Jeff-Qwen3.5-0.8B v1.2 + 9 LoRA adapters: put it in front of Qwen3.8-27B for 38× faster decisions and +8.7 points accuracy, for under 2 GB extra memory

Jeff-Qwen3.5-0.8B v1.2, a small "System 1" model, when combined with 9 LoRA adapters and placed in front of Qwen3.8-27B, significantly improves performance. This setup achieves 38x faster decisions and an 8.7-point increase in accuracy, reaching 95.3% compared to Qwen3.8-27B's 86.6%. It also reduces wrong answers from 13.4% to 4.7% while adding less than 2 GB of memory. The goal is to provide a fast, calibrated decision layer for future local models.