 I'd like to thank Julian and Andre for inviting me to present in this seminar series. It is my great pleasure and I also appreciate the opportunity to share some of my current research with people from different fields. So the paper that I'm presenting today is about whether the geek platforms should decentralize the dispute resolution. And this is a joint work with my doctoral student at Cornell, Vickette Lee. So this is a working paper. We are actually revising it right now. So any feedback will be very helpful for us to push the model further. So in this paper, I'm going to focus on a particular but major type of geek platforms, and that is the online labor platforms. So here I'm showing you some examples. So usually, basically any of these geek platforms will specialize in a particular type of tasks that you can find on their website. For example, the Task Revit specializes in home services and Upwork and freelancer.com specialize in the more creative tasks such as web design or marketing tasks. And there's also the Amazon M-Turk. In there, we can basically do simple tasks like filling out a survey or tagging an image. More recently, these online labor platforms have actually been growing very fast. And one important reason that workers may want to work on those platforms as a freelancer is that they get to work whenever they want. So they actually get to enjoy the flexibility in their work schedules. And also they can get access to a broader group of clients. So basically you can work for anyone around the world. And that's particularly true if the tasks can be done online. And there's also some data that showed recently the online labor platforms account for more than 90% of the net employment growth in the United States. And it is also projected that by the year of 2027, more than half of the U.S. workforce will participate in those online labor platforms. And another factor that kind of contributed to the recent growth of this gig economy is the COVID-19 pandemic. So with the pandemic, a lot of people actually switched to working online. And it's just so much easier for them to find jobs on those gig platforms. So overall, what we do see here is these online labor economy has really become an important part of the overall economy. But at the same time, there are also problems that could arise and there are also challenges that these platforms need to manage. For example, quality could be a primary concern for these platforms. Because usually the workers there are freelancers. So many of them may not be professionally trained employees like the company employees. So the quality of their work could be a question mark. And then because of that, typically, these platforms will allow the client to reject paying the freelancer after the work is done. So if the client thinks the work is not worth getting paid, then the client can actually reject to pay. But in that case, the freelancer may not be happy. So the freelancer may actually file a dispute over the payment. So that's basically how a dispute could occur in this case. So if you look at these online forums where people share their experience from using those online platforms, there were actually quite some complaints about the quality of work and also about the disputes. So many of the clients actually complained that the work quality from the freelancers is actually not sufficiently good. So they think many of the freelancers have failed to meet the required standards. And at the same time, many of the freelancers were also complaining that the clients were being overly picky. So their services are being undervalued to a point where they were practically working for free. So there's apparently disagreement over the task assessment between the two sides of the market. And when dispute occurs traditionally, the platform is going to be the one that handles this dispute case because the platform is making centralized decisions. In this case, in the paper we call this traditional model, the centralized dispute system. However, there has been discussions and even criticism about this centralized decision-making mechanism because people were saying that the platform may have a conflict of interest. Maybe the platform would want a particular site to win. And if this keeps happening and if the unfair judgment keeps happening, then there could be a long-term damage to this market. And because of that, more recently there's actually a number of emerging online platforms that promise to handle disputes differently. So the idea that these emerging online platforms were putting forward is that instead of letting the platform decide who wins the dispute case, maybe we should ask the other platform users to vote on the dispute case. So that's exactly the idea that they want to implement with these new business models. And of course, many of these platforms here are startups and several of them have already been in operation, such as CryptoTest. And next, let me basically introduce how this decentralized dispute system works. So in this case we consider it as a decentralized dispute system because the dispute decision is not made by the platform, but it's being made by the individual platform users through a voting mechanism. So in this case, suppose that the client wants to reject paying the freelancer. And then the freelancer is happy about it. So the freelancer filed a dispute case. Under the decentralized model, the platform will basically form a tribunal consisting of individual platform users. So there are going to be voters in this case. And each of the voters will reveal this dispute case and submit their vote. And who wins will be dependent on the majority rule. For example, in the example that I'm showing you right now, there are three voters in this tribunal. And two of the voters have voted for the client. So that's the majority. So the client will win. And what that means is that the client will not have to pay the freelancer. But if the freelancer wins, then that means the client will have to pay the freelancer. So this is how the voting works in this case. Actually, this whole voting, decentralized voting idea is not new. In 2007, eBay actually did something very similar by creating the eBay community