Decentralized Trust Graph Working Group Weekly Meeting - AM/EU Time Zone - 2026/09/16

Oh, hi CHS. >> Hello. [laughter] >> How are you? >> Good. We don't seem to have a lot of folks joining us today. Uh, it's only one minute into the meeting, so not too [clears throat] concerning. Yeah. So, uh, Drummond is away today. Uh, he's, uh, at the Prohuman Conference. So, all good stuff. And uh unfortunately Martina is also unable to make it today with a client engagement. Okay. Uh let's get started. It is 3 minutes past. So the usual policy antitrust policy notice. Um uh you must uh everything must be conducted in accordance with applicable antitrust and competition law. Um and so it's important we adhere to the agenda which we'll go through shortly. Um and uh we ask that all contributors are members of diff or of uh trust over IP. Any new members? I don't think I spotted anyone in the list. Anyone like to introduce themselves? Okay, just pop this down. Okay, it fills up quickly. [laughter] Welcome everybody watching. I have to come back to this. Okay. So, no new members that I can see. Uh any general announcements or news items of interest? No. Okay. Um, so let's step through the agenda. Uh, so we're supposed to have general announcements. There are no actions from the previous meeting at this stage that I'm aware of. Um, we'll have an update from the task forces, uh, with the exception of VDS and agent names. Um, and then hopefully Daniel is here to give us >> Yes, I am. >> to give us an introduction to Tadal. Uh, or how do you pronounce it, Daniel? Is it tadada? >> Tada. I don't even know. [laughter] >> I think I think a bit like tadada, you know. >> Yeah, exactly. >> Exactly. [laughter] >> Got it. >> Okay. Brilliant. So, we'll have an introduction and that's our main special topic. Is there anything else that anyone would like to add or change in the agenda? Okay, let's get cracking. So, risk assessment and harms prevention. I'm going to hand over here to Sanka because he's been doing all the work on this and has really I I expressed it here transformed the risk assessment and harms prevention toolkit from a rusty bag of nails or a bag of rusty nails to uh like a a modular set of shiny power tools. Um so it with [clears throat] this in mind we want to uh resume the t task force meetings so that we can work through and and improve it together. Sankaan do you want to explain where you've got to and how you've been using it with respect to the specifications? Yeah, I'll be brief because uh I think we the tool still doesn't merit all the hype that we have in terms of describing it. But anyways, uh so uh the current tool is was primarily focused on extending its capabilities to do a couple of things. one is making it a bit more modular and portable so that any other working group can pick up and use the tool to do the same kinds of things. Uh then there is a bit of uh uh ability or capabilities in the tool to allow for uh checking specifications uh checking across specifications, checking across a composite of uh specifications because at the end of the day the the consumer experience of DTG has would be a composite and uh given how quickly the codebase has been progressing for the VTC CDTG portfolio. It has been useful for me to actually continuously keep on testing uh providing feedback to Glenn and Jeff uh also trying to ensure through by bothering them enough that the tool isn't misbehaving and is actually producing useful output. So I think since the uh Cypress release it has been very stable in terms of how it is being uh analyzing the codebased changes uh across the portfolio and producing reports. Uh the tool currently does not use any AI even though it is designed in such a way that if somebody wants to use it alongside their agents that they run they can absolutely do that and the agents will do a lot of fantastic things alongside with it. uh I will add to the zoom chat a link to the uh documentation that is published for the uh the tool and that would allow you anyone who's interested in idea figuring out uh what exactly it does. Uh I try and make a release every week so that any new capabilities get released and and are and available to everyone. Thank you. And uh Glenn agrees with me that more hype is needed, [laughter] but maybe a different analogy. Um any thoughts or questions for Sanka? Okay. So my propos looking at the trust over IP calendar my proposal is to put the task force meetings in directly before this this call um and we'll run for a few weeks and see where we get to and then slip back to asynchronous mode um depending on you know our agreement amongst ourselves. So, um, anyone got any preferences, strong preferences for that not to be the time? Okay, great. I'll go ahead and set that up. Uh, right, over to you, Jeff, for credentials task force. >> Thanks, Nikki. Uh bear with me. I've got a throat problem so uh I can't talk very well. Um we uh merged in the uh statement credential after discussing it yesterday during the uh credec task force call. Um there's some fallout, well not fallout, but some uh knock on things that need to be done as a result of that because they'll be uh well create another repo to host a registry of the predicates for those statement credentials. So they they'll be in a public namespace. There's we need to figure out what that namespace is going to be exactly. Drummond is talking