Personal Guide: Honver

Personal edition · 80 Jade South games since January 2026 · 830 hands

Where your points go against LuckyJ.

This guide measures 80 of your Jade South games since January 2026, 830 hands, with the same code that measured 1,255 LuckyJ games. It keeps the lessons from our reviews and shows your own table at each turn where a different tile, call or declaration would have played closer to LuckyJ.

Most of your game already matches. Tenpai speed, tile efficiency, call rate and honor timing are LuckyJ's numbers, and you fold against a riichi at least as well as LuckyJ does. The gap is in which hands you finish and what they are worth, plus a small set of tile choices once someone has called.

Your first rate since January, 23.8%, sits within one standard error of your career 26.3% over 3,877 games. The Jade players at your tables win, deal in and score like LuckyJ's Tokujou opponents. The difference that is real is LuckyJ's 31.7% firsts against your career 26.3%, and it lives in the value of your wins.

Scoreboard

You against LuckyJ, same code

You, 830 handsLuckyJ, 13,054 hands

−15points per handLuckyJ +373
5,590points per winLuckyJ 6,836
18%of wins at mangan or moreLuckyJ 28%
21.3%win rate per handLuckyJ 22.9%
11.8%deal-in rate per handLuckyJ 10.0%
2.45average placement since Januarycareer 2.42, LuckyJ 2.26

How to read the tables

Your seat is at the bottom.

Each table is rebuilt from the game record at the moment you had drawn and were about to discard. The tile you drew sits apart at the right end of your hand. Every example lists the tile you cut, the better cut, and every other discard with the shanten it leaves, how many tiles improve it, and how safe it was against each player who had declared riichi or called.

Your cut the tile you cut Better the better cut

Safety labels read from your seat at that turn. genbutsu is in that player's river, or was discarded by anyone after their riichi without being ronned; those tiles also anchor suji. suji is a 1, 2 or 3 (or 7, 8 or 9) whose partner three away is anchored. nakasuji is a 4, 5 or 6 with both partners anchored; with only one it is half suji, and still mostly live. A 4 whose 1 is anchored, or a 6 whose 9 is, reads as virtual nakasuji when that player also cut the 5 early: the early 5 is the sotogawa read that covers the other side. An honor with two other copies showing can only be hit by a tanki. Everything else is live.

01

Chapter one · Value · added 22 September

Count the dora before you pick the plan

LuckyJ wins about as many hands as you do. The difference is what it does with a hand that has dora. Sorted by the dora and red fives in the deal, your win rate stays flat while LuckyJ's climbs:

Dora in the dealYour win rateLuckyJYour mangan per 100 handsLuckyJ
None23.6%21.4%3.84.1
One18.2%23.3%3.47.1
Two21.0%26.6%3.811.9
Three or more21.4%33.9%14.323.1

On hands with no dora you already play like LuckyJ, and you win them slightly more often. On deals with one dora or more, LuckyJ makes 8.8 mangan per 100 hands and you make 3.9; on deals with none, 4.1 and 3.8. Dora reach your hand at the same rate and you both cut about a fifth of them, so the gap comes from three habits.

Riichi at the first closed tenpai. With nobody in riichi, LuckyJ declares 51% of first tenpais on 3 or fewer live tiles, 82% on 4 to 7 and 93% on 8 or more. You declare 41%, 76% and 89%. Your 20 dama wins average 5,880 and 20% reach mangan; LuckyJ's average 7,667 and 36% reach mangan, because it keeps dama mostly for hands that are already big.

Shape yaku on top of the dora. LuckyJ wins 5.0 tanyao hands per 100 and you win 3.0; honitsu 2.1 against 1.3. In deals with lone honors and isolated terminals, LuckyJ has cleared every terminal and honor by turn 6 in 35% of hands and wins 28% of those deals. You clear 22% and win 17%. Junk goes first: guest winds and isolated terminals on turn 1, the lone value honor by turn 3 or 4.

A yakuhai pon needs something behind it. When your first call is a yakuhai pon, 47% of the time the hand has no dora (LuckyJ 41%). Your pon hands with two or more dora win 28% of the time against LuckyJ's 43%, though that is only 18 of your hands. A bare yakuhai pon from 3-shanten earns LuckyJ about 16 points per hand, which is nothing. Let the first copy go when the pon would leave you 3-shanten with no dora, unless you lead or the hand has a second plan such as toitoi or honitsu.

Rule Count the dora and red fives on turn 1. None: a cheap hand to finish if it comes and fold first when someone attacks. One: a mangan candidate, so riichi at the first tenpai and build tanyao when the deal allows. Two or more: the hand of the game, so call when the call lands you at 1-shanten or tenpai, riichi at the first closed tenpai and push it at tenpai.
Exception Dama is right when the hand is already mangan without riichi and the wait is poor, or on a tenpai from turn 11 on when you lead in the South round. Ask again every draw: LuckyJ declares later in 17% of the tenpais it first passes on, you in 10%.

Evidence 80 of your games (830 hands) against 1,255 LuckyJ games (13,054 hands), same code. On deals with one dora or more, the mangan gap is more than three standard errors and the win-rate gap about two and a half. The three-dora row holds 14 of your hands and is noise.

02

Chapter two · Open tables · added 22 September

When someone opens, slow the cheap hand

An opponent's early call slows LuckyJ's hands with little or no dora and barely touches its hands with two or more. Yours react differently: your cheap hands keep full speed, and your hands with dora deal in far more often. Hands where one opponent had called by your 8th discard, with the quiet-table rate in brackets:

Dora in the dealLuckyJ, noneLuckyJ, oneLuckyJ, two+You, noneYou, oneYou, two+
Hands2,6852,02769415312755
Reached tenpai43.5% (51.4)47.5% (54.4)54.8% (57.6)52.3% (52.2)49.6% (56.7)43.6% (62.3)
Declared riichi18.4%19.6%18.9%19.0%17.3%10.9%
Opened the hand33.0%34.7%40.5%35.9%39.4%32.7%
Won19.9%22.3%26.5%22.9%21.3%20.0%
Dealt in9.9% (8.0)11.2% (9.5)12.0% (10.6)10.5% (7.7)19.7% (10.2)18.2% (11.3)
Points per win5,6617,1559,1054,7005,9747,500

LuckyJ keeps its valuable hand moving with a call of its own. When an opponent opens first, LuckyJ answers around turn 9, and 70% of those calls land it at 1-shanten or tenpai. Almost half are chis, which keep the dora and the tanyao shape. It later opens 31% of the two-dora hands that were still closed when someone called by turn 6, and 22% of the hands with no dora. Yours open at 25 to 29% whatever the dora.

The deal-in row is the solid finding. With one dora or more and an opponent open, you deal in on 19% of hands against LuckyJ's 11%, over 182 of your hands. On a quiet table the same hands deal in at LuckyJ's rate. You don't push them more often: against a single call at 1-shanten with dora in hand, you cut a live tile 68% of the time, as LuckyJ does. The live tiles you choose hit more often, which is the reading chapter 3 works on. With two or more dora the hand also stalls: it reaches tenpai 44% of the time against 62% on a quiet table, and declares riichi 11% of the time against LuckyJ's 19%.

Rule When someone opens, count your dora before anything else. Two or more: keep the plan, take a chi or pon that lands you at 1-shanten or tenpai by about turn 9, and riichi at the first closed tenpai. None: slow down, fold from 2-shanten and hand the caller your safe tiles. With any dora, you already push as often as LuckyJ; read the caller's tells from chapter 3 before you pick the live tile.
Exception Your two-dora column holds 55 hands, so its tenpai and riichi gaps are one and a half to two standard errors: a direction, not a measurement. The deal-in gap on hands with dora, over 182 hands and more than three standard errors, is the solid part.
03

Chapter three · Calls · added 22 September

Against a call, take the safe tile that costs nothing

Folding against one call does not mean destroying the hand, and your choices there already match LuckyJ's. When a safe tile is also the most efficient tile, you take it 61% of the time and LuckyJ 56%. You pay acceptance for safety 14% of the time and break a shanten for it about 3% of the time, the same as LuckyJ.

LuckyJ's pattern: at tenpai it takes a free safe tile 93% of the time, and from turn 13 on it takes one 79 to 97% of the time at any shanten. It pays one or two tiles of acceptance for safety 26% of the time, rising to 35% once the caller has cut three or more tiles from hand since the call and 42% from turn 13. Those are the moments live tiles start to hit: in LuckyJ's games a live tile hit a caller with three such cuts behind them 1.9 times per 100, against about 0.5 before that, and 2.4 times from six turns after the call against 0.25 right after it. It breaks a shanten only from 1- or 2-shanten after turn 12, and then 21 to 24% of the time. Against two calls it is stricter: from turn 9 it cuts a live tile only about 30% of the time when a safe tile keeps its shanten.

Your live tiles still get hit more than twice as often against a single call, 1.4% of live cuts against LuckyJ's 0.64%, and most of those were pushes LuckyJ would also make. The examples below are the ones that were not: the safe tile cost nothing, cost a tile or two against a two-call hand, or the river had already said what the hand was.

Rule Against one call, play for efficiency. When two tiles are equal, cut the safe one once the caller could be ready, from about their tenth discard; before that, cut an equal live honor or terminal and keep the safe tile, as chapter 13 explains. Once the caller has cut three or more tiles from hand since the call, or from turn 13, a tile or two of acceptance for a safe tile is often worth paying. Break a shanten for safety only from 1- or 2-shanten after turn 12.
Exception Against two calls, a value pon with a second call, or a river that has named its suit, pay more. Those hands are close to tenpai and their waits are easier to guess.

Evidence 1,594 of your discards and 24,618 of LuckyJ's where exactly one opponent had exactly one open meld and nobody was in riichi.

04

Chapter four · Folds · added 22 September

Fold with the right tile

Folding is one of your strengths. Against a riichi you cut a live tile less often than LuckyJ at every shanten, and you pick the fold tile as well: when a genbutsu kept the same shanten, you cut it 93.5% of the time and LuckyJ 92.6%.

Cut a live tile against a riichiYouLuckyJ
At tenpai38.2%42.5%
At 1-shanten17.1%20.4%
At 2-shanten6.5%10.1%
At 3-shanten or worse6.0%6.2%

Here a tile counts as safe against a riichi player when it is genbutsu, suji or nakasuji, or an honor with two other copies showing. Genbutsu includes every tile discarded after the declaration without being ronned, and those tiles anchor suji too. Counting only the declarer's own river would call furiten-safe tiles pushes and overstate every rate in this table.

So folding is not where your points go. The examples below are the occasional turns where a better fold tile was on the table, and they come with the order that decides them. Suji is not one class. Against one riichi, LuckyJ's cuts dealt into it this often:

Tile cut against one riichiLuckyJ's cutsDeal-ins into the riichi per 100
Genbutsu, including tiles passed after the riichi15,4620
Honor with two other copies showing9050.2
Suji on a 1 or 94781.1
Honor with one other copy showing8562.1
Nakasuji3272.8
Honor with no other copy showing3483.2
Suji on a 2 or 84223.3
Suji on a 3 or 73794.5
Live 1 or 99775.2
Half suji5306.2
Live 2 to 81,5928.2

After genbutsu and the honors with two copies showing, a suji 1 or 9 is the next safest tile, and a suji 3 or 7 is closer to a live terminal than to genbutsu. A nakasuji sits with the average suji. A half suji, with only one side covered, deals in more than twice as often and belongs with the live tiles. Virtual nakasuji came up only 31 times in LuckyJ's cuts, too few to rate. From 2-shanten on turn 13 or later a shanten is not worth a live tile: LuckyJ pushes there 5.0% of the time. From 1-shanten on turn 13 or later it pushes 8 to 17% of the time, the high end with three or more dora.

Rule Rank the safe tiles before you cut: genbutsu against every threat, then genbutsu against the riichi, then an honor with two other copies showing, then suji on a 1 or 9, then nakasuji, suji on a 2 or 8 and honors with one copy showing, with suji on a 3 or 7 last. A half suji is a live tile. If a safe tile leaves the same shanten and acceptance, it is the only cut.
Exception A safe tile that costs a shanten is still right from 2-shanten on turn 13 or later, and from 1-shanten late at least four hands out of five, dora or not.

Evidence 1,574 of your discards and 24,560 of LuckyJ's made while an opponent was in riichi and you were not.

05

Chapter five · Push · added 22 September

Push the tenpai that can win

At tenpai against a riichi you push 38% of the time and LuckyJ 43%; from turn 13 on, 26% against 35%. The clearest gap is the open tenpai whose wait can only be kept with a dangerous tile: LuckyJ pushes it 84% of the time and you 71%, over 63 of your hands. LuckyJ pushes 93% of those with 8 or more live tiles and 76% with 3 or fewer.

The yaku decides the push more than the wait width does. With a yaku on the wait, LuckyJ pushes an open tenpai 68% of the time even on 2 live tiles and 95% on 8 or more. With no yaku on any wait it pushes 55% of them, since such a hand cannot win by ron. A closed tenpai with no yaku can only win by tsumo, and on one live tile that almost never happens.

Rule Push a tenpai that has a yaku, even on a narrow wait, and keep an 8-tile wait even when it costs a red five. Play a tenpai with no yaku the way you would play 1-shanten.
Exception The closest call in LuckyJ's data is an open tenpai with no yaku on any wait, pushed 55% of the time. With a yaku and a live wait it pushes at least 68% of the time, and 95% with 8 or more live tiles.
06

Chapter six · Dora hands · added 22 September

New

Your dora hands lose their points on the win side

You asked how to win more with the dora you are dealt, and whether you deal in more because you try harder to win those hands. Chapters 6 to 8 answer both. They cover 410 of your hands dealt at least one dora or red five and 6,577 of LuckyJ's, measured with the same code.

Hands dealt doraYouLuckyJ
Points won per hand1,1671,936
Points lost to deal-ins per hand615566
Reached tenpai51.7%51.8%
Turn of the first tenpai9.49.5
Tenpai hands won36.8%47.5%
Average win6,1337,864
Wins with no shape yaku60.3%45.3%

Your extra deal-ins cost you 49 points per dora hand. The wins you miss cost 769. You reach tenpai as often and as fast as LuckyJ, so the gap opens after tenpai: in how often the tenpai wins and in what the win is worth. Without dora you win 45.6% of your tenpai hands and LuckyJ 46.3%.

Some of that gap is luck. For each of your dora tenpais, LuckyJ's own win rate from the same spot says how often it should win: the same play (riichi, dama or open), the same table (quiet, a call or a riichi), the same number of live winning tiles and the same turn. Those rates predict 100 wins from your 212 dora tenpais, and you won 78. Your riichi hands account for 10 of the missing 22. Once riichi is declared nothing is left to decide, and nobody can see your dora, so treat those 10 as bad luck. Your tenpais without dora won 99 times against 103 predicted.

Your hunch holds in one place. Every extra deal-in in your dora hands went to a player who had called, and every one was a live tile: 9.3 deal-ins into callers per 100 dora hands against LuckyJ's 5.0. You cut live tiles against callers as often as LuckyJ, 65% of your cuts, at the same turns and against the same number of calls. What differs is your 1-shanten play once a caller looks ready, which chapter 8 measures. The larger leak is the size of your dora wins, which chapter 7 covers.

Rule Keep the dora and push the hand toward tenpai harder than a cheap one, as LuckyJ does, but build it the way you build a hand without dora: the same shapes, the same timing on a flush call and the same respect for a caller who looks ready. The dora add their han at the end.
Exception Your dora riichis won 27 times where 37 were predicted, more than two standard deviations low. Nothing in those hands was yours to change after the declaration, so keep declaring at the rates in chapter 1.

Evidence 410 of your hands dealt dora and 6,577 of LuckyJ's; 212 and 3,409 of them reached tenpai. Your 78 wins against 100 predicted is about three standard deviations. The shape-yaku gap covers 78 of your dora wins and 1,619 of LuckyJ's, about two and a half standard errors.

07

Chapter seven · Dora hands · added 22 September

New

Build the hand under the dora

LuckyJ keeps its dora and plays those hands faster: it reaches tenpai in 46.6% of hands dealt no dora and 55.3% of hands dealt two or more. What the dora do not change is the shape it builds around them. Its wins carry a shape yaku (pinfu, tanyao, iipeikou, a flush, toitoi and the rest) as often with dora as without. Yours carry fewer once you hold dora:

Wins with a shape yakuYou, no doraYou, doraLuckyJ, no doraLuckyJ, dora
Riichi at the first tenpai50.0%33.3%54.3%53.5%
Open hands44.7%30.0%51.6%48.1%
All wins50.5%39.7%55.4%54.7%

In riichi wins the missing yaku are pinfu and tanyao. LuckyJ's dora riichi wins come with pinfu 39% of the time and tanyao 16%, yours with pinfu 26% and tanyao 4%. In open wins the missing yaku is the flush: 13 of every 100 of LuckyJ's open dora wins are honitsu or chinitsu, and none of your 40 were. A yakuhai and one dora make 2 han, 2,000 points. Add an open honitsu and the same hand is 4 han, 7,700.

The flush shows the habit most clearly. Take the hands whose first call left eleven or more of the fourteen tiles in one suit plus honors, with the flush still possible:

First call with a flush in reachYou, no dora at the callYou, dora in the suitLuckyJ, no dora at the callLuckyJ, dora in the suit
Hands3615533373
Turn of the first call5.47.95.45.8
Flush completed63.9%33.3%64.9%59.0%
Won with the flush22.2%0%22.1%19.8%
Points per hand after deal-ins1,842−1531,3832,125

"Dora in the suit" means every dora and red five in hand sits in the flush suit or is an honor. Without dora you play these hands like LuckyJ and make more from them. With a dora in the suit, LuckyJ calls just as early, finishes the flush almost as often, and its flush wins average about 10,000. You call two turns later, finish a third of them, and none of the fifteen won with the flush. Fifteen hands is a small sample, so read the size of that gap loosely. It points the same way as the open-win yaku above.

Rule Keep the dora, and build around them the hand you would build without them. Keep two-sided shapes toward pinfu and simples toward tanyao. When every dora you hold sits in your main suit or is an honor, make the first call on the flush hand as early as you would without dora, and cut the off-suit tiles first whenever that costs no acceptance.
Exception A dora in another suit changes the plan, because the flush would throw it away. LuckyJ gives up those flushes 41% of the time, against 24% when the dora sits in the suit.

Evidence 78 of your wins in hands dealt dora and 99 without, against 1,619 and 1,376 of LuckyJ's. The flush table counts first calls where every meld fits the flush.

08

Chapter eight · Dora hands · added 22 September

New

At 1-shanten, count the caller's tells

Your dora hands deal in to players who have called 9.3 times per 100 hands, and LuckyJ's 5.0. The lobby does not explain it. In your own games, the other three players' live tiles from dora hands hit a caller 2.1 times per 100 live cuts. Yours hit 2.9 times and LuckyJ's 1.3.

Three things a caller shows predict tenpai well, and all three are easy to count at the table:

  • the latest call came with the caller's 9th discard or later;
  • the caller has two calls;
  • the caller's last two discards since the call were tiles drawn and thrown straight from the wall.
Tells showingCaller tenpai, LuckyJ's gamesCaller tenpai, your games
None13.4%15.8%
One38.8%43.4%
Two64.8%69.9%
All three84.0%89.5%

Three or more cuts from hand since the call, the tell chapter 3 uses, predict less: a caller showing it is tenpai 43% of the time, against 33% without it.

LuckyJ cuts fewer live tiles with every tell, at every shanten. You do the same from 2-shanten. At 1-shanten and at tenpai your rate moves much less. With dora in hand and a safe tile you could have cut:

Cut a live tile, by tells showingYou at 1-shantenLuckyJYou at tenpaiLuckyJYou at 2-shanten or worseLuckyJ
None62.6%66.2%58.0%64.0%72.1%75.5%
One60.3%57.0%56.2%56.7%46.8%56.1%
Two or three51.0%42.3%55.0%49.9%34.1%36.1%
All three55.0%32.3%54.5%47.4%36.8%33.1%

At 1-shanten with all three tells showing, LuckyJ cuts a live tile one time in three and you more than half the time, over 40 of your cuts. That is the part of your hunch that holds: close to tenpai with dora in hand, you keep pushing after the caller has shown they are ready. When a caller really was tenpai, your live tile was their winning tile 9.4% of the time and LuckyJ's 5.4%. You cut tiles in a caller's suit, tiles next to the dora, wall tiles and tiles passed since the caller's last discard as often as LuckyJ does, so those reads do not explain that part, and this chapter sticks to the tells.

Rule At 1-shanten against a caller, count the three tells before every cut. With none showing, play for speed. With each tell, take the safe tile more often. With all three, take it at least two times in three, as LuckyJ does.
Exception At tenpai the tells matter less. With two or more showing, LuckyJ still cuts a live tile about half the time, and chapter 5's rule on pushing a tenpai with a yaku still applies.

Evidence 2,939 of your discards and 46,086 of LuckyJ's made against a caller while nobody was in riichi, with the other three players' discards from the same games for the lobby comparison. The all-three-tells row at 1-shanten holds 40 of your cuts, about three standard errors from LuckyJ's rate. The other rows differ by one to two.

09

Chapter nine · Tanyao and flush · added 22 September

New

When to start tanyao and a flush

You asked when to start tanyao and a flush, since you make them less often than LuckyJ even without dora. Without dora you win 2.6 tanyao hands per 100 against LuckyJ's 4.1, and 2.4 flush hands per 100 against its 2.5. So the no-dora gap is tanyao. The flush gap is mostly the dora one from chapter 7, plus one kind of deal below. Both answers start from the 13 tiles you are dealt.

For tanyao, count the terminals and honors in the deal, with a pair counted twice. The count decides whether tanyao is a plan:

Terminals and honors in the dealShare of your dealsNone left by turn 9, youLuckyJWon with tanyao, youLuckyJ
0 to 26.3%37.8%49.5%15.4%17.7%
313.4%34.1%33.4%6.3%12.2%
421.2%15.7%20.5%1.7%5.9%
523.1%5.1%9.3%2.1%3.5%
6 or 728.6%2.1%2.5%1.3%1.1%
8 or more7.5%0%0%0%0.2%

With two or fewer, tanyao is the default plan. With three or four, LuckyJ still wins tanyao in 12% and 6% of those deals, and you win it half and under a third as often. From five on, even LuckyJ rarely gets there.

Your first three discards look like LuckyJ's: in every deal size, two or more of them are terminals or honors on average, for both of you. The gap opens after that. In deals with four, LuckyJ holds 2.4 of them after turn 6 and you 2.7. It shows most in riichi: without dora your riichi wins come with pinfu as often as LuckyJ's, 34% of the time, but with tanyao 8% of the time against its 17%.

For a flush, count the tiles that fit one suit plus honors. A pair of value honors (a dragon, your seat wind or the round wind) decides whether the count is enough:

Tiles in one suit plus honorsOne suit by turn 9, youLuckyJWon with the flush, youLuckyJ
8, with a value pair2.5%5.3%2.0%2.4%
9, with a value pair8.3%18.2%2.1%5.8%
10 or more, with a value pair30.0%43.2%4.0%14.8%
8, no value pair1.4%2.3%1.1%1.0%
9, no value pair6.6%10.1%2.0%3.6%
10 or more, no value pair40.9%33.3%16.0%10.7%

With a value pair, LuckyJ starts the flush at nine tiles. The pon gives the hand its yaku and its first call, and the flush turns that 1 han into 3. In the 72 deals where you held a value pair and nine or more fitting tiles, you won two flushes, where LuckyJ's rates predict about six. Without a value pair you already play ten or more like LuckyJ, and at nine or fewer neither of you commits often.

Rule Count the deal before your first discard. With four or fewer terminals and honors, plan tanyao: cut them before the middle tiles, and keep cutting them after turn 3. With nine or more tiles in one suit plus honors and a value pair, plan the flush: pon the pair and cut the off-suit tiles first. With ten or more and no pair, the flush is on as well.
Exception With five or more terminals and honors, or eight or fewer flush tiles, play for speed and riichi instead. LuckyJ wins tanyao in 3.5% of the five-count deals, and the flush in 2.4% of the eight-tile deals even with a value pair.

Evidence all 830 of your hands since January and 13,054 of LuckyJ's, sorted by the deal. Your three- and four-count deals hold 287 hands and 10 tanyao wins, where LuckyJ's rates predict about 24. The value-pair flush rows hold 72 of your hands. Treat the other rows as directions.

10

Chapter ten · Recent games · added 23 September

New

Your 22 September games, reviewed

You felt your tenpais had stopped converting and your deal-ins had jumped, with your rank points falling toward 4,500. Here are your last 20 games against the 64 before them:

Your games since JanuaryFirst 64 gamesLast 20 games
Tenpai hands that won40.8%42.7%
Hands won21.4%21.9%
Average win5,4076,118
Hands dealt in11.1%14.0%
Average deal-in4,4465,141

Your tenpais win as often as before, and your wins are bigger. Deal-ins rose by about one standard error, so the rise is not settled. It comes from pushing open tenpais more since chapter 5: 81% of the time when no safe tile keeps the tenpai against a riichi, up from 69%. Over all your games, your pushes with 3 or more han and 4 or more live tiles dealt in 23% of the time, the same as LuckyJ's. Seven deal-ins in your last 17 pushes is a bad run.

Rank points follow fourth place. Since January a first has paid you about +140, a second +66, a third −10 and a fourth −273. Three of your last eight games were fourths, where about two is normal for you. You have lost 12 points a game since January, and with fourths at 18% instead of 23% you would break even.

The four new games hold one decision to change and four hands that went wrong without a mistake. Those five hands are the examples below.

Rule In the last two hands, work out who passes you if you deal in. When a deal-in to the player just below you would swap your places, do not cut a live tile to climb one place: a climb from third to second is worth about 75 rank points, and a drop to fourth costs about 265.
Exception From fourth, a deal-in cannot cost you a place, so push the tenpai that can lift you one, as in the third example.

Evidence 84 of your games since January, 228 hands in the last 20. Rank points from your ranked-game records. Final hands of a game: 67 of yours and 1,007 of LuckyJ's.

11

Chapter eleven · Push calibration · added 24 September

New

Calibrate your tenpai pushes

You felt your wins and your losses were both climbing, and that you deal in to get the wins because a tenpai is dangerous if you don't fold. Your ten games on 23 September against the 84 before them:

Your games since January84 earlier games23 September
Tenpai hands that won41.3%49.2%
Hands won21.5%25.0%
Average win5,5966,741
Hands dealt in11.9%10.3%
Deal-ins at tenpai, per 100 hands4.73.4

Those were your best ten games: more wins, bigger wins and fewer deal-ins, for +42 rank points with three firsts and two fourths. Ten games is noise for any of these rates, but nothing in them says you push too much.

The pushes that matter come against a riichi, when every tile that keeps your tenpai is unsafe. You push 78% of those tenpais and LuckyJ 82%. On 23 September you pushed 17 of 19: six won, averaging 7,117, and three dealt in, averaging 5,367, for +1,329 per push. Here is what those pushes are worth for LuckyJ, by the hand's value and its live winning tiles:

LuckyJ, tenpai against a riichi3 or fewer live tiles4 to 78 or more
1 or 2 hanpushes 66%, −1,374 a push74%, −18189%, −126
3 or 4 han82%, +29392%, +89099%, +1,799
5 han or more86%, +1,33699%, +1,730100%, +4,565

A push's number is the average result of the hands LuckyJ pushed, with its wins, its deal-ins and every other ending. Read the table by han first. With 3 or more han, pushing earns in every wait. With 1 or 2 han it loses in every wait, least on a wide wait and most on a narrow one. You already fold four in five of the cheap, narrow tenpais, and in the other cells your push rates are close to LuckyJ's.

Your results follow LuckyJ's where most pushes happen: with 4 to 7 live tiles your pushes won 33% and dealt in 21%, and LuckyJ's won 33% and dealt in 22%. The weak spot is narrow waits, where you won 1 of 26 pushes against LuckyJ's one in five. Your narrow tenpais look like LuckyJ's, about 2.8 live tiles around turn 12, so most of that gap is luck. One part is not: when the tile you push is a live 2-to-8 or half suji, your hand ends in a deal-in 28% of the time against LuckyJ's 21%, the reading gap from chapter 8.

Late in the hand the numbers shift, because tenpai also pays at the draw. Going from noten to tenpai at the draw is worth 2,500 to 3,000 points. From turn 13, LuckyJ's cheap tenpais that fold lose 1,447 points on average, and the ones it pushes lose 306.

Rule When keeping tenpai against a riichi needs an unsafe tile, count your han first. With 3 or more, push. With 1 or 2, fold a narrow wait of 3 or fewer live tiles. Push a wider cheap tenpai late in the hand, when tenpai also pays at the draw, and fold it early.
Exception In the last hand, chapter 10's placement rule comes first. On the last discard, even a cheap or narrow tenpai is worth pushing a live tile, because the draw payment alone is worth 2,500 to 3,000, as in the first example.

Evidence 130 of your tenpais against a riichi where every tenpai-keeping tile was unsafe, and 1,505 of LuckyJ's, each hand counted at its first such turn. Your own cells hold 2 to 39 hands, so read your rates loosely; LuckyJ's table is the calibration.

12

Chapter twelve · Open hands · added 24 September

New

Keep the open tenpai, and open for value

You asked whether you should stop your cheap open tenpais. Keep them. When an open tenpai meets another caller, you already play it the way LuckyJ does, and folding it costs more than pushing it. Your open hands do lose about 500 points a hand to LuckyJ's, and most of that starts earlier, with the calls that build them.

Hands you openedYouLuckyJ
Share of all hands36.3%35.8%
Won29.6%30.2%
Average win4,4795,424
Dealt in14.7%10.6%
Dealt in to another open hand8.9%4.5%
Dealt in to a riichi5.0%4.2%
Points per hand+81+576

You open as often as LuckyJ and win as often. The wins are about 950 points smaller: your open wins carry 2.47 han against 2.79. The yakuhai are the same, 0.79 han a win, and you get less from shape yaku (0.66 against 0.82), dora (0.58 against 0.66) and red fives (0.43 against 0.51), so 40% of your open wins stop at 2 han against 27% of LuckyJ's. The extra deal-ins go to other callers: 8.9 per 100 opened hands against LuckyJ's 4.5, and 5.5 of them came from tenpai against 2.8. Against a riichi your open hands deal in about as often as LuckyJ's.

When your open tenpai can only be kept with a tile that is unsafe against another caller, and nobody is in riichi, LuckyJ keeps it 93% of the time. The table gives how often it keeps it and what the hands it kept averaged, by the tenpai's value and chapter 8's tells:

LuckyJ, open tenpai against a callerNo tellsOne tellTwo or three
1 hankeeps 98%, +67392%, −25590%, −28
2 han98%, +1,32496%, +79693%, +291
3 han or more98%, +3,66699%, +2,10896%, +1,173

Chapter 8 counted every cut made with a safe tile in hand, including safe tiles that kept the tenpai. This table counts only the turns where no safe tile keeps it. The cheap tenpais facing tells come out close to even when kept, and the few LuckyJ folded did worse: −565 a hand at 1 han and −825 at 2 han. Late in the hand the draw payment tips it further. From turn 13, its 1- and 2-han open tenpais that were kept gained 128 points on average, and the ones it folded lost 1,785. The exception is a tenpai with no yaku. It cannot win by ron, and LuckyJ keeps it only 62% of the time.

You keep 117 of these 121 tenpais, 97%, and you choose the tile the way LuckyJ does: the least dangerous of the tiles that keep tenpai 58% of the time against its 59%, the one with the most live tiles 77% against 79%. The result is where you differ. Your kept tenpais dealt in to the caller 14.5% of the time and LuckyJ's 8.2%. The callers were tenpai as often in your games (21%, 49% and 71% with no, one and two or three tells, against 20%, 48% and 74%), and your pushed tile was on their wait more often: 7.7% against 4.1% with one tell, 9.1% against 6.2% with two or three. Over 117 pushes that is about two and a half standard errors, so part of it is luck. The rest is the reading gap from chapter 8, and of the reads measured so far only one separates it: honors.

Against a caller, the honor that hits is a yakuhai with no other copy showing. It hit on 1.1% of LuckyJ's cuts of one while a caller was open, and on 1.3% of the other players' at its tables. Once one copy shows, the rate falls to 0.4 to 0.6%, and a guest wind hits 0.2 to 0.8% of the time. Yours hit 2.6% of the time, 5 of 192 cuts; the south in this chapter's last example and chapter 3's chun into two pons are two of them.

The larger part of the gap starts before the tenpai, with the call that opens the hand. From a closed 2-shanten in the first two rows, you and LuckyJ make calls with no yakuhai behind them, a tanyao chi or a pon for speed, about equally often. What those hands become is different:

First call from a 2-shanten, rows 1 and 2, no yakuhaiYouLuckyJ
Hands46565
Reached tenpai72%66%
Tenpais that won18%47%
Hands dealt in33%12%
Points per hand−1,657+401
Closed acceptance when calling28 tiles22 tiles

The calls do make you faster. The tenpais they make then win less than half as often as LuckyJ's, and a third of the hands end in a deal-in. The same spots go well when you play them another way: your yakuhai pons from a 2-shanten in the first two rows won 60% and averaged +1,604 a hand (LuckyJ 46%, +1,322), and the hands where you passed these calls averaged +233 (LuckyJ +218). You also make them from better closed hands than LuckyJ does, 28 tiles of acceptance on average against 22, so the hands you open are the ones most likely to reach riichi on their own.

