The comparison & result
One policy folded whenever an opponent declared riichi. The other policy assessed its own hand before deciding whether to continue.
The report recorded about +44 net points per hand for the comparison policy, with win shares of 50.9% and 49.1%. The model did not enforce furiten, which is central to Riichi defence. Use this as a question for further testing, rather than a real-table rule about when to fold.
At your table
- Two-plus tiles from ready when the riichi lands — fold fully, safe tiles only
- NOT when you're ready with a decent wait — that hand has earned the right to fight
Check your table’s scoring rules before applying the finding. Read the Japanese Riichi reference.
How it was tested
Tested in the Hé Fēn Lab: 100,000 hands — reflex-folders against players using their judgment, seats swapped so the deal is fair.
The players were software policies, not human participants. This report describes the published experiment; it does not establish the best move for every possible hand or opponent.
Results without a published uncertainty interval should be read as point estimates. Similar results do not prove two choices are equivalent.
Read the Lab’s method and limitationsPublished by Hé Fēn Mahjong. Ask about this finding or suggest a correction.