court. But in that case, the participation of voters is entirely based on a sense of community contribution. But there is no monetary incentive involved. And what happens is that it basically didn't work for long. So after some time, eBay actually terminated that program. And the way that these new platforms are doing this voting model is actually different because they want to create a monetary incentive to drive participation of the voters. So they want to give the people some monetary incentive so that they will want to consistently participate in this voting process. And the way that it works is every time a voter, in order for a voter to participate in the voting process, each voter will have to deposit a participation fee. And all the participation fees from the voters will form the reward pool. And then after the voting is done, the people who have voted for the majority side will split the total reward pool in this case. So here you can imagine that every time a dispute case is filed, we will form a tribunal consisting of a number of voters. And then there will be a reward pool. And then the reward pool will be shared by the people who actually voted for the winning side. And for the more to make sure that the payment transfer is credible and trustworthy, these platforms basically rely on blockchain to build this whole system. So that's basically the business proposal of these emerging online platforms. And I guess by now, you might have seen a concern, right? Because in this case, the voters may be participating on those tribunals for the purpose of earning more money, right? Then it is actually not clear whether the voting outcome would be a good one, right? So whether the voting outcome will be a fair resolution to this field. And that is indeed one major concern that these online platforms have thought really hard about, right? So one thing that they want to prevent is the voter's collusion, right? So something that should not happen is that, let's say one voter actually knows many other voters and then this voter basically says, hey, let's all vote for the freelancer this time. So we're going to win and then we're going to split all the money, right? So these cannot be happening in this case. And these platforms have actually taken several measures to prevent the voter's collusion from happening in the first place. For example, in terms of how to select the voters for each dispute case, these platforms actually doing selecting the voters in a random way. And the reason to do that is that each voter will not know which dispute cases I will actually vote for in the future, right? So that's going to make it very difficult for people to collude with each other. And for the more the platforms strictly forbid any communication channel between the voters. And finally, they actually, the voting will actually happen within the limited time window, usually just 24 hours, right? So all these measures are intended to make sure that the voters are not going to find out about each other so that they will not be able to collude, right? So let's say that these basis model will actually work in terms of preventing the voter collusion. However, voters will still vote strategically, right? Because each voter will still need to guess what the other voters will vote. And then the voters may naturally have a tendency to vote for the side that they believe will also win the dispute case, right? So if the voters vote strategically, then will these decentralized models still work? In other words, will the decentralized dispute system will achieve justice? And then what will be the condition for this new business model to actually work? So in this paper, so this is basically how we got really interested in this problem. And what we want to do in this paper is we want to build a theoretical model to capture the strategic voting behavior of the voters. And then we want to find out, under work circumstances, will this voting model work? And furthermore, we will compare this new decentralized model to the traditional centralized model. And then we will see which model will be more profitable for the platform. And finally, we will take the perspective of the social planner, and we will find out which dispute system will be better for the social welfare. Okay, so that's basically how a little bit of background for how we got interested in this research. And what are the research questions we want to address in this paper? And if there is no questions at this point, I'm just going to continue. So in the interest of time, I'm going to be very brief about the literature review. I just want to mention that our paper is related to several streams of research. In particular, in recent years, there's actually a fast growing literature in OM, which is the field that I am in. And that literature is about the platform operations. And of course, there's also the huge economics literature that look at platforms. And our paper is also related to other modeling works, such as the voting games and also the quality contracting models. All right, so now let me introduce our model. So I will first talk about the centralized case. And then so I will show you the model and then show you the equilibrium. And then I will move on to the decentralized model. And finally, I will show you the results from comparing the equilibrium under these two models. All right, so the centralized model. So in this case, there's no voting or tribunal, right? So the platform is going to make the dispute decision. So basically here we consider a typical platform setting. There is a client and there is a freelancer and they're going to contract with each other. And then the platform is going to intermediate the transaction between the two sides of the market. So what I'm showing you on this slide is the sequence of events. I'm going to go over this sequence of events to show you how we model the strategic interactions between the different parties in this setting. And we are going to assume that everyone makes rational and forward-looking decisions. So first, the platform is going to decide on the fees to charge. So F is the dispute fee. That is basically the fee that the freelancer will need to pay if the freelancer