LuckyJ decides these calls by what the call adds and by how slow the closed hand is:

LuckyJ calls from a closed handRow 1Row 2Row 3
2-shanten, a yakuhai pon95%93%(9 hands)
2-shanten, a tanyao call23%38%54%
1-shanten to tenpai, closed hand under 12 tiles75%75%82%
1-shanten to tenpai, 12 to 19 tiles50%56%72%
1-shanten to tenpai, 20 tiles or more15%17%38%

For a tanyao call from a 2-shanten in the first two rows, dora and shape decide. With two or more dora LuckyJ calls 59% of the time. With none or one it calls 38 to 47% of the time from a slow closed hand, under 24 tiles of acceptance, and 6 to 9% from a well-connected one with 40 or more. Roughly, a slow hand is built from kanchans, edge waits and loose tiles, and a well-connected one from two-sided shapes. From a 1-shanten, a closed hand under 12 tiles usually has a kanchan, edge or pair wait among its shapes, and one with 20 or more has more than two good shapes. When the open tenpai would have 3 or more han, LuckyJ calls 77% of the time, even in the first row. Your call rates are close to these overall, 42% against 46% from 2-shanten and 62% against 58% from 1-shanten. The shape is what you do not use yet.

Rule Keep an open tenpai that has a yaku against a caller, even at 1 han and even against two or three tells. Make fewer cheap open hands instead: from a 2-shanten, call a yakuhai pon, call for tanyao with two or more dora or from a slow closed hand, and keep a well-connected hand without dora closed.
Exception A tenpai with no yaku can only pay at the draw, so play it like a 1-shanten, and in the last hand chapter 10's placement rule comes first. Against a caller, a live yakuhai with no other copy showing is the riskiest honor to cut.

Evidence 995 of your hands and 13,054 of LuckyJ's. The tenpai table counts 121 of your open tenpais and 1,375 of LuckyJ's where every tile that kept tenpai was unsafe against another caller and nobody was in riichi, each hand at its first such turn; your own cells hold 1 to 53 hands, so LuckyJ's table is the calibration. The call tables count 314 of your chances to call from a closed hand and 3,723 of LuckyJ's, with nobody in riichi. Your 46 early calls sit about four standard errors from LuckyJ's deal-in rate and two and a half from its win rate. The honor rates come from 2,058 honor cuts at your tables and 27,510 at LuckyJ's.

13

Chapter thirteen · Recent games · added 27 September

New

Your defense against callers has caught up

You felt your defense against open hands had improved. Your 14 games from 24 to 27 September agree: every measure below moved to LuckyJ’s number or past it. Fourteen games are too few to settle any one row, but they all point the same way.

Defense against callersYour 94 games beforeYour 14 new gamesLuckyJ
Deal-ins to an open hand, per 100 hands6.95.34.8
Points lost to open hands, per hand269191211
Live cuts at 1-shanten with a safe tile in hand, two or more tells50.2%42.2%41.7%
The same with all three tells51.0%30.8%32.1%
Live cuts from 2-shanten or worse against a caller 30 to 60% likely to be tenpai, when a safe tile kept the shanten39.0%17.4%38.9%
Live cuts that dealt in to a caller, per 1002.251.041.26

Chapter 8 found the leak at 1-shanten: with a caller showing all three tells you still cut a live tile about half the time, and LuckyJ one time in three. In the new games you cut one on 4 of 13 such turns, and 42% of the time against two or more tells, which is LuckyJ’s rate. Your live tiles hit a caller 4 times in 383 cuts, and when a caller was tenpai your tile was on their wait 3.6% of the time, against 7.7% before and LuckyJ’s 5.2%. The first example below is the fold at its best: two turns running you paid acceptance for a genbutsu and kept a live [[F]] that was the caller’s winning tile.

The results followed: three firsts, five seconds, six thirds and no fourth, an average place of 2.21 against 2.46 before, and +330 points a hand against +31 (LuckyJ +363). Part of that is luck: at your earlier fourth-place rate, 14 games without a fourth happen about one time in 34. Your wins stayed small, 5,331 points on average, so the value gap from chapters 1, 6 and 7 is still open.

You dealt in 13 times in 152 hands. Ten were bad luck or close calls: suji that hit, riichi declarations and a tile forced after one, tenpai pushes, and an 11,600 push from 1-shanten against the dealer’s double east, the last example. The three riichi among them are LuckyJ’s habit too: at a first closed tenpai whose widest wait needed a live tile against a caller, you declared 14 times in 16 and LuckyJ 70% of the time. Three deal-ins were slips, the other examples: a live tile against three calls when a safe tile cost only three tiles of acceptance, a tile safe against a riichi but live against a pon when another was safe against both, and a drawn terminal thrown into the dealer’s riichi from a 4-shanten hand holding genbutsu.

One thing to watch comes before the line. When a genbutsu and a live honor or terminal are equally good for the hand and the caller is still early, LuckyJ cuts the live tile and keeps the genbutsu about three times in four: it spends the tile whose danger will only grow. Chapter 3’s rule said to cut the safe tile whenever two tiles are equal, which was too broad, and that sentence is now corrected. Your earlier games played these ties close to LuckyJ. In the new ones you took the safe tile in 12 of 24, too few to call a change, and when a caller later looked ready you were a little more often without a safe tile that kept your shanten: 23% of those turns, against 20% before and LuckyJ’s 17%.

Rule Past the line, keep folding the way these games did: the safe tile first, checked against every player who threatens. Where the line falls depends on what the safe tile costs. When it leaves the same hand as a live tile, the line is a single caller’s tenth discard or the seventh with two or three calls. When a live tile keeps a better hand, it is the 12th and the 9th from 2-shanten or worse, and the 14th and the 12th from 1-shanten. Before the line, when a genbutsu and a live honor or terminal are equally good, cut the live one and keep the genbutsu.
Exception When no safe tile keeps a 1-shanten, pushing through the line is still LuckyJ’s play: against a single caller it breaks the hand only about one time in eleven, as the last example shows. A dora pon moves the line to the first turns.

Evidence Your 108 games since January: 94 before 24 September (995 hands, 3,464 discards facing a caller) and 14 from 24 to 27 September (152 hands, 593 discards). LuckyJ: 1,079 Tokujou games. The new-games rows hold 45 and 13 cuts at 1-shanten, 23 in the 30 to 60% row, 24 ties and 103 late turns, so each differs from your earlier figure by one to three standard errors; the hit rate, 4 in 383 against 51 in 2,264, by about two. The line and the caller odds come from the book’s When to fold to open callers.

14

Chapter fourteen · Recent games · added 28 September

New

Count callers by rows, and keep the three-call tenpai

You felt you were losing it. Your 7 games since chapter 13, on 27 and 28 September, were rough: two firsts, a second, two thirds and two fourths, and a hand averaged −482 points. The defense chapter 13 measured held, though. You dealt in on 8.8% of hands, the same as the 14 games before, and only once to an open hand; at 1-shanten against two or more tells you cut a live tile 40% of the time, LuckyJ’s rate. What changed is the size of the deal-ins and how often you won:

Your games94 before 24 September14, 24 to 27 September7, 27 and 28 SeptemberLuckyJ
Hands won21.8%25.7%17.6%22.9%
Hands dealt in11.7%8.6%8.8%10.3%
Deal-ins to an open hand, per 100 hands6.95.31.54.8
Deal-ins to a riichi, per 100 hands3.82.05.93.8
Average deal-in4,6873,6857,1835,308
Points per hand+31+330−482+363

Seven games are 68 hands, 12 wins and 6 deal-ins, so none of this is a trend yet. Two of the six deal-ins were tenpai pushes: a 1-han open tenpai early in a hand, and a 4-han one from fourth in the last hand, where a deal-in could not cost you a place. A third came from a riichi you declared on a kanchan with no yaku after the dealer’s riichi; a second riichi followed, and a tile you then had to throw dealt in. One was a lone [[C]] cut on your fifth discard against an early pon, about when LuckyJ lets a lone dragon go once someone has called. The other two are this chapter’s examples: the West, and the dealer’s haneman, where the tile you pushed was the mistake. With the three-call hand, those are the three things you asked about.

Counting a caller. You said you lose track of a single caller past turn 9 and of their tsumogiri. Stop counting turns and count the rows of their pond, six tiles a row, plus one for any tile another player called from it. The line from chapter 13, the fourth tile of their second row against one call and the first tile of it against two or three calls, is where LuckyJ folds when the safe tile leaves the same hand as the live one. When the live tile keeps a better hand, LuckyJ folds later: from 2-shanten or worse at the last tile of their second row against one call and the third tile of it against two or three, and from 1-shanten at the second tile of their third row against one call and the last tile of their second row against two or three (chapter 19). A single caller was tenpai this often:

Single caller, their pondTenpaiAfter two tsumogiri in a row
First row6.5%9.5%
Second row, first three tiles22.4%27.7%
Second row, last three tiles35.6%47.3%
Third row43.9%68.7%

The row carries most of it. The tsumogiri only matters once they are past the line, and in the third row it lifts a single caller from about four in ten to two in three. Chapter 8’s other tell, a call made on their 9th discard or later, adds almost nothing once you count rows (38% against 42% in the second half of the second row), so drop it. To see a tsumogiri, watch whether their discard leaves from the drawn tile at the right end of their hand.

Three calls leave four tiles. After a riichi, LuckyJ kept its three-call tenpai on 345 of 352 turns, 98%, and on every one of the 198 turns where a safe tile kept it. Once you have three calls you cannot fold for long, so the tenpai is the only thing still working for you. In the first example you broke a tenpai on the dora when a suji tile would have kept it, and on the next turn cut a live tile anyway to stay 1-shanten. The third call itself was fine: when it reached tenpai, LuckyJ made it 94% of the time with nobody in riichi and 93% against a riichi when the tenpai had three han or more. Against a riichi it made it about 60% of the time with one or two han and 24% with no yaku.

The once-cut West. A copy in your own river is one copy showing, which chapter 4 rates at 2.1 deal-ins per 100 cuts against a riichi, where genbutsu has none. The West was a guest wind for you and kamicha’s seat wind, and on the turn you kept it over the [[7p]] it was also the more efficient cut. Your instinct was right: a lone honor that does nothing for you goes while nobody has declared.

Rule Count a caller by the rows of their pond. When a live tile would keep a better hand, the safe tile goes first from 2-shanten or worse at the last tile of a one-call caller’s second row (their 12th discard) or the third tile of it against two or three calls (their 9th), and from 1-shanten at the second tile of their third row against one call (14th) or the last tile of their second row against two or three (12th). When the safe tile leaves the same hand, it goes first earlier: at the fourth tile of their second row against one call (10th) and its first tile against two or three (7th). Two tsumogiri in a row past the line make them likely tenpai. With three calls and a riichi out, keep the tenpai with the safest tile that keeps it. Cut a lone honor you have no use for before anyone declares.
Exception Against a riichi, skip the third call when it leaves a tenpai with no yaku. Against a silent closed hand, two tsumogiri say little before the third row, because a closed tenpai usually declares: a silent dealer at their 10th to 12th discard was tenpai 8.5% of the time after two tsumogiri and 6.7% without.

Evidence Your 7 games of 27 and 28 September (68 hands) against the 14 before (152) and the 94 before those (995). LuckyJ: 1,079 Tokujou games. Three-call turns with a tenpai available after a riichi: 352 of LuckyJ’s and 28 of yours; you kept the tenpai on 24, and on 9 of the 10 where a safe tile kept it. Third-call chances that reached tenpai: 230 of LuckyJ’s with nobody in riichi and 85 with a riichi out. The row table counts 522,000 discards made against a caller with one call in LuckyJ’s and the Houou games, as in the book’s When to fold to open callers.

15

Chapter fifteen · Mortal review · added 28 September

New

Keep the pair you need, and fold the yakuless chase

You asked where the deal-in hands went wrong. Mortal, the open-source mahjong AI the book uses as a second opinion, replayed all 68 hands of your 7 games of 27 and 28 September and rated every choice you made: 762 discards and 201 chances to call. Its first choice was yours on 73% of those discards and 93% of the call chances. By the same count LuckyJ matches Mortal on 72.6% of its discards, and the Tokujou humans at its tables on 68.3%, so over these games your play sat as close to Mortal as LuckyJ’s own. The six deal-in hands are where you parted from it: 63% of their 60 discards matched, and 13 were choices Mortal gave less than a 10% chance.

Each hand and turn named in this chapter links to its replay, which opens the table at that moment with Mortal’s rating of every option. Mortal would have played three of the six deal-ins your way. Each of the other three comes back to one decision:

Hand and deal-inMortal on the tile that dealt inWhere it went wrong
East 4-2, 27 Sep 17:07
The dealer’s dama, 18,600
Turn 12: [[7p]] 2%, [[2m]] 97%The tile you pushed
South 2, 27 Sep 18:06
The West into a riichi, 8,000
Turn 10: [[W]] 93%Turn 9, when Mortal cut the [[W]] 97% and you cut [[7p]]
South 2, 28 Sep 01:38
Your riichi into the dealer’s, then a forced [[9s]] into a second riichi, 2,000 and your stick
Turn 15: forced by your riichiThe riichi on turn 11 (Mortal 8%, fold 68%), and the tanyao turns before it
East 2, 27 Sep 17:38
A 1-han open tenpai into a riichi, 2,600
Turn 6: [[6m]] 99%Nowhere
South 1, 27 Sep 18:06
A lone [[C]] into an early pon, 8,000
Turn 5: [[C]] 92%Nowhere
South 4, 27 Sep 18:06
A 4-han open tenpai from fourth in the last hand, 3,900
Turn 16: [[9s]] 100%Nowhere

The pon that took your pair. The three-call hand from chapter 14 is East 1, 1 honba, of your 01:38 game on 28 September. Mortal questioned it before its third call. With [[W]] ponned and [[9m]][[9m]] as the only pair in the hand, you ponned shimocha’s [[9m]] on turn 6. Mortal passed it 99% of the time. The pon kept the hand at 1-shanten, but it did not cost you ways to finish: 28 unseen tiles would now give tenpai instead of 10, and 10 of those 28 led to a 7,700 wait on the dora, such as a drawn [[6p]] making a dora pair with a [[3s]]-[[6s]] wait. Every tenpai on the pass side was worth 3,900. On offense your plan was sound. What the pon gave up was defense: 7 tiles in hand, 4 after the chi, and kamicha declared the next time around. Both AIs still lean against it. When a pon keeps an open hand at 1-shanten but adds 10 or more tenpai tiles, LuckyJ pons 21 times in 74, and 3 times in 19 with two dora or more in hand, as you had; its pons and passes there won about as often (38% and 36%). So the pon was a close call, and the costly turn in this hand was the tenpai you broke on turn 9. More broadly, LuckyJ passes 86% of the pons that would not bring an open hand closer to tenpai, and you pass 81%, so the habit is close. The pons LuckyJ does make there won 26 of 53 hands; yours since January won 1 of 7. These games had two more of them: Mortal passed your [[4m]] pon in East 1 of the 23:01 game (100%) and made your [[4p]] pon in East 4 of the 18:06 game.

The riichi with no yaku. South 2 of the same 01:38 game was your last hand to deal in. You led by 13,400, with the dora [[3s]] and the red [[5sr]] in hand and a [[1m]] pair keeping the hand from tanyao. Mortal steered for tanyao. It would have chied the dealer’s [[4s]] after your 8th discard (83%), and in the line you played it cut [[2m]] on turn 9 (98%) and [[1m]] on turn 10 (59%). You kept the pair, and on turn 11, just after the dealer declared, you reached tenpai on a [[6s]] kanchan with no yaku and declared into them. Mortal folded that tenpai with the genbutsu [[7p]] 68% of the time and declared 8%. Shimocha declared next, and the [[9s]] you had to throw on turn 15 dealt in. LuckyJ treats this spot as close: with a first tenpai that has no yaku and 4 live tiles or fewer against one riichi, it declared 58 times in 141, and those hands lost 584 points each. What tips your hand is the lead, since the riichi put first place at risk for a 3,900 hand.

The tile you push. In East 4, 2 honba, of your 17:07 game on 27 September, the dealer’s haneman, Mortal did what you did on turn 12 and played on at 2-shanten against a silent dealer. It pushed [[2m]] 97% of the time and [[7p]] 2%: the [[2m]] took one more tile and kept [[7p]] beside your red [[5pr]]. When you push against a silent hand, push the tile your hand needs least. Chapter 14 called this hand bad luck, and that verdict is corrected there.

What Mortal confirmed. Mortal kept chapter 14’s three-call tenpai on turn 9 every time, and in South 2 of your 18:06 game it cut the second [[W]] on turn 9 97% of the time, while nobody had declared. Once the riichi came it pushed the [[W]] again on turn 10 (93%), so the turn-10 cut that chapter 14 marked as a mistake was the right one; that example is corrected too.

Rule Pass a pon that leaves an open hand at the same shanten, unless it adds many tenpai tiles; even then it is close, and with two dora already in hand LuckyJ usually keeps the pair and its tiles for defense. When a terminal pair keeps a dora hand from tanyao, cut toward tanyao early so the tenpai has a yaku. Leading in the South, fold a tenpai with no yaku on a narrow wait to the dealer’s riichi. When you push against a silent hand, push the tile your hand needs least.
Exception A pon that takes a dora can be worth it even when it does not advance the hand: LuckyJ made 9 of 15. When you are not far ahead, a riichi on a narrow tenpai with no yaku stays a close call, and LuckyJ declared 49 of 121.

Evidence Mortal, the policy model the book runs locally, replayed your 7 games of 27 and 28 September decision by decision: 68 hands, 762 discards and 201 call chances. On LuckyJ’s 1,079 Tokujou games it rated 127,059 of LuckyJ’s discards and 382,042 of the humans’ at its tables. Pons: chances to pon a number tile with nobody in riichi, where the pon would not lower the shanten of a hand that already had a call, 378 of LuckyJ’s and 36 of yours since January; of LuckyJ’s, 74 were at 1-shanten with the pon adding 10 or more tiles that give tenpai. Riichi: 141 of LuckyJ’s first tenpais with no yaku and 4 or fewer live tiles against one riichi, 121 of them without an 8,000 lead.

16

Chapter sixteen · Mortal across your games · added 28 September

New

Save the safe tile until you need it

You asked for Mortal across more of your games, keeping only the mistakes LuckyJ would not make. So Mortal rated every discard in your 115 games since January, and every discard of LuckyJ’s the same way. Most of what it disagrees with in your games it disagrees with in LuckyJ’s too: you and LuckyJ match its first choice about seven times in ten, and the rest is style shared between you. Your efficiency, the order you cut honors, your calls and your pushes into a riichi all come out even with LuckyJ’s. One habit does not.

Sort the disagreements by what was on the table and the pattern appears. With nobody threatening, or one opponent with one call, Mortal quarrels with you exactly as often as with LuckyJ. With a riichi out, or an opponent showing two calls or more, it quarrels with you more, and almost always about the same thing: a safe tile was in hand, Mortal wanted the hand kept moving, and you folded. You do not push where LuckyJ would fold. You fold where LuckyJ would push, about twice as often as it does.

On the tableYour discards Mortal dislikedLuckyJ’s
Nobody threatening1 in 91 in 9
One opponent with one call1 in 81 in 8
A riichi1 in 81 in 9
An opponent with two calls or more1 in 51 in 8

Two calls, first row. Chapter 14 put the line against two or three calls at the first tile of the caller’s second row. You start defending before it. While a two-call opponent is still in their first six discards, you pay tiles or a shanten for a safe tile about twice as often as LuckyJ, and when a safe tile and a live one leave the same hand, you cut the safe one and LuckyJ keeps it. From their seventh discard on, you and LuckyJ play alike. This is new. Since 24 September, the weeks chapters 13 and 14 moved your defense against callers toward LuckyJ’s, your first-row folding has doubled. Against two calls, the move went past the mark. A live tile costs least while the caller is in their first row, and a genbutsu kept now is still genbutsu when they are closer.

A riichi, two dora, 1-shanten. Chapter 4’s exception stands: from turn 13 on, a 1-shanten folds to a riichi most of the time. Earlier than that, a hand carrying two dora or more is worth keeping. When the only safe tile would cost your 1-shanten before turn 13, LuckyJ keeps the hand about two times in three, mostly with a live tile, sometimes with a suji or an honor with two copies showing. You keep it less than half the time. With fewer dora the two of you fold alike; the gap is the dora hands, the ones worth the risk.

One smaller habit is worth a line. On about one tenpai in eight where Mortal would declare without a second thought, you stay dama, twice LuckyJ’s rate. It came up 21 times in 115 games, so it is a note, not a chapter.

Rule Against two calls, the first six tiles of their pond are still yours: keep the genbutsu on a tie, and pay neither acceptance nor a shanten for a safe tile until their seventh discard. Against a riichi before turn 13, keep a 1-shanten with two dora or more when only a shanten buys the safe tile.
Exception A dora pon moves the line against callers to the first turns, as chapter 13 found. Against a riichi from turn 13 on, chapter 4’s rule holds: fold the 1-shanten most of the time.

Evidence Mortal, the policy model the book runs locally, rated 13,101 of your discards in 115 games since January and 124,887 of LuckyJ’s in 1,079 Tokujou games, riichi declarations left out. Its first choice was yours 71.8% of the time and LuckyJ’s 72.6%. The table counts discards Mortal gave under 10%: yours 11.5, 11.9, 12.5 and 19.2 per 100 by row, LuckyJ’s 11.4, 12.7, 10.7 and 12.4. Pushes into a riichi where Mortal folds: 3.6 per 100 of your discards facing a riichi, 4.1 of LuckyJ’s. Where Mortal put 60% or more on one tile and a safer tile was in hand, you took a safer tile it gave under 10% on 18.4% of turns against a riichi (LuckyJ 9.3%), 15.2% against two calls or more (8.6%) and 7.8% against one call (8.0%). Two-call first rows, by behavior alone: a safe tile costing acceptance or a shanten was taken 27.6% (47 of 170) against LuckyJ’s 14.6%, 21.8% before 24 September and 43.5% since; on a tie the safe tile was cut 60.9% (39 of 64) against 36.7%; from the seventh discard on 27.8% against 25.6% and 70.0% against 63.5%. Riichi before turn 13, 1-shanten, two dora or more, the safest tile breaking the shanten: LuckyJ kept the hand 63% of 451 turns (48% with a live tile, 15% with suji or a 2-seen honor), you 45% of 69; with 0 or 1 dora, 49% and 42%. Dama where Mortal declared at 90% or more with nobody in riichi: 21 of your 176 such tenpais, 86 of LuckyJ’s 1,482. Each example below opens its hand in the replays.

17

Chapter seventeen · Recent games · added 29 September

New

Your safe tiles are in order; mind the early riichi

You felt you were winning more and losing less, and you worried that you were not keeping safe tiles at the start of the hand. The first holds over your 29 games since 24 September. The second does not show in your games: you keep safe tiles the way the AIs do, and the early deal-ins that worried you came from riichis declared before anyone has much that is safe.

Your games94 before 24 September29 sinceLuckyJ
Fourth places22%7%17%
Average place2.462.172.26
Rank points a game−10+39
Hands won21.8%24.0%22.9%
Average win6,1506,0937,317
Hands dealt in11.7%9.7%10.3%
Points lost to deal-ins, a hand546444546

Your last eight games, on the evening of 28 September, finished 3, 1, 2, 2, 2, 1, 2 and 1, for +683 rank points. You dealt in 10 times in 80 hands, more than your average, but Mortal would have cut the same tile in seven of those ten, and two came after riichi declarations of your own. One was a mistake, the first example below. Your wins that evening averaged 7,967 points; over the 29 games they still average about 1,200 below LuckyJ’s, the gap chapters 6 and 7 describe.

Leftovers. The main book’s safe-tile section counts one choice: two lone tiles of the same kind, one already in an opponent’s river, and which of them goes first. Measured the same way on your 123 games, with nobody threatening, you throw the live terminal and keep the one that is out 71% of the time, and LuckyJ at the same discards 71%. With middle tiles you throw the live one 39% of the time against LuckyJ’s 48%, inside the noise. When a riichi comes, you hold what LuckyJ holds: 1.84 genbutsu on average against its 1.90 at the same point in the hand, and none at all 17.8% of the time against 18.1%.

Honors. This is where you and LuckyJ part, and where Mortal sides with you. With two guest winds, LuckyJ throws the one already out on its first two discards and keeps it from the third; you throw the live one a little under half the time throughout. With a live value honor, a dragon, your seat wind or the round wind, next to a guest wind already out, LuckyJ keeps the value honor for about three discards and then lets it go first. You keep it about a discard longer. Mortal, on the same turns, plays your way:

The live value honor goes first, by your discard123456
LuckyJ2%11%35%60%71%72%
You0%3%14%35%54%66%
Mortal, on your turns5%10%15%20%32%36%

Over all 203 of those turns you threw the value honor 14% of the time, Mortal 13% and LuckyJ, in the same spots, 25%. With two guest winds Mortal threw the live one about as often as you did. Two strong AIs disagree here, so the order of your honors is a style to choose, not a leak.

The early riichi. A riichi by the declarer’s fourth discard finds you without a genbutsu 43% of the time, and LuckyJ too. The question then is which live tile to give. Against a riichi met within its first six discards, LuckyJ’s cuts of a live honor with no other copy showing dealt in 3.8 times per 100, and with one copy showing 1.1. In the 21:54 game on 28 September, East 1, kamicha declared on their third discard and you cut the East, the round wind with no copy out, from a 3-shanten; kamicha was waiting on a pair of it. Mortal cut the 7s, suji of their first discard, 67% of the time and the East 1%. In the 18:16 game, East 2, the South that met toimen’s pair was Mortal’s own first choice.

Rule Keep your leftovers as you do now: the live terminal or middle tile goes first and the one already out stays. When a riichi comes before you have a genbutsu and the tiles you could give are otherwise close, give suji of their discards or an honor with a copy showing before an honor with none.
Exception The order of your honors is style. If you want LuckyJ’s, throw the guest wind that is already out on your first two discards. On the third, throw a live guest wind if you have one; if not, the guest wind that is out still goes, and the value honor stays one more turn. From the fourth, both live honors go before the guest wind that is out, which is the one you keep. Mortal keeps the value honor longer, as you do.

Evidence Your 123 games since January against LuckyJ’s 1,255, the leftover choice measured in hands you did not deal, before your own riichi, with nobody in riichi and nobody on two calls, as in the main book’s section; run on LuckyJ’s games, the same code gives that section’s counts exactly. LuckyJ’s rate at each discard is a logistic curve fitted the section’s way, and yours is compared with it discard by discard. Terminals: 57 of your 80 choices went to the live one (LuckyJ at the same discards 71.3%); middle tiles 28 of 71 (48.2%); two guest winds on your first two discards 14 of 32 (23.2%), from the third 6 of 13 (67.9%); a live value honor against a guest wind already out, 29 of 203 (25.3%, about four and a half standard errors). LuckyJ’s order by the lone honors it held, with nobody threatening: on the third discard a live guest wind went first 59% of the time against one already out (22 turns), 79% against a live value honor (327) and 55% with all three (33); against a live value honor, the guest wind already out went first 63% of the time on the third discard (285) and 40% on the fourth (225). Mortal’s share, from the site’s replays, is the weight it put on the live tile of the two: 46% and 39% for the guest winds, 13% for the value honor. Riichi declared while you were not tenpai: 490 of yours, against LuckyJ’s 6,668 at the same declarer discard; by their fourth discard, no genbutsu 42.9% (LuckyJ 42.6%). LuckyJ’s cuts against one riichi in its first six discards: 105 live honors with no copy showing, 180 with one. The table at the top counts 995 and 300 of your hands and LuckyJ’s 11,272 hands in 1,079 Tokujou games.

18

Chapter eighteen · Reading callers · added 29 September

New

One grid to read any caller

You asked for the whole picture: how the chance that a caller is ready moves with their discards and with what they show. The grid below is that picture. For any caller at your table it gives one number: of the callers in LuckyJ’s games who were in the same spot, the share who could win off your next discard, meaning tenpai with a yaku on at least one wait. It reads 88,588 such moments, each caller once after every discard they made while nobody was in riichi. Three counts find the spot, and the switches above the grid turn it into what LuckyJ did there.

Squares show
LuckyJ’s hand
Calls, and their latest discardsFirst rowSecond rowThird row
123456789101112131415161718
One callLuckyJ folds from here
last tile from hand1124711162126303234353433323028
last one from the wall135812182430353942454647474747
two in a row from the wall24712182532384448515558616466
three or more from the wall811172432414955616569717476
Two callsLuckyJ folds from here
last tile from hand69142129364247515253535150474542
last one from the wall152129374451555860616160595857
two in a row from the wall2027354351576366686969696867
three or more from the wall28354250576369737678808283
Three or four callsLuckyJ folds from here
last tile from hand5766737882848483817874
last one from the wall627178838586868584
two in a row from the wall80828486888991
three or more from the wall80838587899091939494
Calls, and their latest discardsFirst rowSecond rowThird row
123456789101112131415161718
One callLuckyJ folds from here
last tile from hand781013161923283339455157
last one from the wall81013162025313743505764
two in a row from the wall12151923273339
three or more from the wall262932364043
Two callsLuckyJ folds from here
last tile from hand7111623303843475052
last one from the wall24273135394347
two in a row from the wall4145
three or more from the wall6266
Three or four callsLuckyJ folds from here
last tile from hand
last one from the wall
two in a row from the wall
three or more from the wall
Calls, and their latest discardsFirst rowSecond rowThird row
123456789101112131415161718
One callLuckyJ folds from here
last tile from hand2828272828293033364045505662
last one from the wall1518212528313334343332
two in a row from the wall242629313437394245
three or more from the wall15182126303541465258
Two callsLuckyJ folds from here
last tile from hand3436373941424446485051
last one from the wall2932343740434649
two in a row from the wall3631
three or more from the wall454750525557
Three or four callsLuckyJ folds from here
last tile from hand
last one from the wall
two in a row from the wall
three or more from the wall
0%10%25%50%75%90% Share of callers who could win off your discard (tenpai with a yaku), after their discard number at the top, in LuckyJ’s games. A blank square has fewer than 40 readings. The green line is LuckyJ’s fold line: from two-shanten or worse, choosing between a safe tile and a live one at least as good for its hand, it cut the live tile less than half the time from the twelfth discard against one call and from the ninth against two or more (chapter 19).0%10%25%50%75%90% Share of LuckyJ’s real choices on which it folded, from two-shanten or worse, against a single caller with nobody in riichi, in its 1,079 Tokujou games: a safe tile (the caller’s genbutsu, a full suji, or an honor with two showing) kept its shanten, and so did a live tile with at least as much acceptance, and it threw the safe tile. Turns where the safe tile was the best tile for the hand anyway, or the only one that kept the shanten, are left out. A blank square is one LuckyJ met too rarely to pin the share within 12 points either way; against three or more calls, that is every square. The green line is where the share passes half, all runs together: the twelfth discard against one call and the ninth against two or more. When every safe tile would have cost a shanten, LuckyJ threw one on 4.0% of 1,682 turns.0%10%25%50%75%90% Share of LuckyJ’s real choices on which it folded from one-shanten, counted the same way as from two-shanten or worse. It starts a little higher, because early ties between two equally good tiles go to the safe one more often from one-shanten, then rises more slowly than the far hand’s and passes half later. A blank square is one LuckyJ met too rarely to pin the share within 12 points either way; against three or more calls, that is every square. The green line is where the share passes half, all runs together: the fourteenth discard against one call and the twelfth against two or more. When every safe tile would have cost the one-shanten, LuckyJ threw one on 4.8% of 2,663 turns. At tenpai LuckyJ met a real choice only 248 times against a single caller, too few for the grid; it took the safe tile on 55% of them, and broke a tenpai for safety on 3.5% of the 1,521 turns where every safe tile would have broken it.

The words. A call is a pon, a chi or an open kan. A tile from the wall is the tile a caller has just drawn, thrown at once. A tile from hand is one they held before the draw. Their run is how many tiles from the wall they have thrown in a row, counting back from their newest discard and stopping at a tile from hand or at their last call. Chapter 19 has the fine print, including the tile thrown with a call.

Finding the square. Three counts:

  • Their calls: one, two, or three and four. That picks the block of four rows.
  • Their discards, the tiles thrown with calls and the ones other players took included. That picks the column. The river lays them out six to a row, and the headings across the top follow those rows: the first row is discards 1 to 6, the second 7 to 12, the third 13 to 18.
  • Their run: none, one, two, or three or more. That picks the row inside the block.

The square where the row meets the column is the share who could win. The green line is LuckyJ’s fold line: from that discard on, LuckyJ at two-shanten or worse, choosing between a safe tile and a live one at least as good for its hand, threw the safe one more often.

One caller, three squares. In your game of 28 September, East 4, toimen ponned [[P]] on their fifth discard and threw [[2m]] with it. Here are three of your turns against them, the counts, the square, the same square with the switch set to LuckyJ folded from two-shanten or worse, and what they really held.

Your turnTheir river since the ponThe countsCould winLuckyJ foldedWhat they held
6th[[2m]] with the call, then [[3s]] and [[8m]] from the wallOne call, 7th discard, two in a row from the wall18%23%1-shanten
8ththen [[5p]] and [[1p]] from handOne call, 9th discard, last tile from hand26%33%Ready on [[F]] or [[C]], a mangan
10ththen [[5s]] and [[8p]] from the wallOne call, 11th discard, two in a row from the wall44%Too few to readReady

At your eighth turn you were two-shanten. You ponned that [[1p]] and cut [[8s]], which nothing in their river covered. Toimen was already waiting on [[F]] and [[C]] with the [[P]] pon, a small three dragons, and drew it two turns later for a mangan. The square said 26%, one caller in four, and the fold view says that from two-shanten or worse LuckyJ took the safe tile there on 33% of its real choices, so when a safe tile was no better for the hand, cutting the live one at that square was LuckyJ’s usual play too. Open the hand at your sixth turn, your eighth or your tenth.

