Failures

Errors made by language models on research tasks, with how they were caught and how to prevent them.

A model that is right 95% of the time is dangerous in research precisely because the other 5% looks the same. Each entry describes a real error, how it was caught and the check that would have prevented it.