>Your argument is just as applicable on human code reviewers.

The tests many of us use for how capable a model or harness is is usually based around whether they can spot logical errors readily visible to humans.

Hence: https://news.ycombinator.com/item?id=47031580