The LuckyJ Playbook

What this book is · A table-first LuckyJ playbook

Digital points for Playing Like LuckyJ

I reviewed 1,255 unique NAGA-reviewed hanchan, 13,054 hands, and 147,337 LuckyJ discard decisions, then grouped them into habits you can use at the table.

Scope before style

Start with ordinary efficient play.

LuckyJ matches Nishiki's top discard in about 73% of eligible reviewed decisions. This book is about the other 27%, because the splits make good review questions, not because deviating is LuckyJ's default. Copy a split only when its value, safety, opponent, or placement condition is visible before the result.

Training versus test: LuckyJ learned through self-play, not by imitating Tokujou players. Tokujou was its public live test room. Tencent's description.

This book covers 1,255 unique hanchan from Apr 19–May 30, 2023: 1,079 Tokujou, 127 Upper, and 49 General. It contains no Houou or Mahjong Soul sample. Teaching examples below are restricted to Tokujou.

Transfer limit: shanten counting, yaku legality, genbutsu, and target-specific safety carry over to any room. Riichi, calling, keiten, safety-inventory, and placement thresholds have to be recalibrated for your room and point system.

Saint 3 calibration (four-player South): at an even 25,000-point finish, the current nominal placement spread is about +125 / +60 / -5 / -255 in Jade and +135 / +65 / -5 / -255 in Throne. The final-score adjustment moves those totals, but one fourth still costs roughly two firsts. So a named exit tile and late fold discipline carry over. Slow, untargeted safety hoarding does not: stronger opponents give you fewer spare turns, and lost first and second equity still counts. Jade table; Throne table (checked July 2026).

The points

The LuckyJ Operating System

LuckyJ does not play from a tile-efficiency chart alone. Efficiency is one input to a placement engine. The pattern that repeats across the corpus: decide what the hand is for, keep the routes that serve that job, hold enough defense to keep making choices, and drop a stale plan as soon as the table changes.

Decision method

How to read every discard.

1. Job

Decide what the hand is for before you look for the best shape. A hand has to matter to the table before it can buy much risk.

2. Routes

Count real routes: closed riichi, open yaku, value upgrade, safe tenpai, and clean fold.

3. Table Texture

Read the opponents' rivers and calls before trusting a raw danger number. One open hand, one riichi, and two live threats are three different games.

4. Future Danger

Price the tiles you will have to discard later. A tile that is fine now can become the discard that ends the hand.

5. Point of No Return

Decide in advance which draw, call, riichi, or second threat ends the attack. LuckyJ's aggression always has a stop condition.

6. Human Override

Check LuckyJ's third-row choices against NAGA, Mortal, and your own point arithmetic. That part of the style is the hardest to copy.

Prescriptions

Simple switches to remember.

These are the habits worth copying first: what LuckyJ does with a tile or a call before anything special happens, and what makes it change its mind. They come from its non-dealer hands across the whole review, so they describe how it usually plays, and the table can always overrule them. As dealer it plays a little closer to the textbook line, faster and more willing to push.

Single value honors leave early

A lone dragon or value wind is the first tile in a new hand that LuckyJ questions. If nothing needs it and nobody has called, it usually goes on the third discard: late enough to see whether it pairs, which it does about 12% of the time, and early enough that giving it away costs little.

Picture your third turn at a quiet table, with [[3m]][[4m]][[5m]], [[5p]][[6p]][[7p]], a few loose souzu and a lone [[C]]. The [[C]] is your only honor and it does no work. LuckyJ lets a tile like that go now, before an opponent has a hand that wants to pon it.

Once someone has called, the same tile tends to stay until around the sixth discard. Treat that as timing, not a schedule: let it go while the caller shows no yaku and no progress, and hold it once their hand starts moving. It goes sooner when your own hand is already two-shanten or better by turn 3 or the tile is the round wind, and it stays longer when it is your double wind. With two such honors, the first goes early and the second a couple of turns later.

Keep it once it pairs, when it is your only realistic yaku, or when an open hand may still need that exact tile. Point 12 prices that last case.

Evidence 8,461 dragon and 5,248 value-wind starts in LuckyJ’s non-dealer hands. Median cut: turn 3 on a quiet table, turn 6 once someone had called; about 12% paired first. As dealer, 76.6% were gone by turn 6.

Guest winds are cheap until they defend

A wind that is neither your seat wind nor the round wind does almost nothing for your hand, so it is usually the first tile out. LuckyJ’s median is the second discard.

Its value changes when someone declares. Now make it turn 8, with Shimocha in riichi and a [[N]] in your hand that nobody has discarded. It looks like a free tile, but a live honor is as good a tanki or shanpon wait as any, and LuckyJ threw one like that only 13% of the time. Once a copy had shown up in a river, it went 33% of the time.

Distance from tenpai decides most of it. From one-shanten or better, LuckyJ spends the live honor 29% of the time; from two-shanten or worse almost never, because holding it is part of folding. It lets go sooner when it has no genbutsu at all against the riichi, and a live honor drawn late is held a little more often than cut.

Keep the wind when it is your safest answer to a specific opponent, or when every cleaner exit is gone. Point 16 labels each honor by the job it is doing.

Evidence 6,027 guest-wind starts in non-dealer hands, median cut turn 2. Against a riichi on turns 7 to 9, a live honor went 12.9% of the time with no copy seen and 32.5% with one; 29.2% from one-shanten or better, 4.9% from further back.

Pon value pairs by default

With a pair of dragons, the round wind or your seat wind, the pon is the default, and the first copy is the one to take. LuckyJ called it 76% of the time. A later copy was called only about half the time.

Take [[F]][[F]] with a dora in hand and a two-shanten shape on turn 4. The pon gives the hand its yaku and a step toward tenpai for the price of one open meld. That is the easiest call in the game, and this early LuckyJ makes it nearly nine times in ten.

The brake is the table. When an opponent is already threatening, the same pon drops to 36%, because opening leaves fewer safe tiles for the turns that matter. It also cools in the middle of the hand, when a closed path may already be worth more than a cheap open one. Dora in hand pushes the other way.

Skip the pon when it leaves a cheap, still-distant open hand with no safe next discard, or when your closed line is already better. Point 03 is about calls that change the hand.

Evidence 2,220 yakuhai pon chances in non-dealer hands: 75.9% on the first copy, 55.8% on a later one; 82.3% on a quiet table against 36.2% with a threat.

The first call has to earn its price

Opening a closed hand gives up riichi, much of the hand’s value and several defensive tiles, so LuckyJ’s first chi has to buy something specific. From a closed hand it took 7% of its chances. Once the hand was already open that price was paid, and it called 27% of the time.

Kamicha throws [[4p]] and you hold [[2p]][[3p]] in a closed three-shanten hand with no yaku in sight. The chi moves you from three steps away to two and leaves a cheap open hand to defend with. LuckyJ passes that call almost every time. When the same chi reaches one-shanten, it calls seven times as often.

Late in the hand the price drops. With 12 tiles or fewer left, a first chi from a closed hand is about three times as common, because a tenpai at the draw is worth having even without a win. Junk calls it mostly passes, but a junk hand holding a value pair still takes the pon, since the pair brings its own yaku.

Take the call when it gives a yaku, tenpai, a real one-shanten, denial, or a safer path. Points 03 and 04 follow what happens after the call.

Evidence 19,131 chi chances in non-dealer hands: 6.7% of 16,284 from a closed hand, 26.6% of 2,847 once open; 17.0% when the chi reached one-shanten against 2.4% when it did not.

Riichi wants a real wait

With a wide wait, LuckyJ declares at once: 95% of the time when eight or more winning tiles are left, and 78% with four to seven. With three or fewer it declares only 34% of the time.

Your tenpai waits on a [[3s]] kanchan with two copies left, and you lead by 10,000. A riichi pays 1,000 to announce a two-tile wait to three players who are free to fold. LuckyJ stays quiet about two times in three there, and a riichi it passes up usually stays a dama tenpai instead of turning into a declaration later.

Two things push it back toward declaring: a thin wait on two tile types instead of one, and sitting in fourth. Two things hold it back: a riichi already on the table, since the declaration starts with one more tile that has to pass right now, and a hand that already carries two dora. As dealer it declares more often everywhere, thin waits included.

Do not riichi by reflex when the wait is bad, the declaration tile is the danger, or your lead already has enough value. A cheap hand with a good wait is the one LuckyJ declares most often (Against the humans, 3). Point 07 covers the pressure a riichi puts on the table.

Evidence 2,381 first riichi chances in non-dealer hands: declared 95.3% with 8 or more tiles, 78.3% with 4 to 7, 33.7% with 3 or fewer. Of 821 declined first chances, 58.1% stayed dama.

Cross danger mainly with a real hand

Against a riichi, what decides a push is the hand you are pushing with. From tenpai LuckyJ still threw a dangerous tile more often than not. From one-shanten it did so 32% of the time, from two-shanten 20%, and from further back 14%.

Its first answer to a riichi runs the same way in reverse. From tenpai the next discard was genbutsu only 35% of the time; from three-shanten or worse, 69%. Late in the hand everything tightens: with 20 tiles or fewer left, a tenpai crossed danger 52.6% of the time against 63.4% earlier, and a one-shanten dropped to 22%.

Some of that danger is forced. A one-shanten with no genbutsu in hand crosses danger about half the time, simply because it has nothing safe to throw, which is a good reason to keep genbutsu while you can. As dealer the rates climb, to about two in three from tenpai.

Do not fold a tenpai that has value, and do not throw a safe tile that destroys your only winning route. Price the push again on every draw; Points 08 and 09 walk through it.