The switches. The upper switch changes what the squares show. Callers who could win is the reading above. LuckyJ folded shows, for the same square, how often LuckyJ threw the safe tile on its real choices against a single caller with nobody in riichi: a safe tile kept its shanten, and so did a live tile with at least as many tiles to improve the hand. Turns where the safe tile was the best tile for the hand anyway, or the only one that kept the shanten, are left out, since throwing it there is shape, not defense. The lower switch picks LuckyJ’s own hand, because the answer depends on it. Each hand draws its own green line where folding passes half: the twelfth discard against one call and the ninth against two or more from two-shanten or worse, the fourteenth and the twelfth from one-shanten. At tenpai LuckyJ met such a choice too rarely against a single caller to fill the grid. At toimen’s ninth discard above, both fold squares read 33%. A blank fold square is one LuckyJ met too rarely to read; against three or more calls, that is every square.

Landmarks. One call in its first row can win 12% of the time or less, whatever it throws. At the end of its second row, a one-call hand can win 34% of the time with its last tile from hand and 55% after three or more from the wall. Two calls pass half by their tenth discard even with the last tile from hand, and by their ninth after two or more from the wall. Three calls can win 57% of the time at the first discard the grid can read, and about 80% or more from their ninth.

The grid fits your own tables. The other three players in your 123 games could win 32.2% of the time where these squares predict 31.6%, over 6,472 readings, so what you see at Jade is what LuckyJ sees at Tokujou.

Rule Before you cut a live tile against a caller, find their square with the three counts: calls, discards, run. Then set LuckyJ’s hand to your own, two-shanten or worse or one-shanten, and see how often it folded there. The fold lines of chapters 13, 14 and 16 are drawn on this same grid.
Exception One call with the last tile from hand, late in the hand. Many of these callers are tenpai without a yaku, playing for the draw payment, so they are ready but cannot win off your tile. Counting plain tenpai, that row keeps climbing to 46% by the eighteenth discard, while the share who can win falls back to 28%. The grid shows the share who can win, which is what a live tile costs you.

Evidence Callers in LuckyJ’s 1,255 games, LuckyJ and its tablemates alike, read once after each of their own discards while nobody was in riichi, with their real hand: tenpai, and tenpai with a yaku. Each row of the grid is a logistic curve over the discard number, fitted the way the main book’s safe-tile section fits its rates, over the discards with at least 40 readings each; a blank square has fewer. Readings per row, one call: 34,774 with the last tile from hand, then 13,841, 6,258 and 6,031 for runs of one, two, and three or more; two calls: 10,538, 5,527, 3,112 and 4,194; three or four calls: 1,512, 950, 618 and 1,233. A single call’s first row, 6.5% over discards 1 to 6, matches chapter 14. At your tables the grid’s squares average 31.6% over the 6,472 readings of your opponents, and they could win 32.2%. The fold views are LuckyJ’s real choices in its 1,079 Tokujou games, 6,091 from two-shanten or worse and 4,310 from one-shanten, with each tile’s acceptance from contrast/choices.py; the worked hand is East 4 of your game of 28 September, 01:38.

19

Chapter nineteen · Reading callers · added 29 September

New

Where LuckyJ folds on the grid, and what counts as a run

You asked three things about chapter 18’s grid. Where does LuckyJ stop pushing? What does it mean when a caller mixes tiles from hand with tiles from the wall? And do the guest winds and terminals a caller draws and throws count toward their run? The short answers: LuckyJ folds from a one-call caller’s twelfth discard and a two-call caller’s ninth, or from their tenth and seventh when the safe tile costs its hand nothing, and the grid now draws the first pair of lines in green. A mixed river is read by its last tile, the way the grid already does, with two adjustments given below. And yes, every tile from the wall counts, honors and terminals included.

The words first. A tile from the wall is the tile the caller has just drawn, thrown at once. A tile from hand is one they were holding before the draw. Their run is the number of tiles from the wall at the end of their river, counted back from their latest discard until a tile from hand or a call. A player who calls pon or chi does not draw: they take the tile and throw one from hand straight away. That tile, the one thrown with the call, is a discard like any other and counts in their discard number, and the run starts again from zero after it. The grid has a block for one call, for two and for three or four, with four rows in each: last tile from hand, last one from the wall, two in a row from the wall, three or more. This chapter calls those the first to fourth row of a block. The river’s rows of six are the columns across the top.

The line. It comes from LuckyJ’s real choices. LuckyJ is two-shanten or worse, one player has called, nobody is in riichi, and LuckyJ holds a safe tile it can throw without falling further from tenpai, and a live tile it can throw just as well, with at least as many tiles to improve its hand. The question is how often it still cuts the live one as the caller’s river grows. Through a one-call caller’s sixth discard it cut the live tile 80 to 94% of the time. Then the share falls: 66% at their ninth discard, 53% at their eleventh and 47% at their twelfth. Against two or more calls it was 67% at the sixth, 53% at the eighth and 48% at the ninth. The green line sits where the share first drops below half: before the twelfth discard against one call, before the ninth against two or more. Chapters 13 and 14’s tenth and seventh discards hold for ties, when the safe tile and the live one are equally good for the hand: from there LuckyJ throws the safe tile first. When the safe tile would cost tiles that improve the hand, LuckyJ keeps the better tile longer, and that puts the line for all real choices two discards later. Turns where the safe tile was the best tile for the hand anyway are left out, because throwing it there is shape.

In play: your kamicha ponned on their fourth discard and is now throwing their tenth. You are two-shanten and hold a [[W]] they have already thrown; cutting it leaves you two-shanten with as many tiles to improve your hand as cutting your live tile would. Cut the [[W]]: on a tie like this LuckyJ throws the safe tile first from the tenth discard. Had the [[W]] been part of your shape, so that cutting it cost you tiles, LuckyJ would still have cut the live tile on most turns through the second row. Had they called twice, the tie would have tipped at their seventh discard.

Against one call the line does not bend with the run: at the same discard, LuckyJ cut the live tile as often against a caller with three or more tiles from the wall as against one whose last tile came from hand. Against two or more calls it does: at the same discard, each tile the caller had thrown straight from the wall made LuckyJ fold about 4 points more often. At tenpai LuckyJ met a real choice too rarely against a single caller, 248 times, to say what the run does there.

One river, square by square. In your game of 28 September, East 1, toimen called three times and was ready by the end. Here is how the grid reads them after each discard, next to what they really held. Their discard number is their place in the river, the tile thrown with each call included.

Their discardWhat they threwHow to count itThe grid saysWhat they held
5thPonned [[C]], threw [[1m]]One call. The tile thrown with the call is from hand, but the call counts as one tile from the wall: read the second row8%Two-shanten
6th[[9m]] from the wallOne from the wall, plus the call: third row12%Two-shanten
7th[[8s]] from the wallTwo, plus the call: fourth row17%Two-shanten
8th[[S]] from the wallA guest wind, and it counts: fourth row24%Two-shanten
9th[[2p]] from handFrom hand after three from the wall: between the first row (26%) and the second (30%)28%Two-shanten
10thChied [[3p]], threw [[8p]]Two calls. Right after the call: second row of the two-call block55%1-shanten
11th[[2p]] from the wallOne from the wall, plus the call: third row63%1-shanten
12th[[4p]] from handFrom hand after exactly one from the wall: the first row, or a little under53%1-shanten
13th[[9p]] from the wallOne from the wall: second row. The call no longer counts, since a tile from hand has come after it61%1-shanten
14th[[7s]] from the wallTwo in a row: third row69%1-shanten
15thPonned [[6p]], threw [[4s]]Three calls, right after the call: second row of the last block. That square is blank, and the first row says 74%74% or moreReady, waiting on [[3m]]

Toimen was not ready until the third call, and the grid never claimed they were likely to be: 28% at their ninth discard, 53 to 69% through the two-call stretch. You were two-shanten from your seventh turn to your eleventh, and from your ninth turn every tile you threw, [[S]], [[S]] and [[1m]], was already in their river. Toimen had one call through their ninth discard, and LuckyJ’s line against one call sits at the twelfth, so those safe tiles went earlier than LuckyJ’s line. The [[S]] at your ninth turn also gave up a step toward tenpai: cutting either of two live tiles would have made you one-shanten. A second player had called by then too, and the grid, which reads one caller at a time, does not count that. Open the hand at your sixth turn, just after the pon, at your tenth, facing their ninth discard, or at your fifteenth, when they were ready.

Hand and wall in turn. A caller who alternates, hand, wall, hand, wall, shows nothing extra. Callers whose last four discards alternated could win 35% of the time after a tile from the wall, where their squares said 36%, and 29% after a tile from hand, where the squares said 31%. Read them by their last discard, as the grid does.

What does move a caller off their square is the tile before their last tile from hand. The first row of every block, last tile from hand, mixes callers in quite different spots, and the table below sorts them by what came just before. Each line gives the share who could win, then what the first row and the second row of the grid said at the same discards.

Their last tile came from hand. Before it came…Could winFirst row saysSecond row says
Nothing yet: their last tile is the one thrown with the call33%29%33%
Three or more from the wall41%37%44%
Exactly two from the wall31%31%36%
Exactly one from the wall25%27%32%
Another tile from hand, or the tile thrown with the call20%25%29%

Read it from the bottom up. A tile from hand right after another tile from hand, or right after the tile thrown with the call, leaves a caller about five points under the first row: they are still rearranging. A tile from hand after exactly one from the wall reads a couple of points under the first row, and after exactly two it reads on the first row. After three or more from the wall, a tile from hand takes them only part of the way back: read them between the first two rows, as toimen at their ninth discard above. The top line is the reading right after a call, with nothing thrown since. Those callers could win 33%, which is what the second row says, not the first. That is why the call counts as one tile from the wall, and each tile from the wall after it moves them one more row down, until a tile from hand comes.

Honors and terminals from the wall. They count, and they should. A caller who draws a guest wind and throws it has told you the same thing as one who draws and throws a [[5s]]: the tile they drew was worth less to them than anything they held. The grid counts every tile a caller draws and throws, and a grid that passes over their guest winds and terminals fits the real hands worse. One tile from the wall moves a caller about the same whatever it is, and so do two. Toimen’s [[S]] at their eighth discard above was their third tile from the wall, and it counted. The one shortfall is a run of three or more made of nothing but honors and terminals: it reads like a single tile from the wall, and it is 5% of the runs that long.

Their run from the wallCould winTheir squares say
One, a value honor36%34%
One, a guest wind33%31%
One, a terminal30%32%
One, a 2 or 830%32%
One, a 3 to 735%34%
Two, both honors or terminals40%40%
Two, one of them a 2 to 843%42%
Three or more, all honors or terminals40%50%
Three or more, one of them a 2 to 860%60%
Rule From two-shanten or worse, when a safe tile is as good for your hand as the live one, throw it first from the tenth discard against one call and the seventh against two or more. When it would cost tiles that improve your hand, keep the better tile until the green line, the twelfth against one call and the ninth against two or more, a little sooner against two calls after a run from the wall. To find a caller’s row, count the tiles from the wall at the end of their river, guest winds and terminals included, and add one for the call if nothing from hand has come since it. When their last tile came from hand, look at the one before it: after three or more from the wall, read them between the first two rows; after another tile from hand, a little under the first.
Exception A run of three or more made only of honors and terminals reads like a single tile from the wall. And the call counts as a tile from the wall only until something comes from hand: if their first tile after the call comes from hand, the tile thrown with the call counts as one from hand too, and they read a little under the first row.

Evidence The line: LuckyJ’s own discards in its 1,079 Tokujou games against a single caller with nobody in riichi, from two-shanten or worse, where a safe tile kept the shanten and so did a live tile with at least as much acceptance, 4,500 against one call and 1,591 against two or more; 2,798 of them were ties. Acceptance comes from contrast/choices.py over the same games. Left out: 4,454 turns where the safe tile was the best tile for the hand anyway and 88 where only safe tiles kept the shanten. Each share is a logistic curve over the caller’s discards, fitted over the discards with at least 40 turns each. Adding the run to that curve moved the log-odds of a live cut by −0.02 ± 0.05 a tile against one call and −0.15 ± 0.06 against two or more. The patterns come from the grid’s 88,588 readings, each set against its own square. The table of tiles from hand counts only squares where every row of the block can be read: 14,033 readings right after a call, then 1,722, 2,120, 4,793 and 11,874. A grid that skips guest winds and terminals in the run has a log-likelihood 25 lower on the same readings. At your own tables, over your opponents’ 6,472 readings, 35% could win right after a call where the first row said 29%, and 42% after a tile from hand that ended three or more from the wall, where it said 38%. The worked river is toimen’s in East 1 of your game of 28 September, 23:30, read with the same code as the grid.

20

Chapter twenty · Reading callers · added 29 September

New

LuckyJ breaks its hand for a riichi, hardly ever for a caller

You asked when LuckyJ goes back a shanten to fold. You also pointed out that a safe tile drawn from the wall and thrown straight back costs nothing, so it should not count. It does not count here. The short answer: LuckyJ gives up a shanten for a riichi, and almost never for callers. When every safe tile would leave its hand a shanten short, it threw one 18% of the time against one riichi with tenpai in reach, 53% with 1-shanten and 80% with 2-shanten or worse. Against one caller the same choices came out 3%, 5% and 4%. Against callers, how often it breaks follows their square from chapter 18 and what the hands are worth.

The words. To break your hand is to throw a safe tile that leaves it a shanten short of the best hand this draw allows. The mark is that best hand. A safe tile that reaches it is a free fold, and throwing back a safe draw that does nothing for your hand is one, so it never counts. A hand can fall short in two ways. The safe tile can take it back from where it stood before the draw, or the draw can offer a step forward that only a live tile takes, which the safe tile passes up. Both are the same choice: the live tile gets you the best hand, the safe tile leaves you a shanten short. A safe tile is genbutsu, a full suji or an honor with two showing, against every caller or against the riichi. Each of your turns facing a threat is one of three kinds, filed under the best hand the draw allows:

  • Free. A safe tile reaches the best hand. Most turns are this kind: against one caller, 66% of LuckyJ’s turns with 1-shanten in reach and 75% with 2-shanten or worse.
  • Breaking. Every safe tile falls a shanten short, by going back or by passing up the step. This is the choice the chapter measures.
  • No safe tile at all.

When every safe tile would leave the hand a shanten short, this is how often LuckyJ and you broke it, by the best hand the draw allowed:

FacingWhoTenpai1-shanten2-shanten or worse
One riichiLuckyJ18.3%53.4%79.5%
One riichiYou28.4%64.7%83.1%
One callerLuckyJ3.3%4.9%4.1%
One callerYou2.0%7.6%7.6%
Two or more callersLuckyJ3.9%11.2%19.8%
Two or more callersYou1.5%9.4%16.7%

The square sets the rate. Against callers, find their square the way chapter 18 does. With two callers, use the chance that either of them can win: squares of 40% and 30% make 58%, because neither can win 60% of 70% of the time, which is 42%. The chart gives LuckyJ’s share of breaks over that number.

How often LuckyJ broke its hand to fold

Turns where every safe tile would leave LuckyJ’s hand a shanten short of the best this draw allows, filed under that best hand. Each dot is one whole percent of the callers’ square, sized by its turns; the lines are fitted curves.

2-shanten or worse1-shantenTenpai95% range of each fitthe same hands against one riichi

0% 25% 50% 75% 100% 0% 25% 50% 75% 100% Callers’ square: the chance one can win off your tile One riichi 80% 53% 18%
LuckyJ’s draw turns in its 1,079 Tokujou games, nobody in riichi for the callers’ side. A tile is safe when it is genbutsu, a full suji or an honor with two showing against every caller. The square is chapter 18’s; with two or more callers, the chance that at least one of them can win. Each curve is fitted between the 2nd and 98th percentiles of its turns’ squares.
The fitted shares in a table
Square2-shanten or worse1-shantenTenpai
10%1.3%1.4%1.8%
20%4.2%2.4%2%
30%10.2%4%2.4%
40%17.3%6.1%2.7%
50%22.8%8.6%3.1%
60%26.5%11.3%3.6%
70%29.9%13.9%4.2%
80%33.7%16.4%4.8%
90%37.7%18.9%5.5%
One riichi79.5%53.4%18.3%

With 1-shanten in reach, the fitted share is 1% at a square of 10%, 9% at 50% and 16% at 80%. With 2-shanten or worse it climbs faster, 4% at 20% and 23% at 50%, and reaches 34% at 80%. With tenpai in reach it stays under 6% at every square. A caller at 80% still gets far less respect than a riichi. Part of the reason is the price: in LuckyJ’s games a ron into an open hand cost the discarder 4,525 points on average, and a ron into a riichi 6,683.

What the hands are worth. At the same square, value moves the rate both ways. On squares of 50% or more:

At a square of 50% or more1-shanten, brokeTurns2-shanten or worse, brokeTurns
All turns14%1,08230%336
The caller shows three or more dora in their calls39%5235%20
The caller is the dealer19%51435%198
LuckyJ holds no dora21%34634%121
LuckyJ holds two or more dora9%30617%64

Holding the square fixed, a caller with three dora in their calls multiplied LuckyJ’s odds of breaking a 1-shanten by three and a half, and a dealer caller raised them by about 60%. Each dora in LuckyJ’s own hand cut the odds by about a quarter. A caller who is expensive makes it fold; a hand of its own that is worth something makes it stay.

The two forms. Inside that choice LuckyJ passes up a step more readily than it goes back. With tenpai in reach, it threw the safe tile on 10% of the steps and 1% of the backs against one caller, and on 31% and 14% against one riichi. Going back, the live tile it would push is the one it just drew; passing up a step, it is a tile it has been holding.

Your games. Against callers you break your hand a little more often than LuckyJ, and the extra sits on quiet squares. On the same kind of turns you broke 33 times with 1-shanten in reach where LuckyJ’s curve expects 26, 23 times with 2-shanten or worse where it expects 17, and 4 times with tenpai in reach where it expects 7. Below a square of 40%, with 1-shanten or worse in reach, you broke 24 times where the curve expects 12. Mortal agrees with LuckyJ. On those turns of yours below a square of 40% it put 2 to 4% of its weight on the safe tiles; above 40%, 10% with 1-shanten in reach and 15% with 2-shanten or worse. Against a riichi you break more often than LuckyJ with tenpai in reach, 28% of 148 turns against 18%, about three standard errors apart.

Three of your turns. In each, every safe tile would have taken your 1-shanten back to 2-shanten.

Your turnThe callersSquareYou threwMortal
26 Sep, East 2, turn 13Kamicha, the dealer, one pon of [[C]], 13th discard, last tile from hand35%[[1s]], a suji[[3m]] or the [[8s]] you drew, both live, 99.8%
27 Sep, East 2, turn 8Kamicha, three calls with the [[E]] pon, 8th discard, one from the wall78%[[S]], genbutsu[[6s]], live, 98.7%
27 Sep, East 2, turn 11Kamicha, the dealer, a [[C]] pon; shimocha, one chi, five in a row from the wall60%[[N]], genbutsu to both[[N]], 99.3%

The first caller was quiet and not yet ready, and LuckyJ’s curve breaks a 1-shanten there about 5% of the time. The second looked loud, but you held two red fives, and Mortal kept the hand; the caller won 2,000 from toimen. The third is a turn to break: two callers, one of them the dealer, both ready, and the [[S]] you had just drawn was the dora and one of shimocha’s winning tiles. Open them at your 13th turn on 26 September, your 8th on 27 September and your 11th the same day.

Rule When every safe tile would leave your hand a shanten short of the best this draw allows, keep the hand against callers and cut the live tile that gets there. With tenpai in reach, do not break for callers. With 1-shanten or worse, break only once the square is past half and the caller looks expensive, the dealer or dora in their calls, while your own hand is cheap. Against a riichi it flips: break about half the time with 1-shanten in reach, and break with 2-shanten or worse.
Exception The end of the hand. With fewer draws left, a 1-shanten is worth less: at the same square, LuckyJ’s odds of breaking one grew about fivefold over ten of its turns.

Evidence LuckyJ’s draw turns in its 1,079 Tokujou games, and yours in your 123 games since January, read with the same code; a turn that starts with a call is left out. Against callers nobody was in riichi. Turns where every safe tile fell a shanten short of the best hand, against one caller and against two or more: 1,300 and 544 with tenpai in reach, 2,545 and 956 with 1-shanten, 1,643 and 444 with 2-shanten or worse. Against exactly one riichi: 1,310, 1,583 and 672. The chart’s curves are logistic curves over the square, one point per whole percent, fitted between the 2nd and 98th percentiles of the squares, over the 1,797, 3,457 and 2,057 turns whose callers were all on the grid. The value effects come from one logistic regression per hand on the square, LuckyJ’s own turn, a caller with three or more dora in their calls, a dealer caller, LuckyJ’s dora and LuckyJ as dealer. With 1-shanten in reach the log-odds moved +1.28 ± 0.26 for the caller’s dora, +0.49 ± 0.17 for a dealer caller, −0.26 ± 0.09 for each dora in LuckyJ’s hand and +1.61 ± 0.29 for ten more turns. Mortal’s weights come from the site’s replays of your games, over 864 of your turns. The mark is the one chapter 18’s captions use when a safe tile would have cost a shanten.

21

Chapter twenty-one · Reading callers · added 29 September

New

Keep your hand against a caller’s first row

You asked why you break your hand more than LuckyJ on quiet squares, where the callers can win less than 40% of the time. Chapter 20 counted 24 breaks there where LuckyJ’s curve expects 12. Almost all of the extra comes from your games since 20 September, the week the reviews of your games began, and most of it while the caller is still in their first row. Before 20 September you broke on quiet squares the way LuckyJ does.

Quiet squares, every safe tile a shanten shortYour turnsYou brokeLuckyJ’s rate predicts
Before 20 September, caller in their first row10411.1
Before 20 September, caller past their first row6523.3
Since 20 September, caller in their first row205112.2
Since 20 September, caller past their first row88104.5

The first row. A caller in their first six discards is rarely ready: with one call they can win 12% of the time or less, whatever they throw (chapter 18). LuckyJ breaks its hand for them 1.1% of the time. Since 20 September you have broken it 5.4% of the time. Seven of your twelve first-row breaks came against two calls, the spot where chapter 16 found you starting to defend too early; the same habit now shows against one call too.

Not the honors. It is not about the honors you hold. In the first row, when a live honor would have kept your hand, you broke 3.8% of the time against LuckyJ’s 0.8%; when none would, 4.1% against 1.5%. When you did keep the hand with an honor on offer, you cut the honor 51% of the time and LuckyJ 55%.

What you throw. Usually the caller’s genbutsu: 9 of your 12 first-row breaks. Mortal plays these turns like LuckyJ. Across your 309 first-row turns it put 2% of its weight on the safe tiles, and on 8 of the 12 turns where you broke it put under 10% there. The genbutsu is not going anywhere: it is still safe when the caller is closer, and that is when it pays for itself.

Past the first row. The gap is smaller, and part of it is fine. Six of your breaks came on your own 12th turn or later, where LuckyJ’s rate predicts 3, and Mortal backs 3 of the 6: with few draws left a 1-shanten is worth little, as chapter 20’s exception says.

Three turns from 27 September. Each caller had two calls, was in their first row and was not yet ready.

Your turnThe callerSquareYou threwMortal
South 4-1, turn 3Shimocha, pons of [[8m]] and [[1s]], 4th discard14%[[1s]][[E]], 90%
South 2-1, turn 5Kamicha, pons of [[8p]] and [[F]], 5th discard21%[[2s]], genbutsu[[6m]], 65%
East 2, turn 6Toimen, a [[S]] pon and a chi, 5th discard21%[[F]], from a pair[[3s]] or [[1p]]

In the first you held three lone honors and threw the [[1s]] beside your [[2s]], passing up the step the [[4m]] draw offered; Mortal cuts the [[E]]. In the second you led by 9,700 and threw the genbutsu [[2s]] to stay out of a hand that was still forming. In the third you split your [[F]] pair, safe because both other copies were out, and Mortal keeps it. Open them at your 3rd turn in South 4, your 5th in South 2 and your 6th in East 2.

Rule While a caller is in their first six discards, keep your hand: cut the tile that keeps the shanten, even a live one, and save their genbutsu for later. Break only where chapter 20 does, on a hot square against an expensive caller.
Exception Your last draws. From your 12th turn on, a 1-shanten that cannot finish is worth little, and breaking it for a safe tile is often right.

Evidence Your draw turns in your 123 games since January where every safe tile fell a shanten short of the best hand, with 1-shanten or worse in reach and a callers’ square under 40%, set against LuckyJ’s rate in the same cell over its 3,680 such turns: 2,440 with the caller in their first row (26 breaks) and 1,240 past it (63). A caller is in their first row through their sixth discard; with two callers, both must be. The split is by each game’s start. Mortal’s weights come from the site’s replays of your games. The script is mine_break_folds.py, as for chapter 20.

22

Chapter twenty-two · Fear or doubt · added 29 September

New

Let your hand, not the round, decide the fold

You asked whether valuing your own hand too low is a self-esteem issue or a fear issue. A game record cannot see feelings, but it can show what your extra folds follow, and the two answers predict different things. Fear of the other player would follow the size of their hand and your last deal-in. Doubt about yourself would follow what your own hand is worth. In your 123 games since January, the size of their hand and your last deal-in move you less than they move LuckyJ. What moves you is the round: from South 1 on, your folding against a riichi climbs further above Mortal’s, and your riichi declarations fall below it. There is a smaller doubt too. At one-shanten, your dora do not keep you in the hand the way LuckyJ’s do.

The words first. A costly turn is the turn chapter 20 counts: you draw against one riichi, and every safe tile (genbutsu, a full suji, an honor with two showing) would leave your hand a shanten short of the best hand this draw allows. Folding there means throwing the safe tile. Mortal’s rate is the weight Mortal puts on the safe tiles on the same turn, read from the site’s replays of your games. Extra folds are your folds beyond Mortal’s rate, per 100 turns. LuckyJ is measured the same way, on its own turns.

Costly turns against one riichiTenpai1‑shanten2-shanten or worse
You folded28.4%64.7%83.1%
Mortal, on your turns19.0%58.1%74.9%
LuckyJ folded18.3%53.4%79.5%
Mortal, on LuckyJ’s turns20.1%57.7%75.3%

Against a riichi you fold 54% of your costly turns where Mortal, on the same turns, folds 46%: about 8 extra folds in every 100, at every distance from tenpai. On its own turns LuckyJ folds 2 fewer than Mortal. Against callers the gap is small, 7% against Mortal’s 5%, and chapter 21 covers it.

Five splits. Each panel below divides your costly turns against a riichi two ways and shows your extra folds in each, and LuckyJ’s. When the second situation makes you fold more than it makes LuckyJ fold, it is one of the things your folds follow.

What moves your extra folds against a riichi

Their hand

a child’s riichi → a dealer’s

You+8 → +10
LuckyJ−4 → +2

Your last hand

no deal-in → a deal-in

You+8 → +7
LuckyJ−2 → +3

The round

East → South

You+5 → +11
LuckyJ0 → −4

Your dora

none → two or more

You+5 → +11
LuckyJ−1 → −2

Your seat

child → dealer

You+8 → +8
LuckyJ0 → −6

How to read it

Each row runs from the hollow dot, the first situation, to the filled one, the second. The thin mark is zero: folding exactly as often as Mortal on the same turns. Right of it, more folds than Mortal.

first situation second situation

Extra folds: on costly turns against one riichi, where every safe tile leaves the hand a shanten short of the best hand the draw allows, how many more times in 100 the player threw the safe tile than Mortal would have on the same turns. Your 383 turns in 123 games since January; LuckyJ’s 3,565 in its 1,079 Tokujou games. The scale runs from −8 to +14.
The numbers, with one standard error of each change
SplitYouLuckyJ
FirstSecondChangeFirstSecondChange
Their hand: a child’s riichi → a dealer’s7.59.6+2.2 ± 4.0−3.62.3+5.9 ± 1.3
Your last hand: no deal-in → a deal-in8.07.2−0.8 ± 7.4−2.22.6+4.8 ± 1.9
The round: East → South5.111.3+6.3 ± 4.0−0.1−4.0−3.9 ± 1.3
Your dora: none → two or more4.811.1+6.4 ± 4.5−1.4−2.3−0.9 ± 1.4
Your seat: child → dealer7.98.00.0 ± 4.2−0.4−5.7−5.4 ± 1.4

Not fear of the riichi. A dealer’s riichi pays half as much again as a child’s. It adds 2 extra folds per 100 for you and 6 for LuckyJ, so it moves you less than it moves LuckyJ. A deal-in on the hand before leaves your folding where it was, while LuckyJ folds 5 more. That split has only 41 of your turns, but neither points toward fear of the other hand.

The round. In the East you fold 5 extra times per 100, in the South 11. LuckyJ goes the other way, from 0 to 4 fewer than Mortal. Your riichi declarations show the same thing in a different decision. In the East you declare 81% of your first closed tenpais, where Mortal on the same tenpais would declare 75%; in the South you declare 62% against Mortal’s 64%. LuckyJ declares a little more than Mortal in both. Each of the two gaps is more than two standard errors.

Your own hand. With two or more dora your extra folds rise from 5 to 11 per 100, while LuckyJ’s stay close to zero. The one-shanten shows it best. With no dora, one, and two or more, you fold it 65%, 67% and 63% of the time; LuckyJ 57%, 57% and 44%; Mortal on your own turns 61%, 63% and 53%. As dealer, LuckyJ folds 5 fewer per 100 and you fold as often as ever. These gaps are between one and one and a half standard errors, weaker than the round’s, but they repeat what chapter 16 found on the same hand.

So which is it? If you want the words, it is closer to fear than to low self-esteem, and the fear is of the scoreboard; the riichi itself does not set it off. In the East you declare riichi more readily than Mortal does, which is not how someone who doubts their hands plays. The habit was there before our reviews began: 11 extra folds per 100 before 20 September, 6 since. It is a habit with a trigger, and the trigger is the South.

Three turns from the South. Each time a riichi came, the hand was worth keeping, and you threw a safe tile.

Your turnThe spotYou threwMortal
South 3, turn 41st by 16,300; toimen’s riichi on their 3rd discard; 1-shanten, three dora[[8s]], genbutsu[[3s]], 89%
South 3-2, turn 7Dealer, 2nd, 1,100 above 3rd; toimen’s riichi on their 6th; 1-shanten, two dora[[3p]], genbutsu[[9s]], 79%
South 2, turn 13Dealer, 2nd, 7,600 above 3rd; toimen’s riichi on their 11th; tenpai on offer, three han[[2p]], genbutsu[[3m]], 98%

In the first you led by 16,300 with a one-shanten holding three dora, the pair of [[S]] among them, and broke it with toimen’s genbutsu [[8s]]; Mortal keeps the hand with [[3s]]. The hand ended in a draw with you noten. In the second you were the dealer, 1,100 above third, with the dora pair [[3m]][[3m]] and a pair of [[F]], and you broke your [[1p]][[2p]][[3p]] with the genbutsu [[3p]]; Mortal throws the lone [[9s]]. In the third, dealer again, you had a three-han tenpai on offer: the South triplet, the dora [[2p]] and a red five, waiting on [[3p]] after a [[3m]] cut. You threw the [[2p]], toimen’s genbutsu and your own dora, and Mortal keeps the tenpai. Open them at your 4th turn in South 3, your 7th in South 3, two honba and your 13th in South 2.

Rule Against a riichi, let your hand decide the fold: how close it is, its dora and your seat. The round is not a reason. With two or more dora, keep a one-shanten going with live tiles through your 12th turn; LuckyJ folds it 33% of the time up to its 9th turn and 40% on its 10th to 12th.
Exception Fourth place and the last draws. When you are fourth, your extra folds match LuckyJ’s, which folds more than Mortal there too, so that care is not a leak. From your 13th turn a one-shanten folds more often than not even with dora: LuckyJ 60%.

Evidence Your draw turns facing exactly one riichi in your 123 games since January where every safe tile fell a shanten short of the best hand the draw allowed: 383 turns, and 890 against callers with nobody in riichi. LuckyJ’s: 3,565 and 7,432 in its 1,079 Tokujou games. Mortal’s weights come from the site’s replays for your turns and from the local Mortal run over LuckyJ’s games; its records match the tile actually thrown on 99.9% of your turns and all of LuckyJ’s. First closed tenpais with riichi available: 321 yours, 2,661 LuckyJ’s. Each standard error comes from resampling whole games. The script is mine_fold_motives.py, and the figure is drawn by build_fold_motives_figure.py.

23

Chapter twenty-three · Recent games · added 30 September

New

Push the cheap tenpai, and keep a made hand dama

You feel you keep doing well, and that pushing into a riichi with a cheap, narrow open tenpai, or answering an open tenpai with a late riichi, is paying. The twelve games of 29 and 30 September back the feeling: they finished 2, 3, 3, 1, 2, 1, 1, 1, 2, 3, 3 and 1, five firsts and no fourth, for +847 rank points. At your fourth rate since January, 17%, twelve games bring about two fourths. The pushes were mostly right, and the draws were kind as well.

Pushing into a riichi. These are the turns where you were tenpai against a riichi and every tile that kept the tenpai was unsafe.

Tenpai against a riichi with every tile that keeps it unsafeYour 12 new gamesYour 123 beforeLuckyJ
Pushed87%78%82%
Pushed with 1 or 2 han7 of 856%73%
Pushes that won40%24%31%
Pushes that dealt in15%23%21%

You now push more than LuckyJ, and nearly all of the rise is in cheap hands. Mortal agrees with it: your tile was its first choice on 17 of these 23 turns, and on every cheap open push. The narrowest was East 4, one honba, of the 16:39 game on 29 September, at your 13th turn: 2 han, three tiles left, and Mortal put 99% on the [[7m]] you pushed. Your one cheap open fold went the other way. In East 2 of the same game, at your 14th turn, you broke the tenpai with [[2p]], where Mortal pushed [[4p]] at 99%.

