Developer releases sokudan, an open 310M Japanese decision model

A developer benchmarked System One decision models (Laya, Lev, kev) in Japanese and found a position prior: laya-multilingual never picked the first level of an ordinal question (0 of 300). In response he built sokudan, an open 310M-parameter model (Apache-2.0) scoring 0.880 choice accuracy and 0.075 RPS on 300 Japanese messages.

• laya-multilingual picked the first ordinal level 0 of 300 times in Japanese and 0 of 290 in English
• sokudan: 314.6M parameters, trained on 31,243 pairs, 1.3 GB VRAM, runs on CPU
• Japanese bench: choice acc 0.880, score RPS 0.075, bool AUROC 0.844
• Choice transfers to English nearly intact (0.872), but yes/no collapses to the majority class