Shows LLMs hallucinate API names and arguments when calling large API collections from memory, and that retrieving documentation reduces this.

Disagreeing is the most useful thing you can do here. Both sides of every contested claim in this catalog were assembled by the same person, which is its weakest point.