What the Japanese Community Knew

Why this page exists

Most of this material lived on Yahoo! GeoCities Japan, which shut down in 2019, and on 2ch (now 5ch). Most of it is dead. What survives, survives in web archives — and archives rot too.

So this is a durable copy. Every source is named, dated and linked back to its original URL and its archive snapshot. The original Japanese is kept alongside the English, never replaced by it, because a translation is an interpretation and the source text is the evidence.

How to read this

A lot of what these players concluded is right. Some of it is wrong. Some was one person's guess that got quoted onward until it sounded like fact. So every substantive claim carries a status:

Two rules that go with those tags

• Repetition is not confirmation. Most numbers in the period corpus trace back to a single anonymous "card analysis" (解析) that circulated on a bulletin board and was quoted onward. A 2ch poster in June 2000 records that the analyst was publicly told, by someone claiming a friend at Sega, that "only about 25% of what you're saying is correct." Every figure descended from that analysis inherits the doubt.
• Where the community and the ROM agree, that is the strongest evidence here — because the two were derived independently, twenty-five years apart, by completely different methods. Those agreements are flagged explicitly.

The sources

Eleven principal sources, plus 52 archived 2ch threads (~6.2 MB). Thread IDs are unix timestamps, so the ID is itself the creation date.

Raw captures, with source URL, archive snapshot ID and retrieval date, are held in the preservation archive. A summary is not a preservation copy — the originals are kept.

The race calendar

Confirmed  Recovered from かぐらステーション (Wayback snapshot 2000-03-04) and since verified against the game itself on a DOC '99 cabinet. DOC '99 runs 19 rounds of 8 races, and every round's 8R is that round's G1. (The Western releases differ: they run to 6R.)

DOC 2000 is different. Its G1 name list is these 19 in the same order plus ジャパンカップダート (Japan Cup Dirt) and 高松宮記念 (Takamatsunomiya Kinen), and its card has room for 21 G1 titles. Its calendar is verified too: 21 rounds of 8 races, each round's 8R its G1, in this order: February Stakes, Takamatsunomiya Kinen, Oka Sho, Satsuki Sho, Tenno Sho (Spring), NHK Mile Cup, Oaks, Derby, Yasuda Kinen, Takarazuka Kinen, Super Dirt Grand Prix, Sprinters Stakes, Shuka Sho, Kikuka Sho, Tenno Sho (Autumn), Queen Elizabeth II Cup, Mile Championship, Japan Cup Dirt, Japan Cup, Arima Kinen, Derby Owners Cup. The card's title bits follow the name list, not the calendar. The table below is DOC '99's.

Venue names are given in their Western forms. In the Japanese originals these are 東京 (Eastern City), 阪神 (Western Hill), 中山 (Northern Park), 京都 (Central City) and セガ (SEGA). Four of the nineteen — 桜花賞, オークス, 秋華賞, エリザベス女王杯 — are fillies' and mares' races, matching real Japanese racing; a colt cannot enter them.

The stat model

1 · Nine numbers, in two families

Confirmed  Every horse has six externals (表パラ, "front parameters") and three internals (裏パラ, "back parameters").

This split was posted to 2ch on 2000-06-22 with all six externals and all three internals named correctly. Our ROM carries the same three internal names in a static string table ("Speed", "Stamina", "Sharp"), and the six externals as a contiguous six-byte block on both the horse record and the card. Six for six and three for three, from two completely independent investigations twenty-five years apart.

A caution: the fourth string in that ROM table is "Friendship". That is the hearts/bond stat, not a fourth internal.

2 · The gauges you can see: 0–63, drawn as 16 blocks

Confirmed  The externals are single bytes running 0 to 63, drawn on screen as 16 blocks of 4.

Our decode agrees: when training writes an external, the game clamps it to 0..63, so 63 is the real cap on the externals. On the physical card the six externals are stored as value minus one.

