"Dragos is a top-European player, who used to have GoR close to 2700 at some point. He can accidentally lose to a 4-dan (it can happen in approximatelly 10-15% cases) of course, but definitely not in this way. Maybe it's not so evident from the game, but it was a complete domination of white stones on the board and I don't know anyone in Europe, who would be able to smash Dragos so badly like this (not even EGF pros)."


Here just to mention that Dragos was "smashed so badly" (or even worse imho) in the last round of PGETC by a 2316 GoR (3 dan) serbian player.

Maybe you need to recalibrate your judgment not to be so extreme, maybe go games abilities can fluctuate more than you think, or maybe Go ranks are slightly different online, it seems that strong player got "smashed" by weaker players more often in online tournaments than in real ones.

Thanks for your post Lukan. How much of your view (and those of other strong players you talk to) is based on the initial announcement of Carlo's conviction, and how much did you independently conclude from looking at the games yourself before that? I didn't interpret the appeal announcement as having a subtext of "he was probably cheating but there's just not strong enough evidence to convict".

I see three main aspects which people are citing as evidence of Carlo's guilt:
1) statistical arguments of similarity to Leela's choices, the famous 98% match of top 3 in moves 50-150 of the game vs Reem.
2) history of player rating and comparison to recent offline games (WAGC): Carlo's been 3d for about 6 years, up to 4d in 2017 and game results suggestive of 6d strength in the 2017-18 online league.
3) analysis of the moves in the league games themselves by skilled go players, do they look like a 4d human on good form, 6d human, bot etc.

The initial conviction announcement and related facebook post solely used arguments of type 1 (perhaps there were others in the unpublished report). I was disappointed at the lack of details about a control group and suspected poor use of statistics (plus the 98% figure is dangerously misleading) so did my own analysis. Unless the referees report was both significantly more thorough (which I hope but doubt) and had different results to my analysis I think it was unsuitable to support a conviction. I should clarify that the type of cheating I am considering here is using Leela to play almost all the moves, as that's what the 98% figure suggests. Using Leela to play just a few moves in a game would also be cheating but would be much harder to detect with statistical methods (at least the crude similarity matching approach).

As far as I know the conviction was based solely on the game vs Reem, rather than looking at other games this season or type 2/3 arguments. Indeed, if other games were considered it would have raised doubts about using the similarity metric as evidence, as Carlo beat stronger players than Reem with rather normal Leela similarity (I've reproduced my table at the end of this post). Based on Lukan's post, it seems that these type 2/3 aspects are what he and others and primarily considering so I think they deserve attention (Stanislaw's analysis was a type 3 with reference to Leela and he considered it was not suggestive of cheating).

Type 2: improving from 4d to 6d is hard so doing so in about a year is impressive/unlikely, more so if you had been 3d for ages. Just checking some top Europeans, Lukanwent from 4d to 6d in about 2 years, but he had been improving about that rate from 1k. Tanguy was also about 2 years. So for someone to improve that fast, coupled with bot-like play could naturally arouse suspicion, but set against that it would be terribly sad if someone did work hard to improve fast and was incorrectly accused and then convicted of cheating. There are some fast improvers on the UK team so whilst unlikely I think it's possible.

As for the WAGC games, they are not 6d level results so it's not so easy to quash these allegations. However, although it's making excuses I could well understand that with all the pressure and stress of this case Carlo couldn't play his best, plus you've got jet lag possibilities. So I'd say they weren't evidence of his innocence, but aren't great evidence of guilt either.

Another question is do you really need to be 6d in offline games to get the results of 6d+ 5d+ 6d+ 4d+ (Lukan's suggestion is he stopped cheating here) 6d- 3d+ 7d- 5d- 5d+? I'm a European 4d (and weakish one at that, never over 2400) and have beaten some strong players in the league e.g. Victor Chow EGF 6-7d (in a fairly dominant way from opening through middlegame, Leela Zero gives me 95% win but then he played strong and I weak endgame so ended up only half a point) or was also leading big against Pavol Lisy 1p (LZ gives 90%+) but then ballsed up a ko in overtime. Or I beat Dinerstein 3p on OGS but that's correspondence so I played out variations. So perhaps the non-cheating hypothesis should be more like Carlo is 5d, and getting two 6d wins in a row is nice but not so amazing.