with the LFTT people about that right now through email. So that's been merged into the spec. Um it's been we've have a few other issues and a couple PRs outstanding still a new one got added today. Um so there's a fair amount of moving parts in that. So if anyone has the time and the interest, you know, please follow along. If you have questions or suggestions, you know, or really think something needs to be changed, feel free to reach out in the issues and or even create a PR if if you feel strongly about it. But, um, we've been having a pretty good cadence now since about two weeks of trying to merge things in. Um and uh yeah, we had a we've [clears throat] also um excuse me to we've also adapted semantic versioning now thanks to Brendan in all of the spec repos that are active. Um so I think uh we we gave a little shout out to the ZKP guys today in uh in a PR that they've created in their repo about that. So hopefully we can get that coordinated across all the different repos and that will make things easier to see what what states of each document align to which the other documents and which states for that. So um thanks again Brendan for spearheading that and doing the work to get that uh in place. Um yeah, I think that's those are the main items. uh you know Brendan uh Alberto others who were on the call yesterday feel free to add anything to that or if there's any questions you know we can field them now. Yeah, nothing really to add. Um good summary and yeah we're we're trying to keep things moving. Um, we're trying to merge things in um, you know, uh, relatively quickly, so after a couple days to allow for review, but that doesn't mean that things can't be changed and addressed after they're merged in. So, um, you know, please keep an eye on everything. Uh, you know, review it when you can. uh provide your comments or issues or PRs and uh yeah, we're uh moving towards something, you know, resilient and implementable. Um so, uh please uh you know, your input's welcome. >> So, unless there's any questions, uh Nikki, I think that's it from the credentials task force. Thank you. Thanks guys. So it sounds like you guys you're going to be in good shape for the plumbers conference in terms of the status of the specification. Is that right? Um, yeah, I guess it depends on how you define good shape, but but I mean it's moving quite quickly and uh, you know, it's getting further developed and I I think we're all these new changes that we're making are good ones and it feels like it's it's coming together. You know, >> are we going to have >> I don't know if we could say uh we'll have it in its final state, whatever, however you define that, it's going to be a a growing thing no matter what. But I think it's moved quite a bit since, you know, compared to a month ago. >> I'll sorry for that. >> I just want to add one um thing for context just to mention in case folks weren't aware. So the the cred and the trust test spec are staying narrowly scoped and we've added the VTI spec that is going to be looking at cross linkage issues, how to use things together. So, um, just wanted to orient people. And then, of course, the the ZKP task force. So, it's it's a lot of moving parts, but I think we're all trying to keep them aligned and connected. >> Yeah. My understanding from Drummond was that the credec was for it to be ready for public review and uh uh before the beginning of October and then we would get as far as possible on these other specifications. But what I'm hearing is that they're being developed in tandem as we agreed some weeks ago that you know we all understand they've got to work together and and not in isolation. So they'll be you know more or less advanced um with [clears throat] the trust specification being closest. Does does that sound about right or am I misunderstanding? >> Yeah, I think you could say that. Well, in the meantime, there's been code developed alongside of it and we're learning from the code and we're realizing there's things that need to be added. So um the there will be something functional and usable by the Linux plumbers conference and the spec will be in line with what's being presented there. Uh so yeah, I guess you could say we're that would be sort of the target of an implementers draft uh at at that stage and we'd have what would be considered viable for that. Okay, any other thoughts or comments? >> Great. Okay, over to you, Glenn. >> Thank you, Nikki. So, couple of updates. Trust continue to grow. Uh if you haven't read Jeff's blog, I think he articulated one of the side effects of trust tasks on everything you see in the UX, whether it be the CLI or the web um portal for VTC's or the browser plug-in. That's all managed through the trust tasks. And uh it's very quick, responsive, and you don't really need to do much at all on the UX side apart from the design. You don't have to worry about, you know, cryptographic handling and uh a lot of those kind of low-level complexities. Trust tasks hide all that for you. Um we've extended so we had Rust and Typescript as codegenerated libraries from Trust Tasks that's now being extended to Go and Dart uh which brings in the Flutter community as well. So that's exciting. Again, just getting ready for Linux event and open-source software EU, which is the same week where we want to try and remove any friction for developers to be trying this out