Dannibal/mcr-final2026-testset

MCR Final-2026 Engine Correctness Test Set

A correctness test dataset for Chinese Standard Mahjong (MCR, 国标麻将) engine and judge implementers, built from the 12,288 official IJCAI-2026 Mahjong AI Competition Final Stage-2 games (Botzone, contest finished 2026-07-09; 512 duplicate walls x 24 seat/wind permutations, 4 finalist bots, zero protocol errors).

The product is simple: any new MCR judge/engine (e.g. a JAX implementation) must reproduce these games exactly — every deal, every draw, every claim resolution, every win/fan decision, and the final four-seat score of all 12,288 games. If it does, it is protocol- and rules-equivalent to the official judge on ~590k decision points. The games were played by four independent, strong, mutually adversarial policies, so the stream exercises claim priority, kong edge cases, robbed kongs, 8-fan boundary wins and monster hands far better than random play does.

Files

file	contents
mcr_final2026_full.jsonl.gz	all 12,288 games, one JSON record per line (~14 MB)
mcr_final2026_golden.jsonl	221 curated edge-case games (uncompressed, browsable) — every record also appears in the full file
validate_engine.py	reference validator: dataset self-test + engine harness (pure stdlib)
TAGS_SUMMARY.json	per-tag counts, request-grammar census, judge-vs-MahjongGB scan details, golden game ids
LICENSE	usage note (competition game logs, research use)

Status at publication: validate_engine.py --self-test passes 12,288/12,288 games; the built-in demo engine reproduces 221/221 golden games through the engine harness.

Record schema (one game per line)

{
  "game_id": "6a4eba831ba515095a1f155b",   // Botzone match id
  "srand":   2143175072,                   // judge RNG seed that generated the wall
  "quan":    0,                            // prevalent wind 0-3 (0=E); 3072 games each
  "players": ["player152/player2", ...],   // seat 0..3 -> "user/bot" (metadata only)
  "wall":    ["W4","F3", ... 136 tiles],   // full ordered wall — see deal order below
  "turns": [                               // the complete lockstep protocol stream
    {
      "request":  {"0": "2 T8", "1": "3 0 DRAW", ...},  // judge -> each seat (verbatim)
      "display":  {"action": "DRAW", "player": 0, "tile": "T8",
                   "canHu": [-3,-4,-4,-4], "tileCnt": [20,21,21,21]},
      "responses": {"0": "PLAY T9", "1": "PASS", "2": "PASS", "3": "PASS"}
    },
    // ... terminal turn: display.action == "HU" (fan/fanCnt/score) or "HUANG",
    //     no request/responses on the terminal turn
  ],
  "expected": {
    "ending":    "ron",          // "zimo" | "ron" | "draw"  (draw = 荒庄)
    "winner":    2,              // seat, or null for draw
    "discarder": 3,              // seat that dealt in (ron only; for qianggang = the BUGANG player)
    "qianggang": false,          // true = robbed kong (HU immediately after BUGANG)
    "fan":       [["清龙",16,1],["平和",2,1], ...],  // [name, value, count]
    "fan_total": 20,
    "scores":    [-8,-8,44,-28]  // final duplicate-format game scores, sum = 0
  },
  "tags": ["ron","multi_claim", ...],
  "known_discrepancy": { ... }   // only on gb_discrepancy / gb_fan_diff games (see below)
}

Tiles: W1-W9 characters (万), B1-B9 dots (饼), T1-T9 bamboo (条), F1-F4 winds (东南西北), J1-J3 dragons (中发白). No flowers. Seat s has seat wind s; seat 0 is always the dealer (first to draw) regardless of quan, which is the prevalent wind (uniform 0-3 across the 24 permutations of each wall).

Wall / deal / draw convention (reproduce this or nothing matches)

Each seat owns a private 34-tile wall segment: seat s draws from wall[34*s .. 34*s+33], consumed from the back (wall[34*s+33] first). The deal gives each seat the last 13 tiles of its segment in back-to-front order: hand[s][i] == wall[34*s + 33 - i]. The dealer's first draw is wall[20], i.e. wall[34*s + 20] for s = 0. Kong replacement draws come from the drawer's own segment, same order — there is no separate dead wall. tileCnt[s] in every display is the number of tiles remaining in seat s's segment. The game is a draw (HUANG, 荒庄) when the seat that would draw next has an empty segment. All 12,288 games were verified to obey this convention exactly. srand is informative (the judge's C srand seed) — you never need to re-derive the wall from it; the wall is given explicitly.

