According to God of Prompt on X, two independent researchers built a bilevel autoresearch system where an outer loop reads the inner loop’s source code, diagnoses bottlenecks via structured analysis, generates replacement Python, hot-swaps it at runtime, and restores on failure, yielding a 5x improvement in validation bpb over a standard inner-loop baseline. As reported by the same thread, baseline autoresearch loops repeatedly proposed increasing TOTAL_BATCH_SIZE and became trapped by design-time biases; the AI-generated outer loop introduced a Tabu Search Manager and Systematic Orthogonal Exploration to prevent revisiting regions and to diversify search dimensions, discovering that reducing TOTAL_BATCH_SIZE from 2^19 to 2^17 drove the largest gains. According to the post, parameter-only outer loops produced no reliable improvements, while code-rewriting outer loops delivered −0.045 val_bpb improvement per run vs −0.009 for baseline, with 5 of 6 generated mechanisms importing successfully and automatic rollback on one sklearn-dependent failure. The analysis underscores a business opportunity for LLM-based code synthesis frameworks that dynamically refactor optimization architectures in MLOps and AutoML pipelines, as reported by the X thread.