Evidence Closed-hand decisions against a riichi in non-dealer hands (986 tenpai, 4,361 one-shanten, 4,810 two-shanten, 2,889 further): a discard above 5% on NAGA’s danger proxy 58.9%, 31.7%, 19.6% and 14.1% of the time. The proxy shows risk taken, not intent.

Open hands ask for tile control

An open hand never announces that it is ready, so LuckyJ neither folds to one nor ignores it. Against two open sets it kept throwing tiles NAGA rates as dangerous about half the time early in the hand, and a little over a third of the time late. Three open sets tightened it further.

Suppose Toimen has chi’d twice in manzu, and you hold a live [[5m]] and a [[9p]] that Toimen has already discarded. If your hand is one-shanten, the [[5m]] may be the tile that moves it, and LuckyJ often pays for it. If you are two-shanten or worse, the [[9p]] is the answer and the [[5m]] waits.

Answer the specific player. A tile that is safe only against someone else does not count, and a cheap, distant hand has no business pushing a tile the caller could want. Points 04 and 13 are about keeping that answer in hand.

Evidence Discards against two open sets in non-dealer hands (1,228 early, 6,497 middle, 3,478 late): above 5% on NAGA’s danger proxy 52.7%, 42.6% and 37.6% of the time; 47.4% from one-shanten or better against 26.9% from further back.

Endgame tenpai has a price

In the last 12 tiles a tenpai is worth something even if it never wins, because it collects at the draw instead of paying. LuckyJ calls for it accordingly: 31% of the time from one-shanten, 20% from further back, and almost never once it is already tenpai.

Nine tiles are left, your hand is one-shanten, and Kamicha throws the [[6s]] that fills your [[5s]][[7s]]. If the discard after the chi is safe, the call can turn a noten payment into a tenpai receipt. That is the kind of call LuckyJ makes.

Against a riichi, a tenpai hand in the last 20 tiles kept crossing danger 67% of the time, and still 60% over the last five tiles. A hand two steps away crossed it only about one time in four and mostly stayed safe.

Skip the tenpai fight when the payment cannot change your placement, when the tile is too dangerous for what a draw pays, or when a safe fold keeps more. Points 09 and 10 work through the late count and keiten.

Evidence Last-12-tile call chances in non-dealer hands: 30.8% of 465 from one-shanten, 20.1% of 670 from two or more away, 4.5% of 335 already tenpai. Against a riichi, tenpai discards above 5% on NAGA’s danger proxy ran 60 to 67% in every stretch of the last 20 tiles.

Against the humans

Where LuckyJ beats strong humans.

The prescriptions above describe how LuckyJ plays. These five answer a different question: where did LuckyJ gain on the strong humans it actually played? Each of its 1,079 Tokujou games was replayed with all four hands visible and set against the 3,237 human seats at the same tables, and against 2,714 Houou games from the same year.

LuckyJ finished 484 points a hand ahead of the Tokujou players at its tables. About half of that is deal-ins avoided, and almost all of that half comes from hands that were not tenpai. Once a hand was tenpai, LuckyJ and both groups of humans dealt in at nearly the same rate. The Houou players, the strongest human room, closed about half of the gap in hands that were not tenpai and none of the gap in wins.

The same shape runs through every decision below. The humans’ choices move a little with the factor that should decide them. LuckyJ’s choices move a lot: more risk where the factor says the move pays, less where it says it does not.

Evidence Deal-ins per 100 hands while not tenpai: 7.4 Tokujou, 6.0 Houou, 4.3 LuckyJ; while tenpai 6.5, 6.2, 6.0. LuckyJ also won more hands, 22.9% against 21.0% for both groups, but most of that margin is the table: LuckyJ always sat with three humans, who deal in more often than it does. Per opponent, a human dealt into LuckyJ 4.77 times per 100 hands and into another human 4.67 times. LuckyJ’s wins were larger, 6,592 points per ron against 6,147.

Live tiles cut against a caller

From two-shanten or worse, no dora in hand, holding a safe tile

Tokujou71 → 43
Houou70 → 37
LuckyJ75 → 33

no tells two or more tells

Unsafe tiles cut against a riichi

Share of discards, by the hand’s distance from tenpai

Tokujou41 → 13
Houou39 → 9
LuckyJ42 → 6

tenpai three-shanten or worse

Tanyao calls from two-shanten

Share of chances taken, by dora in hand

Tokujou24 → 49
Houou28 → 50
LuckyJ23 → 59

no dora two or more dora

Percentages on a scale from 0 to 80. Each row runs from the hollow dot to the filled one. The humans at both levels move in the right direction; LuckyJ moves further at both ends.

Answer the caller’s tells, not the call

A single call tells you little. What shows that a caller is ready is what comes after it: a second call, a call on their ninth turn or later, or two drawn tiles thrown in a row since the call. With none of those showing, a caller was tenpai about one time in seven. With two, about two times in three.

Toimen chis [[3p]] on turn 5 and discards from hand. Your two-shanten hand holds a live [[6m]] you would like to cut and a [[9p]] that Toimen has already thrown. Cut the [[6m]]: one call and nothing else is a hand still being built. Four turns later Toimen pons [[F]] and then throws two drawn tiles in a row. All three tells are showing now, so the [[9p]] goes and the [[6m]] waits.

LuckyJ plays that difference in both directions. Against a quiet caller it kept cutting live tiles from far hands a little more freely than the humans did. Against a caller with two tells it cut them one time in three, while the humans went on cutting them about four times in ten, Houou players included.

With two tells showing and your hand one-shanten or further, discard as you would against a riichi. From tenpai the choice is closer, and LuckyJ still pushed about half the time. Point 04 covers the brakes on your own open hand.

Evidence Live cuts while holding a safe tile, from two-shanten or worse with no dora in hand: no tells 70.7% Tokujou, 70.4% Houou, 74.5% LuckyJ; two or more tells 42.8%, 37.3%, 33.1%. The caller was tenpai 13.5% of the time with no tell, 66.1% with two and 85.4% with all three (Houou tables 14.3%, 70.2%, 87.1%). Deal-ins into open hands from hands that were not tenpai cost 172, 126 and 90 points a hand.

Fold a far hand at the first discard

When a riichi comes and your hand is two or more tiles from tenpai, the hand has stopped racing. After a riichi, a two-shanten hand went on to win about one time in twenty-five and a three-shanten hand about one time in fifty. Every unsafe tile it cuts pays for a win that almost never comes.

Shimocha declares on turn 6. You hold [[2m]][[3m]] [[5p]][[7p]] [[4s]][[5s]][[6s]] [[8s]][[9s]] [[E]][[E]] with a [[1p]] and a [[W]] that Shimocha has thrown. The hand is two-shanten on 16 tiles and still looks alive. The [[1p]] and the [[W]] are your next two discards, and after them you keep cutting the safest tile you hold.

Humans lose most of these points early in the hand, when a riichi seems to leave time to catch up. From two-shanten or worse, the Tokujou players cut an unsafe tile on a quarter of their discards against a riichi made by their eighth turn, some of them when a safe tile would have kept the same shanten. The Houou players have learned about half of this.

A far hand keeps a tile only when that tile is also safe, or when every safe tile is gone. Folding from one-shanten is a separate price, which prescription 6 above and Point 08 work through.

Evidence Unsafe cuts against a riichi from two-shanten: 16.2% Tokujou, 13.0% Houou, 10.3% LuckyJ; from three-shanten or worse 13.0%, 9.4%, 6.4%. Deal-ins per 100 of those discards from three-shanten or worse: 1.11, 0.47, 0.21. On the seat’s eighth turn or earlier, from two-shanten or worse: 25.4%, 21.5%, 16.1%. Deal-ins into a riichi from hands that were not tenpai cost 157, 109 and 66 points a hand.

Riichi the cheap tenpai

A closed tenpai that pays 1,000 or 2,000 without riichi is the hand that most wants the declaration. Riichi roughly doubles it, and in the humans’ own games the declared hands won as often as the quiet ones or more. LuckyJ declared about four of these tenpais in five. Both groups of humans kept about a third of them quiet.

Turn 8, nobody has called. You hold [[2m]][[3m]][[4m]] [[5p]][[6p]][[7p]] [[6s]][[7s]][[8s]] [[3s]][[4s]] [[5m]][[5m]], tanyao and pinfu, waiting on [[2s]] and [[5s]]. Dama pays 2,000. Riichi pays 3,900 by ron and 1,300/2,600 by tsumo before ura. LuckyJ declared hands like this nearly every time.

Staying quiet keeps an escape route, and the humans’ own results put a price on it: on these hands, the quiet ones finished 900 to 1,700 points a hand behind the declared ones at every wait width. LuckyJ kept dama for the other kind of hand, three han or more on a wait of three tiles or fewer, where it declared only about one time in five.

A thin wait still holds LuckyJ back, as prescription 5 above describes; a cheap hand does not, and prescription 5’s closing line is corrected to match. Stay quiet when the hand already pays 3,900 on a thin wait, or a mangan without a wide one.

Evidence First closed tenpai with one or two han and a yaku on every wait, non-dealer, nobody in riichi. Declared with three or fewer live tiles: 34.0% Tokujou, 41.5% Houou, 52.9% LuckyJ; four to six: 62.3%, 64.5%, 82.6%; seven or more: 78.9%, 82.9%, 95.8%. The humans’ riichi finished ahead of their dama by 1,094, 895 and 1,730 points a hand in those three bands. Mortal, run on the same human spots, preferred riichi in 76.7% of them.

Let dora decide a two-shanten call

From two-shanten, a call toward tanyao gives up riichi, the closed hand’s value and part of its defense. Without dora it buys a 1,000 point hand that still needs two tiles. With two dora the same call makes a 3,900 point hand that needs the same two tiles. LuckyJ passed the first kind about three times in four and took the second more often than not.

