Gemma4-31B worked in an iterative-correction loop (with a long-term memory bank) for 2 hours to solve a problem that baseline GPT-5.4-Pro couldn’t