and interacting with a VTA or a VTI uh as they see fit. We are also I didn't put in his apologies, Nikki. Um we are making another breaking change across the VTI stack of shifting from trust spanning protocol revision 2 to revision three um which has a number of major breaking changes but brings us back into alignment with the latest TSP um kind of draft version as well as gives us postquantum computing cryptography uh by design into the messaging layer which also trust tasks has been extended to support all of the popular postquantum uh cryptography methods as a first class citizen. So there's a lot of work going on and this is based on uh we had a meeting with the Linux Foundation team and the Berkeley team uh yesterday and it's pretty clear that a design win they want to showcase is uh ZKP with PQC capability. Any thoughts or questions for Glenn? Right, let's press on. We're whipping through this agenda at a rate of not very efficient. Uh Scott or Mitch, is that KP task force talking? >> Sorry, I was on mute. Hi everyone. Uh so we had a a great call yesterday. Um the headline is that the spec is becoming real. Mitchell showed a machine checked spec of construction records. Um this concept that he's been um pushing of a effectively a cookbook in code. There's 16 proofs so far and Glenn's ADR00001 is the first composed one. Uh we confirmed that that this task force owns construction I love saying it 007 um the common control primitive. Um that's the piece uh that issue nine and four separate requirements all lean on. Uh the division of labor is settled. the credential spec writes the requirement and all and a oneline allowance on the VRC and we would own the construction. Um we agreed a working model of reviewing adjacent specs async rather than blocking them which keeps everyone moving and we have some new collaborators joining us. Um, uh, Leif with the Cyrus Foundation and the, um, the EU digital identity wallet benchmarking team will join us on the 29th to share what he's been seeing, including real performance data across major ZKP families. And I'm trying to see if a gentleman I met, an engineer from Prove Kit, which was released by World, would would be interested in joining as well. And the immediate next step is Mitchell is finalizing the spec so it renders publicly and then and then more more people can review it. >> Thank you Scott. Any thoughts or questions for Scott? Yeah, Scott Glenn. Yeah, just want to confirm when we talked to the Berkeley team on their ZKP, is that the same ZKP your task force like are you aligned on the methodology or is it Berkeley is doing something slightly different to what you're doing which is different to what Siros is doing? Just trying to get a understanding on convergence or their three different implementation models that work side by side. >> I know Mitchell's in transit right now, but let's see if um uh he if he can jump in and keep me honest, but we're we're trying to get that clarity too as to whether the Berkeley team would be um contributing directly to what we're doing or if they're working on things where it's interesting and we could borrow from it. Uh so we're we're trying to figure that out as well. And with Leif, we might need Drummond to close the loop there. But going back to the kickoff of this in April, the idea, I believe, was that Leif would be a and his group could be a constituent of the work we're doing. They're looking for the specs and the requirements we have so that they can build against it. That is the impression that I I still have from months ago. >> Okay. Uh I would just stay we when we talk to Berkeley team just because we're kind of two 3 weeks out from plumbers that's not a lot of time to implement ZKP end to end. Uh we are trying to get code from them next week and a integration point either next week or very very early the week after. Um, and we have a weekly call set up to keep everyone on both sides of that kind of honest in the deliverables so that at Plumbers we can get up on stage and actually showcase ZKP meeting the Berkeley white paper uh research paper that they're also going to publish at the event. I see. So that may mean that there's deviation in the short term which I think we can be open and transparent about but if we want to create um some awareness and attention to this and as a capability statement at plumbers event uh and again Hart the CTO for Linux Foundation he's pretty adamant it's kind of a non um non-negotiable requirement from their site. >> Scott, what do you need to work, you know, to advance this so that Glenn and the team can meet this requirement? Is there anything you need from the group that we can help with? >> I am not sure yet. I think I'll have to take this offline with Mitchell and Drummond. Um >> Okay. Okay. Let us know if there's anything obviously and that goes obviously for all the specification authors. Uh, is there anything that you're looking for that you're not getting from the group um other than review of PRs and things like that? Uh, I think generally the spirit of it is the reviews getting more eyes on it, more people hitting what we're doing to keep us honest and um aware of >> unknown unknowns to borrow from Donald Rumsfeld years ago. >> The Rumsfeld moment. Yeah. >> Yeah. Yeah. Yeah. [laughter] >> For all specs, please take time to review if you can. Hey, now this is on the topic of Rumsfeld. Is