Measuring AI answer instability due to typos, plus 2 other topics

Even if a model gets a perfect score on an evaluation question set, its answers fluctuate when faced with text typed by ...

It's just luck that it works — 74.5% of 455 AI libraries were importing dependencies not listed in requirements.txt

Yet, it's a ModuleNotFoundError in CI. When you ask an AI agent to 'set this up and run it,' it stops at the very first step.