wants to file a dispute case. And in addition to the dispute fee, the platform will also be earning the commission revenues from each transaction. And then we're going to assume that the commission rate, which is the percentage commission, as an exogenously given parameter. So it's basically determined by the industry norm. And after the platform announces its fees, the client is going to offer a contract to the freelancer. So the client will choose a price P and then if the freelancer wants to accept the contract, the freelancer will decide on his work quality queue. And then we assume that the freelancer can incur an effort cost to improve his quality. And this effort cost is a convex, quadratic increasing function in the work quality queue. So here the alpha parameter is going to be an important parameter in a model because it captures the inverse of the skill level of the freelancer. So if alpha is bigger, that means it will be more costly for a freelancer to achieve the same quality level. So that means the freelancer pool actually has a lower skill level. All right. So after the work is done, the client will decide whether she wants to accept or reject the freelancer's work. We assume that the client's valuation of the work is equal to the quality level queue in the main model. And we also have a model extension where we allow the client's valuation on the work to be different from the work quality. But all the insights actually carry through in that case. So in this case, the client will actually be forward looking. So she will know that if she rejects and for the more the freelancer initiates a dispute, then the platform is going to make the decision and then both sides will win with a certain probability. So the client is going to take all that into consideration when she decides whether she just accepts the work or actually rejects to pay. And then if the client rejects, then the freelancer will initiate, will decide whether to initiate a dispute or not. And in the case of dispute, finally, the platform is going to decide which side wins. And then on the next slide, I'm going to focus on the last stage. And I'm going to show you more details about how we model the platform's decision for the dispute resolution in this case. And I think there's a couple of questions. So I'm happy to just answer those before I move on to the next part of the model. Do you need me to read them or are you happy to? Please do. Thank you. So Lyric Cabral asked, are you modeling platform participation and network effects from your extensive form? It seems like you're not. Okay, right. So in this case, we basically look at a single transaction because this game actually has multiple stages. So in that sense, we do not explicitly capture the network effect. But we believe the insights will carry through if we kind of consider a multiple transactions or even a repeated game setting. Oh, yes. And Andre responded to that already. Yes. Okay. All right. So now I'm going to continue with the centralized model. So basically to solve this game, we are going to do a background induction. And then the first step of analysis is going to be the last stage, which is the platform's decision on the dispute resolution. So let me show you how we actually model the decision making in this case. So we're going to assume that the platform chooses a quality threshold to evaluate the freelancer's work. So basically if the work quality is higher than this threshold, then the platform will let the freelancer win. And then we also assume that after the platform reviews the dispute case, the platform will receive a noisy signal of the freelancer's true quality. So in this case, the platform does not observe the true quality. And what it observes is a noisy signal X. So basically the platform is going to compare the signal X to its quality threshold, which is endogenously determined. And if X is higher than the quality threshold, the freelancer will win, otherwise the client will win. So here the epsilon is a noise term, which is uniformly distributed between minus one and plus one. And then sigma is the scaling factor. And next, let's actually look at what will the utility function for the platform in the case of dispute happening. So basically based on these decision making rule, we can calculate the probability for the freelancer to win. And that's the H of Q function. So if the freelancer wins, then the client will have to pay the freelancer. And then the platform will be able to extract this percentage commission revenue from the transaction. And at the same time, the freelancer has also paid the dispute fee. So this is the revenue that the platform earns in this case. And then if the client wins, then the client does not pay. So the freelancer does not earn the commission. So he only earns the dispute fee. And from this utility function, you can see that the platform may have a conflict of interest. The platform may naturally have a tendency of letting the freelancer win with a higher probability because if the freelancer wins, then the platform will earn the commission. But if the platform consistently makes biased decisions, then it could actually create a negative reputation damage in the long term. So to capture that, we also have this disutility term in the platform's utility function. And the disutility term is intended to capture this long term reputation loss if the platform actually makes biased decisions. So we're going to assume that the disutility is an increasing and convexly increasing function of the difference between the platform's decision rule and the industry norm. So here, basically, we introduced this parameter y as the industry standard. So what this y parameter captures is that what would the industry norm, what would the industry think that the quality, the how high the quality level needs to be, right? And then the difference between the platform's threshold and the industry standard will basically measure how much bias there is in the platform's decision making. So that's the platform's utility. And then after we analyze this last stage of the game, basically, we can go back to the main game and then we work through the entire background induction process. So next, I'm