Don't Call It a Decision Model

TypeSafe argues that Jev should be understood as a System One model, not a decision model: it makes fast, intuitive judgments, while ordinary code combines those judgments with known facts and rules to determine what a system does. Treating Jev as the whole decision-maker - by giving it all the application state and asking it to choose an action - pushes business logic into lengthy criteria prompts. Those prompts become difficult to maintain as edge cases arise. In a customer-service workflow, for example, code can check refund eligibility, amount, and subscription status; Jev can judge whether a customer wants a refund, reports an unauthorized charge, or is open to a retention offer.

The proposed design separates these parts. Jev supplies specific judgments, and explicit logic routes the case: an unauthorized-charge report goes to the fraud team; an eligible refund of $25 or less can be issued immediately; other eligible refunds go to review; and additional rules handle human handoffs, cancellation offers, or routine replies. This decomposition makes the system clearer because its conditions and paths are visible, and more observable because logs show both Jev’s judgments and the code-derived facts that led to an outcome. When a route is wrong, teams can distinguish a mistaken judgment from a faulty rule and fix the relevant part rather than rewriting one large prompt. The same separation also supports extending the system with additional output paths.