Your code assistant passes all correctness tests. But engineers are quietly switching back to GPT-4. The code is correct. It's just... bad. 200-line functions for 10-line problems. No edge cases handled. Unreadable variable names. Correctness is necessary. Quality is what users actually pay for.