The dora indicator is [[5s]]. You hold [[3m]][[4m]] [[6m]][[7m]] [[3p]][[5p]] [[2p]][[2p]] [[5s]][[6s]][[6s]][[7s]][[8s]], two-shanten with two dora, and Kamicha throws [[4s]]. Chi with [[5s]][[6s]] and cut the [[3p]]: the hand is one-shanten on [[2m]], [[5m]] and [[8m]] and already worth 3,900.

The humans’ own results say the same. Without dora the call was worth nothing against passing, and with two it was worth about a thousand points a hand. The humans did call more with dora, just not enough: about half the time with two, where LuckyJ called six times in ten, and with three more than eight times in ten.

A pair of dragons or value winds is the easy case: from two-shanten LuckyJ pons it more than nine times in ten. Points 03 and 06 price what a call costs and what early dora are worth.

Evidence Two-shanten calls that head for tanyao, nobody in riichi: taken with no dora 24.3% Tokujou, 27.6% Houou, 22.9% LuckyJ; with two or more 48.7%, 50.1%, 59.2%. With three or more dora on any route: 55.7% Tokujou, 59.6% Houou, 83.6% LuckyJ. The humans’ call against a pass, points a hand: no dora −91 (±136), one dora −185 (±141), two or more +1,047 (±219). Value pairs ponned from two-shanten: 83.3%, 87.3%, 93.4%.

Start the flush on the first discard

LuckyJ won a half or full flush almost twice as often as either group of humans. The difference starts in the first three discards. With a flush-shaped deal, LuckyJ throws the off-suit tiles while they are still harmless and keeps the honors that might pair, and the humans throw the honors first.

Your deal is [[1p]][[2p]][[4p]][[5p]][[7p]][[8p]][[9p]] [[P]][[P]] [[N]] [[3m]] [[6s]][[8s]]. Ten tiles fit pinzu and honors, and the [[P]] pair is a yaku. The [[3m]], [[6s]] and [[8s]] go before the [[N]].

The value pair decides it. It gives the hand an early pon with a yaku and lifts a flush toward mangan. In the humans’ own games, cutting two off-suit tiles in the first three discards was worth about 400 points a hand when the deal held a value pair, and less, with wide error, without one.

With eight fitting tiles or fewer, LuckyJ builds the ordinary hand: it was in one suit by its sixth discard only about one time in eighty. Point 02 is about keeping several versions of a hand open.

Evidence Half or full flush wins per 100 hands: 1.17 Tokujou, 1.08 Houou, 2.09 LuckyJ. With nine fitting tiles and a value pair, two or more off-suit tiles cut in the first three discards: 39.1%, 37.9%, 52.2%; with ten or more, 55.8%, 52.0%, 63.3%. The humans’ fast start against a slow one, points a hand: +372 (±98) with nine and a pair, +413 (±138) with ten or more and a pair.

Three differences that are not lessons

The order of the first honor cut

LuckyJ and Mortal keep a lone honor over a lone 1, 9, 2 or 8 more often than humans do, but in the humans’ own hands the order made no measurable difference. Cut whichever tile the hand needs less.

The higher riichi win rate

LuckyJ’s riichis won 51.8% of the time against 48% for the humans. That comes from the table: a human dealt into LuckyJ’s riichi as often as into another human’s, and LuckyJ’s riichi always faced three humans.

Pushing a tenpai

From tenpai, LuckyJ pushed against a riichi at nearly the humans’ rate and dealt in about as often. The points are in the hands that were not tenpai.

Open callers

When to fold to open callers.

A riichi announces the moment a hand is ready. A call does not, so the fold has to come from reading the caller. The four answers below come from the same games as the section before, LuckyJ’s 1,079 Tokujou hanchan and 2,714 Houou hanchan, replayed with every caller’s hand visible. Against the humans, 1 showed that what follows a call tells you more than the call. These go further: when LuckyJ stops cutting live tiles, the caller it respects from the first turns, a free safe tile most players leave unused, and what it does when its own tenpai needs a live tile.

Everything here counts the caller’s own discards. Count the tiles in their pond, the discard right after each call included, and add back any tile another player called from them. Each call adds exactly one discard, the tile thrown right after it.

LuckyJ’s answers keep the pattern of the section before. As a caller gets closer to tenpai, the humans’ choices drift. LuckyJ’s cross a line: before it, live tiles go freely, and after it the safe tile goes first.

Evidence 577,879 discards made against callers with nobody in riichi: 122,582 by the Tokujou humans, 416,744 by Houou players and 38,553 by LuckyJ. A safe tile here is the caller’s genbutsu, a full suji, or an honor with two or more copies showing; every other tile counts as live. Whether the caller was tenpai, and whether a tile was on their wait, comes from their real hand.

Where LuckyJ stops cutting live tiles

Two-shanten or worse against one caller, holding a safe tile and a live tile that both keep the shanten: how often the discard was the live tile, by the caller’s discards so far. Each dot is one discard, sized by LuckyJ’s turns there; the lines are fitted curves.

LuckyJ, per discard and fitted95% range of the fitthe humans at Tokujou and Houou, fittedbelow 50%, the safe tile goes first

One callThe two tiles equally good for the hand

0% 50% 100% 1 3 6 9 12 15 18 The caller’s discards so far 50% betweendiscards 9 and 10 83% 34%

One callThe safe tile costs acceptance

0% 50% 100% 1 3 6 9 12 15 18 The caller’s discards so far 96% 72%

Two or three callsThe two tiles equally good for the hand

0% 50% 100% 1 3 6 9 12 15 18 The caller’s discards so far 50% betweendiscards 6 and 7 68% 25%

Two or three callsThe safe tile costs acceptance

0% 50% 100% 1 3 6 9 12 15 18 The caller’s discards so far 90% 63%
On the left, the two tiles are exactly as good for the hand. There LuckyJ’s line falls at a one-call caller’s tenth discard and at the seventh against two or three calls, and the humans cross in the same places but move less on either side: they spend more safe tiles before the line and more live tiles after it. On the right, the safe tile would cost the hand acceptance, and LuckyJ keeps cutting the live tile past those discards. Counting both, its line is the twelfth discard against one call and the ninth against two or three.
The numbers as a table

Each cell: LuckyJ, then the humans, then LuckyJ’s number of turns. Discards where LuckyJ had no turns are blank.

DiscardOne call
The two tiles equally good for the hand
One call
The safe tile costs acceptance
Two or three calls
The two tiles equally good for the hand
Two or three calls
The safe tile costs acceptance
176.2%58.1%21 turns94.9%92.4%39 turns
279.5%58.0%88 turns96.5%89.7%199 turns75.0%63.0%8 turns100.0%92.1%9 turns
379.4%60.9%204 turns94.1%90.3%341 turns87.0%65.8%23 turns93.3%90.1%45 turns
482.7%59.1%318 turns93.1%90.4%421 turns74.4%62.4%43 turns86.3%86.7%102 turns
577.3%59.8%339 turns92.1%88.9%405 turns62.3%56.4%77 turns85.6%82.2%118 turns
670.7%59.1%317 turns86.8%86.8%355 turns51.7%47.8%118 turns82.9%78.9%123 turns
762.5%58.9%264 turns86.5%85.3%222 turns50.0%47.1%116 turns66.4%73.6%134 turns
866.9%54.6%178 turns80.9%81.7%157 turns35.6%42.6%87 turns60.2%68.3%108 turns
959.1%53.2%132 turns76.5%79.1%98 turns40.0%40.2%70 turns64.7%63.1%85 turns
1047.6%46.5%84 turns74.1%75.5%58 turns28.3%38.5%46 turns60.0%58.6%55 turns
1142.1%42.2%57 turns73.7%61.7%38 turns30.0%33.3%40 turns48.6%53.6%37 turns
1231.0%38.9%42 turns71.4%61.0%28 turns28.6%26.2%28 turns59.3%41.5%27 turns
1325.9%35.0%27 turns53.3%54.7%15 turns7.1%27.0%14 turns31.6%39.1%19 turns
1420.0%31.7%10 turns33.3%52.9%6 turns28.6%24.8%14 turns33.3%36.3%9 turns
1514.3%24.3%7 turns66.7%42.9%6 turns16.7%22.9%6 turns25.0%37.8%4 turns
160.0%33.3%5 turns16.7%48.3%6 turns0.0%20.8%2 turns0.0%30.6%6 turns
170.0%13.2%3 turns14.3%40.8%7 turns25.0%21.5%4 turns60.0%31.9%5 turns
180.0%5.6%2 turns0.0%20.0%1 turns0.0%11.5%4 turns20.0%31.7%5 turns

Count the caller’s discards

How long a caller has gone since the call tells you more than the call. With one call, the caller was tenpai 7% of the time during their first six discards and more than half the time from their thirteenth. LuckyJ plays to that clock. From two-shanten or worse against one call, when a live tile left the hand at least as good as the safe tile, it kept cutting live tiles through the caller’s eleventh discard and from the twelfth threw the safe tile first. Against two or three calls it switched at the ninth. When the two tiles were exactly as good for the hand, it switched sooner, at the tenth and the seventh. When throwing the safe tile would have cost the hand acceptance, it still cut the live tile 72% of the time at a one-call caller’s tenth discard.

Toimen pons [[P]] and discards. Your hand is two-shanten, with a live [[N]] and an [[8s]] that Toimen has thrown. Neither tile helps the hand, so either one can go without costing it anything. On Toimen’s fifth discard the [[N]] goes: a single caller is almost never ready this early, and a live honor only gets harder to let go. By Toimen’s eleventh discard nothing has changed in your hand, and now the [[8s]] goes first while the [[N]] waits.