going to show you the equilibrium. So first of all, for the last stage of the dispute resolution, after we solve the game, we find that the platform's quality threshold is equal to the industry standard minus a positive term, right? So that means the platform is going to set a lower threshold compared to the industry norm because it wants the freelancer to win with higher probability. And here, the bias is also proportional to the amount of the commission revenue because that's exactly why the platform has this bias. And correspondingly, if you look at the probability for the freelancer to win, there's also going to be a bias term. So the freelancer will win with a higher probability compared to what the industry norm requires. And then now if you look at the overall equilibrium of the entire game, so here, basically, we characterize the equilibrium dispute fee of the platform and also the equilibrium contract price that the client offers and also the quality level that the freelancer chooses. So with everything taken into consideration, what we have is that depending on the skill level of the freelancer, one of three cases could happen in equilibrium. So if the skill level of the freelancer is very low, then the freelancer will find it very costly to improve quality even to a minimum acceptable level. So in that case, the freelancer will not be able to participate. So no transaction happens in that case. And if the skill level of the freelancer is moderate, the freelancer will be able to participate. But in that case, the freelancer will not be able to provide a quality level that is sufficiently high. So what happens in the equilibrium is that the client will tend to reject the work and then the freelancer will file a dispute. So eventually it's going to be the platform's decision for who wins or doesn't win. And then it's going to happen in a probabilistic manner. And in the third case, could I ask a question again from the chat? Sure, sure. Which is from Luis again, who asks, how important is the industry standard deviation assumption for your results? And why isn't reputation monotonic in quality standards? Tonic in quality standards. Ah, okay. Yeah, that is a good question. So basically these so we introduced this parameter of the industry standard to kind of capture this long term damage, the negative effect of the decision making bias of the platform, right? So it is an important component of the model, right? So if we don't have it, then basically the platform will be completely strategic in a way in this decision making. And also we model this as a quadratic term, because we want to capture the fact that the damage of this decision making bias can be actually highly convex in terms of the amount of bias that the platform introduces in this decision making. Okay. All right. So I think, yeah, I was here about the three cases under the equilibrium. All right. So these are the three possibilities that could arise in equilibrium. And basically in each case, we have fully characterized the sub game perfect equilibrium. And here you can see in the case where dispute does not happen, the platform does not burn any dispute fee, right? So what it earns is the commission revenues. But if dispute happens, then the platform will earn the dispute fee. And also the platform will earn the commission revenue with a certain probability. And that is the probability that the freelancer will win the dispute. And finally, the minus term here basically is the the disability caused by the bias. So this is the equilibrium of the centralized case. And next, let me actually move on to the decentralized case. So the overall game structure is very similar to the centralized case. And the main difference here is about how the dispute is being handled. So in this case, the platform does not make any decision there. But the tribunal consisting of individual voters will actually vote on the case. So once again, we're going to use background deduction. So we're going to start with the last stage of the game. And here just to kind of recap how this voting mechanism works, right? So we're going to assume there is a continuum of voters with a total mass equal to one. And then each voter will pay a participation fee equal to $1 to participate. And then the total participation fee will form the reward pool. And we also allow the platform to put in an additional amount of money into the reward pool. So eventually, after the voting is done, the voters who have voted for the majority side will actually share the total reward pool. So essentially, what we want to capture in our model is the strategic voting behavior of the voters. So what we have here is that the voters cannot communicate with each other, so they don't know who other voters are. But they still have to coordinate with each other somehow. And to capture that dynamic, we adopted the global games framework to model our voting game. In particular, we're going to assume that each voter has a uniform prior on the true quality of the freelancer. And then after they reveal the district case, they're going to receive an update. So they're going to receive a noisy private signal, which takes a very similar, basically the same form with the platform case. And then they're going to update their beliefs. And then after that, they're going to decide whether to vote for the freelancer or the client. And let me briefly show you how we model the utility of the voter in this case. The structure is actually very similar to the platform's utility. So there's going to be a term for the monetary earnings. And there's also going to be a guild term in this case. So the monetary earning term is basically how much money they can earn in expectation by participating into a tribunal. So the minus one term here is basically the participation fee that they need to deposit. And then they can earn money if they have voted for on the majority side. So there's two possibilities here. If the majority of the tribunal votes for the freelancer. So here, L represents the proportion of voters who vote for the freelancer. In that case, if voter I also votes for the freelancer, so that means AI is equal to one, then he's going to share, he's going to obtain a share