How about leveraging llms to produce deterministic codemods? You can then iterate on this by running the codemods and using other deterministic guardrails, feeding the results back into the llm to improve the codemods?