#japanese

Transliteration libraries store kana-to-romaji as a lookup table, and a single row quietly drifts — a kana gets the wrong reading, or gets dropped. You rarely have to guess the right value, because a sibling in the same table (hiragana vs katakana, じ vs ぢ, the voiced pair beside it) already does it right. The oracle is the round-trip: kana to romaji and back should be stable. When it isn't, compare the broken entry to its sibling.