of the total reward pool and earn more money than the participation fee. On the other hand, if the majority of the tribunal votes for the client, then this voter will only earn the reward if he also votes for the client. So that's the monetary component. But at the same time, we also want to capture the fact that voters may also care about justice. So we're going to assume each voter will incur a guilt if they actually vote for the wrong side. So what we want to capture here is that let's say a voter has observed a very high quality signal Xi. So it's clearly higher than the industry standard Y, but somehow the voter wants to vote for the client. So obviously they can do that to earn more money, but they're going to feel guilty about it. In particular, if the Xi is higher than the industry standard Y, then this term will be negative. And then here we take the maximum between zero and this term. So the first term will be zero, no guilt there. But the second term will be positive if voter i votes for the client. So AI is equal to zero. So in this case, the signal of the quality is actually very high. So the freelancer deserves to win. But if voter i votes for the client, then he's going to incur this utility. And this guilt this utility is also going to increase with how high the quality signal is for the freelancer's work. And then using this utility function, we basically analyze this voting game, and then we characterize the equivalent brand. And in this case, you may have seen that if everyone just always votes for the freelancer, regardless of the quality signal, that could be an equilibrium in certain cases. On the other hand, if everyone just votes for the client always unconditionally, then that could also be an equilibrium in certain cases. So we consider those as the state independent equilibria because the signal will actually not matter in that case. But those will be the type of equilibrium that we don't want to want to have. So for this voting model to work, what we need to have is a state dependent Bayer-Shinash equilibrium. And we actually proved that there always exists a state dependent Bayer-Shinash equilibrium. And in this case, each voter will decide make their decision based on the quality signal. So if their signal is higher than industry standard, then they vote for the freelancer. And otherwise, they vote for the client. And more importantly, we also characterized a condition for this state dependent equilibrium to be the unique equilibrium of the voting game, meaning that if this condition is satisfied, then the state independent equilibrium will actually not arise. So the only equilibrium that survives is going to be the good kind of equilibrium that we want to have. And what this condition requires is that the sigma is sufficiently large. So in this case, the sigma captures the degree of heterogeneity across the voters. So the takeaway from this result is that in order to make sure the voting outcome is going to be fair, the platform should actually make sure that the voter pool is sufficiently diverse. And the intuition is that if the voters are quite homogeneous with each other, then it's going to be very easy for a voter to guess what the other voters are going to vote. So the state independent equilibrium will be more likely to happen in this case. And that's not the kind of equilibrium that we want to have. But if everyone is very homo heterogeneous, then it's going to be a lot harder for the voters to actually converge to the state independent equilibrium. And we also think that this actually does not conflict with how these online platforms are actually implementing this voting mechanism, because like I mentioned previously, they actually select the voters in a random fashion. So that could naturally guarantee that there will be a decent amount of heterogeneity or diversity among the voters. But if the platforms actually select voters based on whether the voter's experience would fit this particular case, then that actually might not be a good idea, because then you would actually make the voters more homogeneous with each other and then the voting mechanism would actually could actually fall apart. Could I put another couple of questions? Yes, please. So firstly, from Jack Krimer, he says surely in practice, platforms pay the voters to provide them with an incentive to examine the case carefully. This could be modelled as a cost of information acquisition. This doesn't seem to be in the model. And could you explain why? And then there's I think a related question from David Salant, who says there does not seem to be any mechanism to induce voters to put much effort into getting good information, only the cost of regret, which could be quite small. And it's not related to the value of a good decision. Could you comment on this? Right. Yeah, those are very good questions. And that's actually one of the things that we are working on right now. Yes. So in the model that I just showed you, we didn't incorporate the effort of the voters to actually, let's say reduce the noise of the signal. And that's actually one thing we are working on right now. But I think something that could be interesting here is because we find for this equilibrium to be a good kind of equilibrium, we actually need the sigma, the degree of noise to be sufficiently large. So that means it may actually not be a good idea that everyone actually has more homogeneous signals after they review the case. So of course, we could also think deeper in terms of whether the heterogeneity among the signals is actually caused by how much effort they incur to review the case or what is their true preference. So those are actually the things that we are looking to capture in the model right now. And then thank you for the wonderful questions. Okay. All right. So I think I'm okay. Yeah. So there's another question. If right. So if sigma is extremely large, and that means, yeah, it will be very, very noisy, right? So the it's not going to be informative, right? That's true. Okay. I see that I don't have much time left. So in the interest of time, let me basically go over the main results, some of the main results when we compare the two systems. Okay. So basically, first of