Both sides of the line matter. Before it, LuckyJ spent live tiles more freely than the humans, mostly honors and terminals that would only get harder to cut. After it, it spent fewer. From one-shanten its rates ran close to the humans’, so the lesson belongs to the far hand, whose win is too unlikely to pay for a deal-in.

Two tsumogiri in a row or a second call move the clock forward, as Against the humans, 1 describes, and a dora pon starts it at once (answer 2). From tenpai the question changes; answer 4 takes it up.

Evidence One caller, two-shanten or worse, a safe tile that kept the best shanten and a live tile that kept it with at least as much acceptance. The first version of this answer also counted turns where the safe tile was the best tile for the hand anyway (4,454 of LuckyJ’s) or the only kind of tile that kept the shanten (88), and put the line at the tenth and seventh discards. On the real choices, LuckyJ’s fitted rate of cutting the live tile fell below half between the caller’s eleventh and twelfth discards against one call (4,500 turns) and between the eighth and ninth against two or three (1,591). When the two tiles were exactly as good, it fell below half between the ninth and tenth (2,098) and between the sixth and seventh (700), and the humans’ rate crossed in the same places. When the safe tile cost acceptance, the rate was still 72% at the tenth discard against one call and 63% against two or three (2,402 and 891 turns). The caller was tenpai 6.7%, 23.5%, 39.5% and 53.6% of the time with one call; 23.8%, 45.9%, 62.1% and 71.8% with two; 63.7%, 79.9%, 88.2% and 89.8% with three. Mortal, on the same turns of LuckyJ’s games, cut live from a far hand 80.5% of the time when the caller was under 15% likely to be tenpai and 30.9% when over 75%; LuckyJ 86.0% and 29.7%, the Tokujou humans 77.8% and 47.2%.

Respect a dora pon at once

A caller who has ponned the dora has three dora in the open, so any yaku takes the hand to a mangan or close to it. Callers showing three dora won 10,377 points on average, more than twice the 4,716 of other callers, and a deal-in into one cost 9,653. LuckyJ does not wait for the clock against them. In the caller’s first nine discards it cut live tiles 44% of the time instead of 62%, and from two-shanten or worse 46% instead of 68%. The Tokujou humans barely moved.

The dora indicator is [[6p]]. Kamicha pons [[7p]] on their third discard. Your hand is two-shanten, with a live [[W]] and a [[2m]] that Kamicha has thrown. Against an ordinary early call the [[W]] would go now. Against this one LuckyJ more often throws the [[2m]] and holds the [[W]].

This early, the caller is only a little more likely to be ready, about 27% against 20%. What changes is the price: a tile that looked free when a deal-in cost 4,000 is no longer free at 9,700. The Houou players moved about halfway to LuckyJ, and Mortal about as far as they did.

Count red fives in the melds the same way. Three dora in view is what matters, and a pon or kan of the dora is how nearly every caller gets there. Point 12 is about reading which yaku an open hand still needs.

Evidence One caller in their first nine discards; one-shanten or further, a safe tile in hand that kept the same shanten. Live cuts against a caller showing three or more dora and against the rest: Tokujou 55.7% and 59.2%, Houou 49.3% and 58.5%, LuckyJ 44.1% and 62.2%; from two-shanten or worse 56.4% and 62.7%, 53.4% and 61.9%, 45.5% and 67.7%. Mortal, on LuckyJ’s games, 48.8% and 59.7%. Of the callers who showed three dora, 98.4% did it with a pon or kan of the dora. They were tenpai 27.2% of the time at that stage, the rest 20.1%.

Use the tile the caller just passed

A tile the caller let go by cannot deal into them until they discard again. If it was one of their winning tiles, passing it left them in temporary furiten; if it was not, it still is not. LuckyJ spends that window. When a live-looking tile that another player had thrown since the caller’s last discard kept its hand’s shape, and the caller was more than 30% likely to be ready by the clock above, LuckyJ cut it 59% of the time. The humans cut it 34 to 38% of the time.

Shimocha has chi’d twice and made eight discards. Since Shimocha’s last discard, Toimen has thrown [[6m]] and Kamicha [[W]], and Shimocha took neither. Your hand is one-shanten, and the [[6m]] in it is the tile you want to cut. On any other turn it would be a live tile against Shimocha; this turn it is as safe as genbutsu. Cut it now, and your genbutsu is still in hand next turn.

The window closes the moment the caller discards or calls, and it covers that caller only. Against anyone else at the table the tile is exactly as live as it looks.

Against a riichi the same pass lasts for the rest of the hand, since a riichi player who passes a winning tile stays in furiten. Point 13 is about knowing which opponent each safe tile answers.

Evidence One caller more than 30% likely to be tenpai by their calls and discards; one-shanten or further; a live-looking tile that another player had discarded since the caller’s last discard kept the best shanten. Cut 34.3% of the time by the Tokujou humans (1,270 spots), 37.9% by Houou players (3,896), 59.2% by LuckyJ (397); Mortal, on the same spots of LuckyJ’s games, 50.3%. Of the 24,095 such tiles cut against a caller, none dealt in.

Give up wait for safety late

Sometimes the widest tenpai needs a live tile and a safe tile keeps a narrower one. LuckyJ’s answer depends on how far into its own pond it is. In its first six discards it took the safe, narrower tenpai only 11% of the time, and Mortal less often still. From the seventh to the twelfth it went safe 37% of the time, and from the thirteenth 47%. The humans went safe less often: about 25% of the time in the second row and 35 to 38% in the third. Mortal went safe about as often as the humans, so the later switch is LuckyJ’s own habit.

Your fourteenth discard. Kamicha has two calls and has thrown two drawn tiles in a row. You hold [[2m]][[3m]][[4m]] [[5p]][[6p]][[7p]] [[7s]][[7s]][[8s]] [[5m]][[5m]] with a [[P]] pon. The [[7s]] is live against Kamicha, and the [[8s]] is Kamicha’s genbutsu. Cutting the [[7s]] keeps the [[6s]] and [[9s]] ryanmen; cutting the [[8s]] keeps a shanpon on [[7s]] and [[5m]], about half as many winning tiles. Here LuckyJ took the shanpon more often than not.

Three more things moved it. The more of the wait the safe discard kept, the more often it went safe: 22% of the time when it kept a third of the wait or less, 41% at about half, and 58% at two thirds or more. Value pulled the other way: with three han or more it went safe only 33% of the time. When the wide wait had no yaku for a ron, it went safe half the time, since a closed hand could still declare riichi for the yaku.

A closed hand has a third answer: declare. Facing a caller with nobody in riichi, LuckyJ declared 75% of its first tenpais even when the widest wait needed a live tile; the humans declared 62 to 69%. When it declared, it threw the live tile for the wide wait 77% of the time, and Mortal 84%. The safe, narrower riichi became common only when the safe tile kept two thirds of the wait or more, where Mortal took it about one time in three.

Keep the wide wait in a closed hand with a yaku: Mortal went safe there only 15% of the time, and LuckyJ in 30% of its 37 such spots. In the humans’ own hands the wide wait finished ahead in every row, but the players who went safe were more often facing two tells, so treat the third-row switch as a close call. Point 09 is the third-row count.

Evidence Tenpai decisions against callers, riichi declarations included, where every widest-wait discard was a live tile and a safe discard kept a narrower tenpai: 740 Tokujou, 2,269 Houou, 202 LuckyJ. Safe tenpai taken in the first, second and third rows: Tokujou 7.7%, 25.2%, 35.3%; Houou 22.7%, 24.9%, 38.0%; LuckyJ 11.1%, 36.8%, 47.4%; Mortal, on LuckyJ’s games, 6.2%, 25.9%, 34.2%. LuckyJ in the second and third rows, by the share of the wait kept: 21.6%, 41.4%, 57.5%; with three han or more 32.5%; with no yaku for a ron 50.0%; closed with a yaku 29.7% (Mortal 15.4%). The live tile was on a caller’s wait 1.9%, 4.6% and 10.7% of the time by row. The humans’ net per hand in the third row: wide wait +1,206, safe tenpai +717. First closed tenpais that could declare, with the same choice: 133 Tokujou, 419 Houou, 44 LuckyJ; declared 62.4%, 68.5%, 75.0%, Mortal 77.5%; with the live tile when declaring 86.6%, 82.3%, 77.4%, Mortal 83.5%.

Safe-tile timing

When LuckyJ starts keeping safe tiles.

When an opponent declares riichi, LuckyJ usually already holds a tile from that player’s river (Point 13). This section asks when it starts keeping such tiles.

The test is a choice between two leftovers of the same kind, single tiles with nothing nearby. Say the hand holds a [[9p]] that Kamicha has already discarded and a [[1s]] that nobody has. Neither tile helps the hand, so the one LuckyJ keeps shows whether it is saving the safe one.

How often LuckyJ kept the safe leftover

Two leftovers of the same kind, one already in an opponent’s river: the share of choices where LuckyJ kept that one and threw the other. Each dot is one discard, sized by its number of choices; the line is a fitted curve.

LuckyJ, per discard and fitted95% range of the fitNAGA, fittedbelow 50%, the safe tile goes first

Guest windsQuiet table

0% 50% 100% 1 3 6 9 12 15 18 LuckyJ’s own discard 50% between discards 2 and 3 18% 61%

TerminalsQuiet table, 1s and 9s

0% 50% 100% 1 3 6 9 12 15 18 LuckyJ’s own discard 74% 61%

Middle tilesQuiet table, 2 to 8

0% 50% 100% 1 3 6 9 12 15 18 LuckyJ’s own discard 50% between discards 6 and 7 78% 20%

Once someone threatensAfter a riichi or a second call, any kind

0% 50% 100% 1 3 6 9 12 15 18 LuckyJ’s own discard 39% 2%
Terminals and middle tiles are saved from the first discard, guest winds from the third. From about the seventh discard the safe middle tile starts to go first, and once someone threatens, every safe leftover does.
The numbers as a table

Each cell: LuckyJ, then NAGA, then the number of choices. Discards with no choices are blank.

