DSS 2026 — A Bug's Life: Why We Ran Our Own Audit Competition Nikita Belenkov Source video: https://www.youtube.com/watch?v=lkbfSzglqUQ Automatic captions — not human-reviewed. Complete spoken-audio coverage is unverified. Source language tag: en (English). Source export SHA256: f1e445104ea22b3425a3abc52c55c16eb0a8828d0302bf78dba0ed8ed348a826 [0:01] Hi, good morning. I'm Nikita. I am I [0:04] work at Anza. And yeah, we're going to [0:06] talk about some bug bounty stuff, [0:09] mainly a bug bounty competition that we [0:10] ran for AlpenGlow, which is the new [0:13] consensus protocol that recently went [0:15] live on Solana, which is very exciting. [0:18] Yeah, let's kick right off. Yeah, so I'm [0:21] Nikita. I do mainly security, anything [0:24] and anything related to security, and [0:26] this talk was written with me by Colin, [0:29] who's another security engineer. [0:30] And yeah, at Anza, we engineer Solana's [0:33] core protocol. So we work on Agave [0:35] mainly, which is the main [0:39] basically the main client along with [0:41] Firedancer. [0:42] And we worked on AlpenGlow, which is the [0:44] new consensus protocol. We work on [0:45] on-chain programs and a bunch of Solana [0:48] SDK stuff like Pinocchio, Solana Care, [0:50] Kudama. [0:52] Yeah, I'll try to answer basically the [0:53] six questions in whatever 30 minutes I [0:56] have. [0:58] Looking at what is AlpenGlow, why we did [0:59] it in house, how did we set it up? [1:03] Slightly controversial having a burn [1:04] wall or a paywall for submissions. [1:08] Agentic triaging of the bugs. What did [1:10] the competition show and like what we [1:13] ended up learning from it and taking [1:14] from our normal bug bounty program to [1:17] not completely drown in submissions. [1:21] Yeah, so what is AlpenGlow? AlpenGlow is [1:23] the new consensus protocol. It's built [1:25] on BLS [1:27] certificates instead of normal votes. It [1:30] replaced Tower BFT along with proof of [1:32] history. [1:33] It went live [1:35] on 21st of October. [1:37] We went through migration. [1:39] And yeah, so the big change was that it [1:41] has a much, much, much lower finality. [1:44] The votes are not sent on chain. They [1:47] are aggregated as BLS certificates and [1:49] then that allows you to do it much [1:51] quicker. As I said, there's no proof of [1:52] history and we also have a higher [1:54] adversarial [1:57] basically adversarial percentage that [1:58] you can sustain. Uh in Tower BFT you can [2:01] sustain 33% adversarial. Here we can do [2:03] 20 + 20, which is 20% adversarial and [2:06] 20% offline stake, which is much much [2:09] nicer. It's been under development for a [2:12] while, and this is kind of talking about [2:14] the final stages of how we got it over [2:18] basically over the end and [2:20] uh yeah. [2:22] Uh this is a [2:24] [sighs] uh [2:25] vibe coded [2:27] a vision of how the migration looked [2:29] like from our Grafana tools. [2:31] [laughter and gasps] [2:32] So we went from on the 21st of October [2:34] the finality drop from about 10 seconds [2:37] down to 0.4 seconds, which was very [2:40] exciting at the same time, but also very [2:42] uneventful, which is kind of what you [2:43] want for a migration. [2:45] Um [2:46] yeah, we got about 25% faster finality, [2:49] which was very very cool. [2:51] Um why competition for this? Why was [2:54] important? I mean this is quite a It was [2:56] This was the biggest change that we've [2:58] had on Solana on par with basically the [3:00] Ethereum merge. [3:02] So we needed to really get [3:03] get this stuff right, and we did a a lot [3:07] of audits. We worked with OtterSec, [3:08] Neodyme, Asymmetric. We did a lot of [3:10] fuzzing. I think Arnau will talk about [3:13] some of the fuzzing later. [3:14] Um [3:16] and uh we did We had a special cluster [3:18] doing public migrations. We did nine [3:20] migration rehearsals, [3:22] and uh yeah, and then the final step was [3:24] the The way our bug bounty operates is [3:26] that the master code is in scope. So [3:28] when Otter