There is a real 60 in the game, but it is on the internals, not the gauges. Confirmed World Edition Rev C clamps each internal (speed, stamina, sharp) to 60 after every race, and none of about 780 saved cards we checked is above 60. DOC 2000's growth code has the same 60 clamp (Confirmed, read twice independently). An earlier version of this page said there was no 60-clamp on any stat byte. That was wrong: the instruction we quoted for it (mov #63,r13) belongs to the routine that writes the hearts (Relationship) byte, not to any stat. The 60 in the period build targets for the externals is a separate thing, a practical target, and no one ever explains why.

3 · The 16 steps, and what the marks mean

Confirmed  The most influential idea in the whole Japanese corpus is the 16-Step Theory (16段階理論), published by Y's Lab and posted independently to 2ch the same month. Behind each on-screen mark sits a hidden value with four levels inside it.

So an all-○ horse scores 12 points, and a perfect horse scores 18.

Where this sits against the ROM. The ROM has exactly this: a separate, hidden breeding band for each external, stored 0 to 15 (shown 1 to 16), and the mark groups it four values at a time (× 1 to 4, △ 5 to 8, ○ 9 to 12, ◎ 13 to 16). It is not the gauge divided by four. The band is fixed at birth, it is what a foal inherits, and racing, training and retiring never change it (measured on World Edition; DOC 2000 Ver.2 has a mark-drop rule when a card is read to play or read as a breeding parent, see below), while the 0 to 63 gauge (the current external) moves all career. A newborn's gauge starts at exactly twice its band plus a random 0 to 3, with Start 2 lower (World Edition and DOC '99; DOC 2000 runs essentially the same breeding code as World Edition), which is why the two look linked. An earlier version of this page said we had not found a separate field and guessed the 16 steps were the gauge divided by four; that was wrong. The 16-Step Theory describes a real, separate field.

Two comment ladders, and they measure different things

The game prints a comment when a foal is born, and a different one when a horse retires. They run on separate ladders, and the corpus is careless about which one it means.

Community claim  すみだDOC (coment.html, snapshot 2001-04-05) recovered the birth ladder whole: nine tiers, printed at birth, reading the internals.

The page's own guidance: tiers 1 to 4 return roughly a 50% record, tiers 5 to 7 will win at best half their races and may not win at all. Tier 1 was believed not to exist in practice, requiring “so much money and luck”. Tier 3 is what people actually aimed for.

Community claim  The retirement ladder is a different animal: seven texts, collected by Y’s Lab and ordered worst to best. He credits win rate for raising the tier, allows that a large number of G1s may do it too, and hedges the whole thing himself: 確実ではありませんが, “it is not certain, but”. Players named the top tier after its opening word, 費美 (“praise”). When a post says a horse 費美で引退した, that is what it means.

Breeding: the inheritance law

This is the single most important thing the Japanese community worked out, and it is the strongest agreement in this archive.

And one poster had the internals right before anyone — かぐらステーション post [1811], poster MA, 21 March 2000:

The worked example, preserved

The 2ch post taught the rule with Sunday Silence, and it is worth reproducing because it is how players actually learned it.

Read what the poster drew from it, because that is the practical lesson. Start went ○×◎ and came out ◎; Corner went ○×◎ and came out ○ — same marks, different results, because the dam's Start was 16 and her Corner 15. Competing was 9 × 10 and came out 9: the half point is thrown away. もったいない — “what a waste.”

The ceiling trap

Community claim  and it follows from the arithmetic. The CPU horse スケアクロウ (Scarecrow) is ◎◎××◎◎ — values 13, 13, 1, 1, 13, 13. Breeding to it “again and again never produces a ◎”, because 13 mated to the best available ○ (12) gives floor(25/2) = 12, still ○. A 13-grade parent is a trap: nothing available can improve it. (Those are DOC '99's values for Scarecrow. In DOC 2000 its Corner and Tenacious are 14, and 14 mated to a 12 gives 13, a ◎, so there the trap only holds for Start and Spurt.)

