I was thinking the same. It looks to me like a much safer process could be:

* use some kind of deterministic transpilation that creates ugly code which for sure reproduces the same behaviors

* add tests to cover all those behaviors

* refactor the ugly code

From my experience with copilot I guess that a LLM could help a lot with steps 2 and 3, but I wouldn't trust it at all for step 1