[Paper] How does information access affect LLM monitors' ability to detect sabotage?

See this post on LessWrong here.