Type 3: I'm interested Lukan chose the game vs Dragos as evidence of cheating, as my impression reviewing that game was that whilst Carlo indeed played well, it was more Dragos who played a terrible opening and herpy-derpy empty triangle, badly dealing with Carlo's 3-3 strategy, and Carlo then managed to maintain his lead for the rest of the game. Here's the game with some comments from me:



- My view, though not shared by Leela classic or Zero and probably many humans, is that 5 shoulder hit is mistake, at least to continue joseki allowing white to jump out on left which reduces black 4-4s top side moyo. At least we can say white has real territory and black has to show in later play how the outside group is valuable.
- r17 3-3, of course Leela classic doesn't consider, it's an AlphaGo or LeelaZero (which was kyu back then) style. Ponnuki joseki I prefer white a bit as r10 odd and maybe q6 became bad direction exchange (b e3 block worse later).
- c14 I just can't understand, surely f17 is loads better (c7 makes left side small, black wants to reduce w top side potential)?! Both Leelas agree (but only -3%, I expected more).
- s17, p15 in sente is normal, then f17 taking advantage of b not playing there, Leela classic prefers e18. I'd have expected black to kick or pincer or something more kiai than just defend corner, is simple territory Dragos's style? I checked the game vs Serbia figgitaly just posted and that seems like both players were having a competition who can play the most boring territory style game ever but then they had some fight and Serbian cut and made miai to attack 2 groups and killed one.
- c9 is odd, it looks bad and unimportant place to me, Leela classic likes it initially but switches to p4 with more sims, Leela Zero doesn't even consider but says not much minus.
- f15 seems kinda slow but solid given the close j16, and Leela classic doesn't even consider (LZ a little) but I suppose it's reinforce top against h17 etc and reduce b left side potential. Both Leelas prefer to play f3.
- n3 attach is try to make sente exchanges to close lower side with e3 block seeing as black didn't play there
- k4 is the first move from Carlo that I think is notably strong for me. I'd also think about k3 or l4 and Leela classic prefers k3.
- Sorry Dragos, but wtf is l4? If cut LZ gives b 40%, but with this ugly bulge and connect and obvious jump jump he plunges to 20%. I wish my 6d opponents gave me such gifts.
- Then he doesn't jump at l9 but makes thick d6 connection in gote (with weird b8 cut, was he really worried about w cut if direct double hane?) so Carlo gets to play there, even a 10k can enjoy playing such a lovely cap.
- Some obvious shape attack from white, but the cut would scare me. Dragos cuts and then o6 defends because white doens't have the ladder.
- Carlo's q15 to start a fight to activate the ladder is skillful, but it's the sort of thing I could find on a good day so I don't think it's indicative of cheating (and neither Leela find it, seems a very human strategy to me though I've seen Zen do things like this). After that the game's pretty much over and nothing much else happens except normal yose.

To visualize the above table:

Wise and good. Especially since the statistical method used is dubious. (Not that you would know that, OC.)



But it turns out that you were influenced by the statistics, considering the 98% match with one of Leela's top three choices to be proof, but not quite convincing. Not to rehearse the problems with the statistics, but the matching-one-of-three statistic is relatively insensitive. Would you have been as impressed by a 72% figure? Carlo's play matched Leela's choice 36 times out of 50. Unusual to be sure, but common sense tells us that it is well within the realm of the possible, without cheating. It may be enough to raise suspicions, but that's all. Carlo played a lot like Leela, but playing like Leela is not the same thing as cheating. So far, no one has put forward a reasonable theory of cheating that would explain Carlo's play.

However, those suspicions would justify analyzing the game record for evidence of cheating. That has been done by some players, with some discussion in this thread.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

There's no doubt that the outcome of this matter divides the Go community and I tend to agree with you on one count: everyone, but Carlo most of all, is caught between a rock and hard place.

