Ggvh Says Jev Offload Cuts Main-LLM Calls 29.6 Percent on 360 Cases
Summary
Developer ggvh reports that offloading bounded decisions to a hosted Jev model cut main-LLM calls 29.6 percent, from 2,537 to 1,785, across 360 SOP-Bench cases run on Qwen 3.8 27B. Estimated cost fell to $4.70 from $6.41. ggvh says the subset result does not establish a general accuracy gain.
Key Points
- ggvh reports mean case latency drops to 6.94 seconds from 8.64 seconds, a 19.7 percent reduction.
- ggvh says completion checking alone saves 23.8 percent of LLM calls and finishes 356 of 360 tasks correctly.
- ggvh reports adding final structured-plan preparation cost lowers the estimated saving to 24.6 percent.