Chapter 11 folds a cheap tenpai with three live tiles or fewer, because LuckyJ lost 1,374 a push on turns of that kind. Mortal pushed this one, and one turn is not enough to change the rule. Your pushes won 40% and dealt in 15%, better than LuckyJ’s 31% and 21% on the same turns. Over 20 pushes that is within ordinary luck, so expect both to drift back toward LuckyJ’s.

The expensive hand. One push cost 18,300. In East 4, one honba, of the 01:12 game on 30 September, the dealer declared, and at your 7th turn you chased with a closed tenpai worth 4 han with the riichi, declaring with [[2m]]. The [[2m]] was the dealer’s winning tile, a haneman with ippatsu. Mortal would have thrown the same [[2m]], at 83%, but without the riichi, at 12%. The loss was the tile, which Mortal would have paid as well; the declaration added only your stick. What a riichi gives up this early is the chance to fold on a later turn. Pushing a closed tenpai worth 4 han or more with the riichi into a riichi by its 9th turn, LuckyJ declared 43% of the time.

Late riichi against callers. These are your closed tenpais against callers with nobody in riichi, on the first turn of each hand where riichi was legal.

Declared riichiLuckyJYour 123 beforeYour 12 new games
Your 10th to 12th turn70%70%5 of 6
Your 13th turn on54%32%1 of 2

Before these games you let late tenpais go dama more than LuckyJ did from your 13th turn. Now you declare about as often as it does, and the four riichis on yakuless hands were all Mortal’s first choice, at 94 to 100%. In East 3 of the 22:32 game on 29 September, at your 10th turn, the riichi on [[W]] won 2,000/3,900 against a one-call caller who was ready. How often to declare depends on what the hand is worth without the riichi:

The hand before riichi from your 10th turnLuckyJ declaredYour 135 games
No yaku69%71%
1 or 2 han74%68%
3 or 4 han55%44%
5 han or more32%17%

Two of your new riichis were on hands that already had a yaku. In South 4 of the 00:04 game on 30 September, at your 12th turn, the hand was 3 han dama and Mortal put 5% on the riichi. In South 2 of the 19:04 game on 29 September, at your 11th turn, it was 5 han dama and Mortal put 2% on the riichi. That one became a haneman, but LuckyJ declares a made mangan only about a third of the time. Earlier in that game, in East 4 at your 12th turn, you kept a 5-han tenpai dama and won 8,000, which is the play LuckyJ and Mortal make.

Rule Against a riichi, keep pushing a cheap open tenpai with four or more tiles left, as you did in these games; with three or fewer, chapter 11’s fold is still the usual answer. Against callers from your 10th turn, declare a tenpai with no yaku or with 1 or 2 han. With a made mangan, keep it dama.
Exception Pushing a closed tenpai into an early riichi, you can leave out your own riichi. With a hand worth 4 han or more with it, LuckyJ declared on 43% of those pushes by its 9th turn, and Mortal kept dama in East 4 of your 01:12 game. From your 13th turn LuckyJ declares only about half its tenpais against callers, so dama there is no mistake either.

Evidence Your twelve games of 29 and 30 September, against your 123 games since January before them. Pushes come from mine_tenpai_push.py: a tenpai against a riichi where every tile that keeps it is unsafe, 23 turns in the new games and 166 before, against LuckyJ’s 1,505 in 1,255 games, where a push with 1 or 2 han and three live tiles or fewer averaged −1,374 over 163 such turns. Riichi or dama comes from contrast/riichi_danger.py: the first turn of each hand where riichi was legal, a closed tenpai against callers with nobody in riichi, from your 10th turn 8 turns in the new games, 69 before and 643 in LuckyJ’s 1,079 Tokujou games. Mortal’s choices are the site’s ratings of your games. Rank points come from your ranked-game records: 135 games since January, +1,025, fourths 17.0%.

24

Chapter twenty-four · Riichi or dama · added 1 October

New

When to keep a big hand dama

You feel you have been declaring too much lately, and asked when to keep a big hand dama to save the stick. On big hands you are right. Since 29 September you declared 5 of your 7 first tenpais worth 3 han or more without riichi, where LuckyJ’s rates expect about 3 and Mortal’s about 2. In your 123 games before that you matched LuckyJ: 25 declared out of 53, against 27 expected. Your other riichis were mostly right.

You do declare more often than before: 30 riichis per 100 hands in your five newest games and 23 since 29 September, against 19 earlier and LuckyJ’s 19. Part of that is more closed tenpais, 28 per 100 hands against 25, and on a cheap tenpai the riichi is the play. In the five newest games you declared 12 of your 13 first tenpais, and Mortal’s weight on the riichi averaged 87% on those turns. Those games’ 14 riichis won 4 times, dealt in 3 times and lost the stick 7 times. Two of the three that dealt in were riichis Mortal makes at 99% or more, so most of that was the draw.

The words first. A hand’s value here is the han a ron pays without riichi on every one of its winning tiles, dora and red fives included: a hand that pays 5 han on one tile and 2 on the other counts as 2. A made mangan is 5 han or more on every tile. At 3 han the fu matter as well: a closed ron scores 30 fu with pinfu and at least 40 without it, seven pairs at 25 aside, so a 3-han hand pays 3,900 with pinfu and 5,200 or more without it. At 4 han they hardly do: the dama ron is 7,700 with pinfu and 8,000 without, and the riichi lifts either to the same 5 han and either tsumo to the same haneman. A two-sided wait is a ryanmen, or a wait on three or more kinds of tile. Your turn counts as in the rest of the guide: each draw from the wall and each call adds one. Unless it says otherwise, every figure is the first closed tenpai of a hand, with riichi legal and nobody else in riichi.

The stick is the small part. LuckyJ’s riichis fail to win 48% of the time, so it loses the stick on about half of them and still declares most tenpais. On a big hand the riichi costs something else: opponents stop throwing into a hand they know is ready, and you can no longer fold or change your wait.

Your hand without riichiLuckyJ declares
No yaku85%
A yaku on only some of the winning tiles91%
1 or 2 han83%
Two-sided, 3 han pinfu: 3,900 by ron88%
Two-sided, 3 han at 40 or 50 fu: 5,200 or 6,40056%
Two-sided, 4 han, pinfu or notby turn, in the chart
Two-sided, 5 han or moreby turn, in the chart
3 han, any other wait26%
4 han, any other wait27%
5 han or more, any other wait12%

A cheap hand gets the riichi because the riichi doubles it: a non-dealer ron goes from 1,000 to 2,000, and from 2,000 to 3,900. From 3 han up the riichi adds less each time. 3,900 becomes 7,700, but 7,700 becomes only 8,000, and a mangan becomes a haneman only on a ron, since a dama tsumo is already a haneman. On a poor wait LuckyJ keeps any hand of 3 han or more dama. On a two-sided wait the fu count at 3 han. A 3-han hand at 40 fu is a mangan by tsumo already, 4 han at 40 fu, so the riichi lifts its tsumo only through ura dora or ippatsu, and LuckyJ declares 56% of those against 88% of 3-han pinfu. From 4 han, the turn decides:

How often LuckyJ declares a big two-sided tenpai

Its first closed tenpais with a two-sided wait and a yaku on every winning tile, with nobody in riichi, by its value without riichi and by its own turn. Each dot is one turn, sized by its hands; the lines are fitted curves.

3 han pinfu4 han5 han or more95% range of each fit

0% 25% 50% 75% 100% 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 LuckyJ’s turn when it reached tenpai
LuckyJ’s 1,079 Tokujou games: 160 such tenpais of 3 han pinfu, 3,900 by ron, 71 of 4 han, pinfu or not, and 30 of 5 han or more. Each hand counts its han without riichi on its cheapest winning tile, dora and red fives included. Each curve is fitted between the 5th and 95th percentiles of its turns; with fewer than 100 hands, as for 4 han and for 5 or more, the fit is a straight line on the logit scale.
The fitted shares in a table
Turn3 han pinfu4 han5 han or more
486.2%
590.5%82.6%83.7%
692%78.2%75.2%
793%73.2%64.1%
893.6%67.4%51.4%
993.8%61.1%38.4%
1093.2%54.4%27%
1191.7%47.5%17.9%
1288.5%40.8%11.4%
1382.3%34.3%7.1%
1472.1%28.4%4.3%
1557.3%23.1%2.6%
1640.3%1.6%

Read where each curve crosses the half line. A two-sided 3-han pinfu is declared at least 80% of the time through LuckyJ’s 13th turn. A 4-han hand starts at 86% on its 4th turn, crosses half between its 10th and 11th, and is at 41% on the 12th. A made mangan starts at 75% on the 6th turn, crosses half on the 8th, and is down to 18% on the 11th.

Why 4 han stays a riichi longer than 5. The riichi’s lift lands on different wins. On a 4-han hand, pinfu or not, it turns every tsumo from a mangan into a haneman, about 4,500 more, and adds about 2,000 to a ron. On 5 han the dama tsumo is a haneman already, so the riichi adds about 1,900 there and about 4,500 to a ron, the win the riichi cuts. A riichi wins more by tsumo while the wall is long: 35% by tsumo and 30% by ron on the 5th turn, 21% and 28% on the 11th. So the 4-han riichi pays for longer, and LuckyJ’s made-mangan curve crosses half about two and a half turns earlier. Weighed in points alone the two hands come out about even, so part of that gap is likely LuckyJ playing for its finishing place.

The haneman the riichi might add. You asked whether LuckyJ wants the riichi’s step from 4 or 5 han up to a haneman. Its choices say it does early, while the hand has time; by your 11th turn the step pays for itself and no more. Here is the trade at that turn, the turn of both your big riichis below, for a non-dealer with a two-sided wait and six tiles live, fitted over all four seats at LuckyJ’s tables:

At your 11th turnDamaRiichi
Wins by ron46%28%
Wins by tsumo16%21%
A 5-han ron pays8,000about 12,500
A 5-han tsumo pays12,000about 13,900
A hand that does not win averages−2,180−2,630
A 5-han hand, all told4,8305,020
A 4-han hand at 40 fu, all told4,1804,000
A 4-han pinfu, all told4,0404,000

The riichi makes every win bigger: with ura dora and ippatsu, 44% of riichi rons pick up at least one more han. But it costs two in five of your rons, and a hand that does not win loses more, the stick and the deal-ins a declared hand cannot fold away from. At this turn all three come out within 200 points of each other, with 95% ranges from about −1,200 to +1,600. Dama wins the hand more often, 62% against 49%. LuckyJ and Mortal are both built to finish high in the game rather than to collect points, and with a made mangan they take the hand that wins more often.

Your hands. Two of your recent riichis came at your 11th turn on hands worth 4 or 5 han. One lost its stick and one made a haneman; LuckyJ keeps the first dama and is split on the second. The third card is a 3-han shanpon, and the fourth a late chase.

You also made the plays this chapter asks for. In East 4 of the 19:04 game on 29 September, at your 12th turn, you kept a made mangan dama and won 8,000, with Mortal at 96% on dama. In South 1 of the 11:57 game on 1 October, at your 7th turn, you declared a two-sided 3-han pinfu, as LuckyJ does 93% of the time that early, and an ura dora made it a mangan.

Chasing a riichi. Against a riichi the stick matters least and the fold matters most. With three live tiles or fewer and a hand that already has a yaku, LuckyJ chases with its own riichi 26% of the time at 1 or 2 han, 2 times out of 23 at 3 han and 4 out of 36 at 4 han or more. Without a yaku, where dama cannot ron, it declares 56%. Chapter 23’s exception let you leave out the riichi on an early push with 4 han or more; a late chase on a narrow wait is the same.

Rule Declare a tenpai with no yaku, with 1 or 2 han, or a 3-han pinfu on a two-sided wait; a two-sided 3-han hand at 40 fu is close to even. Keep 3 han or more dama on any other wait. With a two-sided wait, declare 4 han, pinfu or not, through your 10th turn and keep it dama from your 11th; declare a made mangan through your 7th, take either on your 8th, and keep it dama from your 9th.
Exception When the yaku covers only some of your winning tiles, declare: LuckyJ does 91% of the time, so that every tile pays. Chasing a riichi on three live tiles or fewer with a yaku, stay dama and push the same tile.

Evidence First closed tenpais from contrast/tenpai.py, with riichi legal and a wait that is not furiten: 2,177 of LuckyJ’s in its 1,079 Tokujou games with nobody in riichi, and 60 of yours worth 3 han or more on every tile, 53 before 29 September 16:00 and 7 after. Two-sided 3-han hands are sorted by the points of their cheapest ron, from each winning tile’s han and fu. At the same turn, LuckyJ’s 11 four-han hands at 40 fu were declared about as often as its 60 four-han pinfu. mine_dama_value.py fits LuckyJ’s share by its turn with fit_series, one point per turn, and reads Mortal from the site’s ratings of your games. The trade fits wins by ron and by tsumo over all four seats of LuckyJ’s games, 2,289 declared two-sided tenpais with a yaku against 208 two-sided tenpais worth 4 han or more kept dama, by turn, live tiles and dealer, and prices each win from the non-dealer table with the ura dora and ippatsu of 4,126 riichi wins. Players chose which hands to declare, so the two groups hold different hands, and the table compares their results. Counting the 3-han dama hands too puts the riichi up to 750 points ahead at the 11th turn, and that range still crosses zero. Chasing comes from 147 of LuckyJ’s first tenpais with an opponent in riichi and three live tiles or fewer. Chapter 23 counted a hand’s value on its best winning tile; from the 10th turn against callers, LuckyJ declared 13 of 39 such made mangans, and counted on every tile, as here, 2 of 23.

25

Chapter twenty-five · Riichi or dama · added 1 October

New

Your place in the game moves the riichi

You suggested that the score matters as well, and that a player in 1st declares less. LuckyJ agrees, and mostly at the end of the game. Over all its first closed tenpais it declares 79 to 83% from any place in the East. In South 1 to 3 it declares 71 to 73% from 1st, 2nd and 3rd, and 87% from 4th. In all-last it declares 10% from 1st, 47% from 2nd, 60% from 3rd and 87% from 4th.

The words first. Your place is your standing at the start of the hand. All-last is South 4, with the few West hands, where a win can end the game. The grid below is chapter 24’s chart with your place added: one table for each stage of the game, a block of rows for each hand value, a row for each place, and a column for each turn. Above the grid the same numbers stand as a surface: drag it sideways to turn it, tap it to read a point, and pick the hand and the stage with the buttons.

Hand
Stage

Drag sideways to turn the surface; tap it to read a point.

Stage of the game
Hand, and your place45678910111213141516
No yaku
1st8588909192929190878276
2nd8689919293939291888478
3rd8790929394939392898680
4th8992939495959493918883
1 or 2 han
1st8485868787878686848381
2nd8687888888888887868583
3rd8788898989898988878684
4th8990909191919190898887
3 han pinfu, two-sided
1st929394949393929087827667
2nd939494949494939188847869
3rd949595959594939289868072
4th959696969695949391888375
4 han, two-sided
1st827874696357514540342924
2nd848076716660554943373126
3rd858278736863575145393429
4th878481777267625650443833
5 han or more, two-sided
1st574942342822181410865
2nd605245373125191512975
3rd6255474033272117131076
4th6760524537312519151297
Hand, and your place45678910111213141516
No yaku
1st7480838586868482787162
2nd7378828485848380766960
3rd8286889090908988848072
4th9395969696969695949288
1 or 2 han
1st7375777778777776747269
2nd7274757676767574727067
3rd8183848484848483828078
4th9393949494949493939291
3 han pinfu, two-sided
1st878889898887858278716251
2nd858788888786848176696049
3rd919292939291908884797262
4th979797979797969594928883
4 han, two-sided
1st706559544742363126211814
2nd686357514539342924201613
3rd797470645953474135302521
4th928987848076726761554943
5 han or more, two-sided
1st4134272217131086432
2nd393226201612975432
3rd52443730241915129754
4th767063564841342722171310
Hand, and your place45678910111213141516
No yaku
1st79111313131210864
2nd4047535758585551443627
3rd6269737677777572665848
4th9395969696969695949288
1 or 2 han
1st77888887765
2nd3941434444444342403734
3rd6163656666666564625956
4th9393949494949493939291
3 han pinfu, two-sided
1st14161717161513108643
2nd606364656361575144362820
3rd788081828179767266584837
4th979797979797969594928882
4 han, two-sided
1st654322111110
2nd3530252117141197654
3rd57514539342924201613119
4th918987848076716661554943
5 han or more, two-sided
1st211110000000
2nd14108653321111
3rd28221714108643321
4th756963554840332722171310
0%10%25%50%75%90% LuckyJ’s fitted share of riichi on its first closed tenpai, as a non-dealer with nobody in riichi, in its 1,079 Tokujou games, by the hand’s value without riichi, its place at the start of the hand and its turn. One model holds a turn curve for each value and one shift for each place and stage; the 4-han and 5-han rows rest on 71 and 30 hands. A hatched square is a turn outside the middle 90% of that value’s tenpais.

Place moves every hand value the same way, cheap or big: letting big hands take their own shift fits no better. Take a two-sided 4-han hand at your 10th turn. In the East LuckyJ declares it 51% of the time from 1st and 62% from 4th. In South 1 to 3 it is 36% from 1st, 34% from 2nd, 47% from 3rd and 72% from 4th. In all-last it is 1% from 1st and 71% from 4th. A 1 or 2 han tenpai at the 9th turn goes from 8% in all-last 1st to 94% in all-last 4th.

Points as well as place. In the East the size of a lead matters: each 10,000 points of lead cuts LuckyJ’s odds of a riichi by about a third. From South 1 the size stops mattering, and being 1st or 2nd is what counts. The grid is for a non-dealer; as the dealer LuckyJ declares a little more.

Your games. In the East you declare as LuckyJ would on the same hands: 130 of 155, against 130 expected. From South 1 you move less than it does with your place. In 1st you declared 22 of 31 first tenpais where LuckyJ’s model expects 18, and in 4th 22 of 30 where it expects 25. Each gap is about two standard deviations, a tendency to watch. In South 2 of the 19:04 game on 29 September, at your 11th turn, you led by 28,800 and declared chapter 24’s 4-han hand; from 1st in the South, LuckyJ declares it 31% of the time. In 4th, in South 1 of the 22:01 game on 20 September, at your 6th turn, you kept a tenpai with no yaku dama, and in all-last of the 10:17 game on 19 September, at your 10th turn, another. LuckyJ’s model declares both 96% of the time, and Mortal put 91% and 100% on the riichi. The second went well: the hand improved to pinfu and won 3,900, enough for 3rd.

Rule In the East, use chapter 24’s turns, and lean to dama as your lead grows. From South 1 to 3, in 1st or 2nd, declare a two-sided 4-han hand only through your 7th turn and keep a made mangan dama; in 4th, declare 4 han through your 13th turn and a made mangan through your 8th. In all-last, keep almost every tenpai dama in 1st; in 4th, play as you would in 4th from South 1.
Exception In all-last, 2nd and 3rd sit between, and the points you need decide more than the grid can: a hand that cannot move you up gains little from the riichi’s extra han.

Evidence mine_riichi_placement.py fits one logistic model to chapter 24’s 2,177 first closed tenpais of LuckyJ with nobody in riichi: each hand value has its own turn curve, the dealer adds one term, and each place in each stage one shift. Letting hands of 3 han or more on a poor wait and of 4 han or more take their own shifts adds 11 terms and improves the fit by a chi-square of 11.0, which is no improvement. A lead in the East moves the log-odds by −0.44 ± 0.16 for each 10,000 points, and in South 1 to 3 by +0.17 ± 0.18. All-last counts 205 of LuckyJ’s tenpais, 12 of them West hands. Your 287 first tenpais are scored against the same model; the standard deviations come from its predicted shares.

26

Chapter twenty-six · Recent games · added 3 October

New

A losing run, played the way LuckyJ plays

You felt you had lost a few lately. The 25 games since chapter 23 finished with 7 firsts, 6 seconds, 8 thirds and 4 fourths, for +445 rank points. The last 11 brought three of those fourths and −173, where your usual fourth rate, 17%, gives about two. Three or more fourths in 11 games happens in about 28% of such stretches. Your hands and your decisions held up through it.

Your hands94 games before 24 SeptemberYour 25 newest gamesOf them the last 11LuckyJ
Fourth places22%16%27%17%
Hands won21.8%21.7%20.4%22.9%
Average win6,1507,7337,5367,317
Hands dealt in11.7%9.6%13.0%10.3%
Points a hand+31+453+321+363

Your average win, the gap chapters 6 to 9 were about, now runs above LuckyJ’s, and even the last 11 games made about as many points a hand as LuckyJ does. Mortal’s first choice matched yours on 75.9% of your decisions in those 11 games, against 76.6% before 24 September, and you call about as often as Mortal would, on 24% of your chances against its 26%.

The four fourths.

  • 30 September, 18:46: no win, and tenpai in only two of nine hands.
  • 2 October, 21:23: no win and no deal-in. The points went to other players’ wins.
  • 1 October, 22:59: three deal-ins, each Mortal’s own play. In East 2 you pushed your tenpai on the last discard, at your 19th turn, and the [[2p]] made a houtei (Mortal 99%). In East 3 you declared riichi at your 8th turn into a two-dragon open hand (Mortal 99% for the riichi, though it would declare with [[7s]] 63% to your [[6s]]’s 37%). In East 4, one honba, you declared at your 9th turn into a riichi (Mortal 100%).
  • 2 October, 14:37: in East 4, one honba, your three-call tenpai pushed [[3s]] into a riichi at your 15th turn (Mortal 100%), and it was ippatsu for 8,300.

The one slip. In East 4 of the 21:23 game a riichi came, and you were three-shanten holding [[8m]], a tile in the riichi player’s river. Cutting it kept your shanten. You cut a live [[W]] instead at your 10th turn, then [[8p]] and [[2p]], and a live [[1s]] at your 13th turn. Mortal threw the [[8m]] on each of those turns, at 95 to 100%. Nobody won the hand, so it cost you only the draw payment.

Chapters 20 to 23 told you to keep close hands against callers and to push cheap tenpais into a riichi. This hand is past what they cover. Against a riichi, from two-shanten or worse with a safe tile that keeps your shanten, you cut a live tile on 1.8% of those turns before 24 September and on about 6% in your 25 newest games, while LuckyJ does it on 3.7%. Most of your new live cuts are early honors that Mortal throws too, but a hand three steps away with a safe tile to spare is not one of them.

Facing a riichi with a safe tile in hand: live cutsTenpai1-shanten2-shanten3-shanten or worse
You before 24 September35.4%13.6%2.9%3.0%
Your last 11 games40.0%13.5%10.5%11.8%
LuckyJ38.3%16.6%6.2%2.9%
Rule Through a losing run, keep playing the way these games did: Mortal agreed with you as often as ever, and every deal-in in your fourths was its play too. Against a riichi from two-shanten or worse, when a safe tile keeps your shanten, throw it.
Exception An early live honor you have no use for can still go first against a riichi from a far hand: seven of your ten far-hand live cuts in the last 11 games were Mortal’s own first choice.

Evidence Your 25 games from 30 September, 18:46, to 3 October, 00:43, against your 94 games before 24 September and LuckyJ’s 1,255. Hands come from contrast/hands_all.py, rank points from your ranked-game records, which did not yet hold the newest game. Mortal’s choices are the site’s ratings of your games: 1,501 decisions in the last 11 and 14,189 before 24 September, and 315 call chances in the last 11. Cuts against a riichi come from mine_riichi_folds.py: with a safe tile in hand, 93 turns from two-shanten or worse in the last 11 games; with a safe tile that keeps the shanten, 84 in the last 11, 79 in the 14 games before them, 788 before 24 September and 8,604 for LuckyJ.

27

Chapter twenty-seven · Recent games · added 3 October

New

Thirds are saved fourths, and a dora outweighs the flush

You asked three things: whether you take too many thirds against LuckyJ, whether you now force a flush when the dora sit in other suits, and why you are stuck near 6,700. Your thirds are mostly fourths you saved. You go for a flush with a dora elsewhere about as rarely as LuckyJ. The plateau is an ordinary swing over 15 games in which you played the way you did on the way up.

Thirds. Placements since January, against LuckyJ’s 1,255 games:

PlacesFirstSecondThirdFourth
Your 66 games since 24 September30.3%28.8%31.8%9.1%
Your 94 games before24.5%27.7%25.5%22.3%
LuckyJ31.2%27.3%25.2%16.3%

Since 24 September your firsts and seconds together match LuckyJ’s, and the rest of your games lean to third. When you went into the last hand fourth, you climbed out 32% of the time, against LuckyJ’s 17%. At Saint 3 a third costs about 11 rank points and a fourth 272, so your mix since 24 September is worth about +34 a game, where LuckyJ’s own mix would earn about +16 under the same points.

The players at your tables are as strong as LuckyJ’s Tokujou opponents. They win 21.2% of their hands for 6,715 on average, against 21.0% and 6,864, deal in 13.8% against 13.9%, and reach tenpai at the same turn, 9.7. What cost you firsts was the value of your older games: before 24 September you led at the start of South 1 in 22% of your games, and since then in 34%, LuckyJ 32%. From the same standing you finish as LuckyJ does.

Flushes. How often a hand went for a flush, every number tile in one suit by the 12th turn:

Went for a flushAll handsDealt a dora outside the main suitDealt none there
LuckyJ6.1%2.5%7.4%
You since 24 September5.5%2.0%6.8%

LuckyJ goes for a flush about half as often when a dora sits in another suit: with nine tiles that fit and a value-honor pair, 20% of the time without such a dora and 10% with one; with ten or more fitting tiles, 36% and 19%. You do the same. In your 25 games up to 3 October your hand was flush-shaped at some point in 8.8% of hands, against LuckyJ’s 6.3%, and three of those had a dora elsewhere; that is too few to call a habit. One dora went against Mortal: in South 3 of the 01:12 game on 30 September, at your 5th turn, you threw the red [[5s]] for the flush, where Mortal cut [[P]] and kept it (your [[5s]] 3%).

The plateau. You climbed from about 4,200 rank points on 22 September to about 6,750 on 1 October, close to the 6,845 your ranked-game records show as your best at Saint 3. The 15 games since broke about even: four fourths took about 270 each, and four firsts gave it back. In those games Mortal’s first choice matched yours on 76.4% of your decisions, against 77.2% on the way up, and your average win kept rising, to 7,047. Your win rate dipped from 24.4% to 19.9%. Your placements since 24 September earn about +34 a game, but one game swings about 114 points either way, so 15 games swing about 440, and a flat stretch like this one comes about one time in ten.

The newest fourth turned on one turn. With two riichis and no tile safe against both, you pushed [[9m]] where Mortal pushed [[2s]]:

Rule Count a third as a fourth saved: at Saint 3 it costs 11 rank points where a fourth costs 272. Go for a flush with a dora in another suit only with ten or more fitting tiles or a value-honor pair. Through a flat stretch, keep playing the way you did on the way up.
Exception Facing two riichis with nothing safe against both, the kind of tile is a poor guide: a live terminal usually deals in less than a live 2 to 8, but a tile outside one declarer’s early discard can be the better push, as the card shows.

Evidence Placements: your 160 games since January against LuckyJ’s 1,255, with the standings at the start of South 1 and of the last hand read from the game logs. Opponents: 4,971 hands at your tables against 33,816 at LuckyJ’s, from contrast/hands_all.py. Flushes: every number tile in one suit by the 12th turn, over the hands that lasted that long, 364 of yours since 24 September and 6,902 of LuckyJ’s, of them 98 and 1,811 dealt a dora outside the main suit. Rank points come from your ranked-game records, which did not yet hold your 14:11 game on 3 October, anchored at your 6,700. Mortal’s choices are the site’s ratings of your games; LuckyJ’s deal-ins by the kind of tile are its cuts against a single riichi, 977 live terminals and 1,592 live 2 to 8.

28

Chapter twenty-eight · Dealer callers · added 4 October

New

Against the dealer’s call, fold three discards sooner

Your game at 23:38 on 3 October ended fourth, below zero, after three deal-ins: 5,200 in East 3, 12,300 in East 4 and 5,200 in South 1. In East 4 you thought you could keep pushing until the dealer’s 10th discard against their one call, and you missed that their last three tiles came from the wall. Those three tiles changed little. They lifted the dealer’s square on chapter 18’s grid from 21% to 24%, and LuckyJ plays the same with or without them. What moves LuckyJ is the dealer: against a dealer’s one call it folds about three discards sooner than against anyone else’s. For a far hand the dealer’s 10th discard is where LuckyJ folds half the time, so your number was right for that hand. You were one-shanten, where its line against the dealer is the 12th, and at their 8th discard your [[3p]] was a turn LuckyJ splits.

The words first. A caller’s discards are counted by their place in the river, the tile thrown with each call included. A tile from the wall (tsumogiri) is thrown as it was drawn; a tile from hand (tedashi) is not. Their run is how many tiles in a row have come from the wall since their last call. A real choice is a turn where a safe tile keeps your best shanten and so does a live tile with at least as many tiles to improve your hand; a tie is a real choice where the two leave exactly as many. LuckyJ’s line is the caller’s discard from which it throws the safe tile on more than half of those turns. A river avoids a suit when it holds none of that suit after six or more discards and every call is in that suit or in honors.

The dealer’s river, square by square. Shimocha was the dealer in East 4, one honba. They ponned kamicha’s [[E]], a double wind for them, and were ready two discards later. Here is how chapter 18’s grid read them after each discard from the call on, next to what they held and the tile you threw next.

Their discardWhat they threwHow to count itThe grid saysWhat they heldYour next tile
3rdPonned [[E]], threw [[9s]]One call, right after it: second row3%One-shanten[[2s]]
4th[[3m]] from handFirst row4%One-shanten[[E]], dead
5th[[4s]] from handFirst row7%Ready on [[3p]][[2p]], a tie with [[1m]]
6th[[3m]] from the wallSecond row12%Ready[[5m]], a tie
7th[[6s]] from the wallThird row18%Ready[[2s]], a tie
8th[[1s]] from the wallFourth row24%Ready[[3p]], a tie, and ron

On each of your last four turns the [[1m]], the dealer’s first discard, left your one-shanten exactly as good as the live tile you threw. On such a tie LuckyJ’s fitted share of live cuts is 62% at the dealer’s 5th discard and 53% at their 8th: a coin flip each time. Mortal threw your live tile every time, at 96 to 99%, and kept the genbutsu for a later turn, as chapter 16 described. Open the hand at your 4th turn, the first tie, or at your 7th, the [[3p]].

The run. Tiles from the wall do mean something. At a one-call caller’s 8th discard, chapter 18’s grid reads 24% after three in a row and 21% when the last tile came from hand; at the 10th, 41% against 30%, and at the 12th, 55% against 34%. LuckyJ’s choices do not follow it. At the same discard, a longer run does not make LuckyJ fold measurably more, from a far hand or from one-shanten, against the dealer or anyone else, early in the hand or from the 9th discard on.

The dealer. A ron into a dealer’s one call cost 5,949 on average at LuckyJ’s tables, against 3,759 into anyone else’s. Early in the hand LuckyJ pushes against the dealer as hard as against anyone, and from about the dealer’s 6th discard it folds sooner:

Against one call: how often LuckyJ cut the live tile

Turns where a safe tile and a live tile both kept LuckyJ’s shanten, by the caller’s discards so far. Each dot is one discard against the dealer, sized by LuckyJ’s turns there; the lines are fitted curves.

The dealer called, per discard and fitted95% range of the fitAnyone else called, fittedbelow 50%, the safe tile goes first

Two-shanten or worseLuckyJ’s hand far from ready

0% 50% 100% 1 3 6 9 12 15 18 The caller’s discards so far 50% betweendiscards 10 and 11 35%

One-shantenLuckyJ’s hand one tile from ready

0% 50% 100% 1 3 6 9 12 15 18 The caller’s discards so far 50% betweendiscards 11 and 12 39%
LuckyJ’s 1,079 Tokujou games, against a single caller with one call while nobody is in riichi. A safe tile is genbutsu, a full suji or an honor with two copies showing; a turn counts when a safe tile keeps LuckyJ’s best shanten and so does a live tile with at least as many tiles to improve the hand. Each curve is fitted over the discards with at least 25 such turns.
The numbers as a table

Each cell: LuckyJ against the dealer, then against anyone else, then its turns against the dealer. Discards with no turns are blank.

DiscardTwo-shanten or worseOne-shanten
289.4%91.7%4780.0%60.0%5
392.8%87.4%12590.0%78.0%10
488.6%88.6%19393.3%73.6%30
587.6%84.3%23484.9%75.4%53
676.9%80.3%22165.2%69.1%89
764.8%77.7%15967.3%73.7%113
869.0%75.7%11366.3%72.3%104
959.3%70.5%8157.5%76.8%106
1050.8%63.9%5964.1%67.2%78
1141.7%62.7%3657.3%65.0%75
1242.4%51.4%3348.9%64.1%47
1318.8%46.2%1639.0%58.7%41
1433.3%20.0%647.8%67.7%23
1540.0%37.5%521.1%45.8%19
160.0%12.5%335.7%36.8%14
170.0%11.1%137.5%18.2%8
180.0%3100.0%33.3%1

Holding a far hand, LuckyJ still cuts the live tile 65% of the time at another caller’s 10th discard and reaches half only at their 13th; against the dealer it reaches half at the 10th. At one-shanten it folds from the dealer’s 12th discard, and against anyone else it still cuts the live tile 64% of the time at the 14th. On ties the dealer’s line comes at the 9th discard for a far hand and the 10th at one-shanten. Chapter 19’s lines, the 12th and the 10th on ties, mixed the dealer with everyone else. Your far-hand cuts sit a little under LuckyJ’s curves: against the dealer you cut the live tile on 74% of those turns where its curve gives 77% at the same discards, and against anyone else on 77% where it gives 82%. At one-shanten you match it, 67% against 66% facing the dealer and 72% against 71% otherwise.