all, under the decentralized system, we also characterize the entire the equilibrium of the entire game. The structure is actually similar to the centralized case. So either the transaction does not happen or transaction happens. But dispute also happens or transaction happens, but dispute does not happen. Okay. And then when we compare these two systems, one main effect that we see is that the decentralized model, right? So it's going to remove the decision making bias of the platforms. Okay. So in this state dependent equilibrium that I just showed you, the decision of the voters is actually based on this industry standard, right? So everyone will basically vote according to the industry standard. That means the bias of the platform's decision making is improved, is removed. And what that also means is that the tribunal is going to set a higher quality standard for the freelancers, right? So that will actually incentivize the freelancer to actually work harder and improve the quality in equilibrium. And correspondingly, the client will actually be willing to pay a higher contract price. So the platform could actually extract more commission revenues from the contact price. However, these phenomena will actually mean different things for different freelancers, right? So overall, the change is that the tribunal is going to set a higher standard and then the freelancer will have to work harder. This may not be an issue if the freelancers have higher skill levels, right? So think the Ford to work harder, right? And then offer higher quality and then the platform can also earn more commission revenues in that case. So that is why when the freelancer skill level is sufficiently high, the decentralized case should be preferred by the platform. But if we are dealing with the lower skilled freelancers, right? So having a higher quality standard will actually make them make it even harder for these lower skilled freelancers to participate, okay? So what we find is that under the decentralized model, more of the lower skilled freelancers will actually not be able to participate, right? And even for the ones that do participate, the platform will not be able to extract much surplus from them. But if the platform uses the centralized decision-making model, right? So he can actually control the quality threshold, right? So if the freelancers have lower skill levels, then the platform can also set a lower quality threshold. So in that case, more freelancers will be willing to participate and the platform can also extract more surplus from those lower skilled freelancers, right? So that's why the centralized case will actually be better for the platform if the freelancer has a lower skill level, right? So basically, this is a single threshold result of the result that we obtained regarding which system should be preferred by the platform, okay? And a couple of takeaways is that in order to make the decentralized model work, it will be important for those platforms to make sure that the skill level of their freelancer pool is sufficiently high. Some of the things that the platforms could consider doing is that maybe they could require certifications for professional software, for example, and they could also consider providing training sessions to improve the skill levels of the freelancers. And another factor that's related here is if you look at what's happening over time, one thing that we observe right now is a lot of the professional employees are actually switching to those online labor platforms. And if this keeps happening, then over time, the overall skill level of the freelancer pool will actually keep increasing. So that means in the future, the decentralized model will be even more likely to be the optimal solution. And finally, although I didn't show you the detailed results for the social welfare analysis here, I do want to mention that because the decentralized case is going to set a higher quality standard, what we find is that the equilibrium quality level is actually closer to the socially optimal level. So that means if the platform uses the decentralized model, then its incentive will be more aligned with the social planner, okay? And finally, those gig platforms have been criticized to behave more like a monopoly, although they claim to be running a decentralized marketplace. And we also believe that decentralizing the dispute resolution can be one step to move those platforms closer to being a true sharing economy, okay? All right, so I'm basically going to stop here just to briefly summarize some of the main findings from this paper. We find that the decentralized dispute system indeed can work and then provide a fair dispute resolution, but that only happens if the tribunal is sufficiently diverse, okay? And another thing that the platforms will need to be very careful about is to make sure that the freelancers are sufficiently highly skilled, right? So only by making sure of those two things can the platform make this new decentralized model work. And in that case, the decentralized system can also align the platform's incentive with the social planner, okay? So that is my presentation. Thank you very much for listening and for all the wonderful questions. I think now I should pass it on to Andrei for the discussion part. Yes, right over to Andrei. Sure. I'll try to keep it brief so we have time for Q&A. So first thing, I think it's a super interesting paper. I mean, the topic I just want to emphasize is the topic of decentralization of platforms. I find that like one of the most interesting current topics, both in the real world, but I think it should become like a super like rich and interesting research topic. And as far as I know, I think this is I think Yao is one of the first papers to look seriously with the model at the trade-offs that are involved in decentralizing, let's say, certain governance decisions. So they're focusing on dispute resolution. But of course, you can imagine platforms decentralizing other governance decisions. And the idea of decentralization is pretty simple, right? I mean, so instead of the platform controlling and deciding everything is to essentially like delegate some of those decisions to a randomly chosen subset of