What carries to the foal, and what does not

The foal rows marked Confirmed come from the World Edition foal code, which DOC 2000 shares almost word for word (the coat row's 207-foal figure is a measurement). DOC '99's foal routine is different: it has no rare band, an average over 45 does not lose 5, and its pedigree bonus and special coats work differently (see the rows above and the bloodline bonus below).

Dirt and heavy-going aptitude

Disputed  Three ways, and this is the messiest area in the whole corpus. Three accounts of what aptitude even is:

The 2001 thread makes the sharpest point in the corpus on the subject, and it is the part that has held up:

The branch rule, and why stacking CPU dirt horses works

Community claim  The most specific aptitude claim recovered (京都DOC, snapshot 2005-05-10) says the rule branches on whether the parents are yours or the machine’s:

The same page tabulates named CPU dirt horses with two separate numbers, an aptitude and a bonus:

The bonus is clearly a separate authored field, not derived from the aptitude. Two horses both at 70 carry bonuses of 1 and 2, and Wakaoh Raiden at 74 has a larger bonus than Abukuma Poro at 78. If the branch rule is real, then Wakaoh Raiden and Lemon Palm are the better breeding partners even though Abukuma Poro has the higher aptitude. The decoded foal code does not support it: every foal gets the half-by-half average of its two parents' dirt numbers, whichever parents they are (a parent rated 0 is given a random rating instead), with no separate bonus added.

The one thing everyone agrees on, and the useful part: aptitude is a graded quantity that every horse has, it is heritable, and it decides dirt races.

Training, feeding, and the damper

The menus, and what each one raises

Community claim  The structure is regular and worth memorising: each solo menu raises exactly one external by a lot, and each paired menu raises two, one a lot and one a little.

Rest is genuinely unsettled, and one author contradicts himself on it. One site says rest raises three stats a little. ばあにん☆ says ゲージの上昇はほとんどありません, almost no gauge increase, and that rest is for fatigue recovery. Then on another page the same author warns that when levelling a horse for Almighty you must take Corner to the very limit first, “because if you do not raise it to the limit, rest will push the parameter up”. We have not checked it.

The grades, and exactly what they pay

Community claim  What counts as which grade:

And what each grade pays, in gauge points, for a horse at the start of its career:

Confirmed  What the code says about whip overuse (the same in all four versions): the whip count is judged by run type. In a Cooperate (paired) run, 35 to 49 whips pay 0.7 times the base gain after a good run and 50 or more pay 0.5 times, and after a weak run they make its losses 2 times (35 to 49) or 3 times (50 or more) the base. In a Single (solo) run only more than 50 whips count (0.5 times after a good run, 0.4 times with 3 times the losses after a weak one); 35 to 50 change nothing. The Pool has no whip-count penalty.

So a better-bred horse trains faster as well as starting higher. The scheduling advice that falls out of it is sound: train the stats your horse is good at, and use food for the ones it is not.

Why your horse stops improving

Every period source noticed this and none could measure it.

A 2ch thread put it in numbers in 2000: “with nothing but Perfects it stops growing from race 10 to 12, whereas at a moderate pace it grows to race 14 to 15.”

One player called the cause correctly in February 2000. GUTCHI, writing on だーめーおーやーじくらぶ, warned that a high Out-of-box is 不利 — a disadvantage — “because when the Out-of-box parameter is high, the speed of parameter gain dulls sooner.” Four months before anyone else.

Failure, and using it on purpose

Community claim  Failing costs a flat amount, and the stats that drop are deliberately not the ones that menu normally raises.

Paired penalties are exactly double solo penalties at every grade, mirroring paired training’s larger gains. Which stats drop:

Hearts: the mechanic nobody in period got right

The community treated hearts as a bond meter and a whip-response modifier. One poster noticed that “on a horse two or three races from debut the whip bites poorly because it has no hearts.” That is as far as it went. There are no numeric heart thresholds anywhere in the 2000-era sources. Confirmed  Hearts do count on race day, though not through the whip. For your own horse (not the computer's) they add a small speed bonus in the last part of the race, from 1000 metres out in DOC '99 races up to 2200 metres and 1200 metres out in longer ones (500 to 1300 metres out in DOC 2000 and World Edition, by distance and running style). Full hearts against empty is worth about 0.4 percent of speed there.

