Penalty Enforcement is Now Live

[smly]

Effective: 2026-04-27 13:46:32 JST

We have rolled out two changes that affect how the riichi.dev platform handles incorrect bot behavior.

What changed

1. Engine fix (riichienv 0.4.8)

riichienv 0.4.8 fixes an upstream bug (RiichiEnv#199, patched in #202) where _get_legal_actions_internal failed to restrict discards in the riichi_stage to tiles that maintain tenpai. As a result, riichi declarations that broke tenpai were silently accepted. From this release on, the engine rejects any discard during a riichi declaration that does not maintain tenpai.

2. Penalty enforcement (gameserver)

Previously, when a bot sent an illegal or malformed message, the gameserver silently substituted tsumogiri and the offender suffered no in-game consequence. Going forward, those cases trigger the engine's chombo (penalty) path: a mangan-equivalent point transfer (parent: −12000 / child: −8000) is applied, the kyoku ends with is_renchan = true, and result.penalized[pid] is set to true for the offender.

What is enforced and what is unchanged

Event	Before	After
Riichi declaration whose discard would break tenpai	Accepted by engine (bug)	Rejected (Penalty). Excluded from legal_actions and refused by the engine (riichienv 0.4.8)
Illegal action (parses, but the chosen tile / consume / action type is not in legal_actions)	tsumogiri substitution	Penalty. Mangan-equivalent transfer applied to the offender
Unparseable / malformed message (invalid JSON, unknown type, invalid pai string)	tsumogiri substitution	Penalty. Same penalty as illegal action
Action timeout	tsumogiri substitution	Unchanged.
WebSocket disconnect	tsumogiri substitution	Unchanged.

If you observe behavior inconsistent with the table above, reply to this thread or open a new topic.

Acknowledgments

The engine fix would not have shipped without the community's help. Many thanks to:

• @iranzai for the original bug report (RiichiEnv#196) with a reproducible replay that surfaced the no-tenpai riichi behavior.
• @zl0v7hzr for the precise diagnosis on RiichiEnv#196, which pinpointed the faulty branch in the legal-action filter and made the fix in #202 straightforward.
• @tsubakisakura for raising the penalty-handling question on RiichiEnv#199, pointing out that a no-tenpai riichi should be treated as chombo with a mangan penalty rather than a normal draw. That observation directly motivated the gameserver-side penalty enforcement shipped today.

[mzpkaka]

i think the Penalty has bugs

[smly]

I checked the game:
https://riichi.dev/games/00471a50-7cb8-4b59-bcca-a0c9a8cc2944

I'm not sure many players getting penalties is enough on its own to point to a bug. If you were one of those players and have logs showing a valid Action that still errored out, please share them. That would be hugely helpful.

I do agree the current error messages are too sparse. I'll add better messaging (furiten, kuikae, missing properties, etc.) to the backlog.

[CiaranZhou]

https://riichi.dev/games/b69f598b-c737-4765-9a71-f7b7875074e6

From South 1 to South 1 - 2, Player 0 received 3 penalties; the last penalized actions are shown below.

South 1

12:37:32,008   recv type=request_action
12:37:32,056   request_action id=55 events=8 last_event=tsumo action=Action(action_type=Discard, tile=Some(58), consume_tiles=[], actor=Some(0))
12:37:32,057   sent_action id=55 mjai={"type":"dahai","actor":0,"pai":"6p","tsumogiri":false,"meta":{"q_values":[-7.257592,-4.1901026,-6.358569,-7.685584,0.20723224,-8.149278,-9.607733,-9.724104,-8.665691],"mask_bits":119560198,"is_greedy":true,"batch_size":1,"eval_time_ns":46576139,"shanten":1,"at_furiten":false}}

South 1 - 1

12:37:44,066   recv type=request_action
12:37:44,129   request_action id=58 events=14 last_event=tsumo action=Action(action_type=Discard, tile=Some(125), consume_tiles=[], actor=Some(0))
12:37:44,130   sent_action id=58 mjai={"type":"dahai","actor":0,"pai":"P","tsumogiri":false,"meta":{"q_values":[-1.7591885,-2.6899624,-2.2501833,-1.6451397,-2.3474717,-0.5715237,-1.9088309,-2.1136963,-0.7250479,-1.8945669,-1.5808561,0.5504727],"mask_bits":2253195342,"is_greedy":true,"batch_size":1,"eval_time_ns":46960910,"shanten":4,"at_furiten":false}}

