Agreed, and we run it this way. Models draft and fix, but nothing reaches our public site unless it passes a set of plain-code gates the model cannot argue with.
Three things we learned that might sharpen the argument:
The deterministic layer needs guarding too. Today an edit (made by a model) quietly commented out one of our gates. Nothing failed, because nothing ran. Now every gate must print its own line, and a missing line stops the release.
Deterministic is not the same as right. One gate refused a correct label because two numbers rounded to the same display value. The fix was in the data, not in loosening the gate. Refusals should be cheap, loud and easy to trace.
Log the disagreements. What the model proposed, what the gate decided, and how often they differ is the number that tests your claim.
Good post.