We built a mahjong dangerous-tile predictor calibrated on 4.97M real hands

The article describes the development of OkkanaiPai, a free, offline iOS app that predicts dangerous mahjong discards using a rule-based system calibrated against 4.97 million real hands from Tenhou's top competitive tier. The app achieves an AUC of 0.83 and is designed for quick, glanceable use at the table without disrupting gameplay, though it is limited to standard 4-player East-focused games.

At a real mahjong table, software can't help you mid-hand. The question — "is this tile safe to discard?" — comes up every few turns, and you work it out from memory and pattern recognition alone. We wanted something you could set next to you at the table, pick up in a second, and put down without disrupting the game. Three interactions: The whole flow is designed to be glanceable — you're not staring at a screen mid-game. We didn't ship a live inference model. Instead, we calibrated a rule-based system against 4.97 million discards from Tenhou's Houou-takujo server the top competitive tier , covering 16 days of logged play. The result is a set of coefficients stored as JSON inside the app. AUC on held-out data came to 0.83. Not perfect, but statistically meaningful for a glance-level judgment call at the table. Everything runs fully offline — no network request, no account, no subscription. Being upfront about scope: For standard 4-player East-focused games, it fits. Outside that, weight the output accordingly. Niixo Labs is running a sprint: small, focused iOS tools, shipped fast. OkkanaiPai explores what "calibrated statistics as UX" looks like in a game-assistance context — where a lightweight, offline coefficient table can deliver meaningful signal without any server infrastructure. Free, no ads, no IAP.