I’m experimenting with DeepSeek for small Python utilities. One workflow is to ask it for an implementation and then ask it to write the unit tests. Everything passing feels reassuring, but I’m not sure how much confidence that should give me.

For example, a date-range helper might interpret “through Friday” as excluding Friday. If the tests follow the implementation’s interpretation instead of the intended requirement, both could agree and still be wrong.

Would a separate conversation, given only the requirements and function signature, make AI-generated unit tests meaningfully more independent? Or could the same model still repeat the same assumptions without seeing its earlier code?

• Should expected results for boundary cases come from a human-written checklist?
• Would asking another model to review the tests expose different gaps?
• How useful is mutation testing for Python here, compared with checking a few deliberately incorrect implementations?

I’d like a lightweight workflow for catching shared mistakes—not just a larger test suite that confirms whatever the generated function already does.