DiscardGuest windsTerminalsMiddle tilesOnce someone threatens
117.0%61.1%282 choices75.2%62.0%222 choices80.0%42.9%20 choices50.0%0.0%4 choices
242.5%54.2%113 choices72.7%63.7%227 choices87.2%50.0%39 choices44.0%25.0%25 choices
359.6%64.6%52 choices64.8%60.0%88 choices67.4%56.0%95 choices39.5%26.5%38 choices
481.2%72.7%32 choices75.8%76.6%66 choices65.3%47.4%147 choices27.0%22.1%89 choices
581.2%71.4%16 choices75.0%64.1%36 choices53.1%43.4%130 choices24.5%18.6%106 choices
668.0%62.1%25 choices63.3%62.1%30 choices50.8%40.7%120 choices19.7%21.0%157 choices
786.7%80.0%15 choices58.8%52.6%17 choices56.4%50.0%110 choices21.3%20.2%188 choices
861.5%100.0%13 choices59.1%60.0%22 choices35.9%33.8%64 choices17.1%18.9%210 choices
950.0%40.0%6 choices66.7%66.7%12 choices43.9%42.9%41 choices12.1%12.4%247 choices
1050.0%66.7%4 choices71.4%62.5%7 choices34.5%39.3%29 choices11.5%12.9%217 choices
11100.0%75.0%3 choices40.0%40.0%5 choices26.3%17.4%19 choices10.4%10.9%241 choices
120.0%0.0%3 choices20.0%20.0%5 choices40.0%41.7%10 choices8.1%6.9%210 choices
130.0%1 choices50.0%50.0%2 choices13.3%18.8%15 choices4.4%4.6%203 choices
1416.7%16.7%6 choices6.0%5.7%183 choices
150.0%0.0%1 choices0.0%0.0%1 choices0.0%14.3%5 choices5.3%6.3%152 choices
160.0%0.0%2 choices0.0%0.0%8 choices5.7%6.7%105 choices
170.0%0.0%3 choices0.0%1.0%95 choices
180.0%0.0%3 choices3.2%0.0%62 choices

Terminals and middle tiles are saved from the first discard. In the example, LuckyJ usually throws the [[1s]] and keeps the [[9p]]. NAGA does the same with terminals, a little less firmly. With two middle tiles NAGA starts near even and throws the safe one first from about the fourth discard.

Guest winds are the exception. Given a [[W]] that someone has discarded and a [[N]] that nobody has, LuckyJ usually throws the [[W]] on its first two discards, while NAGA leans the other way. From the third discard LuckyJ keeps the safe wind as well. The discarded wind has one copy fewer left to make a pair, and a live guest wind stays cheap to throw for a few more turns.

The saving stops in two ways. The first is a threat, and threats arrive in the middle of the hand, as the chart below shows. Once someone has declared riichi or called twice, the safe leftover goes first whatever its kind, and NAGA agrees. The second needs no threat. From about the seventh discard LuckyJ throws the safe middle tile first more often than not, since a live middle tile is what a silent tenpai most often waits on.

When threats arrive

Share of LuckyJ’s decisions facing a riichi or a hand with two calls, by its own discard

0% 50% 100% 1 3 6 9 12 15 18 LuckyJ’s own discard 8% 30% 52%
At its fourth discard, 8% of LuckyJ’s decisions face a threat. By its tenth, more than half do.
The numbers as a table
DiscardFacing a threatDecisions
10.2%9,449
21.3%9,250
33.9%9,128
48.1%9,042
513.9%8,835
621.4%8,556
729.8%8,062
837.6%7,413
945.0%6,669
1052.4%5,816
1159.9%4,974
1263.9%4,192
1369.0%3,451
1473.2%2,822
1576.5%2,246
1678.8%1,768
1781.2%1,415
1879.1%851
Rule Between two leftovers of the same kind, throw the live one and keep the one already in a river: from your first discard with terminals and middle tiles, from your third with guest winds. Once someone declares riichi or calls twice, throw the safe one first. From about your seventh discard, start throwing a safe middle tile first even when nobody has.
Exception This is only for leftovers with no other job. A tile next to the dora, a guest wind you want for seven pairs, or a tile that would finish a block outranks either label. Which opponent the kept tile answers is the question of Point 13.

Evidence LuckyJ’s 9,745 hands as a child in the 1,255 NAGA-reviewed games, its own discards 1 to 18, before any riichi of its own. A leftover is a single tile with no pair and nothing of its suit within two ranks. Guest winds leave out the dragons, the seat wind and the round wind. Safe means already in at least one opponent’s river. The three tile panels of the first chart count quiet tables only, with nobody in riichi and no opponent on two calls. Each dot is one discard, sized by its number of choices. Each curve is a logistic fit on a smooth function of the discard number, drawn only over discards with at least 10 choices each, and the shaded band is its 95% range. The tables under the charts and the method notes give every rate with its sample.

Pattern synthesis

What the supporting hand reviews add.

The supporting statistical and hand reviews look at LuckyJ from several angles: every East 1 start, placement-specific hands, last-place-avoidance games, and comparisons against NAGA and Mortal. They shaped the commentary, but the table logic still decides each hand. What they add up to is that LuckyJ's style is a set of switches. Hand value, hand distance, number of calls, river texture, round, and placement all change what the same tile means.

East 1 Reveals the Base Settings

With scores level and no placement pressure, LuckyJ's default target is a hand worth pushing. As a child it wants either a good-shape one-han riichi route or a hand that can reach three han or better. As dealer the value threshold drops, because the dealer turn itself is worth more. Hands below those thresholds start preserving yaku, dora, and safety instead.

Hold Safety by Shanten Count

The farther the hand is from tenpai, the more a safe tile is worth. A far hand can keep honors or visible safe material because those tiles later buy the right to call, push one more tile, or reach a safe tenpai. Being dealer or holding a very high-value hand lowers that premium.

Weak Penchans Can Be Broken

LuckyJ will break a penchan or another bad block when its best future is a riichi-only bad-shape hand. The tiles it keeps instead usually have a second job: a yakuhai pair, a honitsu or chanta route, dora use, or defense. Copy this when the bad block mostly leads to a hand you would rather not win.

Read the River Shape Around the Tile

Several reviewed hands show how much LuckyJ reacts to discard patterns and calls. It will push into one open hand, match-discard live honors early, or stop on an ordinary-looking tile when a river looks wrong. A static danger chart misses that context. The same 6m is a different tile against a normal river, a honitsu river, and a riichi declaration sequence.

Last-Place Attack Still Has Brakes

When it is fighting out of fourth, LuckyJ uses riichi pressure, keiten calls, and value upgrades more freely. The same review hands show it refusing calls or discards that would leave no safe tile after the next dangerous draw. Buy pressure when the opponent has to respect your hand, and keep one exit.

Third Row Is the Training Zone

The main warning is that LuckyJ looks less reliable in the third discard row. Sometimes it gives up tenpai equity too readily. Sometimes it is strained by multiple threats. Treat the late endgame as a separate skill. When NAGA, Mortal, and your own point calculation all disagree with a late LuckyJ move, file it as a review case rather than a lesson.

Danger Labels Need Live Context

Suji, one-chance, kansuuken, genbutsu, and early-riichi heuristics are starting points, not verdicts. Before calling a tile safe or dangerous, update the label with visible blockers, the opponent's route constraints, suit pressure, bad-shape concentration, and your own next discard.

Route Value Beats Surface Shape

The best-looking wait still has to beat the whole route. Wait shape competes with yaku certainty, value ceiling, ron likelihood, callability, future danger, and placement need. Shape is one price inside that route, not the route itself.

Inventory Comes Before Commitment

Riichi freezes your discards, so take inventory before declaring: what value you win, which dangerous tiles you are accepting, and what flexibility you lose. Calling or pushing through danger still leaves later choices, so every leftover tile needs a job: value seed, future block, safe tile, denial tile, or liability.

Synthesis rule LuckyJ's odd-looking tiles usually have a job. Ask which switch changed: value threshold, distance versus safety, bad-block quality, opponent river, wait count, number of threats, or placement pressure.
Evidence addendum Treat ideas from the qualitative reviews as claims to test against the replay data. The ones that held up in general form are live-context danger reading, route over shape, inventory before riichi, and continued role labeling after calls and push decisions.
01

Point one of sixteen

Put Placement Before Ukeire

Look at the scores before you count acceptance. Tile efficiency tells you which discard improves the shape. It does not tell you whether a 1,000-point win changes anything, whether 5,200 moves you up a place, or whether pushing for 8,000 mostly risks handing first to someone else. In the South rounds the same two-shanten hand can be a hand to finish cheaply, a hand to build toward mangan, or a hand to drop the moment a dangerous tile appears.

LuckyJ's good games were not the games where it won the most hands. They were the games where it converted the right hands and declined the wrong losses. In its top-half games it dealt in about half as often as in its bottom-half games, and that gap is where placement came from.

When LuckyJ leads and two discards are close in shape, it takes the one that leaves an easier next turn and avoids a pointless live middle tile. That is a tie-break, not a rule to choke every leading hand. A leading hand with a real route still plays for the win.

Urgency reverses the tie-break. From third or fourth, a 1,000-point hand that leaves the order unchanged is worth very little, and protecting shape only slows the loss. If the order needs 5,200 or 8,000, keep the tile that keeps riichi, dora, or mangan alive, even if it is not the safest one. When a cheap win holds first, do the opposite: do not spend danger on a 5,200 you do not need.

