A model flagging potentially unsafe user content is being tuned. Missing a genuinely unsafe post is far more damaging to the company than reviewing a safe one unnecessarily. Which metric should be prioritised, and what is the consequence?

• APrecision, and the moderation team will see fewer items overall.
• BRecall, accepting that more safe posts will be flagged for review.
• CAccuracy, because it accounts for both kinds of error at once.
• DArea under the ROC curve, because it removes the need to choose a threshold.