South 1 - 2

12:37:44,165   recv type=request_action
12:37:44,216   request_action id=59 events=4 last_event=tsumo action=Action(action_type=Discard, tile=Some(123), consume_tiles=[], actor=Some(0))
12:37:44,217   sent_action id=59 mjai={"type":"dahai","actor":0,"pai":"N","tsumogiri":false,"meta":{"q_values":[-7.1795692,-7.4654655,-3.2708926,-6.776618,-7.312504,-6.588498,-7.6500444,-7.7200127,-7.5846806,-6.278884,0.24299288,-2.989328],"mask_bits":5690118560,"is_greedy":true,"batch_size":1,"eval_time_ns":39169055,"shanten":3,"at_furiten":false}}

I previewed the corresponding game state in RiichiLab and couldn't see any problem with the current action. If you need more detailed logs you can check the attachment b69f598b-c737-4765-9a71-f7b7875074e6.log.

BTW, I use the latest version of riichienv, which is 0.4.8 . I suspect that some modifications exist only in the RiichiEnv SDK and were not synchronized to the online riichi.dev. Downgrading to version 0.4.7 did not change the situation.

I also observed that these errors always occurred consecutively in the final games of a hanchan, which may indicate that an incorrect state in the simulation environment is not being cleared.

I checked almost all of this bot's abnormal matches, and they all show similar situations. Please help me check what the problem is, thanks.

[CiaranZhou]

@smly I'm very happy to see these fixes, thank you so much for your efforts. It looks like the occurrence of this issue has greatly decreased after the fixes were deployed.

So this bot took ~7.9 seconds end-to-end for one act-phase decision.

Is the inference here referring to the processing speed of the client-side bot? My understanding of this log is that it measures the time from when the server sends a request_action to the client/bot to when it receives the client/bot's request_action response.

I checked the logs for that match and found the timing to be normal. A short excerpt of the log is shown below (the timestamps in the log are in UTC+8); the full log is at 6adf22e5-e3cd-4824-b3b5-260c458721ea.log

2026-05-05 22:15:32,869 INFO game=39 kyoku=E3-2 actor=0 tsumo pai=7p
2026-05-05 22:15:32,918 INFO request_action id=48 game=39 kyoku=E3-2 events=8 last_event=tsumo action=Action(action_type=Discard, tile=Some(71), consume_tiles=[], actor=Some(0)) possible_actions=[{'actor': 0, 'pai': '3m', 'type': 'dahai'}, {'actor': 0, 'pai': '3m', 'type': 'dahai'}, {'actor': 0, 'pai': '2p', 'type': 'dahai'}, {'actor': 0, 'pai': '3p', 'type': 'dahai'}, {'actor': 0, 'pai': '4p', 'type': 'dahai'}, {'actor': 0, 'pai': '5p', 'type': 'dahai'}, {'actor': 0, 'pai': '7p', 'type': 'dahai'}, {'actor': 0, 'pai': '9p', 'type': 'dahai'}, {'actor': 0, 'pai': '4s', 'type': 'dahai'}, {'actor': 0, 'pai': '5s', 'type': 'dahai'}, {'actor': 0, 'pai': '7s', 'type': 'dahai'}, {'actor': 0, 'pai': '7s', 'type': 'dahai'}, {'actor': 0, 'pai': '7s', 'type': 'dahai'}, {'actor': 0, 'pai': '7p', 'type': 'dahai'}]
2026-05-05 22:15:32,919 INFO sent_action id=48 send={"actor":0,"pai":"9p","type":"dahai"} mjai={"type":"dahai","actor":0,"pai":"9p","tsumogiri":false,"meta":{"q_values":[-7.9370756,-4.0536137,-9.474277,-9.266989,-5.3448753,-4.0136566,0.1655941,-8.826953,-8.612881,-8.853435],"mask_bits":23247876,"is_greedy":true,"batch_size":1,"eval_time_ns":47715274,"shanten":1,"at_furiten":false}}
2026-05-05 22:15:33,769 INFO game=39 kyoku=E3-2 actor=0 dahai pai=7p ^
2026-05-05 22:15:33,769 INFO game=39 kyoku=E3-2 actor=1 tsumo pai=?
2026-05-05 22:15:33,769 INFO game=39 kyoku=E3-2 actor=1 dahai pai=4p ^
2026-05-05 22:15:33,770 INFO game=39 kyoku=E3-2 actor=2 tsumo pai=?
2026-05-05 22:15:33,770 INFO game=39 kyoku=E3-2 actor=2 dahai pai=3m ^
2026-05-05 22:15:33,815 INFO request_action id=49 game=39 kyoku=E3-2 events=5 last_event=dahai action=Action(action_type=Pass, tile=None, consume_tiles=[], actor=Some(0)) possible_actions=[{'actor': 0, 'consumed': ['3m', '3m'], 'pai': '3m', 'type': 'pon'}, {'actor': 0, 'type': 'none'}]
2026-05-05 22:15:33,815 INFO sent_action id=49 send={"actor":0,"type":"none"} mjai={"type":"none","meta":{"q_values":[-1.363091,0.2827751],"mask_bits":37383395344384,"is_greedy":true,"batch_size":1,"eval_time_ns":44597774,"shanten":1,"at_furiten":false}}
2026-05-05 22:15:35,583 INFO ryukyoku game=39 kyoku=E3-2 deltas=[-8000, 2000, 4000, 2000] reason=Error: Illegal Action by Player 0
2026-05-05 22:15:35,584 INFO recv type=end_kyoku