Rule Before counting ukeire, decide what this hanchan needs from the hand: a cheap finish, a bigger hand, or a fold. Then discard for that.
Exception When the scores are flat and one discard is clearly better for shape with no real danger gap, take the shape. When only 5,200 or 8,000 changes the order, a timid 1,000-point line is not enough.

Evidence Over 1,255 games and 13,054 hands, LuckyJ averaged a placement of about 2.27, won about 23% of hands, and dealt in on about 10%. Top-half games averaged 3.09 wins and 0.75 deal-ins per game; bottom-half games 1.39 and 1.48. The leader tie-break shows up across a little over five thousand leader spots. Each replay below has at least one NAGA head on LuckyJ's side. Mortal backs the three seats that needed something from the hand and rejects the leader's tie-break.

02

Point two of sixteen

Do Not Choose the Hand Too Early

In the first row, the discard with the most acceptance is usually the best discard for one version of the hand. LuckyJ tends to keep the tile that leaves several versions open: a closed riichi hand, an open yaku hand, a dora hand, or a clean fold. That costs a little acceptance now and buys the right to decide later, once the draws and rivers have shown which hand this is.

Take [[2m]][[2m]][[3m]][[4m]], a [[P]][[P]] pair, red [[5pr]], and a weak [[7s]][[9s]] kanchan. Cutting the dragon pair because it is “just an honor” smooths the shape a little. But that pair is also a yaku, permission to call, extra value, and a brake if the table turns. The kanchan is the block that needs a perfect draw to earn its place. If something has to go, it goes first.

Keeping options open is not a goal by itself. Once the hand is a clean two-ryanmen one-shanten, or the anchor no longer makes yaku, value, safety, or a call, cut it and take the efficiency. The mistake this chapter is about is breaking the pair or dropping the dora branch before you know whether the hand needs speed, value, or an exit.

Rule Early, prefer the tile with two jobs, shape plus yaku, value, a call, or safety, over the tile that only improves one narrow version of the hand.
Exception Take the efficient discard when the hand is already a clean one-shanten, when the anchor has no yaku or value, or when keeping it wrecks a strong five-block structure.

Evidence In over 2,300 pair-or-triplet anchor spots and over 600 dora or red-five spots, LuckyJ's splits from Nishiki mostly kept the tile with a second job, and those keeps drew fewer severe disagreements than the average split. LuckyJ's disagreement with NAGA is widest in the first row, about 45%, and drops to about 28% in the third row.

03

Point three of sixteen

Call Only When the Call Changes the Hand

Opening a bad shape is not automatically giving up value. The test is whether the call changes what the hand can do: win, reach tenpai, gain a turn, or end the round. A [[P]] pon that gives you a yaku and a 1,000-point finish before the dealer's riichi arrives can beat waiting for a closed hand worth two more han that may never come. A chi that turns a bad two-shanten into an exposed bad two-shanten pays the closed premium and gets nothing back.

LuckyJ calls a lot, about one hand in three. It also passes a lot of calls that NAGA would take. The lesson is selection. Passing everything and taking everything both lose to opening on purpose.

Before you take the tile, picture the hand after the call. Does it have a yaku now? Is it tenpai, or a real one-shanten? Did the call take a draw from the dealer, or open a path to a safe keiten later? All of that shows up in the next discard. If you cannot pick the post-call discard cleanly, or you cannot say which safe tile you are keeping for the first threat, the call was about wanting to do something.

A good closed hand should be expensive to open. Pinfu with dora and good shape is often a 3,900 hand or better once riichi and ura are counted. Opening that into a fragile 1,000-point hand with no safe tile is impatience.

Rule Call when the call gives you a yaku, tenpai, a real one-shanten, a denied draw, or a safe path to keiten, and that is worth more than the closed premium you give up.
Exception Pass when the closed hand has good value and shape, or when the call would leave you open, cheap, still far from tenpai, and without a safe discard for the next threat.

Evidence About 36% of LuckyJ's hands involved a call, 5.65 calls per hanchan on average, and it passed roughly 2,000 calls that NAGA preferred. Top-half games had 5.80 calls and 2.11 riichi per game against 5.44 and 1.67 in bottom-half games.

04

Point four of sixteen

Keep Brakes on Open Hands

Calling does not remove defense from the hand. It makes defense more expensive. After a call the hand is shorter, the closed value is gone, and there are fewer draws left to find a safe tile by luck. Keeping a genbutsu or a real suji against the live threat is what lets you keep attacking without promising to throw the next live [[5p]] into a riichi.

Cheap open hands show this best. A 1,000-point yakuhai hand still two-shanten cannot spend every safe tile for a few more acceptances, because the push it is setting up may run into 3,900, 8,000, or a dealer. A 3,900 tenpai with a good wait can afford the same discard. Value and distance decide whether a safe tile is fuel or dead weight.

After every pon or chi, ask which player the saved tile answers. Safe against the quiet player is usually dead weight. Safe against the riichi, the dealer, or the open honitsu seat is a resource. When the open hand is already a strong tenpai, or the table is quiet and the value is enough, spend the spare safety. The mistake is spending it by reflex because the hand is now “attacking.”

One saved genbutsu buys one turn, not a permanent fold. After you spend it, the next draw is a new decision. A new genbutsu or a low-danger tile that keeps the one-shanten lets you keep going by mawashi, folding one tile at a time while the shape survives. A draw that makes a good tenpai can justify a fresh push. If the only tile that keeps the shape is live against the threat, mawashi is over: push that tile on the value of the whole route, or break the hand and fold.

Placement changes the price on both sides. When you are fourth and the riichi dealer is third, your win also ends the dealer turn, collects the riichi stick, and stops the gap growing, so a strong one-shanten is worth more to keep. Dealing straight into that dealer is also the worst possible transfer. “Stop the dealer” supports one tolerable continuation when the winning route is real. It does not rescue a cheap hand that still owes a string of live tiles.

Rule After every call, know which tile lets the open hand stop or keep going safely. Treat the last genbutsu as one turn of cover, and decide again from zero once it is gone. If no tolerable discard keeps the hand, mawashi is over.
Exception Spend the safe tile when the hand is already a high-value tenpai, when the tile answers the wrong opponent, or when placement requires winning now. Fourth place against a third-place dealer justifies one tolerable push from a strong one-shanten, not a series of live tiles.

Evidence Across 6,021 LuckyJ/Nishiki open-hand splits, LuckyJ kept a recognized defensive tile in 3,536, and 1,723 of those were tied to a live threat. These are counts of choices, not proof that keeping the tile improved the result.

05

Point five of sixteen

Cut the Tile That Will Get Worse

A floater has two prices: what it accepts now, and what it costs to throw later. Most players count only the first. LuckyJ often pays the second while the river is still quiet, cutting a useful-looking middle tile before it becomes the tile you have to push through in the third row.

Take a slow hand with [[2m]][[2m]], [[6p]][[8p]], [[3s]][[4s]], a lone [[5p]], [[P]][[P]], and loose honors. If [[5p]] is not dora and pinzu is not where the value is, keeping it for the [[4p]] and [[6p]] draws gains little. By the time a riichi comes, that [[5p]] may be the only tile in the hand nobody has made safe.

This is one of the clearest habits in the review, and also one where the engines disagree with each other, so do not defend any single example as settled. The habit to copy is simple: when the hand is slow and a tile has no concrete job, price what it will cost to discard later.

Pre-cutting a middle tile is wrong when that tile is the hand. A red [[5p]], a dora connector, the tile that makes a perfect one-shanten, or the only tile that makes mangan stays. Cut the future problem when it is a floater. Keep it when it is the reason the hand can fight.

Rule In a slow hand, discard the floater whose danger will rise fastest, unless it is needed for value, yaku, or your best one-shanten shape.
Exception Do not use this to throw dora, red fives, your strongest blocks, or the tile that makes the hand pushable. Skip it when the table is already dangerous and the floater is safer than your other options right now.

Evidence In over 3,700 spots where LuckyJ chose a lower-danger tile than Nishiki, the choice drew fewer severe disagreements than the average split. Mortal backs LuckyJ in all four replays below.

06

Point six of sixteen

Buy Value While It Is Cheap

Value planning starts before the hand looks valuable. The first discard is usually the cheapest moment to keep a red-five connection, a dora shape, or a yakuhai pair, because the hand can still reorganize around them. Wait until one-shanten to ask for 5,200 and the price is usually a block, a wait, or a dangerous tile.

A fast 1,000 or 2,000 is right when it ends a dealer turn or protects first. It is wrong when South 3 needs a mangan swing. Keeping the red five or the dora shape can turn the same skeleton into 5,200 or 8,000, and that changes whether a later riichi, call, or push is worth making. As dealer the bar moves again, because dealer hands pay one and a half times.

The trap is fake value. An isolated dora that does not connect, a yakuhai singleton you cannot use, and a honitsu plan that asks you to throw live middle tiles are not value planning. If the realistic ceiling is still 1,000 after you keep the tile, cut it and build the best hand you actually have. Feeding a dealer mangan while chasing your own is bad arithmetic.

Rule Keep dora, red-five, and yakuhai value early while it still improves a real winning line. Do not buy value with isolated danger or decorative yaku.
Exception Cut the value tile when nothing supports it, when it is already too dangerous, or when keeping it breaks the hand's best shape. Being behind raises the value you need. It does not make every risky discard profitable.

Evidence In over six hundred dora and red-five retention spots, LuckyJ kept value material that still belonged to the hand, and those keeps held up well against Nishiki. The broader idea of buying risk because you are behind did not get the same support: the value trades below each have two NAGA heads on LuckyJ's side; Mortal backs two of the four, takes Nishiki's plainer line in one and a third tile in the last.

