Inside the eval gate: how we decide a cheaper model can take over a task

Operant uses an eval gate to safely replace a costly model with a cheaper one, evaluating on holdout data, an LLM judge, and human ratification.