@SakanaAILabs
Use Case 6: Classical Japanese Kana Reading Order It would not be a Sakana AI launch without a uniquely Japanese challenge. Can an AI decipher the chaotic layout of a 400-year-old letter? We tested whether the models could recover the reading order of "scattered writing" (chirashigaki) in a letter from 1610. This letter is held by the Keio Institute of Oriental Classics. Given character bounding boxes and a rough set of rules, the models had to write code to predict the exact order the characters should be read. In the clip below, the green line is the expert ground truth. The red line is the AI’s prediction. The Results: • Fugu Ultra achieved a 0.80 accuracy score, tracing the highly complex path almost exactly. • Models A and B scored a dismal 0.24, jumping wildly and incorrectly all over the page. • Model C failed to produce a working predictor at all. You might not decode 17th-century calligraphy every day, but this proves Fugu’s unparalleled ability to handle extreme spatial reasoning and completely novel, non-linear logic.