and Glow wasn't development [3:30] it was out of scope, so we needed to [3:31] bring it back in. And we were like, [3:33] okay, what would be a really [3:35] Do we just put it [3:37] straight into scope and then see what [3:38] happens or this kind of competition [3:40] style actually makes a lot of sense? [3:42] And then we started looking into what [3:44] are the other options and what we could [3:45] do. There is Immunefi, there's Sherlock, [3:47] there's a lot of cantinas, a lot of [3:49] different providers. We've been always [3:51] been running our own bug bounty and [3:52] GHSAs. We've been self-hosting it, [3:54] self-triaging it. So running it in-house [3:56] was already an somewhat obvious choice. [3:59] Uh a competition does come with its a [4:01] bit more scale and more the promotion [4:04] side, which that's what we lacked. But [4:07] we know, we went, you know what, let's [4:08] let's try it. What do we got to lose? [4:10] The big ones being we define our own [4:12] rules, we control the data, which is [4:14] having the biggest one, trusting an [4:16] external third party with quite [4:18] sensitive findings as a [4:20] becomes a an interesting thing uh which [4:23] we ideally wanted to avoid. [4:25] And then I mean saying that inherently [4:27] running it on Gnosis says you trust [4:29] GitHub for it. So it's it depends. [4:32] Um [4:33] yeah, it's triaged by the people who [4:35] know the code very well and we keep the [4:37] intake. Um yeah, this was the actual [4:39] competition was 50,000 SOL. That was the [4:42] the full price pool. It was a 12-month [4:44] locked SOL. The competition ran for two [4:46] weeks. It required POCs. Uh basically [4:49] the scope was quite [4:51] different. It required it was basically [4:54] anything and anything related to Alpine [4:55] Global in scope. [4:57] Because it's it's like a diff migration [4:59] where it's there is the whole validator [5:02] and the consensus only part of it. The [5:04] most interesting bugs are were either in [5:06] migration itself just because it's such [5:07] a fragile piece of technology. [5:10] And um [5:12] it was much less fragile at the time of [5:13] migration, but still it's there's just [5:15] so many moving parts and so much uh [5:19] so much you had to like plan for. But it [5:22] was basically anything that touched uh [5:24] touched Alpine Global. And that worked [5:26] pretty well. [5:27] Um the this one was quite interesting. I [5:30] don't think I've seen much of this being [5:32] done. This is basically what happens if [5:33] you take the bug bounty program and you [5:36] put it to a competition. Instead of [5:37] having a frozen commit, you end up with [5:39] a moving master. So [5:42] at the time of your submission of your [5:45] entry, you evaluate if that bug is live [5:48] or not. So there were a lot of cases [5:50] where people found the bugs, but at the [5:52] time they came to submit, they were [5:53] actually not live anymore or we already [5:55] publicly disclosed them or they already [5:57] got fixed. [5:59] It is a bit tough on the researcher. On [6:02] the other hand, we also have the ability [6:04] to get our fixes reviewed as well [6:06] because we had a case where some of the [6:07] fixes one of the fixes wasn't [6:09] sufficient. So this was [6:11] exactly kind of what we wanted. [6:13] On the downside that some people had to [6:16] restart a lot of work [6:18] which is um [6:20] just a decision we had to make. [6:22] And to make people's life easier we had [6:24] a live tracker on on GitHub of the [6:28] issues that we are aware about. Though [6:30] we aware so that people don't double [6:32] burn or double submit and just end up [6:35] submitting and we're like oh we know [6:36] about this is getting fixed. [6:39] Uh yeah this is pretty standards. It was [6:41] um [6:42] depending on the severity the pool has [6:44] more unlock. The difference was that we [6:46] don't share the whole pool. The [6:49] vulnerabilities paid out at an [6:51] evaluation. So for example a DOS I think [6:53] was 