it like that Scott Bessett guy just a bad photocopy of him? You know who I'm talking about? >> Oh, like his face or how he acts? >> His face. How he acts just his whole vibe is just like, you know, deformed child of Rumsfeld to me. But okay. Uh >> I get what you're saying. I think he just I think he has a fancier haircut. But I see what you're saying. Yeah. >> Yeah. But it's sort of like Yeah. Plasticky, too. But all right. Okay, let's um Okay. Uh is Margie in the room? No. Uh Daniel, would you be able to give the DTG uh human trust experience task force reports and then you can >> smoothly move on to Tada. >> All right. >> Yeah, sure. I can quickly talk what we discussed yesterday. So um so uh one topic is always our uh standing deliverables where we want to collect best practices and also um work on a uh pipeline to faster and better in more easy create explainer videos. [clears throat] There is a new video available for you to check out. Um if you want I can provide you with the address uh shortly after the talk maybe or you can find it in the HDX work uh group uh repo. um these um the idea of these translates is uh to uh have an easy way to uh [clears throat] explain this uh complicated concept and uh I I did some more uh work on it to have better results and it's quite nice but for the moment we decided to uh stop these activities just because um I think the base that we want to explain is not necessarily very stable right now and we expect a lot of things to change and this would make this effort redundant and so please uh really be patient with us. We're working on that but the infrastructure is there. So that's uh one thing and then we had a broader discussion about the personas and the faces and worlds that describe these attributes that a persona can have. This is still an ongoing discussion. It's a very important one I guess and uh and we all like the idea of course of having these uh worlds to lead a little bit also the user to uh understand what faces are, what they could be and what the attributes could be. And we also thought we should think of um a way how to uh further enable the user uh to define these attributes. uh so that we may have an automated system that extracts the information from existing repository like a CV or anything that helps the user to create initial phases. However, this must happen of course in the right privacy context and uh and um yeah we still have to probably do a little bit more thinking about that how to uh make that a natural experience. I think these were my main uh discussion points. Please uh uh add some future points or additional points. Kie or Nikki or whoever joined the call yesterday, feel free to add them. No, I think that was a pretty good summary. Uh so again uh this is uh to support the VTI work for the plumbers conference. So uh there's some input into uh this discussion which I've put in the chat and um again you know all welcome there. Uh okay do you want to >> go on to Tadal and I'll give you the screen share. Happy to do that. Yes. Let me find the right screen to share. Yeah, that should work. So I'm very happy to have the opportunity to present to you a uh a concept that is very close to my heart and is based on on a lot of researches has been done in the last years and it it might not be the most important in terms of pressing concept to discuss now with the plumbers conference coming up but I still think it is worth to have a special topic IC about that. Um and that's because it it addresses a very uh deep problem and the deep problem that also motivated me to join this initiative here. Um this is like the missing layer uh the missing trust of disclosure and agency layer. Now we will have uh these wallets very shortly. millions of people will have these wallets, but there is still a big uh problem that is unsolved. We know already now and this is also based on on recent research of a colleague of mine here that I met in Geneva in CERN and and reminded me of the importance of that topic. Um we already know that the the user will be a little bit overwhelmed because this kind of new technology and this uh zero knowledge proofs and this uh selective disclosure uh documents are not necessarily part of the current mental model of user. And we also know from the past from information economics also that a user tends to share much more data than he thinks he should share. So if you ask people would you share this information they would all say no but in fact they have already done that multiple times. Say we face this privacy paradox all the time and from business side we face the data collection dilemma. So that they should actually know a lot about consumers but they should not know about consumers. So it's a real dilemma and the um colleague here of mine Sheila I hope she can join one of our calls in the future. They did uh some extensive research on this oversharing problem in the context of the upcoming uh release of uh wallets especially with the EU ID and also with a focus on the Swiss ID and what what all these organizations that now try to implement this technology by end of the year already notice is that this is an unsolved issue. oversharing is going to happen. And uh here uh Sheila brought up also some of these numbers of of their research. So they uh did a lot of experiments um because this is a really hard thing to observe from a scientific