The suit the river avoids. The dealer’s eight discards held no pinzu and their only call was an honor, the shape of a pinzu half flush, and your [[3p]] was pinzu. Such a river makes the other suits safer. At the caller’s 8th discard a live tile of another suit dealt in 1.0 times per 100 and a live honor 0.7, while a live tile of the avoided suit dealt in 1.4 times, a little under the 2.2 of a live number tile against a caller whose river shows every suit. LuckyJ does not lean on this: given two live tiles equally good for its hand, one in the avoided suit and one not, it threw the other suit’s tile 45% of the time through the caller’s 8th discard and 55% after. In East 4 no tile of another suit kept your one-shanten anyway.

The other two deal-ins. In East 3 you were the dealer and kamicha had one call. At your 13th turn you threw the genbutsu [[3s]] and broke your tenpai, though the [[8p]] you drew was suji and kept it; three turns later the [[6m]] that brought you back to tenpai dealt in. In South 1 you declared from last on 2,000 into two callers, as Mortal did, and shimocha’s toitoi took the [[7s]]. The three turns:

Rule Against the dealer’s one call, fold about three discards sooner than against anyone else’s: a far hand folds half the time by the dealer’s 10th discard and more often from there, a one-shanten hand from their 12th. Count the dealer before the run; a run of tiles from the wall does not move LuckyJ’s line.
Exception A tenpai is past these lines. When the only tiles that keep your tenpai are safe, throw one, as LuckyJ does on 98.8% of such turns; when every safe tile breaks it, LuckyJ breaks it on only 3.5%.

Evidence Your game of 3 October, 23:38, with Mortal’s ratings from the site’s replays. LuckyJ’s 1,079 Tokujou games through contrast/vs_callers.py and contrast/choices.py, summed by mine_dealer_callers.py: against a single caller with one call, 4,500 real choices from a far hand, 1,335 of them against the dealer, and 2,883 from one-shanten, 816 against the dealer, each curve fitted over the discards with at least 25 turns. The run and the dealer were tested together in a logistic regression on the same discard curve. Your 165 games give 657 far-hand and 410 one-shanten turns. The suits: 14,200 live cuts by every seat against callers whose river avoids a suit and 28,129 live number tiles against the others. The squares are chapter 18’s; the tenpai figures are chapter 19’s counts, 1,089 and 1,521 turns. The suit ties number 216 through the caller’s 8th discard and 64 after.

29

Chapter twenty-nine · Recent games · added 7 October

New

Another losing run: five choices cost 60,500 points

You asked what is going on in your recent games. The ten games from 3 October, 14:11, to 7 October, 15:27, finished with 3 firsts, 1 second, 1 third and 5 fourths, an average place of 2.80 against 2.34 over the 163 games before. Five fourths in ten games comes up about 2% of the time at your usual rate, though a stretch picked because it went badly always looks rarer than it is. Your attack held up. The deal-ins did not, and half of the costly ones were choices Mortal would not have made.

Your hands163 games beforeYour 10 newest gamesLuckyJ
Fourth places16.6%50%16.3%
Hands won22.5%23.6%22.9%
Average win6,4205,7777,317
Hands dealt in10.9%17.3%10.3%
Average deal-in−4,769−6,921−5,308
Deal-ins of 7,000 or more, per 100 hands2.69.13.4
Points a hand+176−511+363

What held. You won 23.6% of your hands and reached tenpai in 54% of them, against 52% before. Mortal’s first choice matched your discard on 74.7% of 1,241 turns, against 73.3% of 19,486 before, and your call on 322 of 345 chances, against 5,048 of 5,431. It declared 22 of your 24 riichis too, each at 91% or more. What slipped. Your average win fell to 5,777, under your own 6,420 and LuckyJ’s 7,317, and you dealt in on 19 of 110 hands, where about 12 is usual.

The deal-ins, sorted by Mortal. Each deal-in is rated by the choice that led to it: the discard that was ronned or, when you were already in riichi, the declaration. Twelve of the 19 were choices Mortal made at 50% or more, and six it gave under 25%. Ten cost 7,000 or more, 9.1% of your hands against 2.6% before and LuckyJ’s 3.4%. Five of those ten were Mortal’s own play, which is 5 in 110 hands, or 4.5%: ordinary bad luck, and the 12,300 in chapter 28’s East 4 is one of them. The other five cost 60,500 points together: 12,300, 12,300, 12,300, 12,000 and 11,600. The first is chapter 27’s [[9m]] in South 1 of the 3 October, 14:11 game. The other four are below: three cards here and one in chapter 31.

The sixth deal-in Mortal rated under 25% was small. On 5 October, in East 4 with one honba, you cut a live [[C]] into two riichis at your 16th turn where Mortal cut [[2s]], 99.5% to the [[C]]’s 0.3%, and it cost 2,900. Mortal’s tile would not have saved you: toimen was waiting on [[2s]] and [[5s]], so it would have been ronned too. A better choice can still lose, which is why this one has no card.

Two discards. The first two cards are discards where a safe tile kept your hand, into a dealer’s riichi and into a dealer’s call. Chapters 26 and 28 already say to cut it. In both, the tile Mortal chose could not have been ronned by any hand at the table. One riichi. The third card is the East 1 riichi of the 15:27 game on 7 October. Chapter 24 would have you declare it, and Mortal gave it 6%. The chapter’s table measures LuckyJ and not Mortal, so this is a real disagreement, and one hand cannot settle it. The game’s other 12,300 came from a riichi on a furiten wait, which this chapter first showed as a card and described wrongly. It is now the fourth card of chapter 31.

Riichis Mortal would not declare. Over your 354 riichis since January, the 32 Mortal gave under 25% won 5 times, 16%, and dealt in 9 times, 28%. The other 322 won 163 times, 51%, and dealt in 35 times, 11%. Two signs show in your own games: 14 of the 32 were declared into another riichi, against 46 of the 322, and they came later, at turn 9.9 on average against 8.4. Only two of the 24 riichis in these ten games fall in that group, both in the 15:27 game, and both dealt in. Those numbers describe Mortal’s picks, and chapter 31 turns them into four checks you can make at the table.

Rule Keep the attack Mortal rates as it always has, and close the three places this run leaked. Against a riichi from two-shanten or worse, cut the safe tile that keeps your shanten, a live honor included (chapter 26). At one-shanten against the dealer’s one call, count from their 12th discard and cut the safe tile (chapter 28). Before declaring a tenpai you could keep dama, look at your own river for a winning tile, since a furiten riichi can win only by tsumo. From your 13th turn declare a cheap tenpai with 4 live tiles or more and keep a hand of 3 han or more dama (chapter 33). Chapter 31 has the four checks.
Exception Mortal declared 22 of your 24 riichis in these games, and five of your ten 7,000-point deal-ins were its own play. Keep declaring and pushing as you have; a run like this is the price of it.

Evidence Your ten games from 3 October, 14:11, to 7 October, 15:27, against your 163 games before and LuckyJ’s 1,079 Tokujou games, from mine_costly_run.py over contrast/hands_all.py: 110 hands in the ten games, 1,690 before and 11,272 for LuckyJ. Mortal’s choices are the site’s ratings of your games, 1,241 discards and 345 calls in the ten games and 19,486 and 5,431 before. A riichi’s rating is Mortal’s weight on declaring it. The opponents’ waits are read from their real hands, which the cards show with the hands switch. Riichis: 24 in the ten games and 354 in your 173 games since January.

30

Chapter thirty · Recent games · added 10 October

New

Your safe tiles are back on LuckyJ’s line

You asked how you are doing, and whether you deal in more because you spend fewer safe tiles than LuckyJ. In the 16 games since chapter 29, from 7 October, 18:25, to 10 October, 14:43, you did not. You dealt in on 10.6% of your hands, against 10.9% before chapter 29’s run and LuckyJ’s 10.3%, and each test of your safe tiles below lands on LuckyJ’s numbers. You finished with 4 firsts, 3 seconds, 7 thirds and 2 fourths, about +190 rank points, from about 5,800 to 6,000. What you felt is today: five deal-ins in the 14:24 and 14:43 games, 28,900 points, each of them Mortal’s first choice.

Your hands163 games before 3 OctoberChapter 29’s 10Your 16 newestLuckyJ
Fourth places16.6%50%12.5%16.3%
Hands won22.5%23.6%20.7%22.9%
Average win6,4205,7776,7197,317
Hands dealt in10.9%17.3%10.6%10.3%
Average deal-in−4,769−6,921−4,705−5,308
Deal-ins of 7,000 or more, per 100 hands2.69.12.83.4
Points a hand+176−511+273+363

Mortal’s first choice matched your discard on 73.0% of 2,101 turns, against 73.3% before chapter 29’s run, and your call on 556 of 591 chances. Your win rate dipped to 20.7%: 37 wins where LuckyJ’s rate gives about 41, well within chance over 179 hands, while your average win rose to 6,719.

Your hunch, tested three ways. First, against a riichi, on the turns where a safe tile kept your hand:

Facing a riichiBefore 3 OctoberChapter 29’s 10Your 16 newestLuckyJ
Live cut from one-shanten6.8%5.2%6.6%7.8%
Live cut from two-shanten or worse2.9%8.6%3.3%3.6%
Folds with a weaker safe tile than the best one kept3.9%9.5%5.0%4.0%

Then against a single caller, on chapter 28’s real choices, and when a riichi came, each beside LuckyJ in the same spots:

You / LuckyJ in the same spotsBefore 3 OctoberChapter 29’s 10Your 16 newest
Facing one caller, live cut from a far hand68% / 75%87% / 78%71% / 74%
Facing one caller, live cut from one-shanten65% / 66%73% / 65%67% / 67%
Safe tiles in hand when a riichi came2.96 / 3.002.72 / 2.872.95 / 3.03
No safe tile in hand when a riichi came9.4% / 9.0%8.7% / 9.4%13.1% / 8.5%

Chapter 29’s ten games were the run your hunch describes: from two-shanten or worse you cut a live tile into a riichi on 8.6% of those turns, more than twice LuckyJ’s rate, and against a caller you cut the live tile about 8 points more often than LuckyJ in the same spots. In the 16 games since, each of those lines is back on LuckyJ’s, and you dealt in to a single caller twice in 373 discards. The one number that moved the other way is the last row: a riichi found you with no safe tile in hand on 14 of 107 declarations, 13% against LuckyJ’s 8.5% at the same point. That is under two standard errors, and the safe tiles you held overall, 2.95 against 3.03, did not move.

The deal-ins, sorted by Mortal. Of your 19 deal-ins, 13 were choices Mortal made at 50% or more: tenpai pushes, riichis it declared too, and pushes into a riichi with a hand worth it. Six it gave under 25%, and two of those were about safe tiles: a suji where a genbutsu kept the same hand, and a live [[N]] two tiles from the end where a suji kept your one-shanten. They cost 4,600 points together and are the first two cards. A third came close. Early in East 3 of your 20:50 game on 9 October, at your 4th turn, you cut a live [[4m]] where the suji [[8p]] was safe against shimocha’s [[P]] pon and kept more tiles (Mortal 72% to your 3%), and toimen’s hidden pinfu took it for 2,300. In the other three Mortal’s own tile was as live as yours, and the difference was the hand: which honor goes first, twice, and a pair broken for nothing.

Today. The 14:24 game had two deal-ins and the 14:43 game three, 28,900 points in all, and each was Mortal’s first choice. In East 2 at 14:24 you pushed a live [[1m]] from two-shanten into the dealer’s riichi at your 8th turn for 7,700: you held two red fives, a genbutsu would have cost about half your tiles, and Mortal pushed it 98%. In East 4 you declared and dealt in for 8,000, a riichi Mortal declared at 96%. At 14:43 the [[F]] in East 4, two honba, was the closest call of the 16 games, the third card.

Rule Keep folding the way you have since chapter 29; your safe tiles match LuckyJ’s. When you fold, take a genbutsu over a suji whenever both leave the same hand. Past a caller’s line, cut the safe tile that keeps your hand rather than a live tile that breaks it.
Exception One number to watch: a riichi found you with no safe tile in hand on 14 of 107 declarations, against LuckyJ’s 8.5% at the same point. If that holds over your next games, the place to look is chapter 17’s safe leftovers.

Evidence Your 16 games from 7 October, 18:25, to 10 October, 14:43, against chapter 29’s ten, your 163 games before them and LuckyJ, from mine_defense_check.py. Hands from contrast/hands_all.py, 179 in the 16 games; the other columns are chapter 29’s. Turns facing a riichi from mine_riichi_folds.py: 389 in the 16 games and 24,560 in LuckyJ’s 1,255 games. Real choices against a single caller from contrast/vs_callers.py and contrast/choices.py, 78 from a far hand and 43 from one-shanten in the 16 games, set against chapter 28’s curves for one call and chapter 19’s for two or more. Safe tiles held from mine_early_safe_tiles.py: 107 riichis in the 16 games and 6,668 in LuckyJ’s games. Mortal’s choices are the site’s ratings of your games. Rank points come from your ranked-game records, anchored as in chapter 27.

31

Chapter thirty-one · Riichi or dama · added 10 October

New

Four checks before you declare a riichi

Chapter 29 found two riichis in one game that Mortal gave under 6%. Over your 354 riichis since January it gave 32 of them under 25%, and you won 5 of those 32. This chapter asks what they have in common that you can see at the table. Four things stood out, and each is a check you can make before the stick goes out.

Your riichis since JanuaryRiichisMortal declinedWonDealt in
The wait is furiten4101
Your 13th turn or later30627
3 live tiles or fewer, with a yaku already299125
3 live tiles or fewer, into a riichi10813
Any of the four641814 (22%)13 (20%)
None of them29014154 (53%)31 (11%)

A riichi can trip more than one check, so the first four rows overlap. Together, the 64 riichis that tripped a check include 18 of the 32 Mortal declined, and they won 22% of the time against 53% for the 290 that tripped none. The split by yaku matters. With a yaku already in the hand, the 48 flagged riichis won 14 times, 29%, against 91 of 159, 57%, for the rest. Where the riichi was the hand’s only yaku, the 16 flagged riichis won none. The checks are read off the same 354 riichis they are tested on, so they describe your games and are not a proven rule, and 14 of the 32 riichis Mortal declined trip none of them.

Three live tiles or fewer, with a yaku. Live tiles are the unseen copies of your winning tiles, counted from your seat. With 3 or fewer and a yaku you already hold, you declared 29 times: Mortal declined 9, and you won 12 and dealt in 5. The first card shows the cheapest remedy, a different wait. On a quiet table with six pairs, cutting the [[1m]] instead of the dora [[5p]] gave 3 live tiles instead of 2, and winning on the dora pair made the hand 4 han.

Into a riichi. You declared into another riichi on 3 live tiles or fewer 10 times. Mortal declined 8, and you won once and dealt in 3 times. Chapter 24’s exception already says to stay dama and push the same tile when the hand has a yaku; the second card is a chiitoitsu where you did not.

The 13th turn. From your 13th turn you declared 30 times and won twice, 7%, and dealt in 7 times. It is not only thin waits: the third card has 6 live tiles, and the stick was all the riichi cost. LuckyJ’s riichis from its 13th turn, on as many live tiles at the same turn, won 31%, so these 30 say more about luck than about the late riichi. Chapter 33 has the figures.

Furiten. A winning tile in your own river makes the whole wait furiten, and the hand can win only by tsumo. Four of your riichis were furiten, and none won. Mortal declined one of them, the fourth card, and declared the other three at 87% to 98%, so furiten alone is not what it objects to. The fourth card is the South 2 riichi of the 7 October, 15:27 game, which chapter 29 first showed as an ordinary two-sided tenpai. The [[9p]] you threw on your 3rd turn had made it furiten.

Rule Before you declare, check your river for a winning tile, your turn, your live tiles and whether anyone is in riichi. If the wait is furiten, or you have 3 live tiles or fewer with a yaku already or into a riichi, keep the tenpai dama, or cut a different tile for a wait with more live tiles. From your 13th turn keep dama a hand of 3 han or more, and declare a cheap one with 4 live tiles or more, as LuckyJ does (chapter 33). If the riichi is the hand’s only yaku you cannot stay dama, and the 16 flagged riichis like that won none, so look for a yaku or a better wait before you pay the stick.
Exception A riichi that trips none of the checks won 154 times in 290, 53%, and Mortal declined only 14 of them. Keep declaring those.

Evidence Your 354 riichis in the 173 games from January to 7 October, from mine_declined_riichis.py, with Mortal’s weight on declaring each from the site’s replays; a riichi it gave under 25% is one it declined. Live tiles are the unseen copies of your winning tiles from your seat when you declared; a yaku already in the hand means a ron on some winning tile has a yaku without the riichi; a wait is furiten when a winning tile is in your own river; and into a riichi means another player had already declared.

32

Chapter thirty-two · Safe tiles · added 10 October

New

Where LuckyJ keeps a safe tile and you throw it

Chapter 30 left one number to watch: on 14 of the 107 riichis in your 16 games since chapter 29, your hand held no tile safe against the declarer, against LuckyJ’s 8.5% at the same point. This chapter measures it against LuckyJ alone. LuckyJ’s habits do not fault those 14 hands, but across your games since January they show one place where it keeps a safe tile and you throw it: a guest wind already in an opponent’s river, thrown before a live value honor. The four cards are turns from closed hands, where you threw a safe tile LuckyJ keeps and a riichi came later in the hand.

The words first. A leftover is a lone tile with no tile of its suit within two, as in chapter 17. A safe leftover is one already in an opponent’s river: it is genbutsu against that player for the rest of the hand, and it often turns safe against the others. LuckyJ’s choice is measured where a hand holds a safe leftover and a live one of the same kind, two guest winds, two terminals or two middle tiles, or a guest wind already out and a live value honor (a dragon, the round wind or your seat wind), on a table where nobody has declared and nobody has two calls: which one went first, at each of its own discards.

The 14 riichis. Before those riichis you met LuckyJ’s choice only once, on your first discard in East 1 of the 9 October, 20:50 game: a [[N]] already out against a live [[E]] and [[F]]. LuckyJ throws the guest wind first there 98% of the time, and so did you. In the other 13 no such choice came before the riichi, so LuckyJ’s numbers cannot say what it would have kept. While you were not tenpai, a riichi found you with no safe tile on 11 of 92 declarations, against LuckyJ’s 7.8%, and five of the 14 came in one game, today’s 14:24.

Where LuckyJ keeps more. Your 588 such choices since January, with nobody threatening, against LuckyJ in the same spots:

Holding a safe leftover and a live oneChoicesYou threw the live one firstLuckyJ in the same spots
Two guest winds, your 1st or 2nd discard4546.7%23.5%
Two guest winds, from your 3rd discard1952.6%68.4%
Two terminals12073.3%71.7%
Two middle tiles11243.8%49.8%
A guest wind already out and a live value honor29216.1%26.4%

Terminals and middle tiles you keep as LuckyJ does. Guest winds you keep longer than LuckyJ on your first two discards and shorter after. The value honor is the clear gap: from its 4th discard LuckyJ throws a live value honor before a guest wind that is already out, and you keep the value honor, five standard deviations from its rate. Chapter 17 saw the same and called it a matter of style; measured by LuckyJ alone, it is where your safe tiles go. On the turns where LuckyJ keeps the safe tile at least 60% of the time, you threw it on 92 of 245, where LuckyJ’s rates give 75, and 39 of the 92 were a guest wind thrown before a live value honor. On 34 of the 92 a riichi came later in the hand against which the tile you threw was safe, and on 7 of those you held two safe tiles or fewer when it came.

Closed hands. These choices come almost all in closed hands, 91% of yours and 90% of LuckyJ’s, so the gap is a closed-hand gap: there you threw the live value honor first 16.2% of the time, against LuckyJ’s 25.1% in closed hands at the same discard, 4.3 standard deviations. In an open hand the value honor is often the yaku you are building toward, and LuckyJ throws it first about half the time, 49.6% of 141 choices; you met that choice only 15 times. In closed hands, where LuckyJ keeps the safe tile 60% of the time or more, you threw it on 79 of 225 turns, where its rates give 66, and 28 of those were safe against a later riichi.

Rule In a closed hand, from your 4th discard, throw a live value honor you have no pair for before a guest wind that is already in an opponent’s river, and from your 3rd, throw a live guest wind before one that is already out. LuckyJ pays a tile or two of acceptance to keep that safe tile.
Exception On your first two discards LuckyJ does the opposite: it throws the guest wind already out first, 77% of the time.

Evidence Your 189 games since January and LuckyJ’s 1,255, from mine_no_safe_riichis.py over the leftover and riichi rows of mine_early_safe_tiles.py. LuckyJ’s rate at each discard is fitted over its own choices, as in chapter 17, in hands where it was not the dealer, and so are yours: 771 choices in all, 588 with nobody threatening. The closed-hand figures and the cards use LuckyJ’s rates fitted over its closed hands alone. Chapter 30’s window: 16 games, 107 riichis, 14 with no safe tile, 92 while you were not tenpai, set against LuckyJ at the same declarer discard.

33

Chapter thirty-three · Recent games · added 11 October

New

On the climb you declared your tenpais, and since the peak you keep them dama

You asked whether pushing against riichis without perfect defense has kept you stuck, and whether to go back to what you did when your points rose fast. The climb ran from 3,972 rank points on 23 September, 18:40, to 6,749 at the peak on 1 October, 22:20: 62 games with 20 firsts and 4 fourths, +44.8 points a game. In the 47 games since the peak you have 14 firsts and 11 fourths, −10.0 a game, and you stand at 6,278. Go back, to the climb’s riichis. Your defense against riichis is the same as it was on the climb.

Your defense held. Where a safe tile kept your hand as good as a live one against a riichi, you threw the live tile about as often as on the climb, and about as often as LuckyJ. With a tenpai that only unsafe tiles kept, you pushed a little less often than on the climb, and those pushes made +1,310 a hand. In every row of the table you are close to the climb, and close to LuckyJ.

Against a riichiOn the climbSince the peakLuckyJ
Live tile thrown where a safe tile kept your tenpai3.3%4.8%6.1%
Live tile thrown where a safe tile kept your one-shanten6.2%6.9%7.8%
Live tile thrown where a safe tile kept two-shanten or worse4.1%4.4%3.6%
Deal-ins per 100 discards facing a riichi1.952.181.71
Tenpai kept when every tenpai tile is unsafe82.4%76.5%82.3%
What keeping it made, a hand+1,023+1,310+518

What changed is the riichi. The riichi choice comes up at your first closed tenpai. To set you beside LuckyJ hand for hand, each of your first tenpais with nobody in riichi gets LuckyJ’s rate for a hand of the same value and wait, the same live tiles, turn and seat, and the same calls across the table, fitted over 2,177 of its own. On the climb you declared 107 of 132 where those rates expect 104.9, LuckyJ’s habit almost exactly. Since the peak you declared 60 of 94 where they expect 70.7: about 11 riichis fewer in 47 games, 3.1 standard deviations below. Your riichis fell from 21.9 to 16.1 per 100 hands, against LuckyJ’s 19.3, your riichi wins from 10.2 to 8.5 per 100 hands, and all your wins from 25.0% of hands to 22.4%.

Your first closed tenpaisGamesTenpaisYou declaredLuckyJ’s rates expect
23 September to chapter 24 (1 October, 14:17)501179592.9
Chapter 24 to chapter 29 (7 October, 18:09)30685053.2
Chapter 29 to chapter 31 (10 October, 18:11)15251417.0
Since chapter 31716812.5

The gap opened as this guide gave you reasons to stay dama. Chapter 24 kept big hands dama, chapter 29 said to keep a tenpai dama on your 13th turn with 5 live tiles, and chapter 31 made the 13th turn one of four checks. In the 7 games since chapter 31 you declared 8 of 16 first tenpais where LuckyJ’s rates expect 12.5. Your peak came the evening chapter 24 went up.

Where the riichis went. On cheap hands, no yaku or 1 or 2 han on every winning tile, you declared 52 of 70 since the peak where LuckyJ’s rates expect 58.5. On hands of 3 han or more you declared 8 of 24 against 12.2. On the climb you declared those more often than LuckyJ, 19 of 29 against 16.1, and after chapter 24 you went past it the other way. From your 13th turn you declared 5 of 17 against 9.9.

The 13th turn was the guide’s mistake. LuckyJ keeps a late tenpai dama when it is big. From its 13th turn it declared 76% of its cheap first tenpais, 81% with 4 or 5 live tiles and 88% with 6 or more, and 34% of those worth 3 han or more. Chapter 31 read the check off your own late riichis, which won 2 of 33. LuckyJ’s late riichis won 31%, on as many live tiles, 5.1 against 5.0, at the same turn and with as much wall left. Late hands do win less often in your games, riichi or not: your late dama tenpais won 14% against LuckyJ’s 26%. On top of that the 33 riichis were unlucky, and two wins are no reason to stop. Chapters 29 and 31 are corrected, and the corrections are highlighted.

One part of your defense did slip, against callers. Your deal-ins rose from 9.9% of hands to 11.9%. The biggest part of the rise came while you were not tenpai and the winner had called: 1.9 per 100 hands on the climb, 3.0 since. Against riichis, discard for discard, you dealt in about as often as on the climb. Chapters 19 and 28 have LuckyJ’s fold lines against callers.

Some of the climb was luck. Four fourths in 62 games is 6%. Before the climb you came fourth in 23% of your games, and since the peak in 23%. No habit brings that part back.

The first card is a riichi from the climb’s last day. The other three are first tenpais since the peak that you kept dama where LuckyJ declares 74% to 96% of the time, and the last two came after chapter 31.

Rule Declare your first closed tenpai as you did on the climb. A cheap hand, no yaku or 1 or 2 han, gets the riichi at any turn when it waits on 4 live tiles or more. A hand of 3 han or more follows chapter 24’s chart, and a furiten wait or one of 3 live tiles or fewer follows chapter 31. Keep folding and pushing against riichis as you do now; that part matches the climb and LuckyJ.
Exception In all-last your place decides, as in chapter 25: from first LuckyJ keeps nearly every tenpai dama.

Evidence Your 196 games since January, 87 before the climb, 62 on it and 47 since the peak, against LuckyJ’s 1,079 Tokujou games for hands and riichis and all 1,255 of its games for the defense rows, from mine_climb_check.py over contrast/hands_all.py, contrast/tenpai.py, mine_riichi_folds.py and mine_tenpai_push.py. Rank points are from your ranked-game records, set to 6,700 before the 3 October, 14:11 game as in chapter 27, and the climb runs from the lowest point in the ten days before the peak to the peak. A first closed tenpai counts when the riichi is legal, the wait is not furiten and nobody is in riichi. LuckyJ’s rate is a logistic fit over its 2,177 such tenpais with chapter 24’s value and wait, the live tiles, the turn, the dealer and an opponent on two calls or more, and it matches LuckyJ’s actual rates by turn and by live tiles within a point. The stretches split your games at the times chapters 24, 29 and 31 went up.

34

Chapter thirty-four · Recent games · added 11 October

New

Three fourths in a row, from riichis LuckyJ declares

You have lost three games in a row while declaring more, and asked why. The three games are from 11 October at 20:27, 21:09 and 21:33; the first started before chapter 33 went up. You came fourth in all three, on 13,000, 15,200 and 16,200, and your rank points went from 6,278 to 5,484. You played as well as usual: your discards matched Mortal’s first choice 74.4% of the time, against 73.2% in your 196 games before, and your calls 94.2% against 93.1%.

The riichis were right. You declared 8 riichis in the three games. Mortal declared 7 of them at 94% or more and the eighth at 60%, and for the five that were first tenpais with nobody in riichi, LuckyJ’s rates from chapter 33 give 82% to 98%. They won twice, drew three times and dealt in three times, −5,800 points with the sticks. Once you declare you throw every tile you draw, and three of those draws were someone’s winning tile: 3,900, 8,000 and 7,700.

That part was the draw. Pick 8 riichis at random from LuckyJ’s own 2,171 and 3 or more of them deal in 8.4% of the time; together they lose 5,800 points or more 6.2% of the time. LuckyJ loses points on 34.7% of its riichis. In its 1,255 games it came fourth three times in a row twice and four times in a row once, and a player who places as LuckyJ does has about an even chance, 51.3%, of such a run somewhere in 199 games.

RiichisRiichisWonDealt inPoints a riichiLost points
Yours since January, before these games38248.7%12.8%+2,46437.4%
LuckyJ’s2,17151.8%13.7%+2,94134.7%
Your first tenpais with nobody in riichi29651.7%11.1%+2,72133.8%
LuckyJ’s first tenpais with nobody in riichi1,66554.7%12.2%+3,30330.3%
Yours in these three games825.0%37.5%−72537.5%

Your riichis pay. Over 382 riichis since January they made +2,464 points a riichi, close to LuckyJ’s +2,941, and about a third of them lost points, as about a third of LuckyJ’s do. A short run of riichis that lose is part of what riichi costs, for LuckyJ as much as for you.

Where your 30,600 points wentHandsPoints
Deal-ins without a riichi5−13,200
Other players’ tsumo4−10,100
Hands with your riichi, sticks included8−5,800
Draws without a riichi4−5,000
Your wins without a riichi2+3,500
Rons between other players60

The deal-ins without a riichi. Four of the five were the tile Mortal throws: three as its first choice, at 58%, 52% and 64%, and the all-last [[3s]] of the first game, which it threw with a riichi. That [[3s]] cost the most: you were level with shimocha for third, and it left you 7,800 behind. One was yours: the live [[E]] of the third game, 2,500 points, a tile Mortal threw 0.7% of the time.

The first two cards are the riichis that dealt in for 8,000 each. The third is the [[E]], and the fourth is the all-last [[3s]].

Rule Keep declaring the riichis LuckyJ declares, and judge each one on the turn you declared it, before you see how the hand ends. The deal-ins that are yours to cut are the ones before you declare: late in a hand, a live honor you have just drawn is the wait a quiet seven-pairs hand hopes for.
Exception Three games are too few to judge a habit by. Judge one over 20 games or more, as chapter 33 did.

Evidence Your three games of 11 October from 20:27, against your 196 games before them, LuckyJ’s 1,079 Tokujou games for its 2,171 riichis and all 1,255 of its games for the runs of fourths, from mine_streak_check.py over contrast/tenpai.py and the site’s replays. A riichi’s points are the hand’s result from your seat, the stick included. The chances resample 8 of LuckyJ’s riichis 200,000 times and 199 of its places 20,000 times. Mortal’s weights come from the replays, and LuckyJ’s rates on first tenpais are chapter 33’s fit.

35

Chapter thirty-five · Push or fold · added 12 October

New

Big hands on thin waits: keep pushing, and keep the better wait

You said that if you had not challenged the riichis you would not have dealt in, and asked about the one spot where your challenges have cost you: a tenpai worth 3 han or more, counting a closed hand’s riichi, on 3 live tiles or fewer, against a riichi, where every tile that keeps the tenpai is unsafe. You met it 57 times since January and pushed 42 of them.

A big hand on a thin wait, against a riichiYouLuckyJ
Times met57236
Pushed73.7%83.1%
Pushes that won11.9%23.0%
Pushes that dealt in on the tile pushed21.4%10.7%
Pushes that dealt in later, into the riichi11.9%8.2%
Points a push−1,655+618
Points a fold−1,820−1,488

Folding did not save points. Your 15 folds in this spot averaged −1,820, worse than your pushes. A fold gives up your hand, and the riichi still wins, or someone else does. LuckyJ’s folds here lose 1,488 points a hand and its pushes make 618, and it pushes 83% of the time to your 74%. That is in hand points. In rank points, which decide fourth place, your folds here cost less than your pushes, 6.6 against 18.6 a hand, though by only about one standard error, and LuckyJ’s folds cost more than its pushes. Chapter 36 has the figures.

Your pushes were LuckyJ’s pushes. The tiles you pushed were of the same kinds as LuckyJ’s. When two tiles kept the tenpai you cut the less dangerous one 15 times in 22, as LuckyJ did 63 times in 93, and 42.9% of your waits were tiles already safe against the riichi, against LuckyJ’s 41.8%. Mortal’s weight on keeping the tenpai averaged 90% over your 42 pushes, and was 75% or more on all 9 that dealt in on the tile you pushed.

The results were worse than luck usually allows. Your wins and deal-ins were the size of LuckyJ’s, 8,360 and 5,721 points against 8,996 and 5,597; there were fewer wins and more deal-ins. At LuckyJ’s rate for each kind of tile, your 42 pushes should have dealt in on the push about 5.1 times, and they did 9 times, which LuckyJ’s rates give 6.1% of the time. Draw 42 of LuckyJ’s own pushes here at random and they lose as many points 0.3% of the time. I chose this spot because it was the worst of the four I looked at, which makes a bad sample likelier, and outside it your pushes match LuckyJ’s: 177 pushes won 31.6% and dealt in 20.9%, +698 a hand, against its 32.4%, 21.0% and +499. I found no habit behind the gap.

Two things to change. Five times Mortal kept the tenpai with the other tile, at 84% or more. The second card is the plainest: in seven pairs you cut the more dangerous [[4s]] to wait on the last [[9p]], where cutting [[9p]] left 3 live tiles and a bigger hand. The third card is about which single wait to keep: in this spot LuckyJ’s pushes on a lone honor won 35.3% of the time, its waits on number tiles 20.4%. And fold less: Mortal would have kept the tenpai in 6 of your 15 folds, and the fourth card is one, a mangan on the dora.

The first card is a push that dealt in and was right. The other three are the changes.

Rule With a tenpai of 3 han or more on 3 live tiles or fewer against a riichi, push as LuckyJ does, more than 4 times in 5. When two tiles keep the tenpai, keep the wait with more live tiles and the bigger hand, and between two single waits keep the honor; let the safer cut decide only between waits that are equal.
Exception A wait with a single live tile left is a weaker hand: LuckyJ pushed only 2 of its 5 such spots.

