TypeSafe AI released Jev, first System One model — transformer-based, outputs calibrated probability decisions (Noul/Choice/Score) not text. Trained via Reinforcement Learning for Calibrated Decisions (RLCD) on synthetic data. Two orders of magnitude faster than LLMs on System One tasks; cannot hallucinate because outputs are predefined. Early access opened Sep 14, 2026; API demand briefly exceeded capacity.