Lichess4545 Ledger 107

Lichess4545 Ledger

Issue #107 - August 3rd, 2019

State of the 4545 League

by @kostasvl

The 18th season of the team4545 league starts on August the 5th. The Monday that recently passed (the 29th of July) teams were formed and all players will have a week to choose a team name and a team captain and maybe also get to know each other and play some games.

Side-leagues news

by @kostavl

The 7th season of Chess960 officially ended last week. Congrats to everyone for participating! Below you can see top ranked players and detailed summary of the season.

1st place: @Somethingpretentious (With a perfect 7/7 score)
2nd place: @Grzybozbur (Scored 5/7 points)
3rd place: @f1nn33 (Scored 5/7 points too. Note also that @postit also scored 5 points but took the 4th place)
U1600 Winner: @finn33 (Yeap, the same player that won the 3rd place.

Click here for the summary of the season. Make sure you also submit the end of season survey if you haven't already done so. Click here for the survey.

Series season 7 has ended. Thank you for everyone who participated this season. In the Masters division, moistvonlipwig managed to get first again despite some close games. IsaVulpes got second and mronemore ended with third. In the Candidate Masters division, group 1 has jfarah first, Finneri second, and tactix47 getting third with some big upset results. Group 2 had some problems with forfeits, but Aleksandr88_62 grasped first, while Mixalaki2705 got ahead of WarGoblin with better tiebreaks. Isachess was in group 2 and got last with his wins coming from the two forfeits. jfarah, Aleksander88_62, and Mixalaki2705 will be promoting to Masters division if they join next season. In the C division, group 1 had adande1 wining all his games, with Numberman768 only being ahead of Tranzoo by the last tiebreak. Group 2 had H0tblackDesiat0 getting first, Brundd coming in second, and izzie26 being third. adande1, H0tblackDesiat0, Numberman768, and Brundd will be playing in the CM division if they play next season. This season had a little more experimentation, we had divisions to allow faster climbing to the Masters group, and had more people in each group, with some updated rule changes. The rule changes were a success. The new division format was mixed, it worked for its purpose but needs a little tweaking in some of its rules. Finally having more people didn't work, as some people didn't schedule any games, while some people had life happen and needed to withdraw. So the amount of people was still about the same, we may or may not try this again next season. If you want to join Series Season 8 there will be a message in #announcements and #general. In the mean time you can join #series-general.

Project

by @hicetnunc

Why I’m not surprised @Realrobbert was flagged as a cheater

by @hicetnunc

Disclaimer 1 :

The data and conclusions presented here are the result of my personal investigations. They don’t relate in any way to lichess detection team’s methods, which I don’t know. Lichess may or may not use vastly different methods and indicators to reach their own conclusions about who they mark as cheaters.

Disclaimer 2 :

The data and conclusions presented here are my sole responsibility. In particular, none of the league mods endorse any part of this material in any way. As far as they’re concerned, the mods may or may not consider this article as complete bullshit :-)

In this article, I will look at RealRobbert’s recent games before he was marked as a cheater by lichess and banned from the league. I’ll compare RealRobberts games to a random sample of OTB players games using Stockfish 7 selection of moves as a reference point. We’ll look at various indicators which may or may not be indicative of cheating - I’ll let the readers make up their own mind.

About @RealRobbert :

As far as I know, RealRobbert was an adult with limited OTB activity, if any. At the start of season 16, his classical rating was 1810. His lichess blitz rating was also around 1600.

According to ErinYu’s article in ledger #88, 1800 classical would roughly compare to 1500 FIDE, which also fits with the blitz rating.

About the sample :

I’ve selected the most recent 10 games played by 100 different OTB players picked in the ChessTempo database. These games are usually played at 90+30 time control or longer. All the players have a FIDE rating between 1400 and 1599 and have at least 20 games in the database. Their last recorded game was played no earlier than 2017.

The full sample data can be found here.

About the methodology :

I’ve scanned all the sample games as well as RealRobbert’s games with pgn-spy. I’m comparing various indicators from RealRobbert’s last 10 league games with those from the sample.

Part I - Choice of top engine moves

Here we look at two indicators, which show the frequency of choice of engine’s top moves (called %T1) and engine’s top three moves (%T3).

Only unclear positions are looked at - ie. positions where one side is winning, or positions where there’s a single logical move (eg. recapture) are left out.

Here’s a graph showing how @realrobbert’s T1 indicator compares to the top 20 scores from the sample :

Here’s a second graph showing how @realrobert’s T3 indicator compares to the top 20 scores from the sample :

And finally, here is an aggregated graph showing how @realrobbert T scores compare to the whole sample’s :

The raw data for T1 and T3 scores can be found here and here.

Part II - Overall stability and accuracy of play

Next, we look at the blunder rate (%BR) in undecided positions - ie. positions where neither side is clearly winning.

Here is how @realrobbert’s blunder rate compares to the top 20 players from the sample :

The raw data for blunder rate can be found here.

Lastly, there’s the well known average centipawn loss indicator, which measures the difference between the eval of the position when you play the engine top move and the eval when you play your move.

Here, we only look at positions where neither side is winning though, so the numbers may be significantly different from lichess’ output.

Here is how @realrobbert’s pgn-spy acl compares to the top 20 players from the OTB sample :

The raw data for pgn-spy acl is here.

Here is a comparison of @realrobbert’s BR and acl to the whole class C players sample. To help with the graphical representation, I’ve used the following conventions :

%accuracy = 100 - acpl

%stability = 100 - 5*BR

In both cases, results under 0 are ignored

Part III - Breakdown by game - proportion of excellent games

While acl gives a good idea of the overall quality of play in a bunch of games, it’s also interesting to look at the game by game breakdown to see if a given player has significant performance swings.

