787,562 function pairs: AI code is half the size with different defect classes, not fewer.

This corpus pairs human functions with implementations generated from their docstrings by GPT, DeepSeek-Coder and Qwen2.5-Coder across Python, Java and C. AI code is structurally compressed and templated at roughly half the size and branching. Security cuts both ways: more and more severe findings in Python and Java, but fewer high-severity memory-safety findings than humans in C. Once you control for size, complexity metrics carry almost no signal while naturalness still separates authors. CQBench ships with 27,346 issue-prone tasks.