300 until 500 or 800. [6:58] And then the the pool acts as a upper [7:00] cap versus the whole pool being shared. [7:03] So this is [7:04] it's it's it's like a Frankenstein's [7:05] monster between competition and bug [7:08] bounty basically. [laughter] [7:11] Uh but a very nice one. [7:13] Uh this one was also quite interesting. [7:15] So we did not allow duplicates but we [7:19] also did not pay the first submission. [7:22] So we paid the first submission who [7:24] proved the highest severity. [7:26] So we had [7:27] uh not this exact case but we had a case [7:29] where [7:30] the fourth submission actually won [7:32] because the initial baseline was a DOS [7:35] but they were actually able to prove a [7:37] highest severity. [7:38] The idea for this was that [7:40] we didn't want to people to spray a [7:42] bunch of findings across of across like [7:46] surfaces that might contain bugs and [7:49] hope that they hit something [7:51] while [7:52] we were encouraging more [7:54] much more developed attack scenarios [7:57] that may be chain multiple bugs. [7:59] Uh and we had one case where this [8:02] actually worked. So I'm pretty bullish [8:05] on this approach. [8:07] Um [8:09] yeah I mean, this is the has been the [8:11] bane of our existence is um, [8:13] with the with the new models or with the [8:15] frontier models, not even with the [8:16] frontier, with any LLMs, you basically [8:19] can pay 10 bucks and spray 50 findings. [8:23] Uh, cuz the cost to write a report is [8:25] basically near zero. I mean, I put [8:27] asterisks on both of them. Cost to write [8:30] a bad report or a not necessarily a [8:32] well-thought report is near zero. [8:34] They're still like I mean, most people [8:36] in this room are very good researchers [8:38] and people really put effort into it. [8:41] Um, is this a pointer? Yeah. [8:44] Which means that the triage cost doesn't [8:46] really change unless you start doing [8:48] agentic triage, but then still the cost [8:50] is of triage is so so much higher than [8:53] cost of submission. So you have to you [8:54] basically have to come up with ways. How [8:56] do you make sure the submissions you [8:57] receive are actually valid and valid [9:01] [clears throat] [9:02] like [9:03] duplicates are good, uh novel [9:05] submissions are good. It's just like [9:06] things are actually reachable and [9:07] actually have impact. [9:09] And uh you basically end up with this [9:12] options where the most basic one being [9:15] the the working POC, which is basically [9:18] all the submissions had. [9:19] The problem like it it stops speculative [9:22] claims uh and it does take time to [9:25] build, but with LLMs it's pretty [9:26] straightforward. Being like [9:28] 2FA on GitHub and minimum account age [9:31] does help a bit. So we had this three we [9:34] had. This one [9:36] we we ended up I'll talk a bit more, but [9:39] yeah, we ended up requiring a [9:41] half-a-soul burn versus a refundable [9:43] deposit. [9:45] And it helps with some like low-effort [9:47] filings. [9:48] This is I think the most I'm pretty [9:51] bullish on this where I would say it's [9:53] like a step beyond proof of concept [9:55] where it's a proof of impact where you [9:59] have some kind of infrastructure where [10:01] you could provably [10:03] show that your POC actually shows [10:06] impact. [10:07] Because the problem with the POC is POC [10:09] is only as good as what it shows. You [10:11] ask like you ask an LLM, it can write to [10:14] you any POC you want. It can prove [10:16] whatever. [10:17] It's just like you start [10:19] mocking different data structures. You [10:21] start mocking different areas of the [10:23] validator. [10:25] And like, oh yes, this is you have I [10:27] don't know, network exhaustion. Oh yeah, [10:29] but you actually don't have any of the [10:31] data center filtering or you don't have [10:34] any of the other protections that we [10:36] have. You're like, okay, yeah, if you [10:38] take that into account, it doesn't work. [10:39] [gasps] [10:40] I think