Tenhou-trained Transformer AI mirrors humans in yaku pursuit, gap just 0–2pt

The analysis is built on a deliberate isolation: only positions where the efficiency-best discard definitely breaks the yaku shape are counted, so any choice to preserve the shape cannot be explained as efficiency. Within that narrow slice, the pattern is a dose-response one. Below a 0.7 suit ratio for mixed flush, the model's share of shape-preserving choices sits near or slightly above uniform sampling; above 0.7 it is about 1.7× uniform, and at 0.8 or higher about 4× uniform.

The score-situation result runs in the same direction as the author's expectation, but comes with a self-imposed caveat. Observed rank-based rates rise from 0.711 at 1st to 0.812 at 4th for mixed flush, yet rank also correlates with hand quality and with how often the efficiency-best discard is chosen at all. The counterfactual rewrite of score inputs isolates the input change, but includes score combinations that do not occur in real play, so the author treats the deltas as extrapolation only.

The calling results complicate any picture of a single-minded flush chaser. Even in strong-target hands, other-suit tiles are called roughly 4割 of the time, and the model's main-suit call rate is about 45%, so the gap is small. The author's reading is that the model mostly inherited human habits from its teacher labels, not that it discovered an independent yaku-seeking strategy: self-play winning-yaku frequencies remain unmeasured, and the efficiency baseline ignores point value, safety, and wait quality.