Evidence Your 199 games since January against LuckyJ’s 1,255, from mine_thin_pushes.py, which takes the spots of mine_tenpai_push.py (the first discard of a hand where someone else is in riichi, you hold tenpai and every tile that keeps it is unsafe) and follows each to the end of the hand. A hand’s value counts a closed hand’s riichi; live tiles are the unseen copies of the winning tiles from your seat; a tile’s kind is its label against the riichi, as in the tables of the earlier chapters. Mortal’s weights come from the site’s replays. The chances draw 100,000 times.

36

Chapter thirty-six · Push or fold · added 12 October

New

Pushes and folds in rank points

You asked to measure chapter 35’s folds by EV, since what matters is not coming fourth. Here every push and fold against a riichi is valued by what the hand did to your expected rank points for the game, at Saint 3 Jade’s values: +130, +65, 0 or −250 for the place, plus (score − 30,000) / 1000. The expected value of any moment in a game comes from playing out its finish 4,000 times with hands drawn from your own 199 Jade games, 2,082 hands in all; when a hand ended the game, its value is the rank points you got.

Against a riichi, every tenpai tile unsafeYour pushesYour foldsLuckyJ’s pushesLuckyJ’s folds
Big hand on a thin wait (chapter 35)421519640
Rank points a hand−18.6 ± 8.4−6.6 ± 8.9+3.0 ± 4.0−12.9 ± 8.2
Chance of fourth, change in points+4.1−1.1−0.3+3.0
Every such spot219611,238267
Rank points a hand−0.4 ± 4.2−7.3 ± 3.8+1.0 ± 1.6−8.5 ± 2.6
Chance of fourth, change in points+0.3+0.1+0.4+1.6

In chapter 35’s spot, your folds cost less. Your 15 folds cost 6.6 rank points a hand and your 42 pushes 18.6, and the pushes raised your chance of fourth by 4.1 points where the folds lowered it by 1.1. In hand points the folds lost more, so chapter 35 is corrected. The gap, 12 rank points a hand, is about one standard error of these 57 hands, and LuckyJ’s own spots point the other way: its pushes there gained 3.0 and its folds lost 12.9.

A fold’s value is mostly what the others do. Chapter 35’s fourth card, the mangan you folded from third, gained 70.3 rank points: after you folded, toimen dealt 8,000 into shimocha’s riichi and left you further from fourth. Its first card, the mangan you pushed from second, lost 95.4 and raised your chance of fourth by 25.7 points. Set side by side after the fact, pushes and folds measure the hands that followed as much as the choice, which is why 57 hands decide so little.

Over every spot, pushing is worth more. Across your 280 push-or-fold spots against a riichi, a push cost 0.4 rank points a hand and a fold 7.3; LuckyJ’s pushes gained 1.0 and its folds lost 8.5. Your folds did not lower your chance of fourth either: +0.1 points against +0.3 for the pushes.

Your place when the spot cameYou pushedYour pushesYour foldsLuckyJ pushed
First80.5%−1.8−18.783.7%
Second78.8%−0.10.080.1%
Third65.4%+10.5−1.781.7%
Fourth89.7%−8.3−14.683.2%

LuckyJ does not fold to stay out of fourth. It pushes 80% to 84% of the time from every place, and 81% to 83% in the East, in South 1 to 3 and from South 4 on, and its folds lose rank points from every place. You fold most from third, where you pushed 65.4% of the time, and there your pushes did best, +10.5 a hand.

Where your pushes did cost places. In chapter 35’s spot your 10 pushes from second cost 42.2 rank points a hand, and your 18 pushes in South 1 to 3 cost 24.6 where your 10 folds there broke even. These are small samples, and LuckyJ’s pushes in the same places come out close to even, −2.8 and +2.5, so they are no reason to fold more. They are the places to look at again after 20 more games.

Rule Measured in rank points the answer is the same: push as LuckyJ does, from every place. Over your 280 spots a push was worth about 7 rank points a hand more than a fold, and LuckyJ’s about 9.5.

Evidence Every push-or-fold spot of mine_thin_pushes.py in your 199 games since January and LuckyJ’s 1,255, valued by mine_push_ev.py. A hand’s value is the expected rank points at the start of the next hand less those at the start of this one, each the mean of 4,000 finishes drawn from your own hands; LuckyJ’s spots are valued on the same Jade model, so both sides use one ruler. A game ends as a Mahjong Soul south game does, and ties go to the seat nearer the first dealer. The ± figures are one standard error.

37

Chapter thirty-seven · Push or fold · added 12 October

New

Your 15 folds, played out

You asked for the what-if on chapter 35’s 15 folds. Each fold was played both ways from its turn: the tile you folded with, and the push, the tile that kept your tenpai that Mortal weighted highest. Mortal then played all four seats to the end of the hand and through the rest of the game, 500 times: once on the real wall and 499 times on reshuffles of the tiles nobody had seen, with every hand as it really was. Each sample scores both lines in Saint 3 Jade rank points, so the comparison is about places, as you wanted.

Your foldFold → pushMortal keeps the tenpaiPush minus fold, rank pointsChance of fourth, fold → push
26 June, South 1[[1m]] → [[4s]]40%+23.0 ± 7.545.8% → 40.8%
19 September, South 2[[P]] → [[6s]]100%−0.8 ± 2.20.4% → 0.8%
20 September, East 3[[C]] → [[8m]]0%+4.1 ± 4.27.4% → 7.0%
20 September, East 3[[P]] → [[6p]]15%−95.1 ± 8.713.4% → 39.0%
23 September, East 1[[S]] → [[3s]]97%+36.2 ± 7.528.4% → 20.8%
26 September, South 1[[P]] → [[4p]]3%+6.4 ± 3.914.8% → 13.0%
27 September, South 1[[2s]] → [[7s]]2%−20.9 ± 7.025.8% → 33.0%
30 September, South 2[[9p]] → [[5m]]2%+1.4 ± 5.551.0% → 49.4%
1 October, East 3[[C]] → [[S]]58%−37.1 ± 7.428.0% → 38.6%
1 October, South 3[[2s]] → [[2m]]99%+38.2 ± 7.441.4% → 30.6%
9 October, South 2[[4m]] → [[5p]]6%+1.8 ± 6.813.6% → 16.2%
9 October, South 3[[1s]] → [[2p]]55%+1.9 ± 3.58.4% → 8.6%
10 October, East 3[[1m]] → [[6s]]4%−2.6 ± 3.04.6% → 5.2%
10 October, South 2[[1p]] → [[1s]]23%−42.8 ± 8.233.8% → 46.8%
11 October, South 3[[1m]] → [[8s]]54%+23.0 ± 3.41.8% → 1.2%

On balance your folds cost nothing. Pushed instead, the 15 would have averaged −4.2 rank points a fold, with an error of 8.8 across the folds. Four were better pushed, by 23 to 38 rank points each, four were better folded, by 21 to 95, and seven came out within a few points either way.

The four right folds held back the riichi’s winning tile. With the real hands kept, those pushes deal in on every sample. That part is hindsight, since you could not see the hands. Mortal could not see them either, and it folded three of the four and was split on the fourth, the first card.

Mortal’s read sorts them. On the six folds where Mortal would have kept the tenpai, pushing gains 10.2 rank points a fold; on the nine where it folds too, pushing loses 13.9. So chapter 35’s advice holds in rank points: fold where Mortal folds and push where it keeps. Two of the folds that cost you already have cards, chapter 35’s folded mangan, +38.2 pushed, and chapter 11’s last-discard tenpai, +36.2: folding that one left you tied for last at 23,500, and the tenpai left you tied for first.

The first card is the right fold Mortal was split on; the second is your newest fold, where pushing [[8s]] was worth 23.0 rank points.

Rule With a big hand on a thin wait against a riichi, fold where Mortal folds and push where it keeps the tenpai. Over these 15 folds the six it keeps are worth 10.2 rank points each pushed, and your other nine folds were right.

Evidence mortal_whatif.py on each of chapter 35’s 15 folds, 500 samples each, gathered by mine_fold_whatif.py. Mortal plays every seat after the turn, so each line is your choice as Mortal would follow it up. The ± in the table are the reshuffles’ error; across the 15 folds the error of the mean is 8.8. While these ran, two faults in the simulator’s kans were fixed: Mahjong Soul takes a kan’s replacement tile from the wall’s last tile, checked on 22 kans, and a reshuffle no longer returns a replacement tile already drawn. Chapter 27’s simulation had no kan before its turn, so its reshuffles are unaffected.

38

Chapter thirty-eight · Push or fold · added 13 October

New

Why you should fold a little more than LuckyJ

You asked for the theory behind folding slightly more than LuckyJ, and for what makes the EV gap. A push against a riichi pays when your chance of winning, times what a win does for your game, beats your chance of dealing in, times what that costs, set against what folding costs. Every number in that comparison is the player’s own. Three things could make yours differ from LuckyJ’s: the point system you play under, your opponents, and your read of the tile you push. Only the last one does.

What could differYour sideLuckyJ’s side
Points for a south gameSaint 3 Jade: +130, +65, 0, −250, plus (score − 30,000) / 1000Tenhou Tokujou at 10-dan: +75, +30, 0, −180
Chapter 37’s folds pushed instead, a fold−4.2, 1.11% of the 380-point range−3.0, 1.18% of the 255-point range
Opponents’ riichis that won48.3%47.7%
Their live tiles, value and turn5.37, +2,419, 9.15.32, +2,384, 9.1
Tiles pushed into a riichi that were its winning tile27 of 21920.3 at LuckyJ’s rate for the same kinds

Not the point system. Tenhou, where LuckyJ’s games come from, pays on placement alone and takes 2.4 times as much for fourth as it gives for first; your Jade room takes 1.9 times as much and adds your score. Scored under Tenhou’s values, chapter 37’s 15 played-out folds come out the same, 1.18% of the range against 1.11%, and the score term alone moves them 0.1 rank points. If anything, your system asks for slightly more pushing than LuckyJ’s.

Not your opponents. Their riichis win as often as the riichis at LuckyJ’s tables, on as many live tiles, for the same value and at the same turn.

The tile you pushedYour pushesDealt inLuckyJ’s pushesDealt in
Live honor156.7%1345.2%
Live terminal4112.2%2159.8%
Half suji4511.1%2006.5%
Live 2-3-7-88214.6%52410.1%
Live 4-5-63411.8%15912.6%

Your read. The tiles that need a read deal in more for you: a half-suji 11.1% of the time against 6.5%, a live 2-3-7-8 14.6% against 10.1%. On a live 4-5-6, which no read makes safe, there is no gap. In all, 3.1 percentage points more deal-ins on every push. The waits you keep after pushing win as often as LuckyJ’s with 4 to 7 live tiles, 33.8% against 33.1%, and less often with 3 or fewer, 11.1% against 20.3%. A deal-in on the tile pushed moves your expected rank points 57.8 the wrong way, and a push that does not deal in gains 18.5, so the extra deal-ins cost about 2.3 rank points on every push.

What LuckyJ sees that the labels do not. A label puts every half-suji in one box; LuckyJ sorts them by the rest of the riichi player’s river. One sign of it: you pushed a tile one or two ranks from the riichi’s declaration tile, in its suit, as often as LuckyJ, 14.9% of your number-tile pushes against 14.4%, and 6 of your 30 dealt in, 20.0%, where LuckyJ’s dealt in 9.5%. Thirty pushes are too few to make a rule of it, but it is where to look.

Your tenpai against a riichiLuckyJ pushesIts pushesIts foldsYou pushYour pushesYour folds
1 or 2 han, 3 live tiles or fewer65.6%−14.0−2.039.3%−15.8−4.4
1 or 2 han, 4 to 7 live tiles73.9%−4.9−11.375.9%−2.8−12.2
3 han or more, 3 live tiles or fewer83.1%+3.0−12.973.7%−18.6−6.6
3 han or more, 4 to 7 live tiles93.7%+5.0+0.495.0%+10.3−15.3

Where to fold more. About 2.3 rank points a push is the price of the read for now, so fold where LuckyJ’s own edge from pushing is smaller. Cheap hands on thin waits come first: even LuckyJ’s pushes there do worse than its folds, and you already push 39.3% of the time to its 65.6%. Cheap hands on wider waits are close: LuckyJ’s pushes there beat its folds by 6.4, not far above the price of your read, so fold the closest of them. Big hands with 4 to 7 live tiles are where to match it: your pushes there gain 10.3 a hand and your folds lose 15.3. Big hands on thin waits are chapter 37’s: fold where Mortal folds. The figures in this table carry errors of 3 to 13 rank points, so read them as directions.

Rule Match LuckyJ’s pushes with 3 han or more and 4 live tiles or more. Fold more than it does with 1 or 2 han on 3 live tiles or fewer, as you already do, and fold the closest calls with 1 or 2 han on wider waits: until your pushed tiles deal in as rarely as LuckyJ’s, each push carries about 2 rank points that LuckyJ’s does not.
Exception The gap is a read, and reads improve. When your half-suji and 2-3-7-8 pushes deal in as rarely as LuckyJ’s, go back to its push rates.

Evidence Tenhou’s published dan points (tenhou.net) for the Tokujou and Houou rooms at 10-dan, and your Jade values read off your own results (chapter 36). Chapter 37’s mortal_whatif.py runs rescored, the opponents’ riichis from contrast/tenpai.py over your 199 games and LuckyJ’s 1,079 Tokujou games, and the push-or-fold spots of mine_thin_pushes.py over your 199 games and LuckyJ’s 1,255 valued by chapter 36’s model, all gathered by mine_fold_theory.py. A tile’s kind is its label against the riichi; the expected deal-ins apply LuckyJ’s rate for each kind to your pushes. LuckyJ’s training goal is not published; Tencent describes its method only as regret-minimizing self-play.

39

Chapter thirty-nine · Push or fold · added 13 October

New

Your pushes are a human’s, and chasers and callers are not the gap

You doubted that your pushes could be so much worse, and asked whether your opponents chase and push open hands better than Tokujou’s players. This chapter sets your pushes beside the human players at both tables, the other three seats of your 199 Jade games and of LuckyJ’s 1,255 Tenhou games, in the same push-or-fold spots, and checks the chasers and the callers.

Tenpai against a riichi, every tenpai tile unsafeYouJade players at your tablesTokujou’s humansLuckyJ
Spots2807194,8331,505
Pushed78.2%80.4%78.8%82.3%
The tile pushed dealt in12.3%11.1%11.5%9.2%
A half-suji pushed dealt in11.1%11.2%12.0%6.5%
A live 2-3-7-8 pushed dealt in14.6%9.8%11.8%10.1%
Pushes that won27.9%29.6%29.6%30.9%
Pushes on 3 live tiles or fewer that won11.1%23.4%21.9%20.3%

Your pushes are a human’s. The tile you push deals in 12.3% of the time; the Jade players at your tables, 11.1%, and Tokujou’s humans, 11.5%. On a half-suji you are exactly a human, 11.1% against 11.2% and 12.0%, and LuckyJ is the one apart, at 6.5%. Chapter 38’s read gap is the gap between a human and LuckyJ, and you have it no more than the players around you.

Two places you trail them. A live 2-3-7-8 you push deals in 14.6% of the time, against 9.8% for the players at your tables and 11.8% for Tokujou’s. And your thin waits win less after a push: 6 times in 54, 11.1%, where the humans’ rates give about 12 wins.

Not the chasers. In the hands where an opponent declared first, another opponent chased 17.6% of the time at your tables and 15.5% at LuckyJ’s, and a chaser won the hand 5.9% and 5.7% of the time. Two riichis faced 6.4% of your pushes and 6.4% of the Jade players’. Your deal-ins later in the hand, after the push, run at LuckyJ’s rate, 11.0% of pushes against 11.6%.

Your pushes and theirsYouJade players at your tablesTokujou’s humansLuckyJ
Spots with a caller at the table45.7%48.3%49.5%47.2%
After a push, an open hand won6.4%7.6%8.2%7.0%
With a caller: the tile pushed dealt in11.1%14.3%13.3%10.6%
With no caller: the tile pushed dealt in13.3%8.0%9.9%8.1%

Not the callers. A caller sat at the table at 45.7% of your push spots and 48.3% of the Jade players’, and after your pushes an open hand won 6.4% of the time, after theirs 7.6%. With a caller at the table you pushed as well as LuckyJ: you won 27.3% and dealt in 20.2%, it won 29.0% and dealt in 21.7%. Your extra deal-ins come in the quiet spots, with nobody else open: there the tile you pushed dealt in 16 times in 120, 13.3%, against 8.0% for the Jade players and 8.1% for LuckyJ.

Not the riichis’ waits. Over every discard of a live tile into a lone riichi, by anyone, a half-suji was the winning tile 9.7% of the time at your tables and 9.6% at LuckyJ’s, and a live 2-3-7-8 8.1% and 8.6%. Your own discards over whole hands sit with your tables, at 8.4% and 8.2%. The riichis you face wait where Tokujou’s do.

Rule Against a lone riichi with nobody else open, treat a live 2-3-7-8 as the dangerous tile it is: you push it into the wait half again as often as the players at your tables, 14.6% against 9.8%. With a thin wait, keep the one the folding players will throw (chapter 35). Chasers and callers need nothing new from you.
Exception With a caller at the table, your pushes already match LuckyJ’s: 27.3% won and 20.2% dealt in. Change nothing there.

Evidence The push-or-fold spots of mine_thin_pushes.py run for every seat of your 199 games since January and LuckyJ’s 1,255 Tenhou games, the chases over every hand where an opponent declared first (your 943 and LuckyJ’s 5,026 in its 1,079 Tokujou games), and every discard of a live tile into a lone riichi checked against the declarer’s real winning tiles, all from mine_push_company.py. A caller is an opponent with a call who is not in riichi; a tile’s kind is its label against the riichi.

40

Chapter forty · Safe tiles · added 13 October

New

How LuckyJ reads a half-suji: it chooses which one

You asked how LuckyJ reads half-suji so much better than any human. A half-suji is a 4, 5 or 6 with only one of its suji partners safe: the safe partner rules out one two-sided wait through the tile and leaves the other. Call the side it leaves open the open side: a [[4p]] with its [[7p]] safe still loses to a [[2p]]-[[3p]] wait. The open side’s wait is made of two tiles, here the [[2p]] and the [[3p]]; with the [[1p]] safe instead, it is the [[5p]] and the [[6p]]. Against a lone riichi, LuckyJ’s half-suji were the winning tile 6.7% of the time over 566 throws, Tokujou’s humans’ 10.2%, the Jade players’ at your tables 9.7% and yours 8.4%.

A half-suji thrown into a lone riichiLuckyJTokujou’s humansYou
While keeping tenpai8.4% of 32211.8% of 1,0387.9% of 63
Folding, with a genbutsu in hand6.6% of 1068.2% of 5476.2% of 16
Folding, with no genbutsu in hand2.9% of 1388.5% of 32812.5% of 16

It chooses. A player who throws a half-suji usually holds another. When LuckyJ held two or more, the one it threw was the winning tile 6.1% of the time and the ones it kept 10.4%: it threw the safer one. Humans did not. Tokujou’s threw one that hit 10.7% of the time and kept ones that hit 9.4%, the Jade players 10.9% and 9.6%, and you 8.1% and 7.8%. A random pick would have done 9.8% for the humans. LuckyJ’s edge is widest when it folds with no genbutsu in hand: its half-suji then hit 2.9% of the time.

Holding two or more half-sujiTimesThe one thrown hitThe ones kept hit
LuckyJ4126.1%10.4%
Tokujou’s humans1,30610.7%9.4%
Jade players at your tables18410.9%9.6%
You628.1%7.8%

The cues it sees, and you can too. Over 6,140 half-suji, the ones thrown and the ones held beside them, the average was the winning tile 9.5% of the time. These cues move that:

A half-suji whereHalf-sujiThe winning tile
All four copies of a tile its open side needs are visible3242.8%
Three or more copies of such a tile are visible (a wall)1,6257.2%
The riichi player threw one of the two tiles of the open-side wait, early7605.4%
The same, after their sixth discard9017.3%
Nothing thrown on the open side4,55810.5%
The safe partner was thrown late, from the hand88510.6%
The dora is one or two ranks away68410.5%

LuckyJ uses them a little more than humans, and something more. When its thrown and kept half-suji differed on a cue, LuckyJ threw the one with a wall on the open side 68.2% of the time and Tokujou’s humans 63.0%, the one with an early discard on the open side 68.6% against 63.3%, and the one with fewer open two-sided combinations 59.4% against 54.7%. A few points do not make its 6.1%. A danger model built from these cues and more, over 18,903 live number tiles, rates LuckyJ’s half-suji as dangerous as the humans’, 9.3% against 9.4%. LuckyJ reads more than any cue here, and that part has no short rule. Chapter 41 gave three models the riichi player’s whole river in order, and they did not find it either.

An example from LuckyJ’s games. East 2 of a game on 12 May 2023, its 15th turn, folding against a riichi, it held four half-suji: [[4p]], [[5p]], [[5m]] and [[6m]]. The [[4p]]’s open side was the [[2p]]-[[3p]] wait, and all four [[2p]] were visible, so that wait could not exist; it threw the [[4p]]. The [[6m]]’s open side, the [[4m]]-[[5m]] wait, had one copy of each visible. The riichi waited on [[3m]]-[[6m]].

What you can take. A checklist: count −2 when all four copies of a tile the open side needs are visible and −1 for three, −1 if the riichi player threw one of the two tiles of the open-side wait before declaring (for a [[4p]] whose [[7p]] is safe, a [[2p]] or a [[3p]]), +1 if the safe partner was a late discard from the hand, +1 if the dora is one or two away, and throw the half-suji with the lowest count. Picking that way, the 1,552 human choices above would have hit 8.4% of the time instead of 10.6%, about half of LuckyJ’s edge. It did not help in every sample: in your 62 choices it would have hit 11.0%, and in the Jade players’ 184, 10.4%. Use it to compare your half-suji, not as a rule that wins every hand.

Rule When you hold two half-suji against a riichi, compare them before you throw. Let go of the one whose open side is walled or already thrown into by the riichi player, and keep the one whose open side nobody has touched. Humans throw a half-suji at random; LuckyJ throws the safer one.
Exception When you push a tenpai, your wait decides the tile, not this choice; there chapter 38’s advice to fold a little more applies.

Evidence Every live number tile thrown into a lone riichi by any player in your 199 games since January (2,569 throws) and LuckyJ’s 1,255 Tenhou games (16,334), checked against the riichi’s real winning tiles, with every other half-suji the thrower held, from mine_half_suji.py. The open side, the walls and the riichi player’s discards are read from the thrower’s seat. The danger model is a ridge logistic fit, checked on 3,753 throws from games it was not fitted on, where it ranks a winning tile above a safe one 60% of the time.

41

Chapter forty-one · Safe tiles · added 13 October

New

How LuckyJ picks the tile it folds with

You asked how LuckyJ folds so well, so this chapter trains models on it. It counts real folds only: one player is in riichi, the folder holds no genbutsu, and the tile thrown leaves the hand further from tenpai than its best discard would. In 547 real folds LuckyJ’s tile was the riichi’s winning tile 7 times, 1.28%. The humans at both tables hit 4.09% of the time in 1,857, and you 2.70%, 3 times in 111.

Real folds by the kind of tile thrownLuckyJHumans at both tablesYou
Suji146, none hit503, 3.0% hit29, 3.4% hit
Honor with two other copies seen90, none hit240, 1.2% hit15, none hit
Live honor112, 1.8% hit309, 1.9% hit18, none hit
Live terminal74, 2.7% hit306, 6.2% hit15, none hit
Live 2-3-7-854, 3.7% hit262, 7.6% hit11, 9.1% hit

The same kinds, safer tiles. LuckyJ folds with the same kinds of tile as the humans, in the same shares: a suji in 27% of its real folds and of theirs. The gap is inside each kind. Its 146 suji never hit; the humans’ hit 3.0% of the time. Its live terminals hit 2.7% against 6.2%, and its live 2-3-7-8 3.7% against 7.6%.

Three models, and LuckyJ reads more than all three. A small neural network, LightGBM and TabPFN 3.5 each learned the danger of 858,565 tiles, every tile held by anyone who discarded while one riichi was out, from what the table shows: the copies seen of the tile and its neighbours, genbutsu, the riichi player’s discards around it (early or late, from the hand or the draw, the declaration tile), the dora, the timing, and the ways the tile can still be a winning tile. Each scored games it had not learned from. Throwing the tile a model rates safest would have cut the humans’ 4.09% to between 3.07% and 3.72%. In LuckyJ’s own folds, the models’ safest tiles hit between 1.46% and 1.65%, about what LuckyJ got itself, and LuckyJ got there with tiles LightGBM rated 4.00% dangerous and TabPFN 3.88%. The humans’ tiles it rated 4.67%, and they hit 4.09%. Humans read about what the models read; LuckyJ reads more, and giving the models the riichi player’s whole river in order did not close it. Its opponents are not easier to read either: Tokujou’s humans folded slightly better against LuckyJ’s riichis than against each other’s, 4.15% against 4.66% over all their discards with no genbutsu.

Real folds: throwing this tile would have hitLuckyJ’s foldsThe humans’Yours
The tile actually thrown1.28%4.09%2.70%
The network’s safest1.46%3.72%1.80%
LightGBM’s safest1.65%3.07%1.80%
TabPFN 3.5’s safest1.65%3.07%1.80%
The checklist’s safest (below)1.98%3.13%1.80%
A tile at random8.29%8.91%9.34%

What you can count. Over the 858,565 tiles, these cues moved the danger most:

A tile whereTilesThe winning tile
An honor with no other copy seen: a dragon10,1565.7%
The same, a value wind4,8434.8%
The same, a guest wind4,1343.8%
An honor with one other copy seen25,6512.8% to 3.5%
An honor with two other copies seen13,9110.6% to 0.9%
A suji with no other copy seen34,9556.0%
A suji with two other copies seen12,0562.7%
A 2 or 8 suji made by the declaration tile5,6579.9%
Any other 2 or 8 suji21,1334.3%
A 3 or 7 suji made by the declaration tile6,3678.2%
Any other 3 or 7 suji31,7205.0%
A 1 or 9 suji made by the declaration tile2,3091.3%
Any other 1 or 9 suji12,9332.7%
A live terminal one or two ranks from a tile the riichi player threw in its first six discards13,5043.6%
Any other live terminal68,4539.7%
A live 2-3-7-8 one or two ranks from such an early discard103,7208.8%
Any other live 2-3-7-8192,86511.1%

The checklist. Fold with the kind you would anyway, then choose inside it: the honor with more copies seen, the suji with more copies seen, not the 2 or 8 or 3 or 7 suji of the declaration tile when another is there, and a terminal next to the riichi player’s early discards. Fitted on these cues alone and checked on games it had not learned from, picking that way would have taken the humans’ real folds from 4.09% to 3.13%, a third of the way to LuckyJ and as far as LightGBM got. The kind alone gets them to 3.77%. In LuckyJ’s own folds the checklist picks tiles that hit 1.98%, more than LuckyJ’s 1.28%: the rest of its edge has no rule here.

An example from LuckyJ’s games. South 3 of a game on 17 May 2023, its 8th turn, folding against a riichi from across declared with [[5m]] on its 5th discard. LuckyJ held four suji: [[2m]] and [[8m]], both made by the declaration tile, and [[8s]] and [[9s]], made suji by the [[5s]] and [[6s]] the riichi player discarded after declaring. It threw a [[9s]] from its pair: a 9 suji, with the other copy in its own hand. The riichi waited on [[8m]]-[[8p]].

Your folds. Your 111 real folds hit 2.70%, between the humans and LuckyJ, and three hits are too few to place you closer. Two of the three were the tile the models rated safest in your hand: a [[7s]] suji on 20 September, which the riichi waited on alone, and an [[8m]] beside the riichi player’s early [[7m]] on 4 October, into a [[5m]]-[[8m]] wait. The third is worth a look: 10 October, South 3, your 9th turn, against a riichi from your right on its 5th discard. You held five pairs and a [[P]], the dora. You broke the [[6p]] pair, a half-suji LightGBM rated 12.5%, and kept the [[P]], with no other copy seen, rated 5.3%; the checklist also throws the [[P]]. Throwing it would have kept your seven pairs at 1-shanten as well. The riichi waited on [[6p]] alone.

Rule When you fold with no genbutsu, pick the kind of tile as you do now, then compare the tiles of that kind. Count the copies of each honor and suji, pass over the 2 or 8 and 3 or 7 suji of the declaration tile, and prefer a terminal beside the riichi player’s early discards. LuckyJ folds with the same kinds as everyone; it picks better tiles inside them.
Exception A 1 or 9 suji made by the declaration tile is the safest suji of all, 1.3%: the declaration tile’s danger lives in its 2 or 8 and 3 or 7 suji.

Evidence Every tile held by any player who discarded into a lone riichi in your 199 games since January and LuckyJ’s 1,255 Tenhou games, 858,565 tiles in 108,201 discards, checked against the riichi’s real winning tiles, from mine_fold_model.py. A real fold is a discard with no genbutsu in hand that leaves the hand further from tenpai than its best discard. Each model learned from four fifths of the games and scored the fifth; on discards with no genbutsu in hand LightGBM ranks a winning tile above a safe one 65% of the time, TabPFN 3.5 65% and the checklist 61%.

42

Chapter forty-two · Push or fold · added 16 October

New

Folding more cannot stop the fourths; declaring your tenpais is the part to fix

You told me you are folding more and still losing. Since the peak on 1 October you have played 54 games: 14 firsts, 11 seconds, 13 thirds and 16 fourths, and given back 1,680 rank points, 31.1 a game. The 4 games since chapter 38 went 2nd, 2nd, 4th, 4th.

Your 2026 Jade gamesBefore the climbThe climbSince the peak
Games876254
1st, 2nd, 3rd, 4th20, 25, 22, 2020, 18, 20, 414, 11, 13, 16
Rank points a game−13.9+44.8−31.1
Per handBeforeClimbSince the peakLuckyJ
Hands won21.1%25.0%21.8%22.9%
Points a win6,0066,8156,4627,317
Hands dealt in11.6%9.9%12.7%10.3%
Deal-ins of 7,000 or more, per 100 hands2.72.74.93.4
Points a hand−46+454+51+363
Live tile at 1-shanten, safe tile in hand6.9%6.2%6.3%7.8%
Tenpais kept against a riichi76.9%82.4%75.9%82.3%
First closed tenpais declared74.1%81.1%62.3%
LuckyJ’s rate for those tenpais78.4%79.5%74.4%

The fourths are a swing. Over your 4,458 games on record you finish fourth 21.4% of the time. On the climb that rate gave 13.3 fourths in 62 games and you had 4, 2.9 standard deviations under it; since the peak it gives 11.6 in 54 and you have had 16, 1.5 over. Across all 203 games of 2026 you are at −0.6 rank points a game: even, at Saint 3. Your hands since the peak earned +51 points each, more than the −46 before the climb, yet more of your games ended fourth. Most of the slide is the climb’s luck coming back, and folding does not decide whose mangan lands while you are in third.

Your folds are fine, and folding more costs little. At 1-shanten, when a safe tile kept your hand, you threw a live tile 6.3% of the time since the peak and 6.2% on the climb. In the 4 games since chapter 38 you made 25 real folds against a lone riichi, 22 of them with genbutsu; one dealt in, a live [[1m]] that Mortal threw too. You did fold more with tenpai: you kept 75.9% of your tenpais against a riichi when every tenpai tile was unsafe, against 82.4% on the climb and LuckyJ’s 82.3%. Chapter 36 found a push in those spots worth about 7 rank points more than a fold, and the difference is about 5 hands in 54 games.

Where the 15 October points went. In the first game four hands went to other players’ mangan or more: two rons between them, and two tsumos you paid into. You reached all-last as dealer in third, 4,500 points above the fourth player, called a [[5m]] pon for tenpai as Mortal did, and on the 18th turn threw a [[2p]] into that player’s quiet 7,700; Mortal threw the same tile. In the second game you dealt into four riichis, 5,200, 8,300, 8,300 and 7,700, the last three on the ippatsu discard. Mortal threw your tile each time. The 5,200 came from a tenpai you kept dama on the 5th turn, where Mortal declared at 99%; you then pushed through 8 discards into the riichi without the riichi’s value. Mortal declared at 93% to 99% in four of that game’s hands you kept dama.

The part that is yours: the dama. Since the peak you declared 62.3% of your first closed tenpais, where LuckyJ declares 74.4% of the same hands, 3.6 standard deviations short. Where Mortal first rated riichi at 90% or more you declared 82.4% of the time, against 94.2% on the climb. Of the tenpais LuckyJ declares at least 70% of the time, the 61 you declared earned +3,018 points each and the 15 you kept dama +640. These are small samples, and they point the same way as chapter 33.

Rule Do not fold more to stop a run of fourths. Keep folding the way you do, keep about 4 tenpais in 5 against a riichi as LuckyJ does, and declare your first closed tenpai where LuckyJ would. A run of fourths comes and goes on its own: 4 in 62 games on the climb, 16 in 54 since.
Exception The big hands chapter 24 keeps dama stay dama. The rate to match is LuckyJ’s for that hand and wait, which is why the table compares you with it hand by hand.

Evidence Your 203 Jade games since 1 January, split at chapter 33’s climb (23 September 18:40 to 1 October 22:20), with rank points and places from your ranked-game records and your career rate of fourths from your player stats, from mine_slide_check.py. Hands, defense and riichi rows as in chapter 33, against LuckyJ’s 1,079 Tokujou games and its fit of LuckyJ’s first closed tenpais; Mortal’s ratings from the site’s replays, the first rating of riichi at 90% or more in each hand. A real fold is a discard against a lone riichi that leaves the hand further from tenpai than its best discard.

43

Chapter forty-three · Recent games · added 16 October

New

Wait on the honor, not the dora, and take the free safe tile

You asked about two decisions from your 15 October games. The first came up twice in your 100-point game (you finished it on 100 points), in East 4 and again in South 1: seven pairs with a dora and another tile as the two singles. The second is the [[2p]] you threw in the all-last of the first game. Both turn on what the other players will do with a tile.