For this comparison, I’m using another sample of 852 games by class C OTB players, as I hadn’t time to extract the information from the current sample, although I wouldn’t expect much difference :

And here is @realrobbert’s breakdown from his last 10 games :

Although the number of games is small, the breakdown is very different, with a large proportion of games with very low acl.

Part IV - Seasonal trend

Finally, here’s the evolution of @realrobbert’s indicators throughout the last 4 league seasons :

Graphique

The general trend is clearly up for 3 of the 4 indicators.

All these indicators may or may not be indicative of cheating, but all in all, I’m not surprised by lichess’ decision to mark @realrobbert as cheater.

Weekly Stats

by @kraaft, @RobUmbra and @Somethingpretentious

Stats for Season 15 Round 1:

The fastest mate was black on move 17 found in Gamelink White: scacchistaimbruttito, Black: min37.
The fastest draw was found in Gamelink White: glbert, Black: lapa87.
The fastest resign was black on move 12 found in Gamelink White: rodeo, Black: alms.
The biggest upset was 484 points in Gamelink White: gallifrey42, Black: robumbra.
The longest game ended with white on move 77 Gamelink White: lgv2018, Black: upgoerfive.
126 was the highest ACPL in Gamelink White: rodeo, Black: alms.
10 was the lowest ACPL in Gamelink White: kyaaron, Black: spaceball.
Combined maximum ACPL was 171 in Gamelink White: rodeo, Black: alms.
Combined minimum ACPL was 21 in Gamelink White: kyaaron, Black: spaceball.
The longest think was 12 minutes 31.0 seconds on move 8 in Gamelink White: rodeo, Black: alms.
The most time left was 42 minutes 4.0 seconds in Gamelink White: badplayer_cm, Black: david-innes.
The most time spent was 64 minutes 21.0 seconds in Gamelink White: artian12, Black: mathijshuis.

-PGNs-

SERIES 90+30:

Overall Season 7:

The fastest mate was white on move 17 found in Gamelink White: tactix47, Black: scarff.
The fastest draw was found in Gamelink White: grizzzly1000, Black: chesspatzerfrance.
The fastest resign was white on move 14 found in Gamelink White: mqll, Black: h0tblackdesiat0.
The biggest upset was 241 points in Gamelink White: numberman768, Black: raitonvsfuuton.
The longest game ended with black on move 93 Gamelink White: numberman768, Black: tranzoo.
138 was the highest ACPL in Gamelink White: wargoblin, Black: mixalaki2705.
9 was the lowest ACPL in Gamelink White: grizzzly1000, Black: chesspatzerfrance, Gamelink White: moistvonlipwig, Black: krsto, Gamelink White: aleksandr88_62, Black: isachess.
Combined maximum ACPL was 248 in Gamelink White: wargoblin, Black: mixalaki2705.
Combined minimum ACPL was 19 in Gamelink White: grizzzly1000, Black: chesspatzerfrance.
The longest think was 56 minutes 19.0 seconds on move 17 in Gamelink White: kostasvl, Black: tranzoo.
The most time left was 93 minutes 4.0 seconds in Gamelink White: brundd, Black: drodnat.
The most time spent was 124 minutes 1.0 seconds in Gamelink White: krsto, Black: isavulpes.

-PGNs-

AWARDS

Accuracy King: with an average of 21 ACPL, moistvonlipwig.
Fast Finisher: with an average gamelength of 41 ply/21 moves, adande1.
Marathon Man: with an average gamelength of 95 ply/48 moves, finneri and drodnat.
Giri Award: 3 draws, isavulpes.
Biggest rating gain: with 55 rating points, darubaru.
Biggest rating loss: with 74 rating points, mekumcu.
Fun fact: The most popular opening played was:
- C00 French Defense and B12 Caro-Kann Defense (4 played each)

Masters
moistvonlipwig (defending his title) IsaVulpes mronemore

CM1
jfarah Finneri tactix47

CM2
aleksandr88_62 Mixalaki2705 WarGoblin

C1
adande1 Numberman768 Tranzoo

C2
H0tblackDesiat0 Brundd izzie26

Database Analysis of Series #07

                   Games     1-0     =-=     0-1    Score
-----------------------------------------------------------
 All report games     81      44      11      26    61.1%
-----------------------------------------------------------

Average Rating and Performance

White rating: 1867; White performance: 1944 (61% vs 1864)
Black rating: 1864; Black performance: 1787 (39% vs 1867)

Frequency of Themes in the First 20 moves of each game

   Same-side castling:             60%    White Isolated Queen Pawn:      15%
   Opposite castling:              11%    Black Isolated Queen Pawn:       9%
   Kingside pawn storm:            15%    White Pawn on 5/6/7th rank:     56%
   Queens exchanged:               53%    Black Pawn on 2/3/4th rank:     37%
   Only one side has Bishop pair:  12%    Open c/d/e file:                58%

Opening Move

    Move   ECO       Frequency    Score  AvElo Perf AvYear %Draws
 1: e4     B00a       51: 62.9%   64.7%  1884  1981         12%
 2: d4     A40a       20: 24.6%   47.5%  1828  1833         15%
 3: c4     A10         5:  6.1%   90.0%  1892               20%
 4: Nf3    A04         4:  4.9%   37.5%  1804               25%
 5: g3     A00t        1:  1.2%  100.0%  1890                0%
_______________________________________________________________
TOTAL:                81:100.0%   61.1%  1867  1943         14%

See the stats page for win percentages for each board by color and ratings distribution.

Please feel free to join #lichessledger on slack if you would like to contribute towards the ledger in any way, or provide any feedback. Both are highly encouraged and appreciated. Thank you for reading.

Thanks to @Forhavu, @hicetnunc, @kostasvl, @kraaft, @okei, @RobUmbra and @Somethingpretentious.