its participants. Now, I think it's important to ask here. So I just want to emphasize this here because it's a theme that I'll come back to in my comments. So it's interesting to ask like, why do platforms do this? So yeah, there's this notion that there's sort of like PR notion that platforms decentralized because they're perceived to have too much control. But I think more practically decentralization also helps, at least in some circumstances, better aligned decisions with the welfare of participants. So I think it's useful to think about this because so yeah, I don't think you necessarily have that mechanism per se in the models. I'll come back to this. And by the way, my job is made a lot easier by the comments in the in the in the chat, which basically kind of what I'm getting at as well. All right. So there are two main comments that I have. And again, they're kind of been hinted by Luis and Jack in the chat that I just want to emphasize though. So the first comment I have is about the voting mechanism. So you're kind of you're assuming that the voting mechanism is this like the so the voters post the bond. And if you're on the winning side quote unquote, then you win. I think it's important to justify because it's not entirely obvious why that mechanism is ideal. I mean, it's not entirely clear that that's the ideal mechanism. In fact, I actually looked out of curiosity, because I was super curious about the examples that you had. It's kind of interesting the examples of decentralized dispute resolution that you have. There's these very obscure platforms. Honestly, I try to go on the websites for a few of them. Some of the websites don't work. Some of them don't have websites or they haven't basically they haven't tweeted since 2018. They look a little bit I don't they look a little bit sketchy, which by the way, I know is the nature of like blockchain decentralized platforms. They're very new. But I think it's interesting. So there's one that I know I can recommend. There's a very successful freelancer platform called brain trust, which you should look at. Now the interesting thing I looked closely at the dispute resolution they have, it is decentralized in the sense that they get like they get members to vote on disputes. However, this is where this is the main point about the voting mechanism I have. It's not like the workers have to post a bond. It's basically they do it out of their own free, like, you know, whatever they want to participate on the platform, it's in their best interest to make sure that you know that disputes are solved in an equitable manner. So more broadly, I do think you're right. I mean, I think it's interesting the mechanism of so, you know, you get so you post a bond and if you if you're wrong, then you lose it. If you write you basically share and the winnings, I can see an argument for that. And the strongest argument is probably like, well, it incentivizes people to take voting seriously, like to consider cases carefully, because you don't want to be wrong. There are two issues with that. Number one, if we take this seriously, so this is Louie's question, I didn't see that you have voter participation, right? So it should be incentive compatible for voters to want to participate in expected value. Now, you don't have that because in your case, like voters have, there's a financial reward, plus there's a subjective element, you know, I don't want to feel guilty about getting things wrong. But I think it's useful to think about like, there should be like, if it's truly monetary, like if it truly matters, then there should be some participation constraints. And the second thing that I wanted to say here, right, so the other thing is, in reality, I think the voters are not exactly posting bonds. I think more likely what's going to happen, especially with blockchain-enabled platforms is they will be paid or rewarded tokens. And I think this is important. In your model, it's basically lump sum payments. I think that the advantage of tokens is it basically aligns voter participant incentives with the long-term incentives of the platforms. When I think this is the sort of the mechanism that I think you want to introduce in the model. And that brings me to the second point, which I think is sort of the crux of the thing I would sort of like focus on the most to sort of improve the paper. So right now, both the platform and the jury members, if it's decentralized dispute resolution, the way you sort of get most of the action is, you know, Luisa's first question is like, well, you have this kind of exogenously given preference for not deviating from some industry standard, right? So like the platform doesn't want to get a file of industry standards and the voters, they just, they don't want to get things wrong for fear of being, for feeling guilty. I think it would be a lot nicer, a lot more realistic and a lot more convincing if you can introduce a mechanism of reputation where basically the incentive to get things right in terms of voting basically comes from the fact that it has the feedback effect on the reputation of the platform. I actually don't think it's that complicated. So I was thinking, how would I do that? We can probably maybe do something like two periods and say something like if the client feels like they're being wronged, they will be less likely to come back in the second period. And vice versa, if the freelancer is wrong in the decision, then they're less likely to come back in the second period. And this way you have this like you will have basically the incentive to get things right is going to be endogenous as opposed to this exogenous parameter. And I think it will make it a lot more realistic and a lot more appealing. Anyway, so I can go on by just in the interest of time. Let me pause here. I'm sure there's going to be a lot of interesting discussion. Honestly, I love the topic. I think this is like it's awesome. This is one of the first papers to do this. There's plenty of things that you have like in the results are very nice. You can make, you know, you can probably make some tweaks to make it more, you know, more convincing. It's like there's really, like really, really good topic to be working on.