this is possible to do. The [10:42] problem is you kind of need specialized [10:44] tooling of how do you define [10:46] a correct exploit? How do you define [10:48] this impact? I think you can do this for [10:50] really big programs like a I think it is [10:53] possible. It's just [10:55] it becomes a bit [10:57] it's tough. Like it's not an easy [10:58] definition to do and also you have to [11:00] cover so so many cases from like runtime [11:02] issues to networking to like [11:05] um [11:07] yeah, to like consensus [11:08] like mapping. It's it's a bit tough, but [11:10] I think this is what we really need to [11:12] get to. [11:13] Um yeah, the next one is like doing KYC [11:16] upfront. I don't think this is a good [11:17] solution. [11:18] It's kind of comes together with a [11:20] refundable deposit. If you want to do [11:21] the deposit, [11:23] as far as I understand, you basically [11:24] have to do KYC at most points. [11:27] This really stops people submitting. I [11:28] think we should only do KYC at payout, [11:31] but that's like personal opinion. The [11:33] next ones are like invite only. Yeah, I [11:36] mean that basically becomes a pseudo [11:37] audit, pseudo like you kind of want new [11:41] people to look into this, get new pairs [11:42] of eyes. And last one is doing like [11:44] penalties and bans and stops repeat [11:47] offenders, but like drawing the line of [11:49] where the penalties and bans go and then [11:51] and then you wake up to like a lot of [11:53] emails people asking to get unbanned and [11:56] like it's it becomes it's a lot of [11:58] management. I 'm not sure it really [11:59] leads to very good results. [12:02] Like [12:03] you probably should ban like really [12:04] really misbehaving people, but like it's [12:06] it's tough. [12:08] Um [12:09] yeah, but I mean fundamentally like [12:10] drawing back to crypto, you either do [12:12] proof of work or you do proof of stake. [12:15] And there's many more, but like in our [12:16] context where proof of work becomes have [12:18] you actually done the work? Like can you [12:20] show the vulnerability is real? And I [12:22] think I mean that is the solution. And [12:25] we want like doing some kind of this [12:27] proof of impact showing that this stuff [12:29] is real in a maybe it's not necessarily [12:32] ZK proof, but like some kind of setup [12:35] where it's hard to dispute versus [12:38] requiring a proof of concept. Everyone [12:40] writes a proof of concept in a very [12:42] different way and it takes a while to [12:44] triage. [12:45] So maybe some like unified tooling for [12:47] proof of work could be the future [12:48] solution. Proof of stake is what we did. [12:51] Um [12:53] it's the best solution we have at the [12:55] time. [12:57] I'm not [12:59] I think eventually [13:01] the best way is you don't charge people, [13:03] but that is just what like that's the [13:07] best solution currently we have. And it [13:10] only works if it hurts more than the [13:12] spam pays. And it fundamentally the [13:14] amount of payouts we do [13:16] the [13:18] half a soul is not a huge huge [13:20] commitment. [13:21] Uh but yeah the that's actually how it [13:23] worked. [13:25] This we had a I I coded like a [13:27] submission portal [13:28] when up and going went live uh when the [13:31] before the competition. [13:32] It required a burn of half a soul, so we [13:35] didn't do custodial. So we didn't care [13:37] about taking the actual payment for it. [13:39] Like the the only purpose of this was to [13:42] add a layer where someone thinks twice [13:45] if this [13:46] think is valid or they will write up a [13:48] better report just because there is some [13:50] stake and money at stake. That also [13:52] means you don't have to deal with refund [13:53] and disputes. [13:54] And um [13:56] yeah it's just like and it also [13:58] technically deflationary, which is nice. [14:01] [laughter] [14:02] Uh and then you basically submit the [14:04] transaction hash, we verify it and then [14:07] it opens a GitHub