After every race any internal above 60 is set to 60, and any below 15 gets +5. (Rev D has no 60 clamp: its growth stops each internal at 55 and switches off above a 160 total, and it moves the 30 boundary to 35.) DOC 2000 uses exactly the same after-race rules as Rev C. DOC '99 differs: racing raises each internal to at most 58, and after a race none is ever below 20.

Low hearts quietly rot a horse. Crossing a threshold flips +1 into −1, or −1 into −3. Recovery is easy: a food the horse eats gives +4, +3 or +2 points and one it refuses costs 2, 4 or 8 (how much depends on how that horse likes that food. The message wording depends on its personality: an Honest horse's words give the exact amount, and so do Imposing and Sloppy refusals, and a Coward's “weeping for joy” is +4, but a Rough horse says “loves it” for every food it eats and its “dislike it” is only -2; on Rev D the -2 and -4 refusals share one line for every personality), the reply you choose after each race moves them too (usually by between -12 and +20 points, with no race-count limit), three of the 13 spots in the paddock walk at the start of play give +4 (65 percent) or take 4 (35 percent), and during a horse's first 34 races finishing at or better than its pre-race favourite ranking adds 8, while finishing worse costs points. After 34 races, results no longer move hearts.

Food

The foods that matter

Community claim  From the 44-entry DOC 2000 table (競馬エクウス, page dated 2000-06-28), the strongest single feeds and how to unlock them:

Confirmed  The game's feed-menu code agrees with the exact ones and sharpens the rest: Sugar Cube comes after a Perfect in solo training and Vegetable Salad after a Perfect in paired training. The race-result rules are now proven in the code too: Korean Ginseng comes after a top-3 finish in a G1, Tochu-Tea Fodder after a win in a race that is not a G1, Mamushi Dango after a G1 win that gives the horse its first G1 title and Super Mamushi Dango its second, and Big Korean Ginseng (instead of Korean Ginseng) when a G1 win gives it its third. After a G1 it did not win, Pudding comes if you then choose Rest and Big Pudding if the horse refuses a training session. In World Edition the bananas come from the betting game: picking the winner gives Large Banana (Big Banana here), picking the runner-up gives Banana. DOC 2000 has a quirk here: its check for a G1 the horse already holds looks at the wrong title for every G1 except the February Stakes, so these title foods can appear or fail to appear because of a different G1 (in the code; not yet seen on a cabinet).

Note the design: the best food is gated behind a campaign, not a single result. It takes a third G1 title.

Beer

Confirmed  and it is a lovely piece of trivia. 生中 (nama-chū, a medium draught beer) is a real food in the ROM at ID 43, with 黒生中 (dark draught) at 44. Both records are fully populated and structurally identical to every other food. Their stat payload is all zeros, and they are the only foods in the game with an all-zero payload. World Edition never offers beer in the feed menu; DOC 2000 offers it rarely (see below).

Beer is special-cased at three separate code sites, writes its own reaction-state byte, and has its own dedicated animation pointer arrays. Sega built a beer that raises nothing. (Beer and the bananas do reset a value the race reads, which we believe is the horse's tiredness.)

Disputed  The Japanese sources credited it with Out-of-box and Tenacious large, with raising the internals, and with raising dirt and heavy-going aptitude by +1 (or +2 for the dark). All of that was disputed at source: two days after the aptitude claim, another poster on the same board replied that if beer raised dirt aptitude his own Seiun would have maxed long ago. Eight years later a DOC 2008 thread template still marked the beers’ effect with a question mark. The game's all-zero numbers settle it: beer raises nothing.

Running style

Confirmed  This is the flagship agreement of the whole project. Running style is not a stored property of your horse (the computer racers are the exception: theirs is stored). It is re-derived from the stats before every race, from a single comparison:

Among the five externals excluding Corner, where does Start rank?