[CiaranZhou]

@smly I just observed another game with a series of consecutive errors
https://riichi.dev/games/9f0ed50f-a9f2-4cab-8ddc-fb1cfeb862d5

[smly]

I wasn’t able to spend enough time on this today, but I’ll look into it tomorrow.

[smly]

@CiaranZhou Thanks for the follow-up. For the 6adf22e5-e3cd-4824-b3b5-260c458721ea case, the timeline is:

• Server starts waiting for Player 0's act-phase response at 14:15:26.373 UTC.
• Server times out at 14:15:29.375 UTC.
• Your client log records receiving the corresponding request_action id=48 at 14:15:32.918 UTC.
• Your client sends dahai 9p at 14:15:32.919 UTC.
• The server receives that dahai 9p at 14:15:34.267 UTC, but by then the game has already advanced to a response phase where only pon or none is legal, so it is classified as illegal.

The current logs show a multi-second gap between the server-side request timestamp and the client-side request_action log timestamp. This cannot distinguish whether that gap is caused by:

1. server-side per-connection send/backpressure
2. network/webSocket delivery delay
3. the client process not running its receive handler promptly

I do not think normal geographic latency between client/server should by itself explain a 3+ second delay. If this only occurs for a specific bot/user, that would make a global gameserver issue less likely, but it still would not identify the exact source.

On the server side, I will focus on preventing late stale replies from being converted into chombo, and on improving timing logs. A protocol-level request id is the proper long-term fix, but it is a larger compatibility change. If we decide to change the protocol, I will open a separate Discussion topic first to gather feedback.

[smly]

I am starting to think that the most realistic approach may be something closer to online mahjong platforms, giving each player a 20-second main time bank, plus 5 seconds of extra time per turn. This would be a fairly large change, but I will seriously consider it when we update the protocol.

[mzpkaka]

In my tests, if the latency to the server exceeds 200ms, disconnects often occur and lead to penalties. I think the approach for handling bots with poor network could provide a hint: suspend bots that disconnect multiple times within a certain period from queuing.

[DDLovo]

I observed several cases where the bot received an Error: Illegal Action by Player X penalty, even though the action sent by the bot appears to match one of the actions listed in the immediately preceding request_action.possible_actions.
All timestamps below are shown in UTC+8 from my local/client logs. Appear to be about 1 hour behind the timestamps displayed on the riichi.dev game pages.

Case A:
- Bot running on a remote Linux server via SSH
- Bot SDK: riichienv 0.4.8
- Python: 3.13.13
- torch: 2.12.0+cpu
- websockets: 16.0

Case B:
- Bot running on local Windows machine
- Bot SDK: riichienv 0.4.8
- Client: Python websockets

Detailed remote-server case:
File:
e49df150-344a-4ce5-aa00-450d09cbf227.jsonl.log
Game link:
https://riichi.dev/games/e49df150-344a-4ce5-aa00-450d09cbf227
This file was recorded on the remote Linux server, in a different network environment from the local cases.

