Asked about Testing a system with a non-deterministic part

What is usually untested in an AI system and should not be?

An answer that works

Prompt assembly. What actually ends up in the prompt after retrieval, history management and compaction is often not what was intended, and the failure is silent. Asserting that the task survives compaction and that the cacheable prefix is stable catches real bugs.

Reading an answer is not being able to give one. The section that teaches this is Testing a system with a non-deterministic part, 6 hours of Testing.

Learn this properly

Also asked about Testing

How do you test code that calls an LLM?What is the difference between a test and an eval?How do you test an agent without calling a model?What is property-based testing and where does it help in AI code?