Key progress goals for turning the auto-interventions side of the work into a paper

[QuintinPope]

• Decide how best to present case studies in the paper

• Specifically: what metrics to offer? How to validate those metrics are meaningful? How to summarize many different validated hypotheses? Which hypotheses to include in the appendix?

• How do you measure hypothesis diversity? (implement some method and add it as a metric)

• Try cross-cluster testing for validated hypotheses (i.e., take hypothesis generated for cluster i and test it on the texts from cluster j)

• Decide which additional data to use (jailbreaking prompts I guess)

• Add those additional data to our experimental configuration options

• Run any additional case studies

• Revisit finetuning facts recall? (low priority)

• Continuously add the above to the paper as we do them