Non-determinism and silent quality drift

Traditional software is deterministic: the same input returns the same output, so a passing test today passes tomorrow. LLM applications break that assumption. The same prompt can return different text on every call, and the quality of that text can shift for reasons you never touched.

Two forces cause this. The first is non-determinism inside a single model: sampling settings such as temperature mean two identical requests can diverge. The second, and more dangerous, is silent drift. A provider updates a model behind the same name, a retrieval source changes, or a small prompt edit ripples in ways you did not expect. Aggregate metrics can stay flat while a specific behavior quietly degrades, so the feature keeps returning answers that look fine and are wrong.

The reliability angle: without measurement, you learn about a regression from an angry user or a compliance incident, not from your own tooling. A confident, fluent, wrong answer is the hardest failure to notice by eye. The discipline this course teaches is to make quality observable, so that a drop shows up as a number you can alert on rather than a surprise in production.

You cannot improve, or defend, what you do not measure.