Protocol stream: requests, displays, responses

turns[k] is one judge step: request is the exact per-seat request string the judge sent (Botzone Chinese-Standard-Mahjong "simple interaction" format), display is the judge's public event record, responses is what the four bots answered. Complete request-grammar census over all 12,288 games (counts in TAGS_SUMMARY.json):

0 <seat> <quan>                    game start (own seat wind, prevalent wind)
1 0 0 0 0 <t1> ... <t13>           deal (the four 0s are flower counts, always 0)
2 <tile>                           your own draw
3 <p> DRAW                         player p drew (tile hidden)
3 <p> PLAY <tile>                  player p discarded <tile>
3 <p> PENG <tile>                  player p pengs the last discard AND discards <tile>
3 <p> CHI <mid> <tile>             player p chis the last discard (run's middle tile =
                                   <mid>) AND discards <tile>
3 <p> GANG                         AMBIGUOUS: melded kong of the last discard, OR a
                                   concealed kong (AnGang) if p just drew — the AnGang
                                   tile is never in the request (only in the display)
3 <p> BUGANG <tile>                player p promotes an exposed peng of <tile> to a kong

Legal responses: PASS, PLAY <t>, PENG <t-to-discard>, CHI <mid> <t-to-discard>, GANG (meld last discard), GANG <t> (concealed kong, after own draw), BUGANG <t>, HU.

The four classic replay traps (all fixed in this dataset's semantics)

These are real bugs found in a published replay harness for this log format. An engine or log consumer that gets any of them wrong diverges within a few games:

1. PENG/CHI events embed the claimer's follow-up discard. {"action":"PENG", "player":2,"tile":"W4"} means seat 2 pengs the live discard and then discards W4. The claimed tile is NOT in the event — it is the last live discard. For CHI, tileCHI is the run's middle tile and tile is again the follow-up discard. Dropping these embedded discards causes silent state drift and later crashes.
2. GANG is ambiguous and must be discriminated by live-discard tracking. A GANG while an unclaimed discard is on the table is a melded kong of that discard; a GANG with no live discard (i.e. right after the player's own draw) is a concealed kong (AnGang) whose tile appears only in the display, never in the request.
3. BUGANG is a first-class event (promoted kong). It kills the live discard, offers every other seat a robbed-kong (qianggang) HU claim, and on PASS the kong player draws a replacement tile.
4. Terminal labeling: a ron is the winner's claim on the live discard (or on the BUGANG tile — qianggang); a zimo is a win on the winner's own pending draw. The expected block encodes the corrected labels: ending/winner/discarder/ qianggang.

Rules the games exercise (and your engine must implement)

• Minimum 8 fan to win (MCR): the judge only granted HU with fan_total >= 8 (flowerless Botzone MCR scoring; fan lists each 番种 as name/value/count exactly as the judge scored it; fan_total == sum(value*cnt) holds on every win).
• Claim priority: HU > PENG/GANG > CHI. On multiple HU claims for the same tile the seat nearest downstream (counter-clockwise) of the discarder wins. CHI is only legal from the left neighbour. Verified on every multi-claim turn in the corpus (1,622 multi-claim games, 140 with competing HU/claim priority on the same tile).
• Scoring (duplicate format, per game, sums to 0):
  • zimo: winner +3*(8+fan), each other seat -(8+fan)
  • ron: winner +(3*8+fan), discarder -(8+fan), bystanders -8 (qianggang: the BUGANG player is the discarder)
  • draw (荒庄): all 0.
• canHu per-step oracle: every display carries canHu[4]; a value >= 0 is the exact fan count that seat could win with on the current tile event; negative values are judge-internal codes (-4: not applicable / own event, -3: cannot win). Verified: whenever a game ends, the winning event's canHu[winner] equals the final fanCnt. Your engine can check its fan calculator at every single step against this field.