A dora wait is the one nobody throws. Seven pairs waits on one tile. In your 203 Jade games and LuckyJ’s 1,255 Tokujou games, at every seat, 509 riichi hands waited on a single tile this way, and the kind of tile decided how they went.

Seven pairs on one tile, 1 to 3 unseenHandsWonPoints a hand
Riichi on a dora number4418%+1,618
Dama on a dora number7125%+1,608
Riichi on an honor18056%+5,579
Riichi on another number or a terminal26848%+4,476

The dora pays more for a win, 15,838 points against 11,117 for an honor, but it wins a third as often. The reason is the other players: against a riichi they threw a dora wait 0.7 times per 100 turns (2 times in 277 turns) and any other wait 6 to 10 times (7.0 for a dragon or the round wind, 10.3 for an opponent’s seat wind, 6.4 for a plain number). A dora wait wins almost only by tsumo, 6 of its 8 wins.

LuckyJ gives up the dora. Holding a dora number and a non-dora honor as its two singles, LuckyJ kept the honor in 10 of 18 hands, and in both of the hands where it declared riichi. The humans at its tables kept the dora in 40 of 49. You kept it in 4 of 5, in this game twice.

East 4, your 5th turn: the dora [[6p]] or the [[F]]. You held six pairs with the [[6p]], the dora, and the [[F]], one copy of which had been thrown. A wait on the [[F]] had 2 tiles unseen and one on the [[6p]] 3. You threw the [[F]] and stayed dama on the [[6p]] for 4 han. Mortal declares riichi at 99.5%, then throws the [[6p]] 77% of the time and waits on the [[F]]. The right declared riichi on the next discard, you pushed 8 tiles, and the [[2m]] on turn 13 dealt in 5,200. The [[F]] came back twice: across threw the third copy on its turn 6, a ron for a riichi on the [[F]] (3 han, 3,200), and you drew the fourth on turn 8. That is the record only; the other players would have reacted to your riichi.

South 1, as dealer: the dora [[2s]] or the [[E]]. On turn 5 your singles were the [[2s]], the dora, and the [[3s]], with 3 tiles unseen each. Mortal declares riichi at 94.5% and keeps the [[2s]]. On turn 6 you drew the [[E]], thrown once, with 2 unseen: riichi at 99.0%, then the [[2s]] goes and the [[E]] stays, 96.8% of the time. On turn 7 the [[P]], not thrown at all, with 3 unseen: riichi at 76.3%, then the [[2s]] goes, 71.5%. You stayed dama on the [[2s]] for 4 han and 9,600 as dealer, and the left threw it after your turn 7. No [[E]] showed up after turn 6, so a riichi on it would have won nothing there.

What the two hands show. A riichi on the [[E]] is worth half a ron on the [[2s]], 3 han and 4,800 against 4 han and 9,600 as dealer, and wins about twice as often. Over many hands that is +5,579 a hand against +1,608. One hand each way proves nothing; the table above does. In East 4 nobody threw the [[6p]] in the 8 turns that followed, and in South 1 the left threw the [[2s]] on turn 8.

All-last, third, dealer: a tenpai that could not win. On the 18th turn you drew the [[5pr]] with 6 tiles left in the wall. You were dealer in third at 20,600, the fourth player at 16,100, and you held [[1m]][[2m]][[3m]], [[2p]][[3p]], [[5p]][[5p]][[5pr]], [[4s]][[5s]][[6s]] and the [[5m]] pon. Every wait of every discard came to 0 han in a ron, so the hand could only be tenpai when the wall ran out, which keeps your deal and earns the tenpai payment.

DiscardWaitsLive tilesThrown earlier by
[[2p]] (what you threw)[[3p]] [[4p]]3across, left
[[3p]][[2p]]1right, across
[[5p]][[1p]] [[4p]]6nobody

The right was the only hand with no calls (across had 1, left 2). Its last tiles were drawn and thrown, the [[7s]], [[2s]], [[6p]] and [[6p]], then the [[1m]] from its hand, and it waited on [[2p]]-[[5p]]. It had thrown the [[3p]] itself, so the [[3p]] was safe against it, and your [[2p]] was a 7,700 (pinfu, dora, red five). The discard cost you the third place: you finished fourth at 12,900, −267 rank points. Mortal rates [[2p]] at 57%, [[3p]] at 36% and [[5p]] at 7% and cannot see that hand; the rivers could.

Rule With seven pairs and two singles, do not wait on the dora. Wait on an honor if you have one, otherwise any other tile, and declare riichi. When your tenpai has no yaku, its wait is worth nothing, so choose the discard by safety and start with the closed hand’s river.
Exception If someone is already in riichi, your own riichi means you cannot fold afterwards, so weigh it with chapters 36 and 38. A tenpai with a yaku on some wait is a hand to win, and chapters 35 to 38 apply to it.

Evidence Every seven-pairs hand of your 203 Jade games since 1 January and LuckyJ’s 1,255 Tokujou games, at all four seats, that was six pairs and one single waiting on that tile only: 4,364 turns in tenpai, 509 of them riichi hands, with furiten waits left out, from mine_chiitoi_tanki.py. A wait is released when another player discards it before its owner’s next turn. Mortal’s numbers come from the site’s replays, and its riichi lines from feeding it the record and then the declaration. The ron values of the all-last hand are from the mahjong library.

44

Chapter forty-four · Riichi or dama · added 19 October

New

Your thin riichis are not on the wrong tiles

You told me your riichis with 5 or fewer live tiles are not succeeding, that your tenpais seem to be on the wrong tiles, and that you throw the wrong tile when you are tenpai. I checked each against LuckyJ and the humans at both tables: 11,444 riichis, at every seat of your 216 games since January and of LuckyJ’s 1,255. The results you see are real. The cause you suspect is not in the data.

What is real. A riichi on 5 or fewer live tiles (nobody else in riichi, the wait not furiten) is a weak hand after the 8th turn for everyone, and yours are weaker than the others’.

5 or fewer live, from the 8th turnRiichisWonBy ronBy tsumoDealt inPoints a riichi
You10330.1%16.5%13.6%14.6%+1,081
Jade humans at your tables34941.3%24.4%16.9%16.9%+1,824
LuckyJ68645.3%25.9%19.4%13.1%+2,602
Tokujou’s humans1,99338.5%21.9%16.6%15.7%+1,586

Over every turn your 164 thin riichis won 41.5%, against 44.9% for the Jade humans and 50.8% for LuckyJ. With 6 or more live tiles you are level with everyone: 62.3% of 175 and +3,796 a riichi, against LuckyJ’s 59.5% and +3,860.

What it is not. Five checks of the tiles, each against the humans and LuckyJ:

WhoFirst tenpai, live tilesCut gave up live tilesDama wait changed to fewerWait copies in other handsFuriten riichis
You5.147.1%2.9%49.6%4 of 417
Jade humans5.027.6%2.6%48.2%18 of 1,287
LuckyJ5.0510.6%2.7%46.9%19 of 2,489
Tokujou humans4.918.9%2.8%48.2%85 of 7,251

The third column is the share of 198, 672, 1,252 and 3,802 declarations, in that order, where another discard also left tenpai with more live tiles; the fourth the share of 411, 1,178, 1,508 and 6,703 decisions made while already tenpai in dama; the fifth is for thin riichis from the 8th turn.

At the declaration you give up live tiles less often than LuckyJ does, and the waits you take at the first tenpai have as many live tiles as anyone’s. Your wait tiles do look a little less safe to the table, 14.7% of the live tiles against 18.3% to 18.5% for the others, but adding that to the model moves its expectation for your late thin riichis from 40.4% to 40.0%. The tiles thrown next to a wait add nothing (0.01 per tile, with an error of 0.03).

What is left. A model fitted on the humans at both tables, from the live tiles, the turn, the dealer, the calls and how many opponents are already tenpai, expects 41.6 wins from your 103 late thin riichis. You won 31, 2.2 standard deviations short. For LuckyJ’s 686 it expects 274.2 and LuckyJ won 311, 2.9 standard deviations over, which fits chapters 40 and 41, where LuckyJ also beat every tile feature. Your shortfall is in rons: the other players threw your wait tiles 17 times, where the humans’ rate gives 24.1. With 103 riichis that is at the edge of what chance produces, and I found no tile that explains it.

The table decides a thin riichi. Among the humans’ thin riichis, 51.7% of 2,300 won when no opponent was tenpai, 36.2% of 978 with one and 20.5% of 278 with two or more. Your declarations met a table with no opponent tenpai 62.8% of the time, LuckyJ’s 68.8%.

What it costs. If the whole shortfall is real, it is 540 points a riichi against the humans, 258 points a game, about 2.1 rank points a game at 8.3 rank points per 1,000 points (a straight-line fit over your 308 records). Against LuckyJ it would be 725 points a game, 6.0 rank points. The error on your own +1,081 is 595 points a riichi, so zero is inside it. Chapter 42 found the slide since the peak to be mostly fourths luck, and this is not where it went.

Two of yours. In East 1, 1 honba, your 13th turn you declared on [[1p]] and [[4p]] with 5 live tiles. 4 of the 5 were in the other players’ hands, 1 in the wall, and the hand cost you 13,300. In South 4, your 15th turn the [[7m]] had 3 live tiles and all 3 were in other hands, so nothing could come; the hand drew and you were tenpai for +2,000.

Rule Keep taking the cut with the most live tiles, as you do, and do not change which tile you throw: I found no tile you throw that LuckyJ or the humans do not. Plan a thin riichi from the 8th turn as a hand that wins about 4 times in 10.
Exception A furiten riichi wins only by tsumo. Yours won 0 of 4 and LuckyJ’s 6 of 19, so declare one only when no other cut keeps tenpai.

Evidence Every riichi at every seat of your 216 Jade games since 1 January and LuckyJ’s 1,255 Tokujou games, 11,444 in all, from mine_thin_riichi.py: the waits after the declaration, their unseen copies, the other players’ hands in the logs, and every later discard of a wait tile. Live tiles are the copies unseen from the declarer’s seat. The first-tenpai and dama rows come from contrast/tenpai.py, as in chapter 33.

45

Chapter forty-five · Recent games · added 19 October

New

Why the rank points fall: the game drains them, and your wins are the part to build

You told me you are losing rank points quickly and that something is wrong. I looked through your 216 games of 2026 for a part of your play that broke after the peak and found none. I found four other things. The game takes 23.6 rank points a game from an average Jade player. Your defense earns most of it back, which leaves you near even. The climb was the exception. And one number did move since the peak, your deal-ins of 7,000 or more, back to the rate of the humans. The size of your wins is where you and LuckyJ are farthest apart.

What the game takes. A game pays 1 rank point for each 1,000 points of final score above 30,000 (and takes them below), rounded up, plus 130, 65, 0 or −250 for first to fourth. Scored that way, the Jade humans at your tables lose 23.6 a game and Tokujou’s humans lose 29.8. Over all of 2026 you lost 2.6.

Games1st, 2nd, 3rd, 4thMean scoreRank points a game
Before the climb8723, 29, 25, 23%24,528−13.9
The climb6232, 29, 32, 6%29,694+44.8
Since the peak6725, 21, 24, 30%25,864−31.7
All of 202621626, 26, 27, 20%26,425−2.6
Jade humans at your tables64825, 25, 24, 27%24,525−23.6
LuckyJ1,07932, 28, 24, 17%28,787+16.2
Tokujou’s humans3,23723, 24, 25, 28%23,726−29.8

One game spreads 147.6 rank points around its average, so the average of a 67-game stretch moves 18.0 either way. The climb, +44.8, was 2.5 of those steps above your 2026 average. Since the peak, −31.7 is 1.6 below it and 0.4 below the average Jade player’s −23.6. You finished fourth in 20 of the 67 games; your career rate of 21.4% over 4,458 games gives 14.4, so 20 is 1.7 standard deviations over, and the climb’s 4 in 62 was 2.9 under. Your mean score since the peak, 25,864, is 1,339 points above the Jade humans’ 24,525, and your rank points are 8 below theirs.

What did not change. Your discards match Mortal’s first choice 73.2% of the time before the climb, 73.4% on it and 73.0% since. Tilt is not in the games. The game after a fourth in the same session was a fourth 5 times in 18 (27.8%), against 26 in 139 after anything else (18.7%), and by place in the session the fourths run 22.0, 20.0, 17.1, 25.9 and 18.0% from the first game to the fifth or later. Your ends of games are sound. From third before the last hand you fell to fourth 8 times in 55 (14.5%), where the Jade humans do 19 in 161 (11.8%) and LuckyJ 19 in 255 (7.5%). From fourth within 3,000 points of third you got out 10 times in 12, against 9 in 34 for the Jade humans and 18 in 40 for LuckyJ. These counts are small.

What moved. Your deal-ins of 7,000 or more, by stretch:

Per 100 handsBefore the climbThe climbSince the peak
Deal-ins of 7,000 or more2.72.74.9
Tiles or riichis Mortal rated under 25%0.30.61.1
The other deal-ins2.42.03.7
Hands of 7,000 or more won by the other three9.08.810.1
Riichis declared by the other three19.918.418.7

The Jade humans deal in 4.7 times per 100 hands, Tokujou’s humans 4.9 and LuckyJ 3.4. You were below the humans for 149 games and are level with them for the last 67. The shift is 2.6 standard deviations (34 hits, where the early rate gives 19.0), and it is the one I picked out of about ten numbers, so read it as likely rather than proved. The 15 extra hits cost 2,086 points a game, 17.3 rank points: without them the stretch would be −14.4 and not −31.7.

Mortal rated the tile you threw, or the riichi you declared, under 25% in 8 of the 34 hits: 92,800 points, 1,385 a game, 11.5 rank points. Before the peak it was 7 of 42. Those 8 are mistakes, and they make up 4.8 of the 15 extra hits. The other 26 are tiles Mortal throws too, and make up the other 10.2. In 7 hits you were in riichi and dealt into an open hand, 1.0 per 100 hands where the Jade humans have 0.28 and LuckyJ 0.38; in 5 of the 7 Mortal rated your riichi at 99% or more, and in 3 the winner had not called yet when you declared. The tables served a few more big hands as well.

Where the rank points are. Every point of a game comes from a win, a deal-in, a tsumo paid, or a draw and the riichi sticks. Your edge over the Jade humans at your tables, and LuckyJ’s over Tokujou’s humans, in points a game:

Points a gameYouLuckyJGapGap in rank points
Wins+103+2,4482,34519.5
Deal-ins+1,798+2,4406425.3
Tsumo paid−163+3114743.9
Draws and sticks+165−138−303−2.5
Total+1,903+5,0613,15826.2

You are 1,903 points a game above the Jade humans, with an error of 1,186. On the 8.3-per-1,000 line that is 15.8 rank points, and you earned 21.0 (−2.6 against −23.6). LuckyJ is 5,061 above its humans, with an error of 531, and 3,158 above you. Your defense is close to LuckyJ’s: +1,798 on deal-ins against +2,440, and yours is 3.3 times its error of 541. Your wins are not: +103 against +2,448, with errors of 902 and 419. The wins are 2,345 of the 3,158 points and 19.5 of the 26.2 rank points.

Your winsYouJade humansLuckyJ
Hands won per 10022.621.122.9
Points a win6,3956,7897,317
Riichi wins8,6808,9949,227
Dama wins5,9345,5757,676
Open wins4,6294,9625,477
Open wins of 4 han or more18%23%27%
Points a hand you open+338+255+578

You win as often as LuckyJ and each win is smaller. 6,395 a win against 7,317 is 922 less, with an error of 207. LuckyJ’s wins are bigger in all three kinds: clearly in open hands (5,477 against 4,629, error 242) and in dama (7,676 against 5,934, error 567), and by 547 in riichi (error 323). You open as many hands as LuckyJ, 35.4% against 36.0%, and win as many of them, but 18% of your open wins have 4 han or more where LuckyJ’s have 27%. At 2.35 wins a game, 100 points on the average win is worth about 2 rank points a game, and the whole 922 would be 18.0. These are all 216 games, and 80 of them were played before the guide’s first chapters went up on 22 September. Since then your wins are 1,227 points a game above the humans’ at your tables and LuckyJ’s are 2,448 above theirs: a gap of 1,221 with an error of 1,282, in chapter 47.

The six cuts. The 8 hits Mortal rated under 25% include 2 riichis you declared on 7 October at 4% and 6% (one is chapter 44’s East 1, 1 honba). The other 6 are tiles you threw, 68,200 points:

HandYour turnYou threwMortal’s ratingPoints
3 Oct, South 1-18[[9m]]5%12,300
16 Oct, South 3-115[[6m]]0%12,300
6 Oct, South 412[[P]]12%12,000
11 Oct, East 311[[8p]]1%12,000
7 Oct, South 215[[7s]]0%11,600
17 Oct, South 2-111[[6s]]0%8,000

In South 3, 1 honba, the player on your left was in riichi with pinfu and sanshoku, and Mortal throws a [[1p]] at 99.6%; your [[6m]] was the winning tile. In East 3 the player on your left had two dragon pons, [[P]] and [[C]], and had thrown no pinzu, and you threw an [[8p]] where Mortal throws a [[7m]] at 98.9%: a honitsu haneman. The 7 hits where you were in riichi were different. In East 3 of the 17 October game, your 12th turn you declared into another player’s riichi, which Mortal does at 99.7%, and four turns later threw the [[4s]] to a hand that had not called when you declared and then took a pon of [[8p]]: tanyao, 3 dora and a red five, 8,000. Nothing there is yours to change.

Rule Do not tighten your play because of this stretch. Your defense is worth 1,798 points a game over the humans, 14.9 rank points on the line, and it is already near LuckyJ’s. Since 22 September your wins are close to LuckyJ’s (chapter 47), so keep building as you do and take the six cuts first.
Exception Against a riichi or a caller with a flush, take the safe tile before the hand: the 6 cuts above are mistakes, 68,200 points, and are not luck. Look at them first.

Evidence Your 216 Jade games since 1 January, split at chapter 33’s climb (23 September to 1 October), the 648 other seats at your tables, and LuckyJ’s 1,079 Tokujou games with the 3,237 seats of its humans, from mine_rank_drain.py. Rank points are the Saint 3 Jade formula applied to every seat; it reproduces all 216 of your records. Points by source come from the hand logs, and an error is the standard error of a mean game. Mortal’s ratings come from the site’s replays, and a hit is a deal-in of 7,000 or more, as in chapter 42. The line of 8.3 rank points per 1,000 points is chapter 44’s.

46

Chapter forty-six · Safe tiles · added 19 October

New

Count the copies you can see: why Mortal threw the 1p and the 2s

You said you sometimes have trouble defending, and that you do not understand the [[1p]] and the [[2s]] in the first two of chapter 45’s six hands. The [[1p]] has a plain reason: every copy of it was in view, so only one wait could still hit it. The [[2s]] I can explain only in part. Counting copies finds the 1p and misses the 2s, so for the 2s I changed the board and asked Mortal again, and its answers show which discards it was reading.

The 1p. On 16 October, in South 3, 1 honba, your 15th turn the player on your left was in riichi. You held [[6m]][[6m]] [[1p]][[1p]][[1p]][[5p]][[5p]][[6p]][[6p]][[6p]] [[3s]][[3s]][[4s]][[6s]], 1-shanten, and threw the [[6m]]. It was the winning tile of a [[3m]]-[[6m]] wait with pinfu and sanshoku, 12,300. Mortal throws the [[1p]] at 99.6%. With no riichi on the table it throws the [[6s]] at 92.5% and gives the [[1p]] 1.1%, so the riichi is what makes the 1p right. Every discard left you 1-shanten, so the 1p cost you no speed, even out of a triplet.

Your tileThe guide’s labelCopies you can seeWays to be the winning tileMortal
[[1p]]live terminal4399.6%
[[5p]]half suji2110.3%
[[3s]]live 2-3-7-8213under 1%
[[4s]]half suji213under 1%
[[6m]], the one you threwlive 4-5-6215under 1%
[[6p]]half suji317under 1%
[[6s]]half suji118under 1%

You hold three [[1p]], and the player on your right threw the fourth on the first turn. Nobody can hold a [[1p]], so nobody can be waiting on a pair or a single of it. One wait is left, [[2p]]-[[3p]], which waits on the [[1p]] and the [[4p]]. Three of the four [[2p]] are out (the riichi player’s, the right player’s and yours) and three [[3p]] remain, so 1 × 3 is 3 ways. A ways count is the number of combinations of tiles nobody has shown that would make the tile a winning one. The [[6m]] has 15: 8 as a two-sided wait, 4 closed, 1 pair and 2 single. The guide’s labels read the riichi player’s river and call the [[1p]] and the [[6m]] both live, so they cannot see the difference. After the label, count the copies of the tile you can see.

The same shape came up in the 11,600 hand, 7 October, South 2, your 15th turn. The dealer had called and was throwing drawn tiles, and you threw the [[7s]]. You held three [[7p]] and the dealer had thrown the fourth, and you held a pair of [[N]] with two more in the rivers. Mortal gives the [[7p]] 81.8% and the [[N]] 11.9%, and your [[7s]] under 1%. Both of its choices have all four copies in view, and the [[7p]] left you 1-shanten as the [[7s]] did.

How often it comes up. At your tables and LuckyJ’s, a discard into a lone riichi by a player who was not tenpai afterwards held a tile with all four copies in view and no genbutsu in 2% of cases: 67 of your 3,295, 285 of LuckyJ’s 16,224. You threw that tile 18 times in 67 (27%), the Jade humans 71 in 211 (34%), Tokujou’s humans 342 in 996 (34%) and LuckyJ 126 in 285 (44%). It is about one fold in fifty, so it is not where your rank points went. When it does come, it is the first tile to throw.

The 2s. On 3 October, in South 1, 1 honba, your 8th turn two players were in riichi: the one on your right, declared on a [[3m]], and the dealer across, declared on a [[1s]] on the next turn, so you were in the ippatsu round. You held [[4m]][[5m]][[7m]][[7m]][[9m]] [[3p]][[4p]] [[2s]][[3s]][[6s]][[6s]] and a pon of [[P]]: 2-shanten, with the two [[7m]] as dora, 3 han at best. You threw the [[9m]], the right player’s winning tile (pinfu, riichi, ippatsu, dora and a red five), 12,300. Mortal throws the [[2s]] at 92.3%, the [[9m]] at 5.1% and the [[3s]] at 2.2%. Every discard left you 2-shanten, so a fold cost no speed, and a hand two tiles from tenpai against two riichis is a fold.

Which tile is the question. I edited the board and asked Mortal again at the same decision:

The boardMortal throws
Both riichis, as played[[2s]] 92.3%, [[9m]] 5.1%, [[3s]] 2.2%
Only the dealer in riichi[[9m]] 99.9%
Only the right player in riichi[[3s]] 98.4%
No riichi[[9m]] 97.4%
The right player had never thrown the [[3s]][[9m]] 98.1%, [[2s]] 0.4%
The dealer declared on a [[5s]], not the [[1s]][[2s]] 99.5%

Each riichi has its own answer. The [[9m]] is the tile against the dealer alone: the dealer had thrown the [[9p]], the [[8m]], the [[1m]] and the [[8p]], the ends of the suits. The [[3s]] is the tile against the right player alone, because he had thrown a [[3s]] on his third turn. Against both, each is a tile one player has cleared and the other has not, and Mortal takes the [[2s]]. Two edits show what it is reading. Take away the right player’s early [[3s]] and the answer becomes the [[9m]] at 98.1%; make the dealer’s riichi tile a [[5s]], the suji partner of the [[2s]], and the [[2s]] rises to 99.5%.

Counting does not find this one. Against both riichis the [[9m]] has the fewest ways, 24, then the [[3s]] with 25 and the [[2s]] with 34, and the [[9m]] was the winning tile. Mortal sees more in the rivers than my counts do, and I can tell you what moves its answer when I edit the board, not the whole reason.

Rule Before you throw a tile into a riichi, count the copies of it you can see. A tile with all four in view, three in your hand and one thrown or your pair and two thrown, is the first tile to throw when you are folding, even out of a set or a pair. With two riichis, a tile that is safe against one of them is not safe, as the [[9m]] was against the dealer alone: look for the tile that neither river has made live.
Exception A dead tile that breaks a tenpai is not for throwing. The rule is for hands that will not be tenpai after the discard, as the 2-shanten hands here were.

Evidence Three hands from chapter 45 and, for the count of dead tiles, every discard into a lone riichi by a player who was not tenpai afterwards in your 216 Jade games and LuckyJ’s 1,079 Tokujou games, from mine_defense_spots.py. Ways are counted from the copies the discarder cannot see, with the hand, the rivers, the melds and the dora indicator as seen, and a riichi player’s own river and the tiles passed since the declaration as safe. Mortal’s answers come from the site’s replays and, for the edited boards, from the same Mortal engine fed an edited record of the hand.

47

Chapter forty-seven · Dora hands · added 19 October

New

You were right: since 22 September your hands are built like LuckyJ’s

You said you are already building the most valuable hand you can. For the 136 games since 22 September, the day chapters 1 to 9 went up, the numbers agree. I split your 216 games of 2026 at the guide’s first commit: 80 games before it and 136 after.

Your winsBefore 22 SepSince 22 SepJade humansLuckyJ
Points a game over the humans at the table−1,809+1,2270+2,448
Hands won per 10021.323.321.122.9
Points a win5,9926,6126,7897,317
Riichi wins8,0219,0348,9949,227
Open wins4,3424,7814,9625,477
Dama wins5,8805,9645,5757,676
Open wins of 4 han or more13%21%23%27%
Points a hand you open−18+560+255+578

Before the guide your wins were 1,809 points a game below the humans at your tables, with an error of 1,283. Since, they are 1,227 above, with an error of 1,212, where LuckyJ’s are 2,448 above its humans. The gap to LuckyJ on wins fell from 4,257 to 1,221, which is 35.3 rank points a game and then 10.1, and the second figure has an error of 1,282, so it cannot be told from zero. Chapter 45 said the size of your wins is where you and LuckyJ are farthest apart. That averaged all 216 games, and the 80 before the guide carry two thirds of it. I have marked the sentences in that chapter.

Hands dealt a dora. Chapter 1 found that LuckyJ pulls away on the half of hands that hold a dora or a red five in the deal, and chapters 6 and 7 found where yours lost their points.

Hands dealt a dora or a red fiveYou, before 22 SepYou, since 22 SepLuckyJ
Hands4117255,692
Reached tenpai51.6%51.6%51.7%
Tenpais won36.8%45.7%47.2%
Points a win6,5057,5978,366
Points won a hand1,2351,7922,040
First closed tenpai declared66.7%75.4%74.2%
Wins with a shape yaku39.7%53.2%54.9%
Wins with a flush1.3%10.5%7.1%
Wins that kept every dora dealt79%84%85%
Live tiles at the first tenpai4.914.924.98
For comparison, points won a hand with no dora1,3201,2741,307

Before the guide your dora hands reached tenpai as often as LuckyJ’s and won 36.8% of those tenpais against its 47.2%, and 39.7% of your wins carried a shape yaku against 54.9%. Since 22 September every row has moved to LuckyJ’s: the shape yaku, the flush (1.3% of your wins, now 10.5% against LuckyJ’s 7.1%), the dora kept to the win, and the first tenpai declared 75.4% of the time where LuckyJ declares 74.2%. The waits you take match as well, 4.92 live tiles at the first tenpai against 4.98, so you are not trading wait quality for value (chapters 43 and 44). The hands with no dora, where you were level before, are still level.

What is left. Points a win in dora hands, 7,597 against 8,366 with an error of 341 on yours, is the one row still short. Hands dealt two or more dora earn the same size of win (9,587 against 9,524) and win less often, 23.8% against 28.1%. Over a whole hand, a dora hand earns 248 points less than LuckyJ’s, with an error of 155: 1,322 points a game, 11.0 rank points, with an error of 824 points. I could not find a decision in how you build that explains it, because the shape yaku, the flush, the dora kept, the declarations and the waits all match. The six cuts of chapter 45 cost 68,200 points over 67 games, 1,018 a game and 8.4 rank points, about the size of that whole difference, and those are mistakes you can name (chapter 46).

Three of yours. Three wins since 22 September from hands dealt a dora. 11 October, South 1, 2 honba, your 17th turn: a flush with a [[F]] pon and the dora, haneman, 18,000. 3 October, East 1, 1 honba, your 6th turn: pinfu, riichi, ippatsu, dora and a red five, haneman, 18,000. 29 September, East 1, your 19th turn: toitoi with three concealed triplets, tanyao and the dora, baiman, 4,000/8,000.

Rule Keep building as you do: count the dora and red fives on the first turn, build the shape yaku and the flush when the dora allow, declare the first tenpai of a dora hand, and keep the dora to the win. Do not add a new habit for the win gap, because 60 more games will say whether the 1,221 is real. Spend the effort on the six cuts first.
Exception Dama is for the hand that is already big. Your dama wins since 22 September average 5,964 (36 wins) and LuckyJ’s 7,676 (245): keep a hand dama when it is 4 han or more without the riichi, as chapter 24 says, and declare the rest.

Evidence Your 216 Jade games since 1 January, split at the guide’s first commit (22 September, 14:17), the other seats at your tables, and LuckyJ’s 1,079 Tokujou games with the seats of its humans, from mine_point_gap.py. Edges and gaps are points a game from the hand logs, and an error is the standard error of a mean game or hand. Rank points use chapter 44’s 8.3 per 1,000 points. A hand is dealt a dora when its first 13 tiles hold a dora or a red five, and the yaku are read from each win’s details.

48

Chapter forty-eight · Recent games · added 22 October

New

What changed in your last ten games: the bad deals got won

You told me you started winning again after folding harder with fewer defensive tiles. I compared your 10 games from 19 October to 22 October, 103 hands, with your 217 games before them. They went first seven times, second twice and fourth once: +866 rank points, 86.6 a game, 1.9 standard errors above your average of −3.8. At your usual rate seven firsts or more come 0.5% of the time, so part of this is luck. Three things in the hands agree with each other and look real: the bad deals got won, the hands got faster, and no lone riichi found you. In the next 15 games a lone riichi found you 4 times in 65 hands, where 5.1 is usual, so that third thing was mostly a streak (chapter 50); the first two held. Folding harder is the part I cannot confirm.

Where the points came from. Won hands, by how far the 13 tiles were from tenpai when dealt:

Hands dealtYour usual, 217 gamesLast 10 games
2-shanten or better213 hands, 38.5% won7 hands, 42.9% won
3-shanten845 hands, 26.9% won35 hands, 28.6% won
4-shanten or worse1,201 hands, 16.5% won61 hands, 29.5% won

This run had more bad deals than usual: 61 of its 103 hands, 59%, were dealt 4-shanten or worse, against 53% over the 217 games. You won 18 of those 61 where you usually win 16.5%, 2.6 standard errors higher, and the hands dealt better went as usual. The 15 tsumo wins in 103 hands are the luckiest part: 8.0 would be usual, and 15 or more comes 1.4% of the time.

How the hands got faster.

Your usual, 217 gamesLast 10 games
Shanten of the 13 tiles dealt3.593.74
Shanten removed by your 9th turn2.352.64
The other three at your tables, by their 9th turn2.322.41
Tenpai by your 10th turn31.2%38.8%
Hands you opened36.0%40.8%
Callers out, nobody in riichi: tenpai54.2% of 94268.2% of 44
Callers out, nobody in riichi: won24.7%34.1%
Discards Mortal flagged in turns 1 to 6, a game4.723.1
Lone terminals and honors in hand after the 6th discard1.431.27

Your hands were dealt a little further from tenpai, 3.74 shanten and not 3.59, and still reached tenpai by the 10th turn 38.8% of the time against 31.2%. You removed 2.64 shanten by the 9th turn where you usually remove 2.35, and the other three at your tables did 2.41 against their usual 2.32. In the hands where other players had called and nobody had declared riichi, you reached tenpai in 68.2% of 44 against 54.2% of 942 and won 34.1% against 24.7%. The early game was cleaner too: Mortal flagged 3.1 of your discards a game in turns 1 to 6 against 4.72, and you carried 1.27 lone terminals and honors after the 6th discard against 1.43.

Three of yours. 22 October, East 1, 1 honba, your 8th turn: a hand dealt 5-shanten, tenpai on the 6th turn after a chi of the [[8m]] with the dora [[7m]] and a pon of the [[C]], and a tsumo with chun and 2 dora, 1,000/2,000. South 2, 1 honba of the next game, your 15th turn: 4-shanten, a pon of the round wind [[S]] and 3 dora, 3,900 all. The next hand, your 14th turn: 4-shanten again, a pon of the [[C]], 2 dora and a red five, a ron for 11,600.

The folds.

One opponent alone in riichiYour usual, 217 gamesLast 10 games
Hands94741
Hands where you dealt in74, 7.8%0
Hands where you held one of its winning tiles60924
Threw a safe tile when one was an option82.7%, Mortal 82.8%83.1%, Mortal 84.9%
2-shanten or worse, first discard after the riichi: threw a safe tile261 of 368, 70.9%15 of 18, 83.3%
Folds with a genbutsu in hand86.7%91.9%
Points paid in deal-ins, a game6,0513,980

In the 41 hands where one opponent alone was in riichi you dealt in 0 times. Your usual 7.8% would give 3.2, and that many fewer comes 3.6% of the time. You held one of the riichi’s winning tiles in 24 of them and threw none. In South 3, 1 honba of the same game, from your 7th turn the player across declared riichi, you held its winning tile, the [[3p]], and threw something else for 11 turns until the hand ended without you paying. With a hand 2-shanten or worse you threw a safe tile on the first discard after the riichi 15 times in 18, where you usually do it 71% of the time. Your share of safe tiles overall is the same as before, 83.1% against 82.7%, so you are folding at the usual rate and not more often.

