Reading up on SOP-Bench

Offloading bounded decisions to a decision model with Qwen 3.8 27B and Jev cuts agent LLM calls by 29.6% across 360 SOP‑Bench cases for referral abuse detection, email intent, and patient intake.