Personally, I believe that an acquittal is an acquittal: it doesn't matter whether Carlo is guilty or not (that's a hidden state that we will never be able to observe) but he was cleared of the charge and so it is wrong to discriminate against him as if he was guilty. Removing him from his role as referee would be wrong.

If I were in Carlo's position, I would probably resign the role voluntarily in order to diffuse the controversial situation but that would be a self-sacrificing course of action and it is terribly unfair for anyone to be forced to take that course, should they be innocent. Carlo has been called innocent. (I don't know how much value Carlo personally places on the role of referee - it could even be an unwanted burden to him - but the principle stands.)

Should the EGC organisers fail to honour the outcome of the appeal? Should Carlo give up some of his own involvement in the Go community? All choices are wrong. The "Cheater" bell has tolled and nobody can truly reverse it.

Once again, this brings me back to my overwhelming conviction that the original judgement was rash and irresponsible. Pandora's box *has* been opened and I am still waiting for someone to find the Hope therein.

Hello, hello,

Well, Carlo beat me in WAGC and he didn't even have to lift a finger. That's some leela sh*t right there!

On a more serious note, more than stats and analysis, I trust (strong) players instincts and feelings (mine too). There are ways of winning, a 4d winning against a 6d isnt that crazy, the 4d can be in a good day, the 6d can make some blunders, etc. But the way Carlo won was like a complete mastering of the games.
The way he performed in WAGC was pretty bad, I saw the games, sure he is a copycat of leela in the fuseki then he collapses. This middle game's collapse didn't happen online.

On a personal note I believe he is guilty but in the end it doesnt matter, something has to be done for online tourneys, if we want to keep the pandanet which is a great tournament, some regulations have to be implemented. If it happens not to be enough to make sure there are no cheating then it will be the end of pandanet (and online tourneys in general)

I always believed that such behavior wouldn't happen in League A because we basically all know each other, we are friends or acquaintances, there is a mutual respect of the game and of the players. Italy goes up to league A and all goes to sh*t ? (ok, no joke about italians cheating in sports...)

Anyway, let's not be blind, the cheating is pretty obvious but we just move on, make some new rules then we will see.

Hello Ben, this must be not what you really think, too much insulting, a team and a Country. I'll never belive a go player could write something like this.

Am I wrong?

Mirco.

The problem I see is well illustrated by the following two quotes:

and


Just assume for a moment, an innocent player played 2 or 3 games that were very strong for his level and suspicions are raised because of that. But now, no matter how he performs the next games, it will always be perceived as evidence towards him cheating before. If he plays "badly"/his usual level, people will think "of course he stopped cheating now, it is such an obvious differnce to before". If he continues to perform exceptionally, it will just be further evidence that he is cheating, cause it's now even more "suspicious games".


This means, no matter how the players' next games go, they will be evidence. But just the fact that he is still playing cant be evidence, can it?
Therefore I suggest, one should stop at the moment were there was a suspicion, collect all evidence up to that point and judge that. Dont judge what happens after if no matter what happens you could always use it against him.

Oh, I did not see the entire game, I just observed the final part after the death of Dragos' group.

Cheers.

Going over the games is important, and necessary when, as here, there is no physical evidence and only weak statistical evidence of cheating. (As my statistics and research methods prof told us, statistics proves nothing. Despite the impressive 98% figure, the statistics was poorly done.) Ken Regan has done a good bit of work about cheating in chess. This note gives some links: viewtopic.php?f=10&t=15538&p=228848&hilit=regan#p228848



Unfortunately, people give far too much weight to confirmatory evidence, and give the 98% figure far too much credit. This is not their fault. Statistical reasoning is only a couple of hundred years old, and is still in flux. Most people are unused to it.



I was amazed to hear that the original verdict of cheating had been based upon a single game, and only on the similarity of Carlo's play to Leela's. You are right to look at multiple games, and to contrast Carlo's play in games where he is suspected of cheating with his play in games where he is not. The contrast in his play between the two types of games is strong evidence, not the similarity of his play to Leela's, as far as cheating is concerned.



