tested deepseek v4 pro, kimi k2.7 and qwen 3.6 on the same messy refactor

had a 400-ish line module at work that everyone was scared to touch. nested callbacks, a couple of silent except blocks, the usual. instead of just picking a model i gave the exact same task to deepseek v4 pro, kimi k2.7 and qwen 3.6 across a bunch of sessions and used whichever felt right each time. v4 pro was the most careful of the three. it traced where the silent excepts were swallowing errors and flagged two i hadn't noticed. the downside is it talks a lot, you wade through a wall of explanation befor