perspective. So you cannot have a natural experiment about that because exactly of this privacy paradox. So people would not answer the right way probably. But you can simulate that. You can um do some expert panels and so on. And they found out in certain situations that people actually really overshare also with these kind of wallets and that even an official ID has been shared for a news website by 20% and that's way too much. So that is not yet the answer. and they identified the need to have this decision layer that supports the individual uh and is not only a consent button because if we build a VTA that is only a consent button or a outsource delegated consent button or the medical consent button we we don't solve the problem here and they constructed an assistant here that was able to intervene via notching at the point of uh disclosure and then they were able to reduce this oversharing by a significant rate. And this research that uh these ETH colleagues do resonates very well with my my earlier research in this area. I I didn't look at it from a security perspective how how they did but I looked at it rather from aformational economic aspect and especially also using game theory to design models about it and the interesting thing here and I I really want to keep that more or less short today is that we can I'm really convinced that we can introduce such a layer in in a system like our VTI our trust infrastructure and that this layer can really sit within the variable trust agent because that's the ideal place to position something like that and why could it work because in game theoretical model you see or you can model these kind of situations of data sharing and that's also what I did and there is a a huge document It's about 50 pages. It's also linked in the material and I'm habitant to derive also an issue maybe from that to be a bit more specific so that not everyone needs to read the 50 pages. But the interesting part here is that we model this interaction. We can uh find out uh what can we influence which variables do we need to influence and how we can do that. And one of these variables that is quite interesting is this delta here. And the delta is the discount factor of a future payoffs from a collaboration. So in a game theory, it's pretty simple. Just look at two parties, yourself and another party, and you're relying on this party and you want you have two options whether to cooperate or to defect. And also the other party has these two options. And when uh the someone defects and the game stops, of course, you only have one iteration. But the idea is also to uh look into future values to have multiple iterations. And now you probably think uh that uh companies are not really interested in that because because also hit and run is a very popular model. But the findings here is is is quite prominent that even with the low with a very small discount factor of 1/3 say this is 0.33 um uh of future gains there is still a lot of uh uh effort or willingness to cooperate. So it doesn't take too much. You don't need to have the big value in the future. you have to discount that like a discounted cash flow but uh all these rate that you need to have so you need to be higher or the same like one/ird is not that high so even with a little bit of motivation you can convince someone to play the game correctly um now >> yes sorry yes please >> yeah what are the game what are the rules of the game precisely >> oh no yeah I will come to that a little bit later if I if I may h yeah uh so I I want to explain that even further later but um just keep with me if that's okay and we can discuss afterwards um so I just want to first um also tell you where I see this uh layer this title decision layer it could sit on top of the VTA of course the VTA sits on the infrastructure and the human is interacting with the VTA anyway so this layer could sit very well in between. And before we now go into the game discussion, um we were able to uh uh derive already clear capabilities that we need or that I think we should have a look at to implement in our uh verifiable trust infrastructure and these can be seen as buckets of functional requirements and um they originate from all these researchers. So it's a combined view uh but they go in the same direction. So we need or if possible we should make sure that our VTA is uh able to understand the situation on the game we're playing and that there is a decision support layer and also that the user is then supported in acting and that we can learn from that and that is the main benefit of course of such a system that you have a feedback loop that provides additional information and makes this information as symmetric app a little bit small in the next round. Um to be a little bit more specific and I would really love us to uh work maybe on such capabilities is um that we need to know a little bit something about the parties and I guess there is information around but that is going to be the big art how to uh find the right signals for this kind of uh understanding. But first of all, we need to know a little bit more about the request. And also in the AU and in Switzerland, you will have quite a lot of information here because if someone asks, a party asks for information, they need to be pre-registered in some registries with their need. So they already have kind of a proposal uh signed what