Logging the human overrides is the smartest part of this. That feedback loop is usually the cleanest way to catch prompt drift before things break silently.

Agreed! The overrides turned out to be more useful than the errors. Errors tell you the model failed; overrides tell you where a competent human quietly disagrees with it, which is the drift that never throws an exception. The pattern I watch for is the same field getting edited...

The override rate is the number I would watch. If it sits near zero for a month the queue has stopped being a review and become a rubber stamp, and it still costs you the seconds.

The other gap I keep hitting is that approval lands on the verb, not the payload. A human approves...