Judgment, Not Generation: Rebuilding Our AI API Classifier on Jev
Harness, Monday, October 5th, 2026
Harness cut AI API classification costs about 1,200x by replacing a generative LLM decision step with its Jev model.
Harness rebuilt a production AI API classification pipeline by replacing the generative LLM judgment step with Jev, a purpose-built decision model.
The change cut per-endpoint classification cost by roughly 1,200x versus the original pipeline and 150x versus a single-shot LLM, and Jev's calibrated probabilities placed all errors in a 0.3-0.8 uncertainty band covering 3.4% of endpoints.
With zero mistakes at the tails and a stable four-field output contract, Harness could swap the decision engine without changing downstream consumers.