It used to but not anymore. There's been something added to training data, I guess to Claude 4.7, that then propagates through traces of Claude executions to all the other models. The solution is usually to scope the review out by section demarcation and then asking it to review each section in subtasks, bring it all together, and do a final scan of the entirety.
This way, you can get a lot of "no findings" on individual sections, which is more reliable. I never instruct literally for a no findings case though, that goes at the cost of completeness because you're giving the bot an out - the "cheating" is trained in too.
If it doesn't have a reasonable stopping rule, you can't really trust the feedback
I disagree with that. In code, you can address a finding with a clarifying comment too.
It used to but not anymore. There's been something added to training data, I guess to Claude 4.7, that then propagates through traces of Claude executions to all the other models. The solution is usually to scope the review out by section demarcation and then asking it to review each section in subtasks, bring it all together, and do a final scan of the entirety.
This way, you can get a lot of "no findings" on individual sections, which is more reliable. I never instruct literally for a no findings case though, that goes at the cost of completeness because you're giving the bot an out - the "cheating" is trained in too.
I disagree with that. In code, you can address a finding with a clarifying comment too.