What I was looking for
Where an LLM loop can run unattended, and where a person has to check.
What I found
Code that runs, maths with an answer, a schema, a field copied from a source: these can be verified, so a model can generate many answers and keep the good ones. Style, advice and subjective quality can't, and an LLM judge stands in for a person with known blind spots.
This is why Cernos runs the original strategy code in a sandbox and makes the LLM's translation match its signals: translation is checkable, so the loop is safe to automate.
The move
Before automating with an LLM, I ask what checks its output. If nothing can, a human stays in the loop.