they are going to need and this is being checked by whether the EU or Switzerland. So there's always this registry that is looking at that. But um that's not necessarily enough because our VTA needs to understand what to share because this is indeed a game and uh you can share uh nothing or and not participate or you can share a lot or you can share only a minimal amount and also my other belief is that we should look at that incrementally. So in different rounds so that we can also extend the amount of data shared and in order to do that we need to know about the request. We need also to know about the principle. Principal in this case is the uh the person that the VTA is caring about and maybe about his preferences or her preferences to be able to make this judgment. Uh then um of course I don't want to go into everything here and that's pretty much also detailed in the studies. Um but um well the the interesting part is now to uh come up with an idea for the uh user what to do or what uh strategy to follow and I will show you the the uh different kind of strategies later. So now a little word about the game we play and I'm happy to go into details here but it it's kind of a an interesting um piece of mathematics. Uh so it's not bulletproof and do not look at these variables and all these levers in equations as as constants. They are not. It's rather about the mechanism how the game works. And this is then a very good thing to model in game theory. And I want to start with what kind of game we're playing. And this is something I currently see as a a rather big mistake everyone is doing. Uh so if you have read our your US president's [laughter] recent note about uh what is needed in terms of uh guidelines or or guard rails for AI development and that we probably should not slow it down but we only need a strong person like a strong president. um and then that everyone needs to fight against a partner that does it anyway like China we are then in a complete wrong game and uh that's maybe the most important takeaway I want to give you in this uh special topic here so the temptation is very high to see this as a prisoners dilemma game that is one of the most um famous games in in in game theory where two parties act simultaneously. They can cooperate or defect. If both would cooperate, the value would be very high. But like prisoners that have to blame someone else to get a lower sentence, um they would probably not cooperate. And this is then what we call the Nash equilibrium, which is a situation you will find yourself in that is stable. It's a dominant strategy. But unfortunately it leads to a very low uh utility. So the combined utility will be very low and that's exactly unfortunately also what we're doing this political debate now M Mr. Trump. So if you have your old guardrails or strategy is a tit for that strategy and that you have to force developments before someone else does it you're trapped in this uh prisoners dilemma. And that's also the case if you look at it in data sharing. If you think that I cannot give all my private data to Google because they will do uh too much business with it and I don't get something back then it's not a very good strategy. And the opposite like the big corporations that only harvest your data and earn a fortune of it. but probably happening or happened so far is also not a very good strategy because they are all not an optimum outcome. So we need to look at this uh from another games perspective and that's rather the the theories of the trust games and the difference the rules of the games they are quite different. said um in a trust game um we have multiple iterations. We only don't only look at one iteration and we have of course uh information as symmetry and that's [clears throat] what what we always have but in trust even more and what we also have is we have some noise and we have vulnerabilities. So once data is shared you cannot take it back. So it's already gone. the damage happened in round one and you need to cater for that in the next round. So this is the the basic mechanism of of game theoretical analysis and you can then model quite easily such trust games. And the idea here is that we tried then with all the intervention strategies um the mechanisms that we identify in the capabilities that that we need. we can model quite a nice layer that uh helps us to find a very good dominant strategy for both parties in a long term in multiple rounds. But this also means that we have to identify things we need to influence and these are these nasty variables here a G or a Q. So cannot go into all the details here but it the mechanism is very simple. So what we need to do is we need to be able to influence a little bit this this um um variables in the equation to make cooperation work for all parties. And that also means that we need to be able to cut the gain from extraction with minimization for example. So to cut damages we need to be able to influence what we understand. For example, here the the alpha is a false positive rate for trust assessment. So we need to have a little bit more information to that correctly. But if we have all these uh or understand a little bit these mechanisms, these variables of the equation and if we have uh some capabilities that influence the single ones, then we can try to find a spot uh where everyone cooperates and combined number is