... 0xbf7... 0x592... 0xbe0... 0x437... 0x4cb... 0x773... wintersun.eth wintersun.eth wintersun.eth wintersun.eth xlootproject.eth zksyncio.eth 0x77f... 0x7c2... 0x9b2... 0xb70... QmURj... 0xec1... gninniw dnoces rehto gninniw dnoces rehto gninniw dnoces rehto gninniw dnoces rehto gninniw dnoces rehto gninniw dnoces rehto gninniw dnoces rehto gninniw dnoces rehto 1.00 0.75 0.50 0.25 0.00 1.00 0.75 0.50 0.25 0.00 1.00 0.75 0.50 0.25 0.00 1.00 0.75 0.50 0.25 0.00 Choice(s) rewop gnitov evitaleR Contributer none other−space same−space Fig.B.2.Decisionsofother-spaceagainstsame-spacecontributors.In30pro- posals, other-space contributors ( ) had the relative majority in the winning choice and overruled the same-space contributors ( ), voting for the second choice. C Network Analysis We present additional material about network descriptive statistics (C.1), an analysis of the small-world properties of the networks (C.2), an investigation 28 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier Network G G G G AA AW TA TW Daos All All Top-100 Top-100 Votes All Winning All Winning Num Nodes 104863 75879 20401 14494 Num Edges 739813062 107374710 19917792 6045065 Max Degree 57304 31664 14456 9039 Avg. Degree 14110.09 2830.16 1952.63 834.15 Contr. Nodes 1.29% 1.45% 3.25% 4.5% Contr. Edges 1.61% 1.76% 3.4% 8.0% Contr. Max. Degree 57303 24346 14196 7755 Contr. Avg. Degree 8795.3 3434.12 2079.8 1490.86 Giant Component 98.1% 98.8% 99.9% 99.4% Assortativity 0.02 0.05 0.11 0.07 Avg. Path Length 2.24 2.58 2.21 2.37 Diameter 32 21 7 9 Clustering 1.55 1.59 1.99 2.06 Small Worldliness x 3.85 4.31 6.41 Louvain Communities x 330 28 12 Largest Community x 42% 43% 34% Table C.1.Network statistics of four co-voting networks.Thetopofthetable defines the four networks as a unique combination of two features: DAOs and Votes. Top-100 DAOs are ranked by total value locked (TVL); Winning votes are votes for the choice that ultimately won the majority of voting power. The rest of the table contains network statistics for the four graphs, the central part referring to all nodes and edges, and the bottom parts only to the giant component. Some statistics could not be computed for the largest network due to memory limitations. of the network of contributors (C.3), additional plots for the centrality and k- core analysis (C.4), and finally, a sensitivity analysis for the choice of the edge threshold (C.5). C.1 Co-voting network statistics InSection5webuiltthefourco-votingnetworksG ,G ,G ,G .Their AA AW TA TW exhaustive statistics are presented in Table C.1. The table is split in four seg- ments, starting on top with the distinction of the networks by combination of their features. Then, it provides descriptive statistics of the network nodes and edges, of the contributor nodes only, and, finally, insights on network measures computed on the giant component. C.2 Small-world properties The co-voting network with all DAOs and all votes (G ) counts 104863 nodes AA and is about 5 times larger than the corresponding network with the top-100 The Governance of Decentralized Autonomous Organizations 29 DAOsbyTVL(G ).Despitethedifferentsizes,thefournetworkspresentsim- TA ilar topological features: a very large giant component (about 99% of all nodes) with clear small world features. Following the procedure in [32], we computed the “small-worldliness” coefficient of each graph, comparing the ratio of average path length (APL) and clustering coefficient (CC) normalized by the values of those statistics in a random network (rnd) with the same number of nodes and degree distribution (test): SW =(APL /APL )/(CC /CC ) (11) test rnd test rnd Intuitively,tobeasmallworld,anetworkneedsAPLtobesimilartothatofa randomnetwork,butCCshouldbemuchlarger,leadingtoa“small-worldliness” coefficient SW considerably larger than one. This is indeed the case for all our networks (SW between 3.85 and 6.41). The non-random structure is also con- firmedbyasmall,yetpositiveassortativitycoefficient,indicatingacertaindegree of homophily. C.3 The contributors network Users might connect spaces through their contributions. We construct the con- tributors network G as a bipartite network of the contribution relation C, SU using spaces S and users U as two node types without taking into account their specific roles T. We utilized the raw data set from Section 3, containing 12294 DAO spaces, and extracted the contribution roles with sources defined in Section A.2. Using these sources, we found a total of 12272 DAOs 17561 addresses. 