GHSA for you. And this [14:10] actually worked really well. [14:12] Uh [14:13] this is what we actually ended up doing [14:14] for the triage. So the the green, the [14:16] green ones are agentic, and then this [14:19] one is human. [14:20] It was [14:21] This was very much a proof of concept [14:23] trial approach where [14:26] uh we used a tool called Tracecat. [14:28] Tracecat is like a agentic SO like SOAR [14:33] tool that allows you It's mainly made [14:35] for like web two [14:38] triaging of large-scale alerts, but [14:42] it actually worked pretty well. It [14:43] allowed us somewhere to host the agents, [14:45] do this um in a in a nice sustainable [14:48] way. We mainly used GPT 5.6 which which [14:52] was pretty good. Yeah, it basically [14:53] pulls in the pulls in the the claim, [14:56] checks the severity, checks the format, [14:58] checks uh some some basic checks that [15:00] it's like [15:01] sane as a submission. Then we do some [15:04] evaluation. We run a um [15:07] no internet sandbox checking the POC, [15:10] matching the claim against uh Alpine [15:13] Glow Agave Master at the time because we [15:15] also wanted to check if these issues [15:17] have already been fixed. [15:19] Uh the next one was This is like where [15:22] the crème de la crème of this triage [15:23] worked was [15:25] an agent that was like the whole job was [15:27] to convince that this is invalid. Is it [15:30] reachable? Is it breaching the 20 plus [15:32] 20 model? Is it intended? Is it like [15:35] Is it just made up? And it did it did a [15:37] very good job. And then you kind of [15:39] draft it, you check duplicates, and you [15:42] rank by similarity, and then you [15:43] basically hand it off to a person to be [15:44] like, does this make sense? [15:46] Final severity payouts, rejection, [15:48] whatever. We were very specific that we [15:51] did not want [15:53] to auto close issues or auto [15:57] like decide things because this is still [15:59] like a supervised approach [16:01] um just because [16:03] this was our first attempt, but it [16:05] worked very well. [16:06] Um [16:07] yeah, this is a timeline like lifeline [16:10] for of a of a of a report. It went from [16:14] like the bug line landing, then gets [16:16] pulled from GitHub, case opened, it like [16:19] agentic starts, pulls in POC, pulls in [16:23] um [16:24] pulls in source code. So this one [16:26] uh [16:27] then [16:28] does like more deeper analysis, does the [16:30] disproving. In the case of this bug, it [16:33] came in as a critical, it went down to [16:35] as a low. [16:37] And this was actually the first of nine [16:39] duplicates, I think. [16:41] And then the human kind of then the next [16:43] ones got bundled up to this one and um [16:48] in the end it was information. [16:50] It went through nearly all the stages. [16:52] Um yeah, this is the the some fun stats. [16:56] It we had 338 submissions over two [16:59] weeks. 67% the agent and human agreed, [17:03] which is pretty good. [17:05] It was nine minutes to an actual agentic [17:07] triage. It took humans about 14 hours, [17:10] which is still pretty good. [17:12] Um [17:13] yeah, it worked for about 90% of the [17:15] real bugs. [17:16] Uh it saved us a lot of time. It [17:19] rejected uh all 21 reports that [17:22] breached the security model. It was very [17:24] good at working out which ones did not [17:26] follow the 20 plus 20. That was very [17:28] good. [17:29] Uh it wasn't so good at [17:31] it was like valid bug, but the impact [17:33] calculations it was not get good at, but [17:36] this is very easy to like this is just [17:38] future solving we can do. This is the [17:40] competition itself. So we had 115 [17:42] researchers, 27 findings, valid [17:45] findings, and we paid out 27795 [17:48] SOL. [17:49] Uh in terms of like timeline, I'll move [17:51] to the to the other side. Um like nearly [17:55] within 10 minutes we got six valid [17:56] submissions. Within the first hour we [17:58] got 12 out of the 27. So nearly like I [18:01] think we kind of