From what I have heard, the original verdict was based upon weak statistical evidence alone, of only 50 moves a single game. Your instincts were right, to examine the play in more than one game. Had the case been built properly, the verdict may well have stood. Or a different verdict may have been reached. You are a better player than I, but in that game it seemed to me that Carlo took the lead without making any remarkable plays, as his opponent showed poor judgement. So there may not be expert agreement that the result was due to the weaker player playing far above himself.

But yes, given the lack of physical or behavioral evidence of cheating, analysis of game records is necessary to render a verdict. The fact of the matter is, a good case was not built for cheating, and it was right to overturn the verdict. A better case might have been built. I feel sure that future cases of possible cheating will be better investigated.



Perhaps you can post your game analysis that leads you to that conclusion.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

Sorry Lukan, but regarding this sequence. I have just done a quick analysis with Leela and Zen and they do not suggest that sequence at all (Is my pc not powerful enough?). It seems to me just a try to use aji / create confusion move by move and let the ladder work, usually AI's play more direct plan and they do not look so hard for some aji like humans do. Am I wrong?

I don't believe I've seen anyone here acting like that. I could have missed a comment or two though.

I'm pretty happy to accept that Carlo might have cheated. In fact, I'd go so far as to say that *someone* in the near future certainly will cheat. The Go community needs to confront cheating in a mature and considered way and sooner is better than later. Whether Carlo is innocent or guilty does not change this fact.

The evidence on which the original judgement was formed and the way in which it was handed down were neither mature nor considered and the arguments flying to and fro in this very thread and the potential impact on future online tournaments as well as real-life tournaments, not only the EGC, serve to illustrate the consequences of mishandling cheating.

Do we want to play in a world of witch hunts and kangaroo courts?

The appeal succeeded and the whole matter should be consigned to the annals of history. Carlo should be treated as if he is and always was innocent and, as a community, we should learn from the incident and open the debate on cheating and how to handle it.

No further judgements should be passed until a proper framework for detecting cheating is put in place. I agree with Bill Spight (and other contributors to this thread) that this will need to involve analyses of several games, comparison to more than one robot player and some sort of analyses of the contrasts and trends within the suspect's play, online and offline. Luckily, the Chess world has already walked this path and we can learn a lot from them.

In the interim, we must endure a dark age of suspicion and, yes, some game results will be marred by cheating. Let this serve as an impulse to bring about the required changes.

Even those who are adamant that Carlo cheated -- whatever their reasons -- are best served by this course of action because, if they are indeed right, two possibilities exist: the new, mature and considered strategy will catch Carlo in the future, should he persist, and, should he never cheat again, we will avoid losing one of our number.

I don't know Carlo at all. I have likely never even met him. I do know that Go players aren't all that numerous and that strong dan players are valuable in our community. I also believe in rehabilitation.

Just a note - somebody asked who the appeals commission was. From the post it appears to be Victor Bogdanov, Ales Cieply, & Frank Janssen.

I think it is more productive now to respond to this part of their statement "We advise the EGF to enforce more strict rules for EGF sanctioned tournaments played on internet"

The EGF AGM is not so far away, time is short.
https://docs.google.com/document/d/1qYqtlX12Ml3pxs5HIInk9bhFdh9_c58h76Etp0gnIgI/edit

_________________
North Lecale

Yup, that was my point that this game was more Dragos lost than Carlo won. So although he is obviously a strong player and used to be almost 2700 this game he didn't play so well so cheating is not necessary to beat him. Maybe he doesn't play so well online? I noticed he resigned early in a fight with Tanguy and losing to Serbia was a big upset, but also he beat Csaba and Fionin so good wins there. About playing online, some people (probably more likely those who learnt offline) find it hard to play full-strength online, concentrate, take it seriously etc. Others, such as myself who mostly learnt online, actually find playing on a real board harder in some ways, e.g. particularly with play in the centre I find it easier to see the "big picture" online which is why in real tournaments I often kneel on my chair or stand up to look at the board from above.