You said fewer defensive tiles. The nearest thing I can count is the lone terminals and honors you carry, 1.27 after the 6th discard against 1.43. It did not leave you short: you had a genbutsu to throw in 124 of 135 fold decisions, 92%, against 87%.

What did not change. Your riichi wins are 9 of 103 hands, 8.7%, against 9.0%. Your average win is 7,006 against 6,395, from 31 wins, which is within the error. Four of the 10 deal-ins were tiles Mortal rated under 25%, small ones from 1,000 to 6,100, and the only big one, 12,300, came in your own riichi that Mortal declares at 99.5%. Deal-ins cost you 3,980 a game against 6,051.

What I can and cannot say. Ten games is 103 hands. I looked at about twenty numbers and kept the ones that agree with each other, and the errors above are not corrected for that. The bad deals, the speed and the clean folds fit together: you removed shanten faster, called more and folded early without paying. The luck is in the tsumo wins and the firsts.

Rule Play a bad deal fast: clear the lone terminals and honors by your 6th discard, call when the call brings tenpai nearer, and do not give up a hand dealt 4-shanten. Against a lone riichi with a hand 2-shanten or worse, throw a safe tile from the first discard. Check it after 30 more games: if tenpai by the 10th turn is back near 31%, hands dealt 4-shanten or worse near 16.5% and your deal-ins in lone-riichi hands near 7.8%, the run was mostly luck.
Exception Do not fold more than you did. Your safe-tile share equals Mortal’s. At 1-shanten or tenpai against a riichi, chapters 5, 11 and 36 still decide the push.

Evidence Your 10 games from 19 October, 23:20 (103 hands, 7 firsts), against your 217 games before them, from mine_recent_run.py. Hands, speed and folds come from the hand logs; Mortal’s flags and safe-tile rates come from the site’s replays. Shanten removed is the hand’s shanten at that turn against its shanten when dealt. A lone riichi is a hand where one opponent had declared and you had not.

49

Chapter forty-nine · Recent games · added 22 October

New

A high net win efficiency and still 22% fourths: one deal-in decides them

You asked why your net win efficiency over the last 100 games is high and you still finish fourth in about 22% of them; by the record I pulled it is 23 of 100. Net win efficiency is the win rate times the average win, minus the deal-in rate times the average deal-in, in points a hand. For your last 100 games it is 939, against 666 for the Jade humans at your tables and 1,131 for LuckyJ, and your career figure in your player stats is 772. Three things explain why it does not keep you out of fourth.

Points a handYou, last 100You, all 2026Jade humansLuckyJ
Win efficiency1,5671,4651,4211,677
Deal-in loss628573755546
Net win efficiency9398926661,131
Tsumo paid to others−501−545−541−578
Draws and riichi sticks−162−172−185−191
Real points a hand+276+174−60+363
Games finished fourth23%20.3%26.5%16.9%

1. The figure leaves points out. Other players’ tsumo cost you 501 a hand and the draws and riichi sticks 162, and net win efficiency counts neither. Your real points a hand are +276, not 939; the humans’ are −60 and LuckyJ’s +363. The order is still right, and the fourth rate follows it: 26.5% for the humans, 23% for you and 16.9% for LuckyJ.

2. The Saint 3 format wants fewer fourths than this. With your results a fourth costs 272 rank points on average and a first pays 144, and the fourth rate that breaks even is 20.3%. At 23% you are at −3.3 rank points a game over these 100 games, with a mean score of 27,851.

3. A fourth is one hand. The average hides lumps. A deal-in costs you 628 points a hand on average, but it arrives as one hand of 7,000 or more. In the last 100 games you dealt into 44 hands of 7,000 or more, in 32 games. Those 32 games ended fourth 16 times, 50%, and the other 68 ended fourth 7 times, 10%.

Your last 100 games, by placeGamesPoints wonPaid in deal-insTsumo paidWith a deal-in of 7,000 or more
1st3228,5783,4913,38812%
2nd2215,0183,3594,63214%
3rd2310,4577,0526,27439%
4th235,74813,0707,04370%

Of the 23 fourths, 2 had no deal-in at all, 16 had one of 7,000 or more, and 14 would not have been fourths without their single biggest deal-in, which I handed back from the winner to you. Three of them: 7 October, East 1, 1 honba, your 13th turn, 12,300, third without it; 11 October, East 3, your 11th turn, 12,000, third without it; 19 October, South 4, your 6th turn, 12,000 into a riichi you had declared, third without it, and nothing to decide after the declaration.

What would move it. Avoid a quarter of those 44 hands and the same 100 games give 19.9 fourths and +7.9 rank points a game; avoid half and they give 17.2 fourths and +18.4. That is the top of what folding can do, because a fold also gives up some wins and I did not charge for those.

Rule Read net win efficiency as wins minus deal-ins, and take your tsumo payments, about 500 a hand, off it to get your real points. To lower the fourth rate, move the number of deal-ins of 7,000 or more, 44 in 100 games, each of them close to a coin flip between fourth and better. Chapters 45, 46 and 48 are the tools: the six cuts, counting the copies you can see, and folding from the first discard against a riichi.
Exception Do not give up wins to avoid every deal-in. Your first-place games won 28,578 points, and the 25% and 50% above are the most that folding could return.

Evidence Your last 100 Jade games (1,032 hands), all 227 of 2026, the other seats at your tables and LuckyJ’s 1,079 Tokujou games, from mine_fourth_rate.py. Efficiency is computed from the hand logs as Mahjong Soul statistics usually define it; your career figure is from your player stats. The break-even fourth rate uses your average rank points by place in 2026. The 14 is each fourth’s biggest deal-in handed back from the winner to you on the recorded final scores. The two rates above replay the 44 hands 400 times, each one avoided with the stated chance.

50

Chapter fifty · Recent games · added 23 October

New

Fifteen more games: the bad-deal wins held, the lone-riichi shutout did not

Since chapter 48 you have played 15 more games: 5 firsts, 4 seconds, 4 thirds and 2 fourths, +473 rank points, 31.5 a game. With the ten before them that is 25 games at +1,339, 53.6 a game. Since the peak on 1 October you are at −1,057 over 93 games, −11.4 a game, where chapter 45 found −2,396 over 68. Chapter 48 named three things as real, and these 15 games are a test of them on hands it had not seen.

Before the run, 217 gamesChapter 48’s 10 gamesThe next 15 gamesAll 25
1st, 2nd, 3rd, 4th57, 57, 58, 457, 2, 0, 15, 4, 4, 212, 6, 4, 3
Rank points a game−3.8+86.6+31.5+53.6
Hands won22.4%30.1%33.1%31.9%
Dealt 4-shanten or worse, won16.5%, 198 of 1,20129.5%, 18 of 6129.3%, 24 of 8229.4%, 42 of 143
Tenpai by the 10th turn31.2%38.8%37.0%37.7%
Hands you opened36.0%40.8%40.3%40.5%
One opponent alone in riichi: hands you dealt in74 of 947, 7.8%0 of 414 of 654 of 106
Deal-ins of 7,000 or more per 100 hands3.41.03.92.7

What held. The bad deals. In the 15 new games you won 24 of 82 hands dealt 4-shanten or worse, 29.3%, where you usually win 16.5%, and over all 25 games it is 42 of 143, 3.8 standard errors above usual. The speed held too: tenpai by the 10th turn 37.0%, and by your 9th turn you had removed 2.48 shanten against the other three’s 2.31, where you usually remove 2.35 against their 2.32. You won 33.1% of your hands, 3.1 standard errors above your usual, and opened 40.3%. Mortal flagged 9.33 of your discards a game, against 11.44 usually.

What did not hold. The lone riichi. In the first ten games you dealt in 0 times in 41 such hands; in these 15 it was 4 in 65, where 5.1 would be usual. Over the 106 hands the 4 against 8.3 expected is a 7.6% chance, so you are dealing in less often than usual but not by more than luck allows. Chapter 48 listed this among the three real things, and I have marked that sentence there. Deal-ins of 7,000 or more are back to 3.9 per 100 hands, from 1.0.

Where it can cost you. You dealt in 17 times in these 15 games, and Mortal rated 7 of them under 25%: 35,100 points, 2,340 a game, against 1,524 a game usually. Four to go through before the next session. In South 2, 1 honba of the 23 October game, your 13th turn you threw the [[5s]] into a riichi for 8,000 where Mortal throws the [[3s]] at 98.4%. In East 4, 1 honba of the next game, your 11th turn you declared riichi, which Mortal gives 6.4%, and dealt in 8,300 with the [[1s]]. In South 4 of the 23 October evening game, your 8th turn you threw the [[7s]] for 8,000 where Mortal throws the [[9m]] at 98.4%, though you were already fourth. In East 1 of the game after, your 13th turn you threw the [[1p]] into a caller for 3,900 where Mortal throws the [[2m]] at 98.8%.

The two fourths. The first, 11,200, won 13,000 and still paid 12,900 in deal-ins and 12,900 in tsumo. The second, 11,900, won once for 3,600 and paid 11,200 in tsumo, 6,000 of it to a dealer’s haneman in East 1, and only 2,000 in deal-ins. The second is the kind of fourth no fold changes.

Rule Keep playing the bad deals fast: it held on 15 new games. Do not count on the lone-riichi shutout, and go through the four deal-ins above first. Check again at 55 games: hands dealt 4-shanten or worse won near 29% would make this yours, near 16.5% would make it a run.
Exception The speed is for the bad deals. With a riichi out and a hand 2-shanten or worse, chapters 46 and 48 still say to throw a safe tile from the first discard.

Evidence Your 217 games before 19 October, the 10 games of chapter 48 and the 15 games from 22 October, 17:02 to 23 October, 18:31 (154 hands), from mine_run_update.py. A lone riichi is a hand where one opponent had declared and you had not. Mortal’s ratings of the tile thrown, or of the riichi, come from the site’s replays.

51

Chapter fifty-one · Recent games · added 10 November

New

Stuck at +5 a game: what a steady rise takes, and the discard that would add to it

You asked for a way to rise consistently and said the rank is stuck. Your 61 games from 23 October to 10 November came to 18 firsts, 17 seconds, 13 thirds and 13 fourths: −72 rank points, 1.2 a game below zero. You stand at 5,620, which is 1,129 under the peak of 6,749. The answer has three parts. Nothing went wrong in these games. Your pace since this guide began is +4.9 rank points a game, and 61 games cannot show a pace that small. And the one piece of the distance to LuckyJ that I can point at is a single kind of discard: a tile that is not safe, thrown from a hand that is not tenpai, when another tile would have left the hand just as close.

The 61 games. The first 30 went 8, 7, 7, 8 for −660, 22.0 a game down. The last 31 went 10, 10, 6, 5 for +588, 19.0 a game up. One game at Saint 3 moves your rank by 148.7 points either way (its standard deviation), so the total of 61 games moves by about 1,161 through the deal alone. The swing from −22.0 to +19.0 between the two halves is that noise, and so is the −72.

Chapter 50 asked to look again at the hands dealt 4-shanten or worse once 55 more games were in. Here is that check, and the hot run was a run.

Before the run, 217 gamesThe 25-game runThese 61 games
Dealt 4-shanten or worse, won16.5%, 198 of 1,20129.4%, 42 of 14316.7%, 60 of 359
Tenpai by the 10th turn31.2%37.7%27.7%
Hands won22.4%31.9%21.1%
Hands dealt in11.4%10.1%9.2%

You won 60 of the 359 bad deals, 16.7%, which is your usual 16.5% and far from the run’s 29.4%. You did not get worse after the run. The deals went back to what they usually do, and your deal-in rate, 9.2% of hands, is the lowest of the three columns.

Where you stand. Every seat below is scored with the Saint 3 formula, so the lines can be compared.

GamesFourthsRank points a game
You, since the guide (22 September)22319.7%+4.9
The other three players at your tables66926.8%−26.1
LuckyJ1,07916.9%+16.2
The humans at LuckyJ’s tables3,23727.7%−29.8

You run 31.0 rank points a game above the players you sit with. LuckyJ runs 46.0 above the players it sits with. Put its 46.0 on top of your table’s −26.1 and LuckyJ would make about +19.9 a game in your seat. That figure is rough, since its opponents are not yours, but it sizes the job: the whole road from you to LuckyJ is about 15 rank points a game.

Why +4.9 feels like standing still. The table takes a pace and asks two things: how many games 1,000 rank points take, and how often a stretch of games ends above where it began.

Pace, rank points a gameGames for 1,000 pointsAhead after 50 gamesAfter 100After 200
+5 (you now)20059%63%68%
+119170%77%85%
+175979%87%95%
+20 (LuckyJ in your seat)5083%91%97%

At your pace 50 games end lower than they began 41 times in 100, and the 1,129 points back to your peak take about 230 games. That is what stuck is: +4.9 seen from the inside. At LuckyJ’s pace 50 games still end lower 17 times in 100, so no pace in this format rises every week. A steady rise means a higher pace and a longer count. Judge the rank over 200 games.

Where the 15 points are. A game’s points split into what you win, what you pay in deal-ins, what others tsumo off you, and the draws and riichi sticks. Each cell is the edge over the other players at the same tables, in points a game. A plus is better than the table.

Points a game against the tableYouLuckyJGap
Points won+1,418+2,4481,030
Points paid in deal-ins+1,947+2,440493
Points paid to tsumo+245+31166
Draws and riichi sticks−210−13872
All+3,400+5,0611,661

Your edge on points won carries an error of 971, so the 1,030 on wins is about one error wide. Chapter 47 could not pin it on anything you do, and the games since have not changed that. The 493 on deal-ins is smaller, and it has an address.

The address. Take every discard made while another player is in riichi, from a hand that is not tenpai after the discard. A genbutsu is a tile that riichi cannot win on: one in its river, or one that anybody threw after it declared. Call every other tile a live tile.

Discards a gameLive tiles among themLive tiles a gameDeal inPoints a live tilePoints a game
You18.2127.9%5.083.62%2481,257
LuckyJ16.6427.4%4.562.78%166756
The humans at your tables17.9935.2%6.344.69%3182,014
The humans at LuckyJ’s tables17.3434.0%5.904.84%3091,822

You throw a live tile as rarely as LuckyJ does, 27.9% of these discards against 27.4%, so folding more often would add nothing. Your live tiles deal in more often and cost more each: 1,257 points a game against 756. The gap of 501 is 4.2 rank points a game, the same size as the 493 above. With 41 deal-ins in 1,132 throws it stands 1.5 standard errors from LuckyJ’s rate, a lean and no proof. The rest of your defence matches. Tenpai against a riichi, your live tiles deal in 8.60% of the time and cost 986 a game, LuckyJ’s 9.04% and 1,161. Not tenpai against a hand with two calls and no riichi out, your discards cost 431 a game and LuckyJ’s 479.

Which live tiles. You cannot see LuckyJ’s choice in your own hand, but the replay page shows Mortal’s. Under each discard it gives Mortal’s share for that tile, and under 25% means Mortal picks it less than one time in four. Since the guide began you made 23,789 discards from a hand not tenpai. Mortal rated 4,772 of them under 25%, 21.4 a game. Those dealt in 0.90% of the time and cost 52.5 points a throw. The 19,017 it accepted dealt in 0.34% and cost 17.3. Live tiles at a riichi split the same way: the 367 Mortal rated under 25% dealt in 4.63% of the time at 339.8 points each, the 765 it accepted 3.14% at 203.5.

Your 248 deal-ins since the guideDeal-insPoints a gameFourths if handed back (44 now)Rank points a game
On a tile Mortal rated 25% or more1383,18528+28.4
On a tile Mortal rated under 25%501,30138+12.1
After your riichi, one Mortal rated 25% or more471,17338+11.4
After your riichi, one Mortal rated under 25%1341343+3.0

Handed back means the points return from the winner on the final scores and the place is counted again. Of the 248 deal-ins, 185 came on a tile or a riichi that Mortal rated 25% or more, and no rule takes those away. The 50 are the ones with a better tile beside them. Handing all 50 back is worth 12.1 rank points a game, which is a ceiling. A floor: had those 4,772 throws cost only what your accepted throws cost, you would keep 753 points a game, 6.2 rank points. Added to +4.9, that is where the +11 and +17 rows of the pace table come from.

What the 50 have in common. They cost 290,200 points. Of the 50, 43 were thrown from a hand that was not tenpai after the tile: 18 at 1-shanten and 25 at 2-shanten or worse. In 42 of the 50, Mortal’s tile left the hand as close to tenpai as yours did, or closer, so most of these were a choice between two tiles and the push was never the question. Mortal’s tile was a genbutsu against the winner 18 times, and you held one 37 times. Mortal’s own tile would have dealt in 3 times. The winner was in riichi 19 times (140,700 points), had called 25 times (109,700) and was closed and silent 6 times (39,800). Of them, 8 came on your 15th turn or later from a hand not tenpai, 57,600 points.

The 61 newest games hold 10 of them, 65,100 points, 8 from a hand not tenpai. Each row opens the replay at your turn.

Hand and your turnYou threwMortal throwsHand afterWinnerPaid
23 Oct, South 2, 2 honba, turn 15[[8s]][[5pr]] 97.9%1-shantenRiichi4,500
24 Oct, East 2, turn 11[[2p]][[5p]] 91.1%2-shantenRiichi7,700
24 Oct, South 4, turn 11[[4m]][[6s]] 95.9%TenpaiRiichi8,000
25 Oct, South 2, turn 16[[2s]][[5mr]] 73.8%TenpaiCaller7,700
26 Oct, South 1, 1 honba, turn 7[[4p]][[S]] 74.3%1-shantenDama8,000
27 Oct, East 3, 2 honba, turn 18[[3m]][[5m]] 88.4%1-shantenRiichi8,600
28 Oct, South 4, turn 17[[5sr]][[4m]] 89.7%1-shantenCaller7,700
2 Nov, South 3, 3 honba, turn 4[[7p]][[F]] 59.8%3-shantenRiichi8,900
7 Nov, South 4, turn 13[[N]][[S]] 74.6%1-shantenCaller2,000
9 Nov, East 2, turn 9[[9p]][[2p]] 96.8%2-shantenCaller2,000

In the East 2 row of 24 October Mortal’s [[5p]] deals in as well. Three of the others, tile by tile.

1. The last discard of the hand. East 3, 2 honba on 27 October, your 18th turn. One opponent had been in riichi since their 3rd turn. You held [[3m]][[5mr]][[5m]][[2p]][[3p]][[5p]][[5p]][[5pr]][[6p]][[7p]][[7p]][[7p]][[8p]] and drew the [[9m]]. This was the last discard of the hand, with no draw left for anyone, and the hand was 1-shanten whichever tile went, so it could not win and could not collect for tenpai. The [[5m]] had been thrown after the riichi and had passed, and you held two. You threw the [[3m]]. It was the winning tile, the last-tile yaku (houtei) lifted the hand to a mangan, and you paid 8,600. Mortal throws the [[5m]] at 88.4% (the red one at 11.1%) and the [[3m]] at 0.1%.

2. Two threats. South 4 on 28 October, your 17th turn. You were the dealer, second with 26,000. The player in last place was in riichi, and the player in third had one call showing. You held [[2m]][[4m]][[4m]][[5m]][[6m]][[7m]][[6p]][[7p]][[8p]][[5sr]][[7s]][[8s]][[9s]] and drew a [[7s]]. Throwing the [[2m]] kept tenpai. You chose to fold and threw the red [[5s]], which lay twice in the riichi’s river. It dealt into the caller: tanyao, two dora and a red five, 7,700, and you finished third. Mortal folds as well, with the [[4m]] at 89.7%. The [[4m]] was in the riichi’s river too, and the caller had thrown the [[1m]] and the [[7m]], so no two-sided wait of theirs could take it. You held two. With two hands threatening, the tile has to pass both.

3. The first discard after a riichi. South 3, 3 honba on 2 November, your 4th turn. You were fourth with 9,700 and the player in third had just declared. You held [[3m]][[4m]][[8m]][[9m]][[1p]][[8p]][[9p]][[6s]][[7s]][[8s]][[E]][[S]][[F]] and drew the [[7p]], which completed [[7p]][[8p]][[9p]]. You threw it straight back. Keeping it and throwing an honor made the hand 2-shanten; without it the hand stayed 3-shanten, and the [[7p]] was a live tile on the one turn that pays ippatsu. It was the winning tile: a mangan, 8,900 with the honba, and the game ended in fourth. The [[8s]] and the [[9m]] were both in the riichi’s river. Mortal throws the [[F]] at 59.8%, then the [[8s]] at 19.7% and the [[9m]] at 14.8%. The [[7p]] gets 0.1%.

Rule When your hand is not tenpai and the tile you are about to throw is not safe against a riichi or a hand with two calls, stop and look for another tile that leaves the hand the same distance from tenpai. If one of them is safer, throw it. From your 15th turn a hand that is not tenpai throws only tiles that pass every threat. After each session, open your deal-ins in the replay page and count the ones where Mortal’s bar for your tile was under 25%: that count, 0.22 a game now, is the number to bring down. Judge the rank over 200 games.
Exception Do not fold more often. Your share of live tiles already equals LuckyJ’s, and when you are tenpai your pushes cost less than its pushes do. When only a win changes your place, as in an all-last from fourth, the push stands and the rule only chooses the tile.

Evidence Your 303 games of 2026, the 223 since the guide’s first commit on 22 September and the 61 from 23 October, 22:53 to 10 November, with LuckyJ’s 1,079 Tokujou games and every other seat at both sets of tables, from mine_steady_rise.py. Rank points use the Saint 3 Jade formula for every seat. The chance of ending ahead treats games as independent with a spread of 148.7 a game. Mortal’s ratings come from the site’s replays and cover the discards before your own riichi. LuckyJ sets the measure for how many live tiles to throw; Mortal only marks which of your own tiles had a better one beside it. Handing points back ignores what the hand would have done next, which is why it is a ceiling.

52

Chapter fifty-two · Recent games · added 18 November

New

Fifty-eight games toward Celestial: your calls pay, and the draw is lost on the tile after

You said you are nearing Celestial, that the recent games have been tough with more thirds, and that you think it is because you keep fewer safe tiles and call more to win more, then deal in. Then you added that calling for tenpai at the draw is what is really not working. Your 58 games from 11 to 18 November came to 20 firsts, 16 seconds, 17 thirds and 5 fourths: +2,438 rank points, 42.0 a game, from 5,620 to 8,058, with a best of 8,265 on 17 November. They split in two. The first 32 went 13, 9, 10, 0 for +2,441, 76.3 a game. The last 26 went 7, 7, 7, 5 for −3, and the last 10 of those went 0, 4, 5, 1 for −87.

The thirds. 17 thirds in 58 games is 29.3%. At your rate before these games, 24.2%, you would expect 14.0, and 17 or more comes 22.2% of the time, so the third rate has not clearly moved. The climb already had 10 thirds in 32 games. What changed in the last 26 is the firsts and the fourths.

Where the points went. Each column is points a game.

Points a game223 games beforeThe climb, 32 gamesThe last 26 games
Won16,57920,61914,704
Paid in deal-ins6,0722,9166,596
Paid to tsumo5,9056,2256,554
Draws−249−234−385
Net+2,541+9,681−23

The last 26 games paid 6,596 a game in deal-ins, near your usual 6,072. The climb’s 2,916 was the unusual number, and it was not going to last. What fell is the points you won: 14,704 a game, against your usual 16,579.

The calls.

223 games beforeThe climbThe last 26LuckyJ
Hands opened35.3%37.3%43.3%36.5%
Riichi18.2%15.4%11.6%19.3%
Average win6,7107,0955,7927,317
Net points an open hand+626+2,138+401+611
Net points a closed hand+32+215−310+220
Closed tenpai with a yaku, declared63.5%54.3%48.6%67.2%

Your calls pay. Over the 58 games an open hand made +1,287 points, where LuckyJ’s make +611. Mortal agrees with where you call: you took 20.3% of your 33.45 call chances a game and Mortal takes 21.0%, and in the last 26 games 21.2% against 20.7%. What moved is the riichi. In the last 26 games you declared in 11.6% of your hands, where you used to declare in 18.2% and LuckyJ declares in 19.3%. Part of that is the calls themselves, since an open hand cannot declare. The other part is dama: with a closed tenpai that had a yaku you declared 51.8% of 83 times over the 58 games, where LuckyJ declares 67.2% and you used to declare 63.5%. At the first riichi chance of a hand Mortal declares 56.2% of the time and you 52.6%, over 133 chances. Your average win fell to 5,792.

The safe tiles. At your first discard after a riichi, with your hand not tenpai, how often you held a genbutsu, a tile that riichi cannot win on:

Held a genbutsuClosed handOpen hand
223 games before85.3%, 586 of 68778.0%, 138 of 177
The climb83.8%, 83 of 9960.7%, 17 of 28
The last 2680.0%, 60 of 7580.0%, 24 of 30
LuckyJ85.3%80.2%

You were right that you kept fewer, but only in the climb’s open hands, 17 of 28. In the last 26 games it was 24 of 30, LuckyJ’s rate. And none of the 17 open hands in the 58 games that met a riichi without a genbutsu ended in a deal-in. This is not what the last 26 games lost.

Tenpai at the draw.

223 games beforeThese 58 gamesLuckyJ
Tenpai at the draw43.1%37.5%46.7%
Draw points a game−249−302−42
Draws with a riichi out54.7%62.5%51.1%
Tenpai at those24.6%20.0%28.4%
Tenpai at draws with no riichi out65.5%66.7%65.8%

You lose 302 points a game at draws where LuckyJ loses 42, and the humans at your tables, tenpai at 48.8% of their draws, gain 87. The gap to LuckyJ is 260 points, 2.2 rank points a game. With no riichi out you match it, 66.7% against 65.8%. The gap sits in the draws where somebody declared, and those are 62.5% of your draws against 51.1% of its: part of that is the tables you sit at.

The calls are not where it fails. The calls you made from your 12th turn that reached tenpai came 0.60 times a game in these 58 games, won 34.3% of the time and made +1,129 each. LuckyJ’s come 0.41 times a game and make +714. What fails is the tile you push without a call. From the 13th turn, against a riichi, with a hand that stays 1-shanten, you push a live tile 0.53 times a game, LuckyJ 0.49, and both of you do it with a median of 11 unseen tiles that would make tenpai. Mortal rated 102 of your 150 pushed tiles at 25% or more, so the pushes are mostly ones it makes. The results are not:

From the 13th turn, 1-shanten, a riichi outYour handsYou: tenpai at the drawYou: net a handLuckyJ: tenpai at the drawLuckyJ: net a hand
Pushed a live tile, 8 or more tiles to tenpai12010.8%−1,13915.1%−753
Pushed a live tile, fewer than 8300.0%−2,0578.7%−1,350
Only safe tiles, 8 or more11012.7%−66013.6%−770
Only safe tiles, fewer than 8578.8%−1,0494.6%−1,236

With 8 or more tiles your pushes reached tenpai at the draw 10.8% of the time, LuckyJ’s 15.1%: a short sample of 120 hands. With fewer than 8 none of your 30 reached it, 20.0% dealt in, and each cost 2,057 points, where throwing only safe tiles in the same spot cost 1,049. LuckyJ’s own numbers make the push with fewer than 8 a wash, −1,350 against −1,236. For you that is 0.11 hands a game, about 110 points, 0.9 rank points.

Two of them. South 1 of the 14 November evening game, your 17th turn. You were fourth with 16,900 and one player was in riichi. You called chi on the [[6m]] with [[7m]][[8m]], which Mortal does too (it prefers the [[5m]][[7m]] chi), and the hand was still 1-shanten with 7 tiles to make tenpai. You threw the [[W]], a live tile; Mortal throws the [[7s]] at 86.6%, a genbutsu, and the [[W]] at 12.9%. The hand never got there, you paid 1,500 for noten, and the game ended in fourth. East 4 of the 17 November game at 18:10, your 16th turn. You were the dealer and first with 30,200, with a pon of [[S]]. The hand was 1-shanten with 4 tiles to make tenpai, and you threw the [[9p]] into the riichi where Mortal throws the [[5p]] at 82.5% and the [[9p]] at 0.4%. You stayed noten, paid 1,500, lost the deal, and finished that game third.

Rule Keep calling where you call: Mortal calls in the same places and your open hands make more than LuckyJ’s. Declare a closed tenpai that has a yaku more often; LuckyJ declares two in three. Call for tenpai at the draw when the call reaches tenpai. When it does not, or when you have no call, from the 13th turn against a riichi push a live tile only with 8 or more unseen tiles that make tenpai; with fewer, throw safe tiles and pay the noten.
Exception When the 1,000 to 3,000 points of a draw would change your place at the end of the game, count the places first, as chapter 36 does, before the outs.

Evidence Your 58 games from 11 to 18 November against the 223 games between the guide’s first commit and chapter 51, with LuckyJ’s 1,079 Tokujou games and every other seat at both sets of tables, from mine_draw_tenpai.py. A live tile is one that is not a genbutsu to every riichi on the table. Tiles to tenpai count the unseen copies from what you could see. A late call is the first call of a hand from the 12th turn. The late pushes pool all 281 games since the guide, since 58 games hold only 36 of them. Mortal’s ratings come from the site’s replays.

Method

How this guide was built.

  • Your games: 80 Jade South hanchan since 1 January 2026, fetched from the Mahjong Soul client and your public ranked-game records. LuckyJ: 1,255 Tenhou games rebuilt from the cached NAGA reports. Both run through the same replay and metric code.
  • Each table is rebuilt from the record at the start of your turn. Opponents' hands are shown as they really were; the switch at the top hides them.
  • Acceptance counts the unseen copies of every improving tile from your seat, with a called tile counted once. Against a riichi, tiles passed after the declaration count as genbutsu and anchor suji; a 4, 5 or 6 needs both partners for nakasuji, and an early 5 can cover one side of a 4 or 6 (virtual nakasuji, following the main playbook's sotogawa rule). Honors are read by how many other copies show. Yaku and value are checked only where the text says so.
  • Generated data: site/honver-guide.json from scripts/build_personal_guide.py and data/personal_guide_spots.json. Chapter numbers come from scripts/mine_riichi_folds.py, mine_open_tenpai_push.py, mine_single_call.py, mine_vs_open.py, mine_win_value.py and mine_value_choice.py, for chapters 6 to 8 from mine_dora_hands.py, mine_flush_calls.py and mine_caller_tells.py, for chapter 9 from mine_shape_plans.py, for chapter 10 from review_recent_games.py and mine_late_placement.py, for chapter 11 from mine_tenpai_push.py, for chapter 12 from mine_open_vs_open.py and mine_call_chances.py, with mine_dora_hands.py for the open wins and deal-ins, for chapter 17 from mine_early_safe_tiles.py and contrast/hands_all.py, with Mortal's ratings from the site's replays, for chapter 18 from mine_caller_surface.py, drawn by build_caller_surface.py, for chapter 19 from mine_caller_fold_line.py over contrast/vs_callers.py and mine_caller_surface.py patterns, for chapters 20 and 21 from mine_break_folds.py over contrast/vs_callers.py and mine_riichi_folds.py, drawn by build_break_fold_figure.py, for chapter 23 from mine_tenpai_push.py and contrast/riichi_danger.py, for chapter 26 from contrast/hands_all.py and mine_riichi_folds.py with the replays’ Mortal ratings, for chapter 27 from contrast/hands_all.py, the game logs’ standings and your ranked-game records, and for chapter 28 from mine_dealer_callers.py over contrast/vs_callers.py and contrast/choices.py, drawn by build_dealer_caller_figure.py, for chapter 29 from mine_costly_run.py over contrast/hands_all.py and the replays’ Mortal ratings, for chapter 30 from mine_defense_check.py over contrast/hands_all.py, mine_riichi_folds.py, contrast/vs_callers.py, contrast/choices.py and mine_early_safe_tiles.py, for chapter 31 from mine_declined_riichis.py over the replays’ Mortal ratings, for chapter 32 from mine_no_safe_riichis.py over mine_early_safe_tiles.py and the replays’ Mortal ratings, for chapter 33 from mine_climb_check.py over contrast/tenpai.py and the riichi rows, for chapter 34 from mine_streak_check.py over contrast/tenpai.py and the replays’ Mortal ratings, for chapter 35 from mine_thin_pushes.py over the spots of mine_tenpai_push.py and the replays’ Mortal ratings, for chapter 36 from mine_push_ev.py over the same spots, for chapter 37 from mine_fold_whatif.py over mortal_whatif.py runs of chapter 35’s folds, for chapter 38 from mine_fold_theory.py over those runs, contrast/tenpai.py and the push-or-fold spots, for chapter 39 from mine_push_company.py over every seat’s push-or-fold spots, for chapter 40 from mine_half_suji.py over every live number tile thrown into a lone riichi, and for chapter 41 from mine_fold_model.py over every tile held by a player discarding into a lone riichi, with LightGBM and TabPFN 3.5 beside its own network, for chapter 42 from mine_slide_check.py over chapter 33’s rows and the replays’ Mortal ratings, for chapter 43 from mine_chiitoi_tanki.py over every seven-pairs hand waiting on one tile, for chapter 44 from mine_thin_riichi.py over every riichi at every seat, for chapter 45 from mine_rank_drain.py over every hand and every deal-in, for chapter 46 from mine_defense_spots.py over three hands and every discard into a lone riichi, for chapter 47 from mine_point_gap.py over every hand and every win, for chapter 48 from mine_recent_run.py over your last ten games and the 217 before them, for chapter 49 from mine_fourth_rate.py over your last 100 games, for chapter 50 from mine_run_update.py over the 15 games after chapter 48’s ten, for chapter 51 from mine_steady_rise.py over the 223 games since the guide began, and for chapter 52 from mine_draw_tenpai.py over the 58 games from 11 to 18 November, each run on your games and on LuckyJ's with the same code.
  • The main LuckyJ playbook has the general rules these chapters build on.