exhausted the search [18:03] space within like a few days. So these [18:05] little [18:06] like dashes are the winning reports. [18:09] And then in the in the whole second week [18:11] we only got 15 submissions, [18:13] uh which was pretty interesting. Um [18:21] and there we go. Uh these uh timings [18:23] between different reports. So [18:26] we kind of [18:27] we didn't [18:28] we didn't expect that we will get so so [18:30] many duplicates within such a close time [18:32] frame where we had a few reports within [18:35] seconds of each other [18:37] at the time of the opening. And about 58 [18:40] of the 85 submissions we had were [18:43] duplicates. [18:45] Um [18:46] and nearly all of the duplicates were [18:48] within 10 minutes of the first one. So [18:50] it was very much like race to the [18:51] bottom. [18:52] And this is something we're going to [18:53] reconsider when we do it next time [18:55] because it becomes a bit unfair where [18:57] like if you are the submission is within [18:59] 5 seconds [19:01] it's like it's this is not a skill-based [19:04] competition. It's just like how quickly [19:06] can you vibe code a submission thing [19:08] which is not ideal. [19:10] Um [19:12] the other question is would we do it [19:13] again? Yes. We would keep the portal and [19:17] we actually ended up adopting the burn [19:18] for our standing bounty. [19:21] Oh [19:22] one sec. [19:23] Um the next time I think spending more [19:26] time on this like [19:28] basically feeding back the triage [19:29] findings and the human verdicts back [19:31] into our like agentic triage I think is [19:33] the solution. It worked extremely well [19:36] and that is [19:37] that is the way we will have to deal [19:39] with scale. [19:40] I think this advanced POC creation [19:43] I'm pretty bullish on it. It requires [19:45] defining it very well, but I think [19:47] especially for consensus bugs [19:49] I think there's a way we can define a [19:52] nice system which [19:54] not necessarily forces the researcher. [19:56] It gives the researcher tools to [19:58] actually validate their findings in a [19:59] very nice way versus them trying to like [20:02] reverse engineer how to do POCs and [20:05] which things that actually supposed to [20:06] plug in. [20:08] And [20:09] that would also allow us to [20:11] quickly validate these things and [20:13] potentially [20:15] do this whole process quicker and save [20:17] people burns for issues that are not [20:19] valid. [20:20] And the duplicate policy [20:23] is very much similar one. I think we [20:25] probably would allow duplicates next [20:27] round just because it becomes a bit [20:29] I didn't realize how much of a race to [20:31] the bottom it will be. So that's just a [20:33] learning and I think better [20:34] communication with researchers. [20:37] We did not do the best of the job [20:40] communicating to people partially [20:41] because GitHub does not have an API to [20:44] write comments and it's just a nightmare [20:46] to manage in notifications. [20:48] And but there's a lot of easy tooling [20:51] you can do and these are pretty easy [20:54] things we can improve on. [20:57] Yeah, so we ended up migrating from [21:01] to our main bug bounty program that is [21:03] now also requiring half a soul burn [21:06] but we also have a white list for [21:08] because our [21:10] researchers like at Asymmetric at [21:12] Neodyme [21:14] they submit their findings also through [21:16] the bounty program. [21:18] They ended up you end up like not [21:20] requiring it and I think there is a [21:22] potential where you expand it [21:25] to potentially new participants you [21:26] require burn for a few for not new you [21:29] might not require burn but it's it's [21:31] it's kind of experimentation stage. [21:35] Yeah, this is actually some stats from [21:38] we for about a month that's this thing [21:39] has been running. [21:43] This so the first two are [21:47] the breakdown of the types of [21:49] submissions. So the first one is the the [21:51] January to [21:54] 