I put a second AI in front of my AI. It answers in probabilities, not paragraphs.

An experiment involved placing Jev, a System One model from TypeSafe, in front of a coding agent. Unlike chat models that generate prose, Jev provides typed judgments and probabilities, such as numbers between 0 and 1 for yes/no questions or scores for ratings. This allows users to branch on answers rather than parsing text. The setup utilizes Claude Code hooks to run scripts during message submissions, transforming how the AI handles decision-making and logic processes.