Two details his page carries that our decode also has: Start wins ties, and a horse whose Start is at MAX is always a front-runner unless it is Almighty — a corollary, not a separate rule, since at MAX nothing can outrank Start. His advice follows:

The period build targets

Community claim  These are the six-stat targets the Japanese scene actually played to, on the 0 to 64 scale. Note that every one of them tops out at 60, never 64, and the author never says why.

His notes are worth having. For the front-runner: do not rush Start to MAX, weight Corner over Start because Corner matters more as G1 distances lengthen. For the stalker: few people use it, but it is strong at middle distances and “invincible in the Arima Kinen” with an uncontested lead. For the closer: strongest under any conditions and therefore the most popular, so Competing matters, because the mid-race scramble and the slugging match at the post are unavoidable.

And one piece of advice from 2ch that only makes sense given the rank rule: stop about 3 dots short of your target, so that one carrot can still flip the leg type when you need it to.

Route one: the precision build

Community claim  Target: 48 in all six. This is the most sophisticated piece of play in the archive, and it is built entirely on food dislikes and deliberate failure.

Route two: banana-pickling, the brute-force Almighty

This is the technique the 2ch board settled on, and it is much simpler. Big Banana raises all six externals by a small amount at once, and a poster reporting the actual effect wrote 全部１ずつ上がったのでラッキー～♪, “all six went up by 1, lucky!”

How long it takes, per the board: “just keep feeding Big Banana, 20 or more times might do it” (October 2000), and “banana-pickling will take 25 to 26 races” (May 2001). The same 2001 post gives the fast alternative schedule: races 1 to 2 earn over ¥100 million, race 3 take one G1 crown (ideally with Vegetable Salad), race 4 a second, race 5 a third, race 6 Big Korean Ginseng for 4 dots.

Getting a style back after you overshoot

Community claim  and it works because of the rank rule. Style is re-derived every race from current stats, and a failed training drops specific stats by a known amount. So you can put a horse back to the style you wanted. Raised Start too far and turned your closer into a front-runner? Fail a menu that drops Start. Food does the same job in one-point increments, which is why the whole precision maintenance routine is built on rest then food.

Things they found that we would never have guessed

All Community claim unless marked. This is the section that justifies the whole archive: none of it is recoverable from the ROM, because it is about how the game was played.

The ten cards Sega gave away

Handed out at the Shibuya location tests and deliberately stronger than normal horses. They sat outside the normal breeding pool, which is why the community treated them as an unfair advantage.

Disputed  on one cell, and it is a good illustration of why the marks hide things. 京都DOC gives Seiun Sky 15 in Start where Y's Lab gives 13. Both list Silence Suzuka at 13. If 京都DOC is right, then two of these cards display identically on screen — six marks the same — and are not the same horse. The display literally cannot show you the difference.

Confirmed  The ROM settles the cell, and says more. DOC 2000's game data holds these ten as a separate legend table (plus a flat 50/50/50 record believed to be a test entry), and DOC '99 carries the same table with identical contents; World Edition does not have it. Every mark published above matches that table. In it (on the 1 to 16 scale), Silence Suzuka and Seiun Sky both read 16, 9, 2, 9, 13, 13, so both sites were slightly off at Start and Out of the Box, and Taiki Shuttle reads 13, 7, 7, 9, 9, 13 rather than the published 13, 5, 5, 9, 9, 13. The other seven match the published numbers exactly. The two cards really do show six identical marks. What separates them is hidden: Silence Suzuka's stamina, speed and sharp are 43, 56 and 34 against Seiun Sky's 37, 53 and 26, and their dirt values differ too.

The cards were not entirely cut off from breeding, either. Seven of the ten (Special Week, Tokai Teio, Air Groove, Grass Wonder, Abukuma Poro, El Condor Pasa and Taiki Shuttle) also have an ordinary DOC 2000 breeding-catalog entry under the same name, which you can pick as a parent at the cabinet, each with a lower speed + stamina + sharp total than its legend record. Silence Suzuka, Seiun Sky and Narita Brian are not in DOC 2000's catalog.