That move 47 sequence I thought was really bad even on my own, but yes having a strong bot to review makes me more confident to criticize it and not think he had some genius plan I didn't understand. Maybe it wouldn't have been so bad if Dragos tried to fight, e.g. play a knight's move instead of one point jump to put more pressure on white (seems like white can get out not too hard though, plus b on left is weak) or not take gote with the thick connect on left side? Ben said Carlo collapsed in middle-game in WAGC, this game Dragos didn't put much pressure on him so few chances to collapse (only m9 cut). [One of my 3d friends has a boring territory style, his "favourite joseki" is 3-4, high approach, attach under, hane, pull-back, connect, one-space jump , and Dragos was his teacher many years ago, maybe that's where he got it from!].


Well, we disagree there as although it is a nice strategy I think some human 4ds can find it. Maybe it's because I watch a lot of pro games so am relatively good at advanced strategies like this (but don't do tsumego so am bad at reading hence only 4d). That's if it is actually a good move, could Dragos have taken sente at some point to extend at m10, suffer a local loss in the upper right but gobble up the 2 stones now cut by m10 or even attack the lower white group? But the important issue is not is this sequence credible for a 4d to play, a 4d who's quickly improved to 5/6d or whatever, it's does Leela suggest this sequence so by playing it we think he cheated? And when I try Leela 0.11 on my PC it effectively doesn't consider it (after 200k simulations it's choice #18 with 5 sims).

P.S. Your English is fine!


I'm not convinced that Carlo didn't cheat, but am fairly sure that the evidence used (assuming it was 98% similarity statistic-based and not that the EGF hired a drone/ninja to peer through his window and spotted Leela on screen during the game) did not justify the conviction in which case the appeal decision is correct. I'm not defending Carlo per se, I'm defending due process and rigorous, quality analysis (misuse of statistics is a particular bee in my bonnet). I do also take the view that incorrectly punishing an innocent is much worse than failing to punish a guilty, so if there is reasonable doubt then not punishing is correct. It's hard to quantify, but at the moment my view is maybe 10-20% chance that he cheated. An example of a move I found suspicious was a reduction move he played against Kulkov, so I posed it as a problem here to see how common it was for dan-level humans to want to play around there.


It's rather funny that Carlo ended up playing Chris, as Chris is actually one factor why I have taken such an interest in this case. He has improved very rapidly: last time I saw him he called himself 1d but was probably 1k, so when he turned up saying he was 5d I was sceptical, and the last tournament he went to he lost to a 1-2d (who I was recently embarrassed to lose to for the first time after 12 straight wins so I teased him a bit about that). But then he beat Daniel Hu 4d (also fast improver, going on 5d?) 4-0 in some practice games online before Daniel beat me in last year's British championship. He's had great results as our "secret weapon" in the league this season winning all his games, somewhat helped by being on a low board with his sandbaggy EGF rating: 2d+ 2d+ 2d+ 5d+ 5d+ 3d+ 3d+ 5d+ 4d+ Carlo-. So given his poor real-life results (2 years ago, he lacks tournament and playing under pressure/overtime experience) one might think his online results are suspicious and suggest cheating. I don't think he is (he's one of these autistic-spectrum maths genius types, he's won mental calculation championships), plus his style is kinda wacky and creative but also careless. This season he's often been winning a lot but then gets complacent and makes a mistake but still wins in the end (like that dumb liberty wasting push against Carlo). So I want to avoid a badly designed cheat detection process that could produce a false positive.

I have done also a quick analysis with zen / leela / leela zero of this game. It seems that he did a lot of mistakes (or at least not suggested moves) according to all of them, especially on top right corner fight.

Do you have different analysis?

Cheers.

Worth repeating.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

As you know, I agree that suspicions are justified. But I also believe that the process that reached the verdict of cheating was flawed, and that the reversal of that verdict was fair and correct.

I also believe that you are right to analyze Carlo's game records. If there is a case to be built, that is where to look. That is why I am interested in your analysis of the Bryant-Metta game.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

The Chess World learnt

• that Stewpot likes cruises
• never to hold qualifiers online
• to ban mobile phones
• to ban headphones
• When things get bad, use the toilet

_________________
North Lecale

Hi Javaness,Noob question: what's this reference ?