11949 of these addresses are voters in the raw data, which would strictly interpreted only be identifiedascontributions.However,weareinterestedinthemostcomprehensive network of contributors to DAOs to find all possible relations. We construct the network G linking 12272 DAOs to all 17561 contribution addresses. Thus, SU on average, each DAO has 1.43 contributors. This network appears fragmented, as it separates into 10651 small components, mostly with ten or fewer nodes and with rather simple relations between addresses and a few larger ones. In Figure C.1 we plot the distribution of the connected component. We limit the networktocomponentssizelargerthan10andcomputethebipartiteprojection of G on DAO spaces. The resulting network in Figure C.2 links DAO nodes SU whentheyhavecontributorsincommon.Weobserveseveralstar-likestructures, but there also exist larger structures created by contributors bridging multiple DAOs together. However, keeping the distribution of Figure C.1 in mind, most of the relation consist only of rather simple relations between DAOs. 30 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier (1,2] (2,3] (3,5] (5,10] (10,100] (100,200] 0 2000 4000 6000 Number of Components ezis tnenopmoC Fig.C.1.NumberofconnectedcomponentsoftheG network.Thenetwork CD mainly consists of small components and a larger component with above 100 nodes. Fig.C.2.NetworkofDAOsconnectedtroughcontributions.Bipartitenetwork projection of contributors network G with more than 10 members. The network SU illustrates nodes that are linked when they have a common contribution. It appears fragmented, with star-like structures and a view larger bridged DAO communities. The Governance of Decentralized Autonomous Organizations 31 A Betweenness B Closeness C Eigenvector 0.04 0.0100 0.00020 0.03 0.0075 0.00015 0.02 0.0050 0.00010 0.00005 0.01 0.0025 0.00000 0.00 0.0000 D Pagerank E K-Core A A l l l l W A i l n l Top A - l 1 l 00To W p- i 1 n 00 3000 9e-05 2000 6e-05 Non-Contributor Contributor 1000 3e-05 0e+00 0 All All Top-100Top-100 All All Top-100Top-100 All Win All Win All Win All Win Fig.C.3. Centrality and k-core statistics in the four co-voting networks. Contributors tend to have higher centrality and k-core across networks and measures. Overall, it suggests that contributors have influence over the entire network. A-D. Centrality measures (some measures for the largest network could not be computed duetomemorylimitations).Allcentralitymeasuresmakeuseofedgeweightsandare applied on the giant connected component. E. Geometric mean of K-core. Error bars are 95% confidence intervals of the means. C.4 Centrality and k-core WereportinFigureC.3alltheplotsofallcentralitymeasuresmentionedinthe main text. Contributors tend to score higher in all centrality measures, but one (eigenvector centrality in the network with all DAOs and only winning votes). In addition, in Figure C.4, we report the density plots for the distributions ofk-coreacrossallnetworks.Contributorsarelessfrequentintheportionofthe distributionswithlowerk-core;however,intwonetworks(G ,G )anoutlier AW TA clusterofnon-contributorswithveryhighk-coreisdetected.Forthisreason,we optedtoreportthegeometricmeaninthemaintext.Econometrictestswiththe arithmetic means are mixed: contributors have on average significantly higher k-core (t-test p<0.001) in two out of four networks. C.5 Edge-weight threshold sensitivity analysis Inthemaintext,wereportresultsfornetworksbuiltwithanedge-weightthresh- old equal to ten, so that an edge between two nodes is added if and only they have have co-voted in more than T=10 proposals. The value of ten has been 32 S. Kitzler, S. Balietti, P. Saggese, B. Haslhofer, M. Strohmaier All DAOs ; All Votes All DAOs ; Win Votes 0.00075 1e-04 0.00050 5e-05 0.00025 0e+00 0.00000 0 10000 20000 30000 0 2500 5000 7500 Top-100 DAOs ; All Votes Top-100 DAOs ; Win Votes 0.0020 9e-04 0.0015 6e-04 0.0010 3e-04 0.0005 0e+00 0.0000 0 1000 2000 3000 4000 5000 0 250 500 750 1000 1250 K-core ytisneD Non-Contributor Contributor Fig.C.4.K-core density plots for contributor and non-contributor nodes in the four co-voting networks. Contributors are less frequent in the lower portion of the distribution in all plots. In two plots (G , G ) an outlier cluster of non- AW TA contributor with very high k-core is detected. conservatively chosen to be low enough to keep as much of the original network structureintact,butatthesametime,allowingustoreducebothdatanoiseand thecomplexityofcomputation.Wetestedhowthechoiceofathresholdaffected themainnetworkstatistics,andtheresultsarereportedinFigureC.5.Formost statistics—i.e., number of edges, number of nodes, average path length—the curve varies smoothly with T; however, assortativity and the pseudo-diameter seemmoresensitivetothechoiceofthethreshold.Furtheranalysesmayprecisely quantify the impact of the choice of threshold on our results. The Governance of Decentralized Autonomous Organizations 33 90000 60000 30000 0 25 50 75 100 secitreV# 1e+08 5e+07 0e+00 25 50 75 100 segdE# 1.00 0.75 0.50 0.25 0.00 25 50 75 100 ytivitatrossA .gvA 0.009 0.006 0.003 0.000 25 50 75 100 ytivitatrossA .dtS 90000 60000 30000 0 25 50 75 100 tnenopmoC tnaiG secitreV# 1e+08 5e+07 0e+00 25 50 75 100 tnenopmoC tnaiG segdE# 6 4 2 0 25 50 75 100 Edge-Weight Threshold htgneL htaP .gvA 30 20 10 0 25 50 75 100 Edge-Weight Threshold retemaiD oduesP All DAOs ; Win Votes Top-100 DAOs ; All Votes Top-100 DAOs ; Win Votes Fig.C.5. Overview of network measures for different edge-weight thresh- olds. Two nodes must have co-voted in more than T proposals, where T is the edge- weightthreshold.ThevalueT =10usedinthemaintextcorrespondstotheleft-most point on the x-axis. All plots are stacked.