“Sean Wang of Cognition cited research from Meter indicating that more than half of the code generated to pass the SWE-Bench benchmark is of such low quality that it is "unmergeable slop."”