What is a System One model?

A System One model answers a question you define about some input by picking from answers you set in advance, and returns a probability with its pick instead of writing text. TypeSafe AI introduced the term with its first such model, Jev, in a launch post dated September 15, 2026, borrowing Kahneman’s name for fast, intuitive thinking.

The difference is the training target. TypeSafe’s AI primer contrasts RLHF, which trains a model toward responses people prefer, with its own method, reinforcement learning for calibrated decisions (RLCD), which aims for answers given 0.8 to be right about 80% of the time. That’s what lets code act on the number: above a threshold, proceed, and below it, hand the case to a person or a reasoning model. An LLM’s self-reported confidence tends to bunch near the top of the scale, which makes the same threshold hard to place.

The primer also names the limit: calibration describes groups of predictions, so any single high-confidence answer can still be wrong. An independent test found a wider gap off-distribution: on a question whose answer wasn’t in the text, Jev’s picks averaged 0.74 probability and were right 44.7% of the time.