Context

StreamSafe uses a binary classifier to score user-generated posts for policy risk and route the highest-risk content to a human safety queue. The current threshold was set six months ago, but policy operations now report that too many harmful posts are slipping through while reviewers are also near capacity.

Current Performance

Validation set size: 200,000 posts. Harmful content prevalence: 2.5% (5,000 posts).

The Problem

Leadership wants a threshold recommendation for routing high-risk content. Missing truly harmful content has regulatory and brand risk, but false positives consume reviewer bandwidth and delay benign posts.

Requirements

1. Recommend the most appropriate threshold and justify it using the table above.
2. Explain the precision-recall tradeoff in this routing setting.
3. Assess whether the model appears well calibrated enough for threshold-based operations.
4. Describe what additional offline and online validation you would run before changing the threshold.
5. Propose follow-up improvements if no single threshold cleanly satisfies business needs.

Constraints

• Human review capacity: 4,000 posts/day
• Estimated cost of a false negative: 20x a false positive
• Benign posts sent to review add moderation delay and creator friction
• Threshold changes can be deployed immediately; retraining takes 10 days