Golden subset & edge-case tags

mcr_final2026_golden.jsonl — 221 games chosen deterministically (lowest game_id per tag; all games of the rare tags). A game carries all tags that apply to it.

tag	full set	golden	meaning
ron	8,436	109	win on another seat's discard
zimo	3,652	92	self-drawn win
draw	200	20	荒庄, wall exhausted, all scores 0
fan8_boundary	1,778	36	win at exactly the 8-fan legal minimum
multi_claim	1,622	57	≥2 seats claimed the same tile (priority resolution)
multi_hu	140	20	≥2 competing HU claims on one tile
bugang	1,204	54	promoted kong occurred
minggang	939	25	melded kong of a discard occurred
angang	519	23	concealed kong occurred
qianggang	28	28 (all)	robbed-kong win (HU right after BUGANG)
fan_qiangganghu	28	28 (all)	抢杠和 scored in the fan list
fan_gangshangkaihua	46	8	杠上开花 (win on kong replacement tile)
fan_last_tile_zimo	9	5	妙手回春 (last wall tile, self-drawn)
fan_last_tile_ron	4	4 (all)	海底捞月 (win on the final discard)
big_fan	20	10	fan_total ≥ 48 (max in corpus: 93)
gb_discrepancy	20	20 (all)	KNOWN-DISCREPANCY, see below
gb_fan_diff	98	20	KNOWN-DISCREPANCY (milder), see below

KNOWN-DISCREPANCY cases (judge vs cai-style MahjongGB replay) — do not "fix" them

Every HU ending was re-scored with the community python MahjongGB (PyMahjongGB) fan calculator by replaying hands/melds through the widely-copied "cai" FeatureAgent (the encoder used by most Botzone MCR RL/imitation codebases). Result over 12,088 wins:

status	games	meaning
agree	11,970	replayed MahjongGB fan == judge fan, fan by fan
cai_replay_fan_diff	98	cai-style replay scores lower but still ≥8 fan
cai_replay_below_8fan_gate	20	cai-style replay scores the judge-accepted zimo below 8 fan — a cai-gated engine would have forbidden a legal win

All 118 disagreements have the same root cause, in the replay layer, not in MahjongGB: the cai FeatureAgent computes is4thTile = shownTiles[winTile] == 4, and the just-drawn tile is never counted in shownTiles, so 和绝张 (Last Of Its Kind, 4 fan) can never fire on a self-drawn win. With is4thTile corrected to shownTiles + (1 if self-drawn) == 4, python MahjongGB matches the judge on all 12,088 wins. The 118 games are tagged (gb_discrepancy = the 20 gate-relevant ones, gb_fan_diff = the 98 milder ones) and carry a known_discrepancy field with both fan lists side by side. The official Botzone judge is the ground truth for engine acceptance. If your engine embeds a cai-style feature/legality layer, these are exactly the games where you will diverge — handle them consciously rather than discovering them as flaky test failures.

Validating your engine

# 1. dataset self-test (no engine needed; re-checks every game against the
#    reference replay semantics: wall order, per-step hand/meld legality, claim
#    priority, terminal classification, exact score arithmetic):
python3 validate_engine.py --self-test mcr_final2026_full.jsonl.gz --jobs 16
#    -> self-test: 12288/12288 games PASS

# 2. interface smoke test with the built-in replay engine:
python3 validate_engine.py --demo mcr_final2026_golden.jsonl

# 3. your engine (start with the golden subset, then run the full set):
python3 validate_engine.py --engine mymodule:MyEngine mcr_final2026_golden.jsonl
python3 validate_engine.py --engine mymodule:MyEngine mcr_final2026_full.jsonl.gz

Your engine implements two methods (see the docstring in validate_engine.py): reset(wall, quan, srand) -> turn and step(responses) -> turn, where a turn is {"requests": {seat: str}, "display": {...}}. The validator feeds the logged responses and requires byte-exact request strings and display events (--loose compares only action/player/tile/tileCHI/hand/fan/fanCnt/score). Acceptance = 12,288/12,288 games reproduced.

Provenance & stats

• IJCAI-2026 Mahjong AI Competition, Final Stage 2 (Botzone contest 6a4eb8c71ba515095a1f1417, finished 2026-07-09). 512 duplicate walls x 24 seat/wind permutations; finalists: kong/shiro, moyu/kdens3, QiuQiuR/丘丘人, player152/player2.
• 12,088 HU endings + 200 荒庄 draws; zero ERROR/TLE/invalid-move events anywhere in the corpus (all per-turn verdicts OK) — every game is a clean, complete protocol trace.
• Harvested from the public Botzone match archive; bot response time/memory/debug fields were stripped, everything else is verbatim.

License / usage

Game logs of a public AI competition (Botzone platform), redistributed for research and engineering use: engine testing, rules verification, imitation-learning research. The bot policies that produced the moves remain the property of their authors; no bot code or model weights are included. If you use this dataset in a publication, please cite the IJCAI-2026 Mahjong AI Competition and link this dataset card.