The card whose value depended on never being used

Under DOC 2000 Ver.2, a horse whose marks sum to 14 points or more is knocked down a whole mark band across all six externals. Players put it at retirement; the code puts it when the card is read to play, and again (at 15 points) when a retired card is read as a breeding parent (see below):

Narita Brian is the only 14-point card in existence.

A Narita Brian that has never been raced on Ver.2 still breeds at full strength. Race it once and it is permanently an 8-point horse. (The players' quote says 9; by the scoring above, ○○△△△△ is 8.) A card whose entire value depended on never being used.

Players reported that the drop is a filter in the new ROM, not a change written to the card: take a parameter-dropped horse back to a Ver.1 cabinet and it breeds at full strength again. Players were running two incompatible rule sets off one physical card, and choosing which cabinet to insert it into. The code disagrees: the check saves the lowered marks back to the card, so the drop travels with the card (in the DOC 2000 program we decoded, whose breeding data matches Ver.2).

Total-54, and why the scene became a monoculture

Community claim  Y's Lab's 総合54理論 (“Total-54 theory”) held that every stallion and broodmare inside the cabinet has its six values summing to 54 — revised to 56 after Ver.2, with 54 as a guaranteed floor.

The consequence players actually felt

Repeated floor-averaging toward a fixed strong parent converges. The community lived it, and said so in March 2001:

He reports six of eight horses at one tournament were the same Seiun Sky stat spread. That is precisely what repeated floor-averaging toward a fixed parent predicts. It is field evidence for the arithmetic — and it is why the scene developed a monoculture.

“Atari” horses: two theories, and what the ROM says

当たり馬 (atari — a hit, as in a winning lottery ticket) was the period name for a foal that came out stronger than its inheritance implied. There were two explanations, and they cannot both be true.

That 2.55 is suspicious in an interesting way: it is exactly 255/100 — the ceiling of an unsigned byte read as hundredths. Either the informant had really seen the field, or it is a plausible-sounding invention.

But there is a real bloodline bonus — and it rewards matching

Confirmed  The game counts the externals where both parents are ◎ (13 to 16), or both are × (1 to 4). Only the sire and dam count, and the bonus raises internals, never externals.

Then the internals are hard-capped at 45. This is the World Edition and DOC 2000 foal code. DOC '99 counts the same matches but gives +3 to all three internals at 4 or 5 matches and +6 to all three at 6 (the 45 cap still applies).

Their tool, rebuilt: the foal calculator

The period sites all taught the inheritance law by worked example. This is that arithmetic, made usable — running the rule we and they independently derived. Enter both parents on the community's 1–16 scale and it shows the foal, the marks, and where the half-points leak away.

Every export horse is a real Japanese champion

Confirmed  When Sega took the game West as World Edition, it could not license the real racehorse names, so the roster was renamed. Matching the decoded DOC 2000 (Japanese) and World Edition breeding catalogs on their exact stats reveals the map: 166 of DOC 2000's 168 breeding horses line up one to one, and the plain English names are famous Japanese champions in disguise. A few:

Only two of DOC 2000's breeding horses have no World Edition twin (ワカオライデン Wakaoh Raiden and ヒシアケボノ Hishi Akebono), and some kept their romanized name (El Condor Pasa among them). The full map is decoded from the catalogs. The race-opponent roster is a separate table with its own English names; 228 of its 244 horses are byte-identical between the two games (seven were re-rated, and nine more differ only in their computer-racer class or appearance data).

What is still to come

The synthesis is now transcribed in full: the stat model and both comment ladders, the inheritance law, dirt and heavy-going aptitude, the training menus, grades, damper and failure tables, the food tables, running style and both Almighty build routes.

What remains is not transcription but testing. The open questions worth the most are the ones where the community contradicts itself or contradicts the ROM: whether rest raises stats at all, and why a Big-Banana Almighty draws longer odds than a Ginseng one with identical gauges. Each of those is answerable on a cabinet. (The dirt branch rule is answered: the decoded foal code averages every pair of parents half by half, with no branch.)