07

Point seven of sixteen

Make Riichi Tax the Table

Riichi is pressure as much as it is a han. The declaration adds ura chances, locks the hand, and often makes opponents fold. A quiet 2,000-point dama and a riichi hand that threatens 3,900 to 8,000 play very differently. The question is not whether you can reach. It is whether reaching makes the table pay.

A good riichi changes what opponents do. Being first to threaten against quiet rivers can make a leader drop a 1,000-point open hand, freeze a dealer's two-shanten, or win you the hand by folding everyone out. It also moves your own bar. A bad 1,300 reach into two live threats is usually a 1,000-point stick plus several forced danger tiles. A 5,200-class reach with a live wait can be worth the lock, because winning and making everyone fold both pay.

Riichi stops being the answer when it stops solving the problem in front of you. If dama already passes the rival in the last hand, or declaring lets everyone fold when you needed direct points from one seat, stay quiet. If two opponents are already live and your wait is poor, most of the pressure lands on you.

Rule Declare riichi when the added value and the forced reactions change what the hand is worth. Stay quiet when the stick and the lock only buy a bad wait with no table pressure.
Exception Do not riichi because the button is there. Dama is better when it already delivers the placement swing, when one specific opponent has to be hit directly, or when several live threats make a low-value bad wait too expensive.

Evidence LuckyJ averaged about 1.9 riichi per hanchan, about 2.11 in top-half games against 1.67 in bottom-half games. Mortal agrees with declaring in all five riichi below. That verdict covers the declaration only, not LuckyJ's declaration discard or the wait it chose.

08

Point eight of sixteen

Reprice the Whole Push Every Draw

A push is not one tile. It is every discard you are committing to make before the hand wins. Price that commitment again after every draw, call, riichi, and new safe tile. The common leak is deciding “this is my pushing hand” once and then treating each later discard as if the table had not moved.

The second and third dangerous tiles in the chain usually matter more than the first one you are staring at. Paying 1,500 at a draw can be cheaper than buying one more live turn that ends at -4,200. A hand with a realistic +8,900 finish can afford a much larger toll. The question is never whether this one tile can pass. It is whether the hand's reward covers every unsafe discard still to come.

Some hands stay on push after all that. A high-value tenpai, a riichi whose pressure and value matter, or a last-hand score need can justify a dangerous tile even after the table gets worse. That is a new decision each time, not momentum from last turn. If the draw added another live middle tile to the path ahead, the old push no longer exists.

The hands below are LuckyJ's most extreme pushes as a non-dealer, every one against at least one riichi. Grade each by the state before the tile was cut, not by the result: value, distance, placement, wall, safe alternatives, and the dangerous discards still ahead.

  • Closed monster tenpai. [[3p]] at 99.99%, every NAGA head agreed, won 10 han 60 fu for +17,000. The clean case.
  • Third place, riichi into riichi. [[1m]] at 95.74% to declare, two heads agreed, won +5,900.
  • Fourth with 2,900 points, still one-shanten. [[5s]] at 99.99% to keep 13 effective tiles instead of nine. One head agreed. Won +6,500, which shows what was bought, not that the price was right.
  • Three melds, five tiles in hand. [[6s]] at 99.99% to keep tenpai. All three heads agreed. Dealt in for -2,900. No riichi locked this hand; the earlier calls had removed the exits.
  • First place with 41,000, five tiles in the wall. [[5m]] at 99.99% to keep an open tenpai while all three heads preferred [[7s]] at 2.97%. Dealt in for -12,000. An overpush, not a model.
Rule After every draw, call, riichi, and new safe tile, count the dangerous discards still ahead. Push only when the whole chain is worth more than folding now.
Exception Keep pushing when the hand can already pay the full toll: a high-value tenpai, a meaningful pressure riichi, or a placement-critical hand that cannot fold. Even then, price the whole sequence again instead of defending last turn's decision.

Evidence The extreme-danger sample is nondealer-only: after removing every hand where LuckyJ was dealer and every locked discard after its own riichi, the Tokujou sample holds 46 chosen discards at 90% or more on the maximum NAGA danger proxy. All 46 faced a riichi; 33 were tenpai, 12 one-shanten, one farther; 20 came from open hands; at least two of three NAGA heads supported 30. Outcomes: 12 self-wins, 11 direct deal-ins, 14 wins by another player, nine draws. Repeated 99.99% readings behave like a capped alarm, not a deal-in probability. Across the wider review, LuckyJ's lower-danger choices (over 3,700) drew far fewer severe disagreements than its higher-danger choices (over 5,000).

09

Point nine of sixteen

The Third Row Is a Counting Drill

The late row is a separate skill, not the opening with fewer tiles. Count three things: the winning tenpai, the safe tenpai, and the fold. By the third row most big plans have expired, and the decisions that remain are smaller, colder, and more exact.

Keiten shows the arithmetic best. LuckyJ was tenpai at a little under half of its exhaustive draws. Noten payments are real value. They do not justify sending a live middle tile through two threats with no winning upside.

Take a third-row spot. You can cut a live [[5s]] for pinfu tenpai on [[3p]][[6p]], cut genbutsu [[9s]] and stay one-shanten, or follow with [[E]] and fold. Do not ask which tile looks best. Ask what the win changes, what tenpai at the draw is worth, how many safe tiles remain after this discard, and whether the next tile is also dangerous. If you have no answer for the next turn, the current push is not a push.

Late precision is not folding everything. If the hand is a dealer mangan tenpai, the wait is alive, and the only discard is a reasonable suji, folding because “it is late” gives away too much. The discipline is separating a real attack from a vague hope: win, take safe tenpai, or fold cleanly.

Rule In the third row, choose among three outcomes only: winning tenpai, safe tenpai for the draw, or a real fold.
Exception Break the rule only when placement sets a clear special goal, such as a forced last-place escape or a guaranteed-top fold. If every tenpai line needs a live center tile against strong threats and the win does not change the table, drop the hand.

Evidence LuckyJ disagreed with NAGA on about 45% of early decisions and about 28% of late ones. In the recorded exhaustive draws it was tenpai in 823 of 1,761 (46.7%); that is 823 of all 13,054 hands.

10

Point ten of sixteen

Treat Keiten as Attack

A drawn hand is not a lost hand. Keiten is a small attack that starts once winning has become unlikely, and the payments are big enough to plan around: 1,000 when several players are tenpai, up to 3,000 when you are the only one. In South rounds that is often the same swing as a small win, because 1,500 or 3,000 changes who has to chase next hand.

Harmless is the whole condition. If genbutsu, dead honors, or visibly exhausted tiles keep you tenpai, the payment belongs in your attack calculation. If the price is a live middle tile into a dealer riichi, you are no longer buying keiten. You are risking a dealer mangan to avoid a small payment. A player who never chases draw points leaves money on the table. A player who chases them through danger has only renamed noten fear.

Forced danger ends the chase. When the only tenpai path passes two unsafe tiles, or an open hand's value has become clear, fold and take the noten. You can still fold actively by choosing the safe discard that leaves the most safe draws for later. Draw points are attack value only while the defense is intact.

Rule In late hands, chase keiten while safe tiles keep you tenpai. Drop it the moment the next required discard carries real deal-in risk.
Exception Do not chase draw tenpai through live danger. If the tenpai path needs unsafe middle tiles against a riichi, dealer pressure, or a clearly expensive open hand, take the safe noten and protect the hanchan.

Evidence LuckyJ reached tenpai in 46.7% of genuine exhaustive draws. Abortive draws are excluded, and the share of all hands is about 6.3%.

11

Point eleven of sixteen

Make Engine Splits Prove Their Case

An engine disagreement is a review prompt, not a verdict. LuckyJ matches NAGA's top discard about three times in four, so the split hands are where this book has something to teach. Not because LuckyJ is always right, but because the split forces you to write down what the move is buying.

Sort each split into one of three boxes. Blunder: the move fails basic value, safety, or yaku accounting, and you correct it. Style trade-off: LuckyJ is buying safety, pressure, or placement that you can explain but should not copy automatically. Genuine insight: the split shows a conditional habit you can reproduce, such as valuing a future safe exit over a little acceptance.

Point values keep the review honest. Copying a split blindly can become a -3,900 deal-in. Flattening every LuckyJ deviation back to NAGA can cost a +2,900 win or a +1,500 draw payment. The number is not proof. It is a check: what did the move risk, what did it buy, and could a human at the table have seen that?

When your explanation is thin and the stronger engines all point the other way, mark the hand uncertain and move on. The goal is not a higher match rate with any engine. It is copying only the decisions you can explain before the result is known.

Rule When LuckyJ and an engine split, sort the hand as blunder, style trade-off, or genuine insight. Copy only the line whose value, risk, and table context you can explain.
Exception Treat the split as a warning, not permission, when your explanation is thin or the engines broadly agree against the move. Fix the ordinary mistake first.

Evidence The full review holds over 36,000 LuckyJ/Nishiki mismatches from about 135,000 split-head decisions, most of them minor. The NAGA heads often disagree among themselves, and Mortal often backs Nishiki or picks a third line, so a disagreement is rarely engine consensus.

12

Point twelve of sixteen

Price the Missing Yakuhai

When an opponent opens and you cannot see their yaku, a live dragon, the round wind, or their seat wind stops being an ordinary honor. It may be the tile that makes their hand legal. Discarding it offers them the pon. Holding it withholds that chance for now. The decision is about timing, not cleanup.

