The Astra system card has a lot of useful info in it about monitorability and alignment; I'd recommend reading. Thanks to the people at OpenAI who worked on this! (I do think it is important to track important limitations of the evals, like limited elicitation for monitorability and difficulty interpreting alignment eval results given meta gaming and uncertainty about what was trained against. But the evals are still informative.) https://deploymentsafety.openai.com/gpt-…