TwIL-LM3-Pro is a 3.66B parameter reasoning model for formal logic that improves in-domain performance by 28% over its base while maintaining general capability, outperforming VibeThinker-3B and competing with larger models like Qwen3-8B on formal logic benchmarks. Built through LoRA fine-tuning, checkpoint fusion, and reinforcement learning on IBM Granite 4.2, it achieves top performance on Track A metrics including macro gate (0.5539) and strict-7 accuracy (0.2879).