TypeSafe AI Pushes Back on “Decision Model” Label: “Jev Makes Judgments, Code Makes Decisions”
TypeSafe AI DevRel Allie Laabs explains why stuffing business rules into a single Choice prompt creates prompt gymnastics—and why code should still own the final branch.
As rivals from OpenAI to Google and Hugging Face coalesced around the term “decision model,” TypeSafe AI published a pointed architectural manifesto on November 3 titled “Don’t Call It a Decision Model.” Its subtitle stated the company’s design philosophy in six words: “Jev makes judgments, code makes decisions.”
Avoiding “prompt gymnastics”
Written by TypeSafe AI developer relations lead Allie Laabs, the post explained a common mistake teams make when first adopting Jev: collapsing an entire multi-rule business workflow—such as customer-service refund eligibility—into a single massive Choice question stuffed with prose caveats in the criteria field.
When teams put deterministic business logic inside model criteria, every new edge case forces another round of fragile wording tweaks—what TypeSafe calls “prompt gymnastics.” Instead, TypeSafe recommends treating a workflow like a circuit: ordinary code evaluates deterministic rules (subscription status, dollar thresholds, order age), while Jev evaluates only the atomic semantic judgments that code cannot check directly (for example, whether a customer is threatening legal action or open to a retention offer).
Why the naming distinction matters
Laabs acknowledged that TypeSafe itself uses the word “decision” in educational materials, but argued that calling Jev a System One model keeps the abstraction boundary honest. When developers keep control flow in deterministic code and delegate narrow, calibrated judgments to Jev, systems remain auditable, testable, and far easier to debug as business policies evolve.