1st of January basically the year coming [21:57] to 2nd October when we turned it off. We [21:59] had about 50% were invalid findings, 31 [22:02] were duplicates and 19% were novel. [22:04] With the burn wall [22:07] 20 we went from 50% invalid to 22% [22:09] invalid which is an excellent reduction [22:13] and then we went from 31 invalid to 52% [22:16] duplicates. So the actual like amount of [22:18] reports didn't go down as much as I [22:20] thought. [22:21] We went down only like 37% submissions [22:24] but I think it'll actually go up because [22:26] this is only a month of data. [22:28] The the most biting one, I think, the [22:30] 52% unique researchers reduction. [22:34] It does limit it to people who can [22:36] afford this burn, and most commonly it [22:38] also limits to um [22:41] researchers who have been paid out [22:43] before in this program. [22:44] I think that's the thing. But the [22:46] interesting thing that the actual novel [22:48] findings, there was basically no change. [22:50] I think it was something like 10%. [22:52] So that means even though we got so 37% [22:55] less findings and 50% less [22:57] participating, the actual amount of new [22:59] issues found hasn't really changed. [23:02] So it kind of implies that this 50% were [23:06] basically finding duplicates or invalid [23:08] findings. [23:09] Uh and we now get about 40% reports per [23:13] researcher, which kind of [23:15] you start drawing parallels that they're [23:16] spending more time looking at the [23:18] codebase and validating these findings [23:19] and finding new things, which is [23:21] kind of what you want. And really the if [23:25] you look at if you put it in the context [23:26] of the the payouts, our largest like our [23:30] if you find a loss of funds in Solana, [23:33] uh it's about starts at about 6.5000 SOL [23:36] and goes up to [23:38] 25,000 SOL, I think. So these are quite [23:40] large payouts. So even without average [23:43] payout, half a SOL is like 0.2% of what [23:47] you receive. [23:48] Um these are some interesting These are [23:51] like our weekly submission rates [23:53] for the last year, and I tried to map [23:55] them to like different models coming [23:57] out. [23:59] Um there is really no major parallel, I [24:02] think. [24:02] [laughter] [24:03] All the models come out so quickly. I [24:05] mean the summer has been very, very [24:07] busy, and I think that's where the [24:09] really interesting models came out. [24:11] And then October at the end there is [24:13] when [24:13] we turned on the paywall. [24:15] And it's definitely been going steadily [24:18] up. And that's just kind of how working [24:23] out how do you deal with this and is a I [24:25] think a problem that everyone has. And [24:27] then if we extend it to like 2025 when [24:32] we didn't have these [24:34] models, [24:35] well, they weren't as good. [24:37] So we got six times more submissions [24:41] than in 2025 than we did in 2026 until [24:44] the paywall. [24:45] Um we got we now have three times more [24:48] researchers, nearly all new. [24:50] And this [clears throat] one is fun. We [24:51] have 10 times more EM dashes per [24:54] thousand words. [24:55] Like this line is EM dashes per thousand [24:58] words. And then very interestingly in [25:01] January it skyrockets. [25:04] [laughter] And then also six times more [25:06] markdown tables. [25:08] Um [25:10] the the thing the line going down I [25:12] assume is more the models are using less [25:14] EM dashes than [25:17] not model usage, but this was an [25:19] interesting But yeah, it basically [25:21] exploded. [25:22] Um but yeah, that's that's basically it. [25:25] Thank you very much. [25:26] Yeah, basically huge shout out to also [25:28] to the Consensus Team Foundation and [25:30] everyone who did the triage for [25:32] competition. Without them this would not [25:34] be possible. And yeah, I'm I'm Belenkov [25:36] on Twitter and Telegram and our [25:39] submission. Yeah, thank you very much [25:40] guys for listening. [25:41] >> [applause]