"agent pipelines that [...] clean a messy [repository]"
This feels like a terrible approach, sufficient to condemn the entire study.
Apparently half of the "minimal pairs" in this work were constructed in this way. I simply am not going to trust any conclusion that requires assuming these AI "cleaned" repos are in any way representative of actually-good codebases.
This feels like a terrible approach, sufficient to condemn the entire study.
Apparently half of the "minimal pairs" in this work were constructed in this way. I simply am not going to trust any conclusion that requires assuming these AI "cleaned" repos are in any way representative of actually-good codebases.