TypeSafe AI released Jev, a transformer-based System One Model that does not generate text but returns typed decisions with probabilities code can branch on, available as a hosted API in early access behind a waitlist. TypeSafe has not published weights, a parameter count, or a self-hosting option. The name borrows from Daniel Kahneman’s split between fast intuition and slow reasoning; TypeSafe argues RLHF tuned models for human preference, producing chat, overconfidence and mode dropping, keeping a human in the loop. Jev uses a new stack: a new architecture, a parallel sampler, and Reinforcement Learning for Calibrated Decisions (RLCD), though the architecture is not disclosed. One endpoint, POST https://api.typesafe.ai/v1/systemone, carries state, model, and a map of questions, with 3 question types: Choice, Score, and Noul. Questions run in parallel and in isolation; a Choice supports up to 255 options. Every Choice and Score answer carries a confidence value from 0 to 1, derived from the probability distribution shape; the docs example shows billing at 0.84 with confidence 0.596. Jev costs $42 per billion input tokens, and in its recorded demo finished in 0.114s for $0.000081, versus GPT-5.6 Terra at 8.566s for $0.013880, which TypeSafe claims is 193.6x faster and 444