Turion Jev 31B

Turion AI's Gemma-4-31B-it fine-tune, tied #2 on the Decision Index's Oct 28 board and #6 on the Nov 6 draw (70.47). A LoRA soup (uniform merge of steps 300/375/450) trained 7.3 hours on one B300 with soft probability targets and annotator distributions as labels; order-averaging and read-twice passes are gated to single-question requests so latency stays inside the board's rule. The card decontaminates against the public suite by 13-gram overlap (4,395 rows dropped) and publishes the full results dataset. Apache 2.0.