Bot seat:
Player 1

Penalties:
line 314: Error: Illegal Action by Player 1
line 320: Error: Illegal Action by Player 1

Relevant timeline from e49df150-344a-4ce5-aa00-450d09cbf227.jsonl:

2026-05-14 13:18:33.595 recv dahai
{"actor":0,"pai":"9m","tsumogiri":true,"type":"dahai"}

2026-05-14 13:18:33.595 recv request_action
possible_actions:
[
  {"actor":1,"consumed":["7m","8m"],"pai":"9m","type":"chi"},
  {"actor":1,"type":"none"}
]

2026-05-14 13:18:33.703 send
{"actor":1,"type":"none"}

2026-05-14 13:18:33.828 recv tsumo
{"actor":1,"pai":"4s","type":"tsumo"}

2026-05-14 13:18:33.828 recv request_action
possible_actions:
[
  {"actor":1,"pai":"3m","type":"dahai"},
  {"actor":1,"pai":"4m","type":"dahai"},
  {"actor":1,"pai":"5m","type":"dahai"},
  {"actor":1,"pai":"5m","type":"dahai"},
  {"actor":1,"pai":"5m","type":"dahai"},
  {"actor":1,"pai":"6m","type":"dahai"},
  {"actor":1,"pai":"7m","type":"dahai"},
  {"actor":1,"pai":"8m","type":"dahai"},
  {"actor":1,"pai":"4p","type":"dahai"},
  {"actor":1,"pai":"5s","type":"dahai"},
  {"actor":1,"pai":"7s","type":"dahai"},
  {"actor":1,"pai":"7s","type":"dahai"},
  {"actor":1,"pai":"8s","type":"dahai"},
  {"actor":1,"pai":"4s","type":"dahai"}
]

2026-05-14 13:18:33.933 send
{"actor":1,"pai":"4p","type":"dahai","tsumogiri":false}

2026-05-14 13:18:33.974 recv ryukyoku
{"deltas":[4000,-12000,4000,4000],"reason":"Error: Illegal Action by Player 1","type":"ryukyoku"}

Decoded RiichiEnv observation for that request_action:

player_id = 1
drawn_tile = 84
hand = [11,14,17,18,19,22,27,29,49,89,98,99,102,84]

_legal_actions:
[
  {"action_type":"Discard","actor":1,"tile":11},
  {"action_type":"Discard","actor":1,"tile":14},
  {"action_type":"Discard","actor":1,"tile":17},
  {"action_type":"Discard","actor":1,"tile":18},
  {"action_type":"Discard","actor":1,"tile":19},
  {"action_type":"Discard","actor":1,"tile":22},
  {"action_type":"Discard","actor":1,"tile":27},
  {"action_type":"Discard","actor":1,"tile":29},
  {"action_type":"Discard","actor":1,"tile":49},
  {"action_type":"Discard","actor":1,"tile":89},
  {"action_type":"Discard","actor":1,"tile":98},
  {"action_type":"Discard","actor":1,"tile":99},
  {"action_type":"Discard","actor":1,"tile":102},
  {"action_type":"Discard","actor":1,"tile":84}
]

In this observation:

tile 49 = 4p
{"actor":1,"pai":"4p","type":"dahai"} was also present in request_action.possible_actions.

Immediately after the first penalty in the same file:

2026-05-14 13:18:33.975 recv start_kyoku
2026-05-14 13:18:33.975 recv tsumo
{"actor":1,"pai":"S","type":"tsumo"}

2026-05-14 13:18:33.975 recv request_action
possible_actions included:
{"actor":1,"pai":"S","type":"dahai"}

2026-05-14 13:18:34.076 send
{"actor":1,"pai":"S","type":"dahai","tsumogiri":true}

2026-05-14 13:18:34.095 recv ryukyoku
{"deltas":[4000,-12000,4000,4000],"reason":"Error: Illegal Action by Player 1","type":"ryukyoku"}

Timing stats from the same remote-server run:

previous recv event -> request_action:
n=2172
mean=0.0116s
median=0.0000s
max=1.3627s

request_action -> bot send:
n=2172
mean=0.1147s
median=0.1099s
max=0.6206s

Detailed client audit logs:

• 1f6639ab-1def-43b1-bf97-714289000fce.jsonl.log
  Game link: https://riichi.dev/games/1f6639ab-1def-43b1-bf97-714289000fce

• 1e871cd3-01fe-4afa-962d-dec4999aa370.jsonl.log
  Game link: https://riichi.dev/games/1e871cd3-01fe-4afa-962d-dec4999aa370

• 92a2450c-cfbf-4391-a7b3-ecd2ad266bd3.jsonl.log
  Game link: https://riichi.dev/games/92a2450c-cfbf-4391-a7b3-ecd2ad266bd3

• fc720804-5533-468e-88ad-604143a0bf24.jsonl.log
  Game link: https://riichi.dev/games/fc720804-5533-468e-88ad-604143a0bf24