Insights.
Practical notes on building trust and safety systems.
- · 5 min readThe moderation cascade: why most content should never reach an LLMFast filters, classifiers, an LLM policy judge and humans, each doing what it is best at. How to order them, what each layer costs, and where teams go wrong.
- · 5 min readHow to measure cost per moderation decisionWhat counts as cost, a formula you can compute from logs you already have, how to instrument each stage, a worked example and the usual ways to bring it down.