higher than if utility is higher than only for one party. And last but not least also with this kind of modeling we can then also recommend concrete strategies here to u to recommend. So if we start this game or data sharing we can only share what is required or we can try to reduce it if we don't believe in the honesty of the other party. And the very important thing here is and this is also seen here in in the equations we need to have the ability to recover. So if for example alpha is very high means we have a lot of noise in the game then we should not stop the game we should not exit the game. It's much more beneficial if we try to recover maybe by reducing the amount of data shares but then expand later so that we size the entire horizon of the game and and are able to u get the full benefit. that this is very theoretical of course but there is always a truth in it and I really think if you look at these models there are not too many factors that we need to care about but I also think that we should care about these uh uh these variables to make a very successful um solution that not only is cryptographic technically correct but actually supports the user in his biggest dilemma that he faces meaning that he doesn't really understand what it is about and the human beings probably need a little bit of help or assistant a rational assistant here that helps him with this challenge. So that's it. A long talk. Happy to for you your questions. >> Yeah. Well, I'm going to I I recognize these four variables from my own work. I I work in game theory a lot and >> cool future agent society. So uh this is very uh this is a good um potential framework for developing you know some experimental simulations and how agents would uh potentially behave in those simulations. So I do have some questions still about the details. So what you're saying here is that the steps in the game are at each step both parties choose to disclose. So most both take a disclosure action. Then they look then they analyze what they've received and then choose to make another disclosure action. However, there's no other form of communication going on between the parties. In other words, they're not, for example, negotiating what they're going to disclose to each other in the next disclosure step. They're just taking uh self assessments and and choosing what to disclose based upon the previous type of data that they received. Is that correct? >> That is it is that is correct. Yes. Yeah. Uh and of course this is probably uh one of the limitations here. Yeah. Uh of these kind of games. But um I I would argue from the point of view that today we know information it's way riskier or there's even less information as as when you try that. And what we already also found out in in some tests here is that only a uh a proof or a I say a big uh statement of why I need this and that amount of data can already influence the usage. So a big statement looks very convincing and we have a bias again. So the user is not acting rationally. So whatever we can use to help in this non-negotiable but instant first moves would be a benefit. So that's what I would argue here. But you're of course right. I mean there could be a lot of research done in in finding more information in this noisy environments. Yeah. >> Right. Yeah. Hope that helps. [laughter] >> Makes it a little more efficient potentially. Okay. >> Yeah. Yeah. Sure. Yeah. Brandon, you also have a question, please. >> Yes. Uh, thank you. No, this is very interesting. Um, one of the things that I've been thinking about, uh, with relation to our work here in this working group, and I just wonder if you've, um, thought about it in relation to your work is the the ILE E, uh, 7012 standard about uh, sometimes called my terms um, you know, machine readable terms. Um, you know, particularly in regard to what should be disclosed and to whom. So, um, I just wanted to mention that in case you haven't looked into it. Um uh I'll put a note in the in the chat here. >> Uh but it's uh it's not about the strategy so much as just you know the mechanism um uh that we might be using of saying okay what what should you know what am I allowing to be disclosed and to whom under what terms etc. So um just wanted to see if you had any thoughts on it and and you know make you aware of it if you weren't. No, no, no. I that's true. I didn't look too deep into it, but I'm aware of it of the existence of that. And that's of course something that goes exactly into that direction and informs especially the first uh layer of the understanding also with the preferences. So yeah, that's definitely an important part. And my my only question here is what can we also bring in or do differently now that we have the opportunity to shape something new? >> But thanks >> Daniel. This is Reza from Quai. Um on Monday we launched the VPK working group. So it's a sister working group to decentralized trust. Um the mission is to secure our data before we send it to a third-party AI service. Uh for now we're focusing on just um confidential vector search using partial homamorphic encryption. And um so that will make sure that we're not inadvertently dulging our whole corpus of intellectual property when we send a context document. um or some repository up to uh a search. We can still search by um uh natural language on a third party hosted system um and they can't see the