There are two deliberate lines. Release the singleton while no yaku and no clear progress are visible, accepting today's pon risk so you do not carry a worse live tile into a later turn. Once progress shows, through more calls, a late turn, or a tenpai-looking river, hold the tile unless your own value, safety, or placement gives a concrete reason to feed it.

Do not read this as a blanket rule. Compare the pon risk now against the danger of carrying the tile, and keep the tile when it is your pair, your value, your only safe tile, or a real honitsu or chiitoi piece.

Rule After an unclear call, mark live dragons, the round wind, and that player's seat wind as possible yaku tiles. Release one early only while no yaku or clear progress is visible. Once progress is visible, hold it. Discarding is not denial.
Exception Keep or even feed the yakuhai when it serves your own value or defense, when a bigger riichi makes it your only safe tile, or in South 4 when a cheap win by that specific opponent is the result you want.

Evidence Tiles eventually cut against an unknown open hand left around turn 8, against tanyao-shaped opens around turn 10, and around turn 7 when the yaku was already visible. Each figure is conditioned on the tile surviving until that context and does not show that the call caused LuckyJ to wait. Each replay below has one NAGA head on LuckyJ's side; Mortal backs two of the four and favors Nishiki in the other two.

13

Point thirteen of sixteen

Name Who Your Safe Tile Defends

A safe tile is safe against a particular player. Genbutsu to a quiet child does nothing against a dealer riichi. Suji to an old river matters only if that player can still punish you. LuckyJ's defensive habit is keeping a tile when you can say which opponent it answers, not hoarding safe tiles.

Put prices on it. Cutting a slightly useful shape tile now costs some speed. If the kept genbutsu prevents one -4,000 turn, it has paid for that many times over. On the other side, a -500 nuisance loss is no reason to spend your only real answer to the dealer. Genbutsu is worth the most, suji less, and both need an opponent attached before they earn a place in the hand.

Most of these decisions come before anyone declares, so the question is who is most likely to attack next. Before a declaration, go by seat: the dealer riichis more often and costs more when you deal in, so the dealer is the default name. A player's whole earlier river becomes safe against them the moment they riichi, which is why the tile you hold now for the dealer usually still works after the declaration. Once someone declares, the declared player replaces the dealer as the name.

The label can be empty. When the table is quiet, the hand is valuable and fast, or the tile is only weak suji to someone who is not threatening you, do not tax the hand for safety that answers nobody. That tile is just a worse floater.

Rule Keep a genbutsu or credible suji only when you can say which opponent it answers and which bad turn it covers.
Exception Skip the keep when the opponent is quiet, the safety is weak or aimed at the wrong player, or your hand is fast and valuable enough that the shape cost exceeds the defensive value.

Evidence Across about 13,000 hands LuckyJ won 22.9% and dealt in on 10.1%; top-half games averaged 0.75 deal-ins per game against 1.48 in bottom-half games. In child rounds LuckyJ already held a tile from the future riichi player's river at 74.5% of 5,504 opponent declarations, and about a third of its first reaction discards were such a tile. The dealer riichied in 21.0% of child kyoku against 17.7% for each child, and a dealer deal-in cost 6,700 points on average against 4,700; early river features moved the riichi rate by only about 3 points. LuckyJ's own quiet-table discards show no dealer bias, 13.3% per turn for dealer-river tiles against 12.4% for child-river tiles, so the dealer preference is a price rule rather than a copied habit.

Safe tiles and push-fold

Safe tiles are timing tools.

The follow-up LuckyJ, NAGA, and Mortal review narrowed the safe-tile idea. Price each safe tile by its target, its timing, and what the hand gives up by keeping it.

14

Point fourteen of sixteen

Spend Safety When Its Opponent Disappears

A tile can be genbutsu and still be the right discard. The question is not whether it was safe last turn. It is whether it protects you from the danger on the table now. When another player becomes the real threat, or the hand gets close enough that the tile blocks tenpai, yesterday's safety has expired.

The cost is easy to see. If an old safe tile kills a live 2,000-point tenpai chance, you have paid real points for insurance that may never pay out. If the alternative is drifting into a -1,500 noten draw, the cost is visible without any deal-in. A stale tile that answers a quiet player is shape damage, not defense.

Do not throw away your last real exit. If the tile still covers the live riichi, your hand is far, or the next draw may leave you with nothing safe, keep it. Spending stale safety is disciplined only after you can say why that tile no longer answers the dangerous player.

Rule Whenever the main danger changes, ask who the safe tile still protects you from. If the answer is nobody relevant and it damages tenpai, spend it.
Exception Protect the last real safe tile when it still covers the live threat, your hand is far from winning, or the next draw could force a dangerous discard with no backup. Do not call a tile stale because you want more acceptance.

Evidence In the model-proxy audit, threat-specific keeps and spending stale or off-target safety were both less often severe Nishiki disagreements, while generic safe-tile retention was more often rejected. That is a review signal, not an outcome measure.

Shape and route selection

Outside cuts can be shape decisions.

15

Point fifteen of sixteen

The Edge Tile Can Be the Attack

Not every terminal cut is defensive. Sometimes the outside tile is the price of keeping the center of the hand alive. A late [[1p]] or [[9s]] can be the attacking discard when it protects the middle connector that actually reaches tenpai, and it lowers the danger of the discard at the same time.

Take [[1p]][[3p]][[4p]][[6p]][[7p]] in a hand that already has value. Cutting [[1p]] leaves [[3p]][[4p]] and [[6p]][[7p]], two normal ryanmen cores. Cutting [[4p]] because [[1p]] looks safer keeps a dead-looking outside tile and damages the part of the hand that wins. In five-block terms, the terminal is not a block. The middle tiles are.

Terminals are not free discards. A [[1m]] pair can be chiitoi material, a [[9p]] can be dora, and an outside tile may be your only genbutsu against a dealer riichi. When the outside tile is value, a pair, a yaku, or safety against a specific player, it is no longer the cheap tile. The principle applies when the edge really is the cost of keeping the winning center.

Rule When choosing between an outside tile and a middle connector, keep the connector if it is the tile that makes your real tenpai or win possible.
Exception Do not cut the outside tile when it is dora, a valuable pair, part of chanta, ittsu, or sanshoku, or your only safe tile against the live threat. Drop the idea when the middle connector is already dead, redundant, or too dangerous to use next.

Evidence In over 5,000 spots LuckyJ cut outside material while keeping the inside line, and the reverse pattern appeared in over 10,000 spots. These are counts of model splits, not proof that the outside cut is better. Mortal support in the showcase is mixed.

Calls, honors, and opponent yaku conditions

Honor tiles need role labels before you judge them.

16

Point sixteen of sixteen

Label Every Honor Before You Move It

“Honor tile” is not a strategy. A dragon pair in your hand, a live dragon for an open opponent, a wind with two copies visible, and a genbutsu honor against riichi are four different tiles. Treat them all as “honors” and you will cut value tiles too early and keep dangerous tiles too long. The first step is naming what the tile is doing right now.

Use four labels. Self value: your yakuhai pair, honitsu material, a chiitoi pair, or a dora-connected plan. Opponent's yaku: live dragons and value winds that could complete an unclear open hand. Dead tile: an honor with little value and enough copies visible that nobody can use it. Defense: genbutsu, safety against several players, or a light tile kept for the first riichi. A self-value dragon can make 1,000 by itself and 3,900 with dora. A dragon that is someone else's yaku can cost you the round.

Labels expire. A lone guest wind is fine early and gets worse late against chanta, kokushi, or toitoi shadows. A dead honor becomes precious when it is your only safe tile against a dealer riichi. Relabel after every call, riichi, visible copy, and score change, then cut or keep the honor for the role it has now.

Rule Before cutting or keeping an honor, label it: self value, opponent's yaku, dead tile, or defense. Then play the rule for that label.
Exception Do not play an old label after the table changes. Keep a “loose” honor when it becomes your only safe tile, and cut a “safe” honor when it becomes an opponent's live yaku tile.

Evidence Loose self-yakuhai with no open hand around were cleared at a median of turn 4; in tanyao-shaped open contexts they waited until around turn 10. That is a role change, not a contradiction.

Practice system

How to train this style

Turn-one job naming

Before every first discard, write one phrase: quick end, value target, open rush, safety hedge, or last-place attack.

Three-route bad-hand drill

For every bad hand, mark the best tile for tanyao, yakuhai, and safety. Discard the tile that contributes least across all three.

Dangerous floater audit

On turn five, find the tile you least want to discard into a riichi. If your hand is slow, cut it now.

Call purpose

Every call needs a stated purpose: tenpai, a real one-shanten, yaku, an ippatsu break, or denial. No purpose, no call.

Opponent yaku-condition audit

After every opponent call, mark dragons, round wind, and that player's seat wind. When their yaku is unclear, clean loose yakuhai before it becomes their yaku condition.

River-safe inventory

Before the middle row, mark every tile in your hand that is already in an opponent river or suji to it. Spend the last safe tile deliberately.

Future push path

Before pushing, name every dangerous tile that must pass before you win. If there are two or more, downgrade the hand sharply.

Draw-tenpai defense

When folding, prefer safe discards that preserve harmless tenpai routes. Count noten-bappu as a real objective.

Penchan autopsy

When you break a penchan, write why: bad riichi-only future, value route, safety route, or river danger. If none applies, keep the block.

Third-row replay

Review every third-row push or fold with NAGA, Mortal, and your own point calculation. This is where copying LuckyJ without checking costs the most.

Method notes

How to read the examples

The examples come from LuckyJ's games and the AI review output, translated back into ordinary riichi ideas. Use the numbers to find interesting hands. Use the table context to decide whether the idea belongs in your own game.

Show audit data

This is the evidence behind each point, for readers who want to check why it stayed in the playbook. The main text stands on its own without it.