Loading field map
Every mahjong player faces the same dilemma at each discard: the tile that advances your own hand may be the very tile your opponent needs to win. Defensive play is the art of managing this tension—deciding when to hold back, what to discard, and how to read the table. Over the past century, five distinct frameworks have shaped how players think about safety, each offering a different answer to the question of what makes a tile safe.
The earliest approach to defense was the Classical Intuitive School. Before formal strategy writing, players relied on pattern recognition and experience. A safe tile was one that had already been discarded by an opponent, or one that seemed unlikely to complete a visible meld. Defense was secondary to hand-building; the primary goal was to assemble a winning hand as quickly as possible, and safety was a matter of gut feeling rather than calculation. This school transmitted its knowledge orally, through apprenticeship and repeated play. Its weakness was inconsistency: what one player considered safe another might not, and there was no way to teach defensive reasoning systematically.
The Codified Defensive Principles School emerged as players began writing down and standardizing defensive heuristics. This framework transformed defense from an intuitive art into a teachable discipline. Its core contribution was a set of explicit rules for reading opponent discards and inferring safe tiles.
Three techniques became central. Suji (the "suji" or "number line" principle) holds that if an opponent discards a 5, then the 2 and 8 of the same suit are relatively safe because the opponent is unlikely to be waiting on a 2-5 or 5-8 sequence. Kabe (the "wall" principle) states that if all four copies of a tile are visible (discarded or in your hand), then tiles one step away become safe—for example, if all 4s are gone, then 3 and 5 cannot complete a 3-4-5 sequence. Furiten awareness uses the rule that a player who has discarded a winning tile cannot win on a discard from another player; knowing which tiles are in furiten allows you to discard them safely.
These rules gave players a shared vocabulary and a repeatable method. But they had limits. Codified principles assumed that opponents always follow standard hand-building patterns, which experienced players could exploit by discarding in ways that mislead. The rules also treated safety as absolute—a tile was either safe or dangerous—without accounting for how likely an opponent actually was to be waiting.
The Mathematical Probability Defense School addressed the Codified Principles School's binary view of safety by introducing probabilistic reasoning. Instead of labeling tiles as safe or dangerous, this framework asked: how likely is it that this tile will cause an opponent to win?
Two methods defined this school. Live-tile counting tracked how many copies of each tile remained unseen, allowing players to estimate the probability that an opponent held a particular wait. Tenpai probability estimation used the number of tiles an opponent had drawn and discarded to infer how close they were to a winning hand. A player who had discarded many tiles was likely still far from tenpai (ready hand), while one who had discarded few might be waiting. By combining these estimates, a player could rank discards by danger level and choose the least risky option.
This school did not reject codified rules; it absorbed them. Suji and kabe became probabilistic shortcuts rather than absolute guarantees. The Mathematical Probability School also borrowed from the emerging field of hand efficiency theory, which analyzed tile acceptance and hand speed. The tradeoff between tile efficiency and safety became a central concern: a tile that was efficient for your hand might be dangerous, and a safe tile might slow you down. The framework provided a way to weigh these factors numerically.
The Push-Fold Strategy Paradigm grew out of the Mathematical Probability School but added a crucial dimension: context. Earlier frameworks treated defense as a static choice—discard the safest tile. Push-Fold recognized that the decision to defend or attack depends on score standing, round urgency, and opponent tendencies.
Consider a concrete scenario: you are in last place with few points remaining in the round. Even a small hand could win you the game, but your opponent is in tenpai with a large hand. The Mathematical Probability School would recommend discarding the safest tile, but Push-Fold says: if the expected value of pushing (attacking) exceeds the expected loss from dealing in, you should push. The fold threshold shifts based on your score relative to opponents, the number of tiles remaining in the wall, and the estimated hand value of each opponent.
Push-Fold also introduced opponent modeling beyond tile reading. Instead of assuming all opponents play optimally, it accounts for individual tendencies: some players fold too easily, others push recklessly. This borrows from game theory's concept of mixed strategies and exploitation. The paradigm remains the dominant human-play framework today because it captures the dynamic nature of mahjong—defense is not a fixed rule but a conditional choice that changes every turn.
AI-Assisted Defensive Analysis represents the most recent shift. Machine learning models, trained on millions of games, can evaluate defensive decisions across the full combinatorial space of possible opponent hands. Unlike human heuristics, AI does not rely on suji or kabe; it calculates exact probabilities for every possible opponent hand consistent with the visible discards.
This framework challenges the Push-Fold Paradigm in two ways. First, AI can identify safe tiles that human rules would miss, because it considers rare hand patterns and non-standard waits. Second, AI opponent modeling is purely statistical—it does not attribute intentions or tendencies but instead computes the likelihood of each hand given the data. This makes AI more accurate in some situations but less adaptable to human psychology.
AI-Assisted Defensive Analysis has not replaced Push-Fold; instead, it serves as a tool for analysis and training. Players use AI to review their decisions and discover blind spots. The framework's insights are gradually being distilled into new heuristics, though the gap between machine precision and human intuition remains wide.
Today, the Push-Fold Strategy Paradigm and AI-Assisted Defensive Analysis are the leading frameworks, but they coexist with earlier approaches. Codified principles remain useful as quick heuristics under time pressure, especially for beginners. Mathematical probability methods provide the foundation for push-fold calculations.
What the leading frameworks agree on: defense is not absolute but probabilistic; the decision to fold or push depends on expected value; and reading opponent discards is essential. Where they disagree: Push-Fold emphasizes human opponent modeling and contextual judgment, while AI-Assisted Analysis prioritizes statistical accuracy over interpretability. The unresolved question is whether AI insights can be translated into teachable human principles, or whether the future of defensive play will always require a blend of machine calculation and human intuition.