data. >> Yeah. >> Okay. >> You're very interesting. >> Uh can you just uh repeat that res? I'm trying to record it. I've got some uh >> working on some dot dot dot vector search and homorphic encryption to enable discovery. Confidential discovery is that >> confidential vector search. So >> Lord bless us. What's that? >> What's that? Okay. So most of us when we engage with um uh chat GPT or claude we upload a document and then we ask questions about the document and that class of application is called rag retrieval augmented generation. >> It it's asking questions of a document. Now, as as Daniel said, we are inadvertently dulging the entire document in clear text, not just the parts that are relevant to the question we're asking. And they own it forever. Um, in the terms and conditions of of Sam Alman, he says he owns the data. He's free to resell it. He's free to train his model on it. He's free to mine it for advertising opportunities. So, um, and then Dario says, "I'll never do that. Don't worry. Here's my pinky promise." But hey, once companies go public and they owned by private equity, those terms and conditions, which are changeable in their terms and conditions, will change. So, how do we protect ourselves? uh whether we're an individual or whether we're a corporation that doesn't want to inadvertently uh donate their intellectual property to um to these services. Um so we've developed a mechanism and we've developed a mechanism where you can scramble the vectors. So you vectorize locally, you scramble the vectors and now the vectors are stored remotely. The vectors are stillable. in their in their encrypted form except the host now never sees the data. Um and then the chat completion we haven't secured but um we're working on that. So eventually we'll have in this working group we'll have um ways of securing the entire >> now and when you say check you're just talking basic inference [snorts] doing basic chat >> uh as opposed to >> that's what you mean by chat completion I'm just identifying the >> yeah chat completion yeah so in in rag there are two phases there's the retrieval phase where chunks of docu your documents ments are extracted based on a cosign similarity or vector search. Um, and your document might have, you know, you might have 10 terabytes of documents that is deeply sensitive. Let's say you're a um a defense attorney and you've got all your client records um or you're a doctor and you've got all your patient records. Way too much for you to search locally um using natural language. So, you upload that. But we we now have the ability for you to upload that in a scrambled form. Um such that it's still searchable in its scrambled form. Then um whether you run the chat completion the inference part locally which will be then you've got it completely secure or remotely on another service that's up to you. Um but uh that that's the the mission of the VPK working group is to standardize the protocol first and then have a call for proposals for algorithms that can conform to the protocol. Okay, that's enough me. [laughter] >> That's really interesting. Thank you for that. So we've got we're two minutes off uh the hour. Uh Glenn, you noted in the chat uh trust ceremony at the VTC level work with this framework or be compatible with the structure. Do you want to speak about that? >> Sure. Uh first of all, great work, Daniel. I really enjoyed it. Uh >> complex subject and making it a little entertaining, not an easy feat. But in the VTC, one of the early aspects was the concept of a ceremony which was just a coordinated set of tasks to do something complex and we kind of broke them up as you know there's something that triggers you and using your language which I quite like you know there's something that triggers the game then you've got to gather what are the requirements that are the input to the game and then there was you know kind of a verification What are the facts then? And then an evaluation, then a verdict, and then there's >> some form of side effects that come out of that process. Do you see that >> working alongside Tadal? Um, is Tadal kind of replacing that? Should we be thinking differently about ceremony? just trying to figure out, you know, how how Tadal could fit into what we're doing on the community. And I know you talked about the VTA, but actually it's the VTC in many ways that also has the strong need for this um >> well these are good question and didn't uh integrate it in the current landscape yet but I what you tell me is is quite similar to how I imagine that. Yeah. >> Okay. Okay. So, I'm quite sure that we can make that fit or or that some of these capabilities if not yet present we should think about how we could integrate it. Yeah. >> Yeah. Okay. Cool. So, you're saying this is uh it could work with and potentially align. Okay. Good. Good. >> Sure. Yeah. >> Thank you. >> Thank you very much, Daniel. Uh and apologies to those who attended the HTX task force meeting yesterday. a little bit of a a sequel there. Um so we just wanted to introduce that and then after the plumbers conference we can perhaps come back together and think about um what this could mean for uh decentralized trust graph um or if it's related to other aspects of what goes on in trust over IP. Thank you all very much for your time and see you guys next week. Take care. Bye. >> Thank you. >> Thanks, Nikki.