LLM's introduces problems, and it finds them in its own internal thinking. But instead of actually modifying the previous generated answer to fix the real issue, it adds another layer to deterministically guard around it, greatly expanding the scope of the fix. This scales with effort, and the result is spaghetti and with a side of bugs.

